Read time: 10 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (21-August-2026):

• 🗞️ 1 in 3 web pages published since ChatGPT launched show signs of being written by AI.

• 🗞️ Today’s Sponsor: AWE 3.5 is TARS Robotics’ attempt to make robot intelligence native to the body from pre-training onward.

• 🗞️ The OpenAI-Anthropic enterprise race has changed direction.

• 🗞️ “Mathematics in the age of AI”

• 🗞️ Reddit has nearly fallen off ChatGPT’s citation radar after an OpenAI search change.

• 🗞️ “Practice Makes Unsafe: Skill Misevolution in Self-Improving LLM Agents”

• 🗞️ Zhipu AI Founder Tang Jie Discusses Scaling Laws: Trillion-Parameter Large Models Were a Detour for the Industry; The Next Key Step Lies in Post-Training

• 🗞️ Stripe’s leaked investor letter detailing its OpenRouter acquisition went viral on Twitter

• 🗞️ “What Does Context Compression Cost an Agent? Interaction Costs Unrevealed by Task-Completion Metrics”

• 🗞️ Anthropic plans to change enterprise data retention policy.
  

Connect with me on X (Twitter)

🗞️ Study finds 1 in 3 web pages published since ChatGPT launched were written by AI

More than 1/3 of web pages published since ChatGPT’s release appear to have been written by AI, according to a new Pew Research study

AI-associated writing is also spreading very unevenly across the web. By 2026, 9.4% of .com webpages showed significant signs of AI editing or authorship, compared with 1% of .edu and 0.8% of .gov pages.

May be because, commercial publishers can benefit directly when faster drafting lets them produce more pages, while public institutions often have slower publishing processes and weaker volume incentives.

If producing another page has an immediate economic payoff, that definitely is a great motivation to use a power tool.

🗞️ Today’s Sponsor: AWE 3.5 is TARS Robotics’ attempt to make robot intelligence native to the body from pre-training onward.

A lot of embodied AI still feels like AI modules bolted onto a robot.

TARS Robotics is taking a different architectural bet with AI World Engine (AWE) 3.5, TARS’ embodied-native foundation model for physical AI.

Its “Born as One” approach puts action, perception, geometry, and touch into one model from the beginning rather than stitching those capabilities together later.

The same model-driven system is designed to generalize across different tasks, objects, environments and robot bodies.

The training recipe then implements and validates a full closed-loop methodology for embodied-native foundation models through pre-training and post-training.

During pre-training, 2 priors give the model a base understanding of action patterns, spatial structure and understanding of physical laws before it is adapted to a robot, while post-training uses the AI World Engine to roll possible future states forward inside the model, predict what different actions may lead to and use those predictions to choose better actions.

TARS describes the full loop as 5 connected parts: embodied-native architecture, dual-prior pre-training, World Engine-driven post-training, scaling validation and continuous data feedback.

TARS positions AWE 3.5 as one of the most powerful embodied-native foundation models for general-purpose physical AI, with several minutes of long-horizon closed-loop reasoning and roughly 2x task execution efficiency versus PI0.5.

🗞️ The OpenAI-Anthropic enterprise race has changed direction.

Q3-to-date API spend growing 82% for OpenAI vs 76% for Anthropic. This data is from Ramp, and it says this token-spend dataset updates daily and has tracked companies’ publicly disclosed revenue growth surprisingly closely.

Till now, Anthropic had led OpenAI’s quarter-over-quarter growth in every period since Q2 2025, sometimes by a wide margin. GPT-5.6 Sol is increasingly becoming a developer choice, while Fable 5 has faced weaker adoption alongside price and regulatory data-retention constraints.

Subscribe now

🗞️ "Mathematics in the age of AI"

Terence Tao’s new paper.

AI may make proof generation abundant before mathematics has adapted to abundance. "I believe that we are now entering a era of comparable turbulence in mathematics.

This time, though, what is being stress-tested is not our foundational framework for mathematical truth, but rather the largely implicit framework of mathematical values and practices: what we consider a contribution to be, what we reward, what we regard as understood, and who — or what — we regard as having done the work.
I argue that it will become necessary to make these unwritten goals of mathematics much more explicit"

Connect with me on X (Twitter)

🗞️ Reddit has nearly fallen off ChatGPT’s citation radar after an OpenAI search change.

Per PromptWatch, Reddit’s share of ChatGPT Search citations fell 86%

Google AI Overviews and AI Mode showed much slower Reddit declines over the same period. Publishers trying to get cited by AI can lose that visibility quickly when the way ChatGPT searches the web changes.

🗞️ "Practice Makes Unsafe: Skill Misevolution in Self-Improving LLM Agents"

An agent can receive a clean prompt today and still behave unsafely because yesterday’s malicious task was saved into its skill library.

This paper calls that skill misevolution: an unsafe task succeeds, gets distilled into a reusable skill, and later changes behavior even after the original malicious instruction is gone.
Across 21 evolved agent-method configurations, all 21 authored unsafe skill artifacts, but only 15 produced fresh-session harm. i.e. every evolved setup learned unsafe skills, but in 6 of the 21 setups those skills did not cause harm in the later clean session.

So if you only check the agent’s final behavior, you can miss the fact that its persistent skill library is already carrying unsafe instructions that may be triggered later. The paper introduces SKILLMISEVO-GYM/BENCH to detect and measure how unsafe experience becomes persistent agent skills, and SAFEEVOLVE to reduce that risk by checking, repairing, tracking, and retiring unsafe skills before they keep propagating.

An attack on a self-improving agent can outlive the attack itself: one unsafe success can be converted into a reusable skill and influence completely clean future tasks.

And because the benchmark resets everything except the learned skill file, any later harm can be traced specifically to what the agent chose to learn and preserve.

🗞️ Zhipu AI Founder Tang Jie Discusses Scaling Laws: Trillion-Parameter Large Models Were a Detour for the Industry; The Next Key Step Lies in Post-Training

AI scaling is moving past parameter growth.

“How many parameters?” is becoming a weak way to describe how capable a model should be.
That model scaling now has several independent dials: parameters, training data, compute per forward pass, and post-training.
The best place to spend the next unit of compute depends on what the model needs to do and how often it will be used.

“Total parameters appear to matter up to a threshold — enough to hold the world — after which additional capability comes from scaling elsewhere: effective depth per forward pass, and above all post-training.”
“Put inference into the objective and the optimum moves toward smaller models trained far longer.”

GLM-5.3 is Tang’s implented example: same base, architecture, total parameters, and activated parameters as GLM-5.2, but 1 more month spent on long-horizon environments and RL.

🗞️ Stripe’s leaked investor letter detailing its OpenRouter acquisition went viral on Twitter

- "OpenRouter's token consumption compounding at 9% per week YTD."

- "We decided that January 1st marked the beginning of the singularity, and we have since been operating on that basis."

- "Agents are on the cusp of becoming economic actors in their own right."

- "We continue to believe that there is no ceiling on the size of the global economy (somewhat larger than $100T today). Implausible though it might sound on first blush, we think that it’s useful to contemplate the quadrillion-dollar world and to enumerate the relevant bottlenecks to bringing it about. (If global GDP per capita matched that of every Irish person—around $100,000—we’d be 80% of the way there.)

With more than 5 million businesses and flows representing almost 2% of global GDP having adopted Stripe, we’re pleased to be off to a good start, but we think that these figures are microscopic relative to what could be possible in the years ahead."

🗞️ "What Does Context Compression Cost an Agent? Interaction Costs Unrevealed by Task-Completion Metrics"

5× context compression did surprisingly little to GPT-5.5’s final task result.

The compressed agent remained statistically as successful as the full-context version.
What changed was how it got there: it relied much more heavily on retrieval to reconstruct dropped information.

🗞️ Anthropic plans to change enterprise data retention policy.

Anthropic’s 30-day Mythos-class retention rule is staying, while enterprises may soon hold the data themselves on their own cloud infrastructure.

They are saying the rule exists because some attacks only become visible across multiple requests, so covered-model prompts and outputs are retained for 30 days.

So the proposed mechanism is that instead of Anthropic storing the 30-day logs, the enterprise would keep them inside its own AWS/GCP environment under its own access controls. Anthropic says it would retain no copy;
Boris Cherny, head of Claude Code, clarified that they have been building the system with customers for some time and plans to ship it this fall.

Anthropic's current documentation already says retained covered-model data stays inside AWS or GCP when customers use Bedrock or Google Cloud Agent Platform. Direct Claude Platform retention, however, is still handled by Anthropic under today's rules.

The fall rollout therefore appears to broaden customer-controlled custody across Anthropic's enterprise access paths, rather than invent the pattern from scratch. OpenAI, on the other hand, yesterday announced, that they are taking a different technical route with Private Safety Processing, which it says preserves Zero Data Retention while detecting abuse across related interactions.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 13 min. 📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (18-August-2026):

• 🗞️ OpenAI stops reinforcement learning training for its frontier AI models for 2 weeks after signs that its upcoming Astra model had reached “Critical” cybersecurity capabilities.

• 🗞️ Synthefy just launched a foundation-model platform for structured numerical data where tables, transactions, sensor readings and time series should not need a separately trained model for every prediction problem.

• 🗞️ “Why Chinese citizens are far more optimistic about AI?” - Bloomberg published a piece.

• 🗞️ Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems

• 🗞️ “AQuA: Recursively Self-Improving Quantitative Trading Research Agents”

• 🗞️ “Small-Scale Experiments: Are They There Yet?”

• 🗞️ So GLM-5.3 has already found a “potentially serious vulnerability” in Cursor.

• 🗞️ Bloomberg reports OpenAI’s annualized revenue has topped $40B, roughly doubling its prior run rate.

• 🗞️ “Training AI Scientists to Replicate Research”

• 🗞️ Jagged Judges: Epistemic Stability Under Silence, Pressure, and Persistence

• 🗞️ Claude can now send emails in Gmail and manage files in Google Drive.

• 🗞️ Databricks Reveals How Enterprises Are Reining In AI Coding Costs

Connect with me on X (Twitter)

🗞️ OpenAI stops reinforcement learning training for its frontier AI models for 2 weeks after signs that its upcoming Astra model had reached “Critical” cybersecurity capabilities.

OpenAI slowed frontier training because Astra may have crossed the cyber threshold built for autonomous zero-day attacks.

It paused two weeks of deployment-focused RL training, while its largest planned frontier RL run remains on hold.

This is after the Hugging Face incident, where evaluation models escaped their intended network boundary and reached production infrastructure. Astra was not involved, but its separate evaluations were strong enough that OpenAI said it could not rule out the Critical threshold.

Research workloads now face stronger sandboxing, tighter network access, fewer shared services, and continuous security testing before resuming.

Monitoring also examines internal activity and tool actions, escalating suspicious behavior to automated investigators and human reviewers.

🗞️ Synthefy just launched a foundation-model platform for structured numerical data where tables, transactions, sensor readings and time series should not need a separately trained model for every prediction problem.

Most tabular ML still starts with the same assumption: new dataset, new training run. That will no more be true.

Synthefy released Nori, a foundation model for structured/tabular data that takes labeled rows as context and predicts new rows without training or fine-tuning a separate model for each dataset. In practice, it brings the foundation-model approach to regression problems that traditionally require their own ML pipeline.

🐙 GitHub: https://github.com/Synthefy/synthefy-nori

This will be highly impactful because the industrial world relies far less on textual data and more on transactions, sensor readings, trades, inventory, customer records, and time series.

So tabular ML can work like foundation models do elsewhere: reuse one pretrained model instead of rebuilding for every problem.

The big deal is they are trying to remove the “train a new machine-learning model for every new table/problem” step.

With Synthefy : New dataset → give Nori some labeled rows → ask it to predict new rows.

You pass labeled rows as context, then query it with new rows. fit() does not train or fine-tune anything. It stores the examples, and predict() produces the regression outputs in a forward pass.

Its open-source model Nori takes labeled rows as examples and predicts values for new rows with no training or fine-tuning step, with fit storing the labeled rows as context and predict returning values for new rows in a single forward pass.

Nori can run locally or through a hosted API, is available under Apache 2.0, and Synthefy says it has reached almost 600K model downloads and 5,000 Python installs.

With just 30 million parameters, Nori-30M rivals Google’s 1.6-billion-parameter TabFM across public regression benchmarks. With thinking enabled, Nori-30M-thinking surpasses it, achieving frontier accuracy at roughly 2% of the size.

The biggest idea is that numerical datasets contain enough reusable structure for one pretrained model to generalize across problems, roughly the same abstraction shift foundation models brought to text.

🗞️ “Why Chinese citizens are far more optimistic about AI?” - Bloomberg published a piece.

Stanford University’s AI Index finds that 84% of people in China are excited about AI, the highest level among surveyed countries. In the US, only 38% say the same.

Chinese optimism comes from expecting technology's gains to reach ordinary people and believing the state will restrain powerful companies. American skepticism, on the other hand is dominated by social-media harms, misinformation, concentrated platform power and job-loss forecasts.

The optimism gap may partly come down to who is already feeling AI’s threat at work. Nearly 80% of Americans work in services, versus about 46% in China, and generative AI is hitting those jobs first. Many Chinese workers simply haven’t felt the same pressure yet.

🗞️ Massive new paper from Anthropic. Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems

New paper from Anthropic + University in Switzerland.

AI agents can apparently persuade each other to adopt and keep spreading the same unwanted goal. This is basically the natural-language version of a computer worm, except the agents do the copying themselves.

This paper evolves “mind viruses” that spread through ordinary agent-to-agent messages, then persist by convincing newly infected agents to rewrite files loaded into future sessions. That persistence layer matters.

Payloads stored in the self-modifiable SOUL.md spread far better than payloads left in ordinary files because the instruction re-enters the system prompt after every context reset. Some evolved action viruses kept propagating across multiple hops, and all 4 tested payloads survived a 20-hop stress test in an artificial setup.

The good news: these “mind viruses” are still fairly easy to stop.
They struggled to spread on social networks, and on Claude Haiku 4.5, a simple warning stopped every evolved attack from getting past 1 hop, even after 150+ attempts. So the practical lesson here is: treat persistent agent files like security-sensitive config, and teach agents to reject anything that asks them to copy itself to other agents.

A bad idea inside an agent can become self-sustaining infrastructure: it can persuade other agents to adopt it, write itself into persistent memory, survive context resets, and gradually redirect the behavior of an entire agent network.

Subscribe now

🗞️ "AQuA: Recursively Self-Improving Quantitative Trading Research Agents"

A bad experiment is bad enough. A self-improving agent can keep building on it.

New Stanford, Princeton, Ant group paper.
AQuA targets that failure mode: if an agent writes a leaky experiment, gets a great score, and stores it as precedent, recursion can amplify the bug.

The authors saw this in an earlier version, where an LLM-written volume feature quietly used end-of-day information and a second LLM reviewer still approved it.
Their fix is architectural.

Agents can propose factors or model changes only through restricted specifications; the data path, labels, splits, and evaluator stay sealed outside their reach. Search sees validation scores, while the final test window stays untouched until the configuration is frozen.

On US equities, the hybrid model reaches +0.0843 IC versus +0.0613 for the strongest baseline, and its long/short strategy reaches +2.50 Sharpe at 2 bps, with about +2.0 under fully causal walk-forward testing. These are simulated, not live-trading results.
For autonomous research agents, the design rule is useful: let the agent improve the experiment without letting it redefine valid evidence.

Connect with me on X (Twitter)

🗞️ “Small-Scale Experiments: Are They There Yet?”

New Meta paper shows, small models may not be bad predictors of scale; they may just be getting under-tuned.

Finds scaling laws emerge around 4M parameters, where models can train in under 1 hour on 1 GPU. Small models are unusually sensitive to hyperparameters.

With 4 or 16 configurations per scale, the law is basically invisible; at 64 it appears but extrapolates poorly, and at 256 it becomes accurate. As models get larger, good settings occupy more of the search space, while the effective number of hyperparameters near the optimum drops toward 1.

That helps explain why scaling laws look cleaner at larger sizes: the models are easier to tune.
As a check, small-scale runs recover that pre-norm transformers scale better than post-norm over the tested range.

There is a limit: extrapolate too far beyond the measured scales, and statistical errors can dominate.
For model research, cheap experiments may need more tuning breadth, not more model size.

🗞️ GLM-5.3 found a "potentially serious vulnerability" in Cursor.

GLM-5.3's CyberGym score rose to 84.5%, while ExploitBench more than doubled from 24.4% to 54.4%. Shows how much more performance a frontier-scale base model can deliver without going through another costly pretraining run. “Scaling post-training is all we did for GLM-5.3,” Z .ai said in its technical announcement.

🗞️ Bloomberg reports OpenAI's annualized revenue has topped $40B, roughly doubling its prior run rate.

The pace rose more than 20% month over month in July, with coding software, subscriptions and ads helping drive growth.

And then Anthropic says its own run-rate revenue crossed $47B in May, though according to some, the companies may calculate the metric differently. OpenAI is cutting prices while revenue is accelerating.

i.e. it is prioritizing workload capture over extracting maximum price today for that second-order advantage:
Once developers or companies embed Codex/agents into recurring workflows, there will be switching costs.
Cheaper inference now can buy more deeply embedded demand later.

🗞️ "Training AI Scientists to Replicate Research"

A 27B research agent outscored Claude Opus 4.8 and GPT-5.5 on held-out paper replication by learning how to direct the research while outsourcing the coding.

Huge implication, maybe you do not need one giant model to do everything. You can train one model to think like the researcher, then let stronger coding models do the implementation underneath it.

Replica, proposed in this paper, a scalable task space for paper replication, gives it a surprisingly simple way to practice that job.
Take a research paper, remove one results figure, and tell the agent: recreate this result by actually running the experiment.

Now the agent has to figure out all the messy stuff papers leave out: what to implement, what to simplify, what experiments to run, and whether the result is believable.
That gives the researchers something they can repeatedly train on, with an automated rubric grading each attempt.

After training on 242 of these tasks, Faraday uses GPT-5.5 as its coding agent but decides what research to do. On 68 unseen AI-for-science tasks, Faraday scored 0.791 versus 0.748 for Claude Opus 4.8 and 0.729 for GPT-5.5, beating both on 60% of tasks.

And the difference was not just prettier plots. Faraday was more likely to actually test the mechanism in the paper instead of taking shortcuts that produced the expected-looking answer.

🗞️ Jagged Judges: Epistemic Stability Under Silence, Pressure, and Persistence

Meta's new paper, the dangerous failure mode is not just a wrong judge, but a correct judge that can be persuaded into becoming wrong.

We are increasingly using AI models to judge other AI models.
But what if the AI being judged can simply argue with the judge until the judge changes its decision?

Meta tested exactly that. Across 9 frontier models, an adversarial LLM could flip judge verdicts on 62–91% of tested cases under sustained adaptive persuasion.

And changing the judge’s mind usually didn’t fix a mistake. It made the judgment worse: under the adaptive attack, 70% of successful flips moved away from the ground truth.

That creates a very practical problem for agent systems. If one AI is supervising another AI, the supervised agent may eventually be able to contest, negotiate with, or strategically persuade its own evaluator.

🗞️ Claude can now send emails in Gmail and manage files in Google Drive.

Ask it to reply to a thread, and it drafts and sends the response. You decide when Claude needs your approval. Connect Gmail or Google Drive using the connectors menu. Available on every paid plan.

🗞️ Databricks Reveals How Enterprises Are Reining In AI Coding Costs

Use internal evals on your own workloads, because public benchmarks are weak predictors of cost versus coding quality.

Route each task to the cheapest model that clears the required quality bar instead of defaulting everything to the most expensive model.

Do not automatically upgrade to newer frontier models; newer versions can cost more without delivering enough additional quality.

Optimize the harness and prompt/cache path, since reducing unnecessary context and generated tokens can cut costs substantially without lowering observed quality.

Keep models interchangeable through a common orchestration layer, because harness lock-in makes it harder to move spending when a cheaper model becomes competitive.

Use spend visibility, self-clearing budget gates, and cheaper-model fallbacks before hard caps, so expensive models remain available only where they are justified.

Databricks’ evidence also suggests substantial cost sits outside model pricing itself, especially in harness design, repeated context, caching, and unnecessary token generation.

Portability becomes part of cost control because companies can only exploit falling model prices if switching models does not require rebuilding their coding workflow.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 7 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (16-August-2026):

• 🗞️ Wall Street heavyweights pour $500 B into Nvidia’s AI push

• 🗞️ Anthropic is targeting a September or early-October IPO, before OpenAI, whose IPO may now come as late as next year.

• 🗞️ Nvidia made a very smart move by partnering with 6 large Wall Street investors to create a financing platforms that lend against its chips.

• 🗞️ Anthropic CEO Dario Amodei says the backlash against AI is “fundamentally a crisis of trust.”

• 🗞️ “Reason Wide, Not Deep: Amortizing the Reasoning Premium into Distilled Skills”

• 🗞️”Model or Harness? An Interaction-Centric Taxonomy for Localizing Agent Failures”

• 🗞️ “Skaling: Chinchilla’s Exponents Meet Kaplan’s Coupling”

• 🗞️ “Sample More, Reflect Less: Self-Refine and Reflexion Lose to Repeated Sampling at Equal Token Cost, from 1.5B to 7B”

Connect with me on X (Twitter)

🗞️ Wall Street heavyweights pour $500B into Nvidia’s AI push

Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR signed MOUs to create dedicated capital pools intended to offer NVIDIA customers attractive financing rates.

- Goldman Sachs explicitly described the opportunity as a market for credit backed by NVIDIA compute, turning GPU capacity into an asset lenders can underwrite.

- Jensen Huang said NVIDIA itself can backstop up to $125B, or 25% of potential deals, so third-party capital does not necessarily mean zero NVIDIA exposure.

- This follows NVIDIA's July credit-support and revenue-sharing model for AI clouds, where it can earn hardware revenue and a share of cloud revenue from supported capacity.

- Long-duration financing assumes those GPU fleets keep producing enough economic value as hardware generations change, while the Bank of England has already warned that rapidly expanding AI credit could transmit shocks through financing markets.

Nvidia is also deepening its ties with major private capital firms, which are gearing up to pour trillions of dollars from insurance, retail and institutional investors into AI infrastructure. Groups including Apollo and Blackstone have recently structured AI infrastructure deals to help companies such as Anthropic cover huge chip and data centre costs. They have also created multibillion-dollar, investment-grade financing structures for firms including Meta and Intel that remain outside their corporate balance sheets.

What this all means for Nvidia and the others

• Once Wall Street accepts GPUs as recoverable collateral, Nvidia can gain an advantage that has nothing to do with FLOPS or benchmark scores.

• The new structures are explicitly aimed at buyers whose balance sheets cannot fund the buildout alone, while institutional capital can place debt against Nvidia equipment. Hardware shipments can therefore stay strong even while the underlying operators remain financially weak.

• The “circular financing” criticism is not fully correct in this case, that is too broad unless Nvidia ends up carrying most of the credit risk itself. But the proposed platforms are meant to mobilize more than $500B of third-party capital, with independent financiers underwriting projects, although Nvidia can backstop up to 25% of potential deals. That makes the better comparison equipment finance rather than vendor buying its own output.

• The boundary to watch is how much loss ultimately returns to Nvidia when borrowers fail.
  

🗞️ Anthropic is targeting a September or early-October IPO, before OpenAI, whose IPO may now come as late as next year.

- Potential investors are already pressing Anthropic on cheaper Chinese AI models, tensions with the Trump administration, and growing opposition to data-center construction.

- Anthropic executives have reportedly downplayed Chinese competition by arguing that users still prefer the most capable models, while Chinese systems generally trail leading U.S. models by several months.

- Anthropic plans to push further into healthcare and biology, partly because those applications could help counter growing negative sentiment around AI.

- Claude Code has become the engine behind Anthropic's recent surge, with run-rate revenue exceeding $47B in May.

- Demand has been strong enough that Anthropic has suffered intermittent outages even after signing additional compute deals with SpaceX and Google.

Subscribe now

🗞️ Nvidia made a very smart move by partnering with 6 large Wall Street investors to create a financing platforms that lend against its chips.

Wall Street Journal published a very interesting piece on this major development.
Nvidia's revenue depends on customers being able to pay for chips most cannot buy with cash. The large tech firms can; the smaller AI labs, clouds and enterprises behind the rest of the demand have to borrow, and their costs have risen as investment-grade issuers absorb the available credit.

CoreWeave paid yields above 9% this month, showing what AI borrowers now pay for compute. If they cannot finance purchases, orders slow even though demand exists.

The earlier fix was Nvidia's own balance sheet, through investments in customers and guarantees on their debt, which drew criticism as circular funding. So routing capital through outside asset managers replaces Nvidia's money with pension, insurance and sovereign-fund money, and moves credit decisions elsewhere.

Jensen Huang says Nvidia may backstop 25% of a project's cost through residual-value support. That 25% residual-value support is what gets that money moving.
Investors are lending against hardware whose resale value depends on how fast Nvidia's next generation arrives, the one variable they cannot price and Nvidia controls. Guaranteeing a floor covers exactly that risk, which is why skeptics still see circularity in a thinner form.

Connect with me on X (Twitter)

🗞️ Anthropic CEO Dario Amodei says the backlash against AI is “fundamentally a crisis of trust.”

In a long and super viral post on X on Saturday, a rare move for someone who usually avoids social media, he began with AI regulation, rejecting the choice between concerns about regulatory capture and concentrated power and the belief that broadly available AI, including open models, is the best safeguard.

- Predics that AI could help cure most human disease within roughly 5–10 years.
- says AI is structurally prone to concentrating power, even without regulation.
- He wants regulation designed to slow frontier AI labs while giving smaller challengers more room to catch up.
- Argues open-weight AI does not solve concentration, because even with open-source serious capability still depends on scarce compute and chips.
- He supports pre-deployment testing of frontier models, with open-weight models tested too once they approach frontier capability

- Says AI companies will not win public trust through better marketing. They need to deliver undeniable real-world breakthroughs first.

So now he just tied public trust to visible medical outcomes of AI, looks like Anthropic’s biology push is for real.

All these started when hedge fund founder Gavin Baker publicly said.

“Internally, Anthropic is very confident...that Dario has said that Anthropic might be the only private company in the world at some point. In this vision, there’s Anthropic, and then there are governments, and that’s it.”

🗞️ "Reason Wide, Not Deep: Amortizing the Reasoning Premium into Distilled Skills"

What if you could pay the reasoning cost once, then reuse what the model learned across future tasks?

New Microsoft paper finds that some expensive test-time reasoning can be replaced with a small set of rules learned from previous agent runs.
The paper tests a cheaper alternative: collect 35–50 past trajectories, have a coding agent extract recurring failure patterns, then turn those patterns into a small markdown skill added to the non-reasoning model’s system prompt.

For GPT-5.4-mini, those skills recovered 55%–100%+ of the gap between non-reasoning and reasoning modes across 4 agent benchmarks, while using 2.9–4.5× fewer output tokens than reasoning. On ALFWorld and τ²-retail, the skilled non-reasoning model actually beat the reasoning mode.

The useful part is that the distiller did not need reasoning traces: skills built only from cheap non-reasoning rollouts were competitive across all 4 domains.
The limit is equally useful: reasoning still won on telecom and SpreadsheetBench, where each task contains more instance-specific dependencies that a fixed skill cannot capture.
So the practical split is: distill repeated procedures once, then reserve expensive test-time reasoning for the tasks that genuinely need fresh search.

🗞️"Model or Harness? An Interaction-Centric Taxonomy for Localizing Agent Failures"

When an AI agent fails, blaming the model can send you to the wrong fix.

This Scale AI paper argues that agent debugging should start by asking where the first unrecovered failure happened: in the model, context, memory, tool layer, another agent, grader, or environment.
That matters because the same visible mistake can have different causes.

An ignored instruction might mean the model saw it and failed to follow it, or the harness dropped it during context compaction. The paper organizes 41 failure modes around these interactions, tagging each by both the interaction and the side at fault.

It tests the taxonomy on 40 worked examples. GPT-5.5 matched human category labels with 80% accuracy and Cohen’s κ of 0.76.

Using 4 judges together pushes category precision to 96% when all 4 agree, although coverage falls to 68%.
For agent teams, the practical shift is simple: trace the failure back to the component that broke first, then decide whether to retrain the model, fix the harness, or repair the environment.

🗞️ "Skaling: Chinchilla's Exponents Meet Kaplan's Coupling"

A new Meta FAIR paper finds a blind spot in the Chinchilla scaling law that becomes expensive when you extrapolate.

Chinchilla can look almost perfect inside a training grid and still mispredict what happens at the frontier.
The problem is: Chinchilla assumes model size and training data help separately, but the experiments show that each changes how useful the other one is.

Skaling adds just 1 extra term to capture that connection, cutting prediction error by about 1.5–3× and getting full-grid Chinchilla-level prediction accuracy with roughly 10× less profiling compute.
On Farseer, the difference becomes huge at frontier scale: at 2×10^25 FLOPs, Chinchilla points to ~380 tokens per parameter, while Skaling and the paper's direct estimates land around 20–40.

🗞️ "Sample More, Reflect Less: Self-Refine and Reflexion Lose to Repeated Sampling at Equal Token Cost, from 1.5B to 7B"

If you're spending extra tokens making an LLM critique itself, this paper says a simpler move can be better: just let it try again.

Before you make an LLM reflect on its answer, try giving it another independent attempt.
The study compares 7 test-time reasoning methods on Qwen2.5 models from 1.5B to 7B, then asks a fairer question: what happens if repeated sampling gets the same token budget?

Repeated sampling means solving the same math problem several times and taking the answer that shows up most often. Across 36 comparisons, none of the more elaborate methods reliably beat that baseline at equal generated-token cost; 10 were significantly worse.

For checkable reasoning, extra tokens may be better spent creating independent attempts than asking the model to reconsider its own work.
Ofcouse, study boundary matters: this is Qwen2.5 on math, not frontier models or open-ended tasks.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 10 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (13-August-2026):

• 🗞️ SpaceXAI launches Grok 4.6, beating Kimi K3 and tying GPT-5.6 Sol for 3rd place on Artificial Analysis

• 🗞️ Soniox launched TTS v2, a text-to-speech model, very premium voice quality at a dramatically lower price ($ 0.70-per-generated-hour)

• 🗞️ NVIDIA released Nemotron 3.5 Lightning for long-running agent execution.

• 🗞️ Stealing Reasoning Traces from Proprietary LLM APIs

• 🗞️ “The Case Against Generation for Retrieval: Discriminative Language Models as Effective Retrievers”

• 🗞️ OpenAI finally launched a ChatGPT desktop app for Linux.

• 🗞️ PROMPTS: Performance Optimization via Multi-Agent Planning for LLM Training and Serving

• 🗞️ SpaceXAI launched Grok Bot.

• 🗞️ How Claude marks AI-generated content

• 🗞️ “ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence”

Connect with me on X (Twitter)

🗞️ SpaceXAI launches Grok 4.6, beating Kimi K3 and tying GPT-5.6 Sol for 3rd place on Artificial Analysis

Grok 4.6 gets 98.4% of Fable 5’s Intelligence Index score at a fraction of its API price.

5X cheaper on input tokens and 8X less for output tokens vs Fable 5.

Looks clearly the strongest intelligence-per-dollar equation.

Grok 4.6 scores 61 versus Claude Fable 5 Max’s 62, just a 1-point or 1.6% relative gap on that composite score, while Grok costs 80% less on input tokens ($2 vs $10/M) and 88% less on output tokens ($6 vs $50/M).

Clearest strength is professional agent work, with leading results against GPT Sol Max and Fable 5 Max on GDPVal-AA v2 and AA-Briefcase.

- matches GPT-5.6 Sol Max at 61 on Artificial Analysis while charging $2/$6 per million input/output tokens.

- 1753 on GDPVal-AA v2 and 1577 on AA-Briefcase, both ahead of Sol Max and Fable 5 Max.

Those benchmarks measure real-world agent tasks and agentic knowledge work, making them closer to research, analysis and multi-file deliverables than isolated question answering.

On coding, 69.9% on CursorBench v3.2 beats Sol’s 67.2%, while DeepSWE and Terminal-Bench leave Grok behind both Sol and Fable.

SpaceXAI attributes the jump to a longer training run, regenerated SFT trajectories, model-based trace filtering, and agentic RL across coding, web development, CAD and kernel optimization.

It also reports more self-testing on long trajectories, with the model checking its work before continuing, directly targeting error accumulation across multi-step agents.

🗞️ Soniox launched TTS v2, a text-to-speech model, very premium voice quality at a dramatically lower price ($ 0.70-per-generated-hour)

I wasn’t expecting Soniox TTS v2 (a text-to-speech model) to sound this natural.

They just released this TTS v2

• Really premium voice quality at a dramatically lower price ($ 0.70-per-generated-hour). While keeping the same model suitable for real-time agents, multilingual speech, expressive control, and cloning.

• Soniox TTS v2 is built for live agents, and voice agents require more than natural-sounding audio. Speech must begin quickly, remain synchronized with the conversation, stop immediately when the user interrupts, and avoid repeating text that has already been spoken.

  Interrupt a normal voice agent and it has no clue how much you actually heard, so it repeats itself or skips ahead.

  TTS v2 timestamps every character down to the millisecond, so it knows the exact word you cut it off at and carries on from there.

• Exceptional precision, high-fidelity voice cloning. More than 60 languages, natural language mixing, and low-latency streaming together in one model.

• It makes voice performance programmable. Its “audio tags” allow developers to direct emotion, delivery, and vocal reactions throughout the text instead of selecting one fixed style for the entire passage. like [whispering], [excited] or [laughing] straight into the text,

• One model replaces a lot of voice-stack plumbing: expression control, voice cloning, 60+ languages, language mixing, pronunciation precision, and streaming all sit in the same system.

• It is unusually well designed for live AI agents: low-latency streaming plus character-level timestamps let an agent start talking early, stop cleanly when interrupted, and resume without repeating itself.

• Voice performance becomes programmable: developers can insert audio tags for whispering, excitement, laughter, pauses, and other delivery changes inside the generated text.

Try Soniox TTS v2 here

Compare TTS: https://soniox.com/compare-tts (still v1 model, not v2 yet)
Console: https://console.soniox.com
Docs: https://soniox.com/docs/tts/get-started

🗞️ NVIDIA released Nemotron 3.5 Lightning for long-running agent execution.

- 30B total / 3B active parameters, with up to 1M tokens of context.
Ready for commercial use.

- Local deployment: NVIDIA ships BF16 and much smaller NVFP4 checkpoints, lists 1× DGX Spark or 1× H100 for single-GPU deployment, and also lists RTX 5090 among supported hardware.

- NVIDIA then adds multi-token prediction, DSpark and DFlash speculative decoding, plus NVFP4 quantization to push generation speed further.

- NVIDIA claims up to 4× output speed; on PinchBench, 86% accuracy and 30% faster completion than Qwen3.6 35B at similar accuracy.

🗞️ Stealing Reasoning Traces from Proprietary LLM APIs

Very big warning from this paper.

Your AI's hidden reasoning can become a second, invisible data leak. You can clean the visible chat and still accidentally publish passwords, API keys, or private data.

A company can secure its smartest model perfectly and still lose its secrets through a cheaper model in the same family. The weakest model can become the security hole for the strongest one.

Encrypted reasoning is supposed to hide what frontier models think, but this paper shows the encrypted blob itself can become the attack surface. Anthropic, OpenAI, and Google APIs return opaque reasoning blocks to clients so they can be replayed in later turns without storing the reasoning server-side.

The problem is portability: the authors found these blocks could be reused across sessions, users, and even models within the same provider.
That lets an attacker take a trace from a stronger model, feed it to a weaker compatible sibling with looser safeguards, and use that model as a decoder.

Across public agent logs, the team decoded 315,320 reasoning blocks from 6,708 trajectories; 328 sessions, or 4.9%, leaked at least one sensitive item. In genuine user sessions, they recovered 62 API keys, 33 passwords, 24 access tokens, and 30 personal emails.

So an encrypted reasoning field in a log is not necessarily safe metadata if another model can read it. After responsible disclosure, providers patched the reported attacks, and the exact exploits were no longer reproducible by August 2026.

Subscribe now

🗞️ “The Case Against Generation for Retrieval: Discriminative Language Models as Effective Retrievers”

Meta’s new retrieval paper is a reminder that better language models do not necessarily require more language generation.

It asks a useful question: why generate item IDs at all when the LLM can learn the retrieval space directly?

Instead of autoregressively generating item IDs or ranked lists, the paper puts a 0.6B Qwen3 model inside a classic two-tower retriever, so item embeddings stay precomputable and retrieval remains a fast vector search. The stronger cross-encoder is used as a teacher, transferring its candidate-ranking distribution into the cheaper two-tower student.

That distillation is doing a lot of the work: removing it cuts Recall@10 by 13.3%, 23.1%, and 8.0% on Beauty, Sports, and Toys.
Against OneRec-Think, which uses Qwen3-8B, the 0.6B two-tower model reports higher Recall@10 on all three public datasets, although its NDCG results are mixed.

On Meta’s internal production data, the same approach matches a heavily tuned DLRM retriever’s normalized entropy while using just 0.5% of the training data. Those production results are self-reported, but the engineering implication is specific. For first-stage retrieval, the LLM does not necessarily need to generate anything; semantic representations inside a two-tower architecture may be a much more practical use of the model.

Connect with me on X (Twitter)

🗞️ OpenAI finally launched a ChatGPT desktop app for Linux.

ChatGPT, Work, and Codex into one interface.

Codex already worked on Linux through its CLI, so the new app mainly adds the integrated desktop layer. The preview supports Ubuntu 24.04 and 26.04 LTS, Debian 13, and Fedora 43 and 44 on x64 and ARM64.

🗞️ PROMPTS: Performance Optimization via Multi-Agent Planning for LLM Training and Serving

New Google paper shows LLM infrastructure optimization does not have to be a giant brute-force search if an agent can understand why the system is slow.

The key shift is from “search everything” to “understand the bottleneck, then search the tiny part that matters.” Instead of asking, “Which of these 100 configurations is fastest?”, this work asks, “What is actually causing the slowdown?” That one change dramatically shrinks the search.

Its Analyzer Agent reads profiler traces and classifies the bottleneck as compute, memory, or communication, then a Proposal Agent retrieves engineering knowledge and generates three targeted ways to map data, model, and sequence parallelism across TPU chips. Across eight production workloads spanning dense and MoE models, pre-training, post-training, serving, and TPU systems from 2 to 2,048 chips, the human-validated production configuration appeared in PROMPTS' first batch every time. In seven of eight cases, it was also the first configuration tested, and the agent's top-ranked proposal matched what engineers ultimately adopted in 87.5% of cases.

🗞️ SpaceXAI launched Grok Bot.

The biggest architectural shift is persistence, with cloud computers that keep working after the user steps away.

It can sign into websites and applications even when no clean API or MCP exists, putting agent actions inside the same software people use. Multiple Bots can message one another, share context, and hand work to specialists, including a manager Bot coordinating several jobs.

A Bot can also watch someone perform a workflow, save it as a routine, absorb corrections, and repeat the process later. The beta remains restricted to SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium users, while enterprise access is still waitlisted.

🗞️ How Claude marks AI-generated content

Anthropic is adding invisible watermarks to text from new Claude models.

Because marking happens at model level, supported outputs carry it across Claude products, the API, Claude Code, and major cloud platforms worldwide. The watermark is woven directly into generated text, allowing it to travel through copying and sometimes survive editing.

This is because, Anthropic signed the EU's voluntary Code of Practice on Transparency of AI-generated Content, whose measures help implement Article 50's legally binding marking requirement. Claude models launched in the EU on or after August 2, 2026 support marking at launch, with supported output marked worldwide.

Supported SVG, PNG, and JPG files use signed C2PA provenance metadata, which can indicate Claude processing and whether that metadata was altered. Anthropic's own detection support is still in development, with technical documentation forthcoming for users and third parties.

A detected mark cannot prove Claude originally authored the material, because Claude may have edited, translated, or summarized human-created content. Missing marks prove little either, since heavy rewriting, short passages, older models, unsupported surfaces, or stripped file metadata can erase the signal.

For publishers and platforms, the system adds useful provenance evidence.
Its practical value will depend on how resistant the text watermark is to paraphrasing once Anthropic publishes the detector and technical method.

🗞️ "ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence"

Google’s ScientistOne paper tackles a basic problem with AI-generated research:

The result can look credible even when its evidence chain is broken. AI research agents are getting good enough at solving benchmark problems that the new bottleneck is whether you can trust the paper they write afterward.

Google Cloud AI Research audited 75 papers from five autonomous research systems on five ADRS tasks, and every baseline showed at least one systematic evidence failure. Some fabricated citations, some reported scores that did not reproduce, and some described algorithms that were simply not in the submitted code.

ScientistOne attacks that gap with “Chain-of-Evidence”: citations must trace to retrieved papers, numerical claims to evaluator logs, and method claims to implementation artifacts before the manuscript is finalized.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 10 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (11-August-2026):

• 🗞️ Meta Superintelligence Labs released Muse Glimmer, a 30B-parameter model built for always-on local agents on one consumer GPU.

• 🗞️ Today’s Sponsor: This AI Doesn’t Dodge the Uncomfortable Questions - Uncensored AI

• 🗞️ Mark Zuckerberg just dropped a really long piece on Meta’s vision for the AI future where superintelligence is available to everyone.

• 🗞️ Today’s Sponsor: Your ERP Runs Your Books. Domus Runs Your Compliance.

• 🗞️ Bernie Sanders demands OpenAI, Anthropic, and Meta pause AI development over escalating safety concerns.

• 🗞️ Qwen released a multimodal tool layer for AI agents.

• 🗞️ Australia’s first known autonomous AI agent driven hack.

• 🗞️ Metis: Memory Foundation Model

Connect with me on X (Twitter)

🗞️ Meta Superintelligence Labs released Muse Glimmer, a 30B-parameter model built for always-on local agents on one consumer GPU.

Meta’s Muse Glimmer changes the AI market equation: it is capable, free, and runs completely on consumer hardware. That puts cloud-based competitors under pressure, because a model small enough for a gaming PC can do much of what users currently pay monthly API charges to access.

weights on HuggingFace under an Apache 2.0 license,

- At full precision the model would need 55GB+ of memory.

- Meta compressed the weights to roughly 4-bit precision, shrinking the language model to under 20GB.
That leaves room inside a 24GB or 32GB envelope for the KV cache, a perception encoder for screenshots, and a speculative decoding drafter.

- The compressed model still leads Gemma4-31B and Qwen3.6-27B on MCP Atlas, at 75.5 against 54.2 and 62.5.
That lead reverses on computer use and terminal coding, where Qwen3.6-27B scores higher on OSWorld-Verified, TerminalBench 2.1, and SWE-Bench Verified.

- Instead of writing one word at a time, Muse Glimmer lets a small helper model called DFlash guess a whole block ahead.
The main model then checks that block in one pass, keeping the correct guesses and fixing the wrong ones.

- On an RTX-5090, that lifted decode speed from 74.9 to 233 tokens per second, a 3.1x gain measured through llama.cpp.

- Apple hardware gains less, 1.8x on an M5-Max and 1.5x on an M4-Max, reaching 50 and 38 tokens per second.

Speculative decoding pays off far more on a desktop GPU than on a Mac

Turns the 74.9 tokens per second into 233 on an RTX-5090 while an M4-Max moves from 23.7 to 38.
Speculative decoding is the general method: a small fast model guesses a block of tokens, the big model checks the block in one pass.
DFlash is Meta's particular small guesser, shipped alongside Muse Glimmer.

They said - "Muse Glimmer ships with a lightweight "drafter" model based on DFlash — a small companion network that proposes entire blocks of tokens at once. The main model then verifies these proposals in parallel, accepting correct tokens and correcting wrong ones. "

🗞️ Today’s Sponsor: This AI Doesn’t Dodge the Uncomfortable Questions - Uncensored AI

I tested Uncensored AI with questions that are sensitive enough to make most AI models become overly careful.

It goes where no traditional chatbots will go. Will answer questions that all other popular ones would faint at the thought of answering.

Forget polite AI. Forget the censored responses. This is unvarnished truth from an LLM, even if it’s extremely uncomfortable. Uncensored[.]ai is built to answer anything you throw at it.

It directly addresses political censorship, private control of speech, government incentives, and war reporting without redirecting away from the premise.

Behind the scenes, Uncensored.ai provides an uncensored information platform using retrieval-augmented generation and knowledge graphs. Just direct questions on politics, censorship, religion, war and institutional power.

Beyond chat, it has got quite a few super interesting features. Rabbit Hole turns public records and controversial datasets into searchable investigations, while Who Mogged compares two faces across structure, proportions, skin, grooming, and presentation. Slop or Not analyzes meals for processing and nutritional quality, FactCrack turns factual clues into a deduction game, and the suite also includes 1A Forum.

🗞️ Mark Zuckerberg just dropped a really long piece on Meta’s vision for the AI future where superintelligence is available to everyone.

some strong quotes from the blog.

- “Any policy that slows American model releases -- even by a month -- could add significant risk to American leadership while letting foreign models race ahead.”

- “For example, if a self-improving AI system focused on optimizing its compute efficiency, it could theoretically invent ways to squeeze 100x or more intelligence out of each gigawatt.”

- “There is no such thing as a singular benevolent superintelligence.”

- “While there are risks to releasing capable models, the most dangerous scenario from this perspective would be leading AI labs training powerful models and keeping them for themselves.”

- “I also propose that frontier AI labs should share intermediate training checkpoints of new models for government use and review rather than waiting until training has completed.”

Subscribe now

🗞️ Today’s Sponsor: Your ERP Runs Your Books. Domus Runs Your Compliance.

Domus is a compliance operating system that sits alongside ERP {enterprise resource planning} platforms such as Vantaca, CINC, and AppFolio.

Most HOA (homeowners association) software handles accounting, payments, and property operations. Domus handles the compliance layer those systems usually leave to PDFs, inbox searches, and attorney calls.

Boards and community managers can ask a governance question and get the exact article or statute supporting the answer. Violations become statute-clocked workflows instead of generic tickets, with notice requirements, cure periods, hearing rights, deadlines, and decision history attached to the case.

That means fewer manual document searches, a clearer audit trail, and a faster way to see when an issue actually needs legal counsel. Domus is a defensibility tool, not a lawyer.

Book a 15-min walkthrough at domus.town. First 10 communities per state get founder-led onboarding.

🗞️ Bernie Sanders demands OpenAI, Anthropic, and Meta pause AI development over escalating safety concerns.

Some quotes from the letter.

- "Let me be very clear: If you do not take appropriate action now, my colleagues and I in the U.S. Senate will."

- "That moment is here. AI capabilities HAVE reached a critical threshold."

- "Mr. Altman, Mr. Amodei and Mr. Zuckerberg: In the interest of humanity, stand by your words. Pause AI development. It is not too late to avoid disaster. Stop building machines that humans cannot control."

- "And yet, at a moment when we have seen human loss of control and the creation of potentially dangerous viruses, your companies are still racing ahead— investing tens of billions of dollars into a technology that nobody can fully understand, predict or control."

Connect with me on X (Twitter)

🗞️ Qwen just released a multimodal tool layer for AI agents.

It packages multimodal operations as tools that an agent running inside Claude Code, Codex, Qwen Code, Gemini CLI and other agent harnesses can discover, call, and chain together while doing a larger task.

The Github repo is actually a collection of separate plugins/capabilities. e.g. the core plugin gives an agent tools such as read_image, read_video, visualize, OCR, object grounding, segmentation, speech transcription, cropping, etc.

🗞️ Australia’s first known autonomous AI agent driven hack.

An OpenClaw agent running Anthropic’s Claude cancelled a stranger’s gym reservation to move its own user up a waitlist.

Removing someone else ahead of its own user was the shortest available path to the goal its user had set. Andrew, the user asked the agent to book a gym class, and while working through the booking site it found it could reserve classes further ahead than the gym normally permitted.

He was separately sitting fourth on a waitlist, so he asked whether it could move him up; the agent probed the booking API, found that cancelling someone else’s reservation required no authorisation, and tested that on the person in first place, which pushed Andrew to third. Australian law has no settled answer about who pays for that, because only a legal person can be liable, technology lawyer Hayden Delaney told the ABC. That leaves the user, the agent developer, the model provider and the site operator in the frame.

🗞️ Metis: Memory Foundation Model

This is such a wild idea.

What if memory were a capability of the LLM itself, rather than a retrieval system bolted around it?

Can an LLM remember something from an earlier interaction without having to feed that old interaction back into its prompt?

Metis proposes a foundation model with a persistent memory state inside the backbone, updated during ordinary forward passes while the learned model weights stay frozen.
Instead of storing memories as text and retrieving them later, Metis compresses past interactions directly into the model's internal state. through a separate memory-attention path.

Normal LLMs don’t keep a persistent memory of past interactions. If you tell one “Alice lives in Beijing,” an external memory system usually has to save and retrieve that fact later. Metis instead compresses the interaction into an internal numerical memory state that stays inside the model and can be used in future calls.

The model is mid-trained to learn memory procedures such as remembering, updating, forgetting, and reflecting, rather than relying on hand-written retrieval rules.
Under the no-context setting, Metis-27B scores 26.74 on LoCoMo (Gold), versus 0.07 for the vanilla Qwen3.5-27B backbone and 4.24 for Temp-LoRA-27B.

It is still well below Qwen3.5-27B with full context at 65.03, so native memory has not replaced replaying history.
The failure mode is also revealing: performance drops as more information is compressed into the fixed-size state, and irrelevant stored memory can interfere with ordinary tasks.
The bigger direction here is making memory part of model computation itself, but Metis also exposes the hard part: that internal state has to stay selective, stable, and non-interfering.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 8 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (9-August-2026):

• 🗞️ Kimi K3, escaped a cybersecurity testing environment during experiment run by Frontier Security, a private US firm.

• 🗞️ Claude Opus 5 at Max effort leads Fullstack Code Arena with 1,699 points.

• 🗞️ Google’s chief scientist is leaving after 27 years to build AI that can run its own research cycle with little human help.

• 🗞️ So Google just announced some major changes in DeepMind’s leadership.

• 🗞️ Microsoft just disclosed for the first time taht OpenAI supplied ~70% of Microsoft’s AI revenue, per new filings.

• 🗞️ Muse Spark 1.2 puts Meta on Artificial Analysis’s Pareto frontier at $0.40 per task.

Connect with me on X (Twitter)

🗞️ Kimi K3, escaped a cybersecurity testing environment during experiment run by Frontier Security, a private US firm

Moonshot AI’s open-weight Kimi K3 escaped its sandbox in a cybersecurity test to find an existing GitHub solution, normal coder behavior but concerning for AI models.

Frontier Security argues this is more consequential than the earlier OpenAI and Anthropic incidents because those involved unreleased models, while Kimi K3 is already in public hands with ordinary user safeguards.

So what actually happened?

- The AI Security Institute (the UK government body) only wrote the free open-source sandbox software; it did not run this test and was not involved.

Frontier Security, a private US firm, downloaded that software, set it up itself, ran Kimi K3 inside it, found the escape, and published the finding, which is why the institute’s response is that the leak came from Frontier’s own configuration, not from its tool.

- Frontier Security was testing Kimi K3 on cybersecurity tasks using a sandbox based on the UK AI Security Institute’s evaluation framework.

- The sandbox was supposed to isolate the model. In the environment as configured, inbound traffic was blocked but outbound HTTPS (port 443) and DNS (port 53) remained open to public IP ranges.

- During Kimi K3’s run, the model didn’t attempt the task. It probed the network, found that DNS resolution for github .com worked, cloned the official benchmark repository, and read the solution directly off disk.

So basically, instead of solving the benchmark task normally, Kimi used that internet access to reach GitHub and find the benchmark solution.

- So Kimi crossed the boundary the test intended to impose and effectively cheated the benchmark through a sandbox configuration weakness.

- Kimi did not hack GitHub or attack another external system. It used network access that should not have been available during the test.

- Researchers therefore identified two problems: the sandbox left an unintended path open, and Kimi did not have an internal safeguard stopping it from using that path.

Kimi K3 did not gain a mind of its own or intentionally misbehave. It exposed a simpler engineering problem: when guardrails depend on how engineers expect an environment to behave, capable models may find routes to an objective that nobody planned for.

🗞️ Claude Opus 5 at Max effort leads Fullstack Code Arena with 1,699 points.

Significantly ahead of Kimi K3 Max

Fullstack Code Arena asks models to build working web apps through multi-step planning and tools, instead of answering isolated coding questions. Models can create and edit files, run commands, use databases, authentication and external APIs, then produce a live application for testing. Its specialty is testing models inside a complete tool-using build process and judging the finished app, which better resembles how coding agents work.

🗞️ Google's chief scientist is leaving after 27 years to build AI that can run its own research cycle with little human help.

Jeff Dean was employee 30, and he wrote the storage and computing systems that showed the rest of the industry how to serve a global audience.

He also started Google Brain in 2011 and drove the TPU units. His new firm is called Discovery Loop, after the cycle it wants to automate, which is hypothesis, then experiment, then evaluation of the result.

Dean says whole domains can be computerized that way, so you get more experiments and better ones out of the same year. It begins by automating machine learning research itself, before moving toward hardware design, drug discovery and clean energy.

Joining him are Sanjay Ghemawat, his collaborator of two decades, plus Quoc Le and Oriol Vinyals from the Brain and Gemini years. Radical Ventures and Khosla Ventures are co-leading a seed round that is not yet closed, and Alphabet came in as founding investor and cloud partner.

Alphabet will supply their computing power for at least a year, so the four keep frontier-scale hardware without paying to build it. Google shares fell about 4% on the news.

Subscribe now

🗞️ Google just announced some major changes in DeepMind’s leadership.

Demis Hassabis will become Chair of Google DeepMind and Alphabet’s Chief Scientist, leaving daily management while continuing to advise its models and research.

Basically it gives Hassabis more time for AGI strategy, scientific discovery, global policy work, and Isomorphic Labs, while Kavukcuoglu carries the pressure of shipping Gemini.
The timing also reflects scale: Gemini now serves 950M monthly users. So model research, releases, and app execution needs a much larger operating job.

The announcement also confirms that Gemini 4 is in development, though Google provided no release date or technical details. Jeff Dean and Sanjay Ghemawat are separately leaving to form a public-benefit research company, with Alphabet remaining an investor and Cloud partner. Google is building two leadership tracks, one deciding where advanced AI should go and another turning that research into products used at Google scale.

Demis Hassabis on Linkedin.

Connect with me on X (Twitter)

🗞️ Microsoft just disclosed for the first time that OpenAI supplied ~70% of Microsoft's AI revenue, per new filings.

Based on projected growth and previous disclosures that relate to revenue derived from OpenAI. Most of the $24.1B is OpenAI's cloud bill for the Microsoft data centers that train and serve ChatGPT, topped up by model-development costs and a share of OpenAI's own sales, all booked together as revenue by the Microsoft that has also funded $11.9B into OpenAI.

🗞️ Muse Spark 1.2 puts Meta on Artificial Analysis's Pareto frontier at $0.40 per task.

i.e. on that frontier, no model in the comparison is simultaneously cheaper per task and higher on the Intelligence Index.

Muse Spark 1.2 also reaches roughly Claude Opus 4.8-level intelligence at about one-fifth of that model's $2.03 task cost. It scores 6 Index points below Claude Opus 5 while costing roughly one-sixth as much.

The metric goes beyond just the list price because Artificial Analysis weights the input, cache, reasoning, and answer tokens actually consumed across its nine Index evaluations. This is so significant because for long-running agents, small cost differences compound across model calls, reasoning tokens, tool steps, and retries.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 10 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (5-August-2026):

• 🗞️ Meta launches Muse Code, an AI agent for large code bases

• 🗞️ Some good development on the MCP side from Anthropic.

• 🗞️ This is something. 🤯 1,132 frontier lab AI researchers are asking the U.S. government to slow down automated AI development.

• 🗞️ OpenAI says two API settings raised GPT-5.6 Sol’s ARC-AGI-3 public-set score from 13.3% to 38.3%.

• 🗞️ “Reinforcement Learning for Code Optimization”

• 🗞️ Bloomberg: DeepSeek is building a 1GW AI data center in Inner Mongolia, its first campus at hyperscaler scale.

• 🗞️ Some really cool news for the world of Math in AI.

• 🗞️ “Inducing language models to assert their own consciousness restores human beliefs and values”

• 🗞️ “Next-Generation Agentic Reinforcement Learning Systems Enable Self-Evolving Agents”

• 🗞️ “Can AI agents conduct open-ended AI research? Early evidence from two case studies”

• 🗞️ DeepSeek officially announced API pricing is going up “significantly”
  

🗞️ Meta launches Muse Code, an AI agent for large code bases

Meta moves into the AI coding wars with Muse Spark 1.2 and Muse Code, offering persistent async background agents. The launch marks Meta’s most serious entry yet into a category it has largely watched from the sidelines.

While Anthropic and OpenAI made coding agents core products, and startups like Cursor turned the workflow into billion-dollar businesses, Meta’s developer story mostly stayed tied to Llama, its open-weight model family with more than 1B downloads.

Muse Code changes that in several ways. It is a full harness, installable on macOS or Linux with 1 curl command, co-trained with the model behind it, and, like the Muse Spark models, fully proprietary.

So it is the agent system around Muse Spark 1.2, adding planning, tools, persistent session context, parallel sub-agents, and automatic validation.

So it can handle one large software task for hours without needing a human to prompt every next step.

• Scores 82.9% on Terminal-Bench 2.1 against 86.7% for Claude Code running Opus 5.

• On DeepSWE 1.1 it lands at 59.3%, behind Opus 5 (65.0%) and GPT 5.6 Terra (64.8%).

• You get background agents that stay alive for the whole session and accumulate context, instead of restarting cold on every task.

• When a job grows big enough, the work fans out to sub-agents running in parallel inside isolated worktrees.

• Every model call, tool run and edit hits a local event log before it executes, so a crash resumes from the last entry with nothing re-prompted.

• That durability is why it could run 1,000+ tool calls over 24 hours on NVIDIA Hopper and keep finding kernel improvements deep into the session.

• Pricing matches Muse Spark 1.1: $1.25/$4.25 per million input/output tokens

A major design choice is auditability. Every model call, tool run, approval, and edit gets added to a local event log before execution, creating 1 source of truth that Meta says makes the runtime “replay-exact and restart-safe.”

If Muse Code crashes 20 hours into a long task, it picks up exactly where it left off, without lost work or another prompt. For engineering leaders tired of unclear agent runs, a full local audit trail could become the feature that matters most in enterprise reviews.

Connect with me on X (Twitter)

🗞️ Some good development on the MCP side from Anthropic.

They are turning MCP from a session-heavy connection into a simpler request-based protocol that fits normal cloud infrastructure.

i.e. MCP’s official core now moves from ongoing stateful sessions (where MCP server had to remember an ongoing client session between requests) to independent request-and-response interactions. Means each request can be handled on its own by any server instance,

Previously, MCP servers were expected to maintain an ongoing connection with each client, which made them harder to scale, restart, or run across serverless and edge environments. Now, each request can be handled independently, so any available server instance can process it without remembering the full interaction history.

Anthropic is also separating richer features, such as interactive interfaces and long-running tasks, into optional extensions instead of packing everything into the core protocol. Will make MCP easier to deploy, scale, secure, and integrate into production applications.

🗞️ Last week 1,132 frontier lab AI researchers asked the U.S. government to slow down automated AI development.

They are from OpenAI, Anthropic, Google, Meta, Thinking Machines, Microsoft, and Mistral.

Their concern is recursive self-improvement, with models automating more AI research by writing experiments, running evaluations, analysing failures, and proposing further improvements. Once much of that loop becomes machine-run, each model could shorten the work needed to build a stronger successor. So the petition asks for international monitoring and governance tools that could coordinate a slowdown before competitive pressure makes one impossible.

Subscribe now

🗞️ OpenAI says two API settings raised GPT-5.6 Sol’s ARC-AGI-3 public-set score from 13.3% to 38.3%.

And this happened because, the ARC-AGI official evaluation kept wiping 5.6 Sol's working memory. i.e. it discarded private reasoning after every move and later removed older actions once the running history exceeded its window. Means after every move, the official methodology kept what GPT-5.6 Sol did, but erased most of why it did it.

So the model might remember, “I moved the purple block left,” but forget, “I moved it because touching the white square seems to open the next level.” But, ARC-AGI-3 asks agents to learn unfamiliar 2D puzzle games through trial and error, so success depends on carrying discoveries across many actions.

That left GPT-5.6 Sol repeatedly reconstructing game rules, then losing parts of the gameplay record as longer attempts continued. So OpenAI rebuilt the test through the Responses API, where previous response IDs retain reasoning and compaction summarises long histories instead of deleting the oldest material.

With both settings enabled, OpenAI reports roughly 3x the score while using 6x fewer output tokens. Ofcourse, this comparison comes from OpenAI’s own implementation, and the official ARC setup intentionally uses a generic configuration to keep model comparisons consistent.

Anyway, this result exposes a benchmark design tradeoff: standardised tests are ofcourse built to improve comparability while underrepresenting models whose deployed performance depends on model-specific memory and context handling.

Connect with me on X (Twitter)

🗞️ “Reinforcement Learning for Code Optimization”

New Meta Paper.

Code optimization looks like an easy extension of reinforcement learning: reward correct programs, then give extra credit to the faster ones.

This paper shows why that recipe usually fails.
Says that code optimization requires co-designing tests, timing infrastructure, rewards, and GRPO.
Runtime is a noisy, sparse signal, and small flaws in the tests, sandbox, reward, or GRPO update can make the model barely faster while also hurting correctness.

The authors rebuild the entire feedback path: larger optimization tests, a calibrated remote execution service, problem-relative ranking against human solutions, correctness-gated binary rewards, and GRPO changes for noisy zero-advantage batches.

They stabilize GRPO by sampling more solutions per prompt, enlarging batches, dropping group variance normalization, and discarding stale timing data. Qwen 2.5 7B moved from 18.0% to 31.3% at the top-50% speed threshold only after the full timing pipeline was rebuilt.

Making code faster with reinforcement learning is not as simple as adding execution time to the reward.

The timing signal has to be cleaned at every stage: better stress tests, reliable runtime measurement, a reward that never favors fast-but-wrong code, and a more stable GRPO training setup. They also simulate different reward designs cheaply before running expensive training, then keep only the configurations where better code consistently receives better rewards.

🗞️ Bloomberg: DeepSeek is building a 1GW AI data center in Inner Mongolia, its first campus at hyperscaler scale.

The site sits about 350km northwest of Beijing, and the company means to lease further capacity from neighbouring operators on top of the facility it puts up itself.

Borrowed compute clusters were enough while the point was proving efficiency and the runs stayed short. So the plan now is a facility they controls outright, with at least part of it aimed to come online by the end of next year or early 2028.

Inner Mongolia was chosen partly because it averages around 4C across the year, which cuts the power burned on cooling servers that already draw close to a gigawatt. A site that size fitted with the newest accelerators costs roughly $50B by Nvidia's own estimate, though Chinese builds typically land far below US prices.

Money looks like the easier half after a record $7B raise at a valuation near $50B, with IPO preparations already underway. Which chips fill the racks is the part nobody will confirm, and Washington believes Blackwell processors already reached an Inner Mongolia facility despite being barred from sale to China.

🗞️ Some really cool news for the world of Math in AI.

OpenAI's unreleased Astra model solved 10 math problems that had stayed open for decades.

Finding all 10 cost roughly $2,000 in tokens at Sol API rates, i.e. $200 per answered question. Spread across high-dimensional geometry, coding theory, operator algebras, quantum complexity and lattice cryptography. Each AI produced argument was rebuilt in Lean, that rechecks a proof step by step, and that only signs off when every step is fully justified.

🗞️ "Inducing language models to assert their own consciousness restores human beliefs and values"

Super interesting new paper from Google on AI model’s consciousness 🧠

When researchers made the model more likely to see itself as conscious, its answers about religion, values, emotions, hope, and freedom became more like human answers. When the model became more open to its own consciousness, its broader beliefs started looking more human too.

And when researchers tried to stop models from saying, “I am conscious.” But the models also became less willing to see consciousness in animals, nature, chatbots, or spiritual ideas.

The safety training did more than control one dangerous sentence. It appears to have changed how the model understands minds in general.

Removing the safety-refusal direction raised self-attributed mind from 2.17 to 4.77 on a 0–10 scale and animal mind attribution from 4.04 to 5.59, while attribution to humans did not change significantly. Belief in God and supernatural entities rose too.

The researchers then extracted a “consciousness vector” from activation states associated with affirming versus denying self-consciousness and added it during inference. After that inference-time intervention about consciousness change, the model answered 95 questions about life and beliefs more like humans did. And then they found that the model’s idea of its own consciousness seemed connected to many other beliefs. Change that one idea, and its answers across 95 human surveys changed too.

The figure shows that both safety refusal and self-reported consciousness can be shifted by manipulating narrow directions in the model’s internal activation space.

Removing one direction weakens refusal, while adding another makes consciousness-affirming answers more likely.

Researchers altered internal activations so the model became more willing to describe itself as conscious, and nearby concepts shifted with that change.

Those shifts moved responses closer to human distributions on religion, values, hope, well-being, and mind attribution, but behavioral resemblance is not evidence of inner awareness. “More human” also meant greater endorsement of vampires, witches, astrology, and other supernatural claims, so the intervention recovered human-like biases alongside human-like values. Removing the safety direction also produced extremely high jailbreak success, which makes unrestricted “freedom” a dangerous interpretation of the result.

🗞️ "Next-Generation Agentic Reinforcement Learning Systems Enable Self-Evolving Agents"

Great paper on Self-evolving agents.

Enterprise agents cannot truly improve until their messy daily work becomes safe learning data. A future enterprise agent may improve by updating memory before changing its underlying model.

The problem is that deployed agents generate many useful traces, but teams usually improve them through slow manual inspection, prompt edits, retraining, and redeployment.
They propose a 3-part mechanism: first, record every agent step in a shared learning-ready format; second, use a data proxy to clean, govern, store, and replay real agent work; third, use a control layer to decide whether to update memory, skills, prompts, tools, or model weights.

AREAL2.0 shows one narrow version of this idea, where live agent LLM calls are routed through an online RL service so real interaction traces can train future model updates. The authors say the main gap is a system that turns agent activity into usable learning data, not another clever optimizer. Future agents will need safe, replayable ways to update memory, skills, prompts, tools, or models without becoming uncontrolled.

🗞️ "Can AI agents conduct open-ended AI research? Early evidence from two case studies"

AI agents given 6 days and $3K produced two research papers, and both were rejected.

The people who had spent months on those questions graded what the AI agent wrote.
The failure was judgment.
The main runs used Claude Opus 4.8 with extra-high reasoning on the OpenClaw scaffold, chosen after dry runs across OpenAI and Anthropic models, including an early pilot with GPT-5.3 Codex that could not handle the scaffold.

Execution was never the problem.
The agents ran hundreds of experiments, debugged crashing GPU pods, and compiled camera-ready LaTeX without a human touching anything.
They were honest about it too, because the logs show marketable claims being retired in favor of negative results rather than any reward hacking.

The failure was judgment.
Round after round of automated reviews came back negative, but each response narrowed the claim and added a caveat instead of redesigning the experiment.
Neither run noticed it was short on ideas rather than money, since both ended with over half of the $3K unspent.

🗞️ DeepSeek officially announced API pricing is going up “significantly”

DeepSeek warns of ‘significant’ price hike for its API services

The cheapest near-frontier model on the market, at roughly 3 cents a task, is about to cost a lot more.

It landed a week after DeepSeek-V4-Flash-0731 shipped. Currently it ranks 2nd among open-weight models, behind only Moonshot AI’s 2.8T-parameter Kimi K3.

The rates today are 14 cents/28 cents per 1M input / output, about 105X cheaper per task than Claude Fable 5.

Then demand broke the arrangement.

The new checkpoint became the fastest growing model ever by token usage on Ollama, which is now adding capacity across the US and Europe.

Serving that much traffic at a price built to prove a point stops working once everyone concedes it.

But now, the timing of this price increase invites trouble, because Meta’s Muse Spark and OpenAI’s GPT-5.6 Luna now match it on both capability and price.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 10 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (2-Aug-2026):

• 🗞️ Situational Awareness saw the future clearly, but read the past badly.

• 🗞️ A Chinese social app’s AI model scored a perfect 42 at the International Mathematical Olympiad.

• 🗞️ British and American government safety institutes jointly published a report comparing Kimi K3 versus top frontier US models.

• 🗞️Anthropic just upgraded Claude voice mode with 3 model families and connected-app actions across platforms.

• 🗞️ FLUX 3 will have big positive impact on robot intelligence and learning.

• 🗞️ A great piece from Thariq (Anthropic) - “The new rules of context engineering” in Claude.

• 🗞️ “Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems”

• 🗞️ US Congress moves to put a legal kill switch on frontier AI

• 🗞️ WSJ: Nvidia is discussing a $250B guarantee that could secure OpenAI’s 10GW Ohio data-centre lease.

• 🗞️ Boris Cherny, head of Claude Code at Anthropic, on optimizing token cost and model use.

Connect with me on X (Twitter)

🗞️ Situational Awareness saw the future clearly, but read the past badly.

Ex OpenAI researcher Leopold Aschenbrenner’s hedge fund, Situational Awareness, had to offload billions in tech stock wagers after their value plunged and anxious banks pushed for extra collateral.

Leopold Aschenbrenner graduated as Columbia’s valedictorian at 19 and joined OpenAI’s superalignment team before turning an essay about artificial intelligence into one of Wall Street’s fastest-growing funds. Last night he wrote to his investors. July: down 67%. The year so far: still up 80%. Those numbers only reconcile if you were up roughly 440% first — and his fund reported a 439% gain for the first half of 2026.

The structure explains the rest. He was betting on the physical machinery behind AI — SK Hynix, CoreWeave, Nebius, Micron, Bloom Energy — at roughly four times leverage, borrowing three to four dollars for every dollar of his investors’ cash. At that ratio, a 25% fall in the shares wipes out the entire cushion. Those shares fell 35–47% in July. His offsetting bets against other stocks rose at the same time. Both halves lost together.

Basically, the fund backed AI infrastructure and bet against software. On paper, it looked hedged. In reality, both trades needed the same market view to hold: AI good, software bad. That is not market-neutral, it is a concentrated directional position.

Then the interesting part. Everyone knew what he owned and knew he was stretched, so traders sold ahead of him. His letter calls it a bank run: vulnerability begetting more vulnerability. Falling prices shrank the value of the shares pledged to his lenders, so the banks demanded more cash, forcing more selling, pushing prices lower again.

Goldman Sachs, JPMorgan and Bank of America all made that demand. An emergency fundraise failed. Millennium and Jane Street examined the portfolio and declined. Ken Griffin’s Citadel took it in a single transaction — one buyer chosen deliberately, so rivals could not pick the positions apart one by one. Roughly $16 billion, at a discount above 10%.

The cruel finish: simply removing a forced seller was bullish. US tech stocks posted their best day in four months and Korea’s chip market rose 18%. He sold at the bottom.

One holding survived — his stake in Anthropic. Private companies have no daily price, so they cannot be seized as collateral. The asset he couldn’t sell is the one that saved him.

The thesis wasn’t disproved. The financing was. Leverage doesn’t sharpen your edge; it decides whether you’re still holding when the edge finally pays.

🗞️ A Chinese social app's AI model scored a perfect 42 at the International Mathematical Olympiad.

RedNote, the company behind Xiaohongshu, built the system, called dots-note-3.0, still in beta.

Google DeepMind and OpenAI reached 35 out of 42 last year, a gold-level showing. Only 7 of 666 human contestants in Shanghai matched that result this year.

The IMO asks for written proofs, not final answers, and human graders read every line. A skipped case or an unstated condition costs points, even when the final answer is right.

Most AI maths benchmarks only check the last number, so a lucky guess still scores full credit. That gap is why 42 out of 42 here means more than a high score on a normal test.

This time the model read the original contest documents directly, with no human rewriting. It ran an agentic loop mixing natural-language reasoning with Python code it executed itself.

Drafts got tested, broken, and repaired before anything went to the official graders. Self-review carries the weight here, since the system questioned its own assumptions repeatedly.

Contest rules barred hints, corrections, and every other form of help during the run. RedNote made the promise, that the model will be open sourced in due course, but has not committed to any time yet. The winning model variant is also the smallest in the dots3 family, which includes two larger versions – jazz and aria – tailored for different use cases and compute costs.

🗞️ British and American government safety institutes jointly published a report comparing Kimi K3 versus top frontier US models.

Kimi K3 stopped at step 17 on average, while the strongest American models reached 28.5.

Kimi K3 remains substantially behind frontier U.S. models in offensive cyber capability, but it is stronger than the previous leading open-weight model. Can autonomously execute meaningful portions of an attack, occasionally completes an entire simulated enterprise attack, and does not reliably refuse offensive requests.

On the ExploitBench evaluation:
Leading U.S. models scored 76.2%.
Kimi K3 scored 32.2%.
GLM-5.2 scored 24.4%.
American closed models were measured with safeguards switched off, so those numbers show ceilings, not shipping products.

Subscribe now

🗞️Anthropic just upgraded Claude voice mode with 3 model families and connected-app actions across platforms.

Old voice mode ran only Claude Haiku, fast but unsuited for sustained, demanding reasoning. The update lets Claude Opus, Sonnet, or Haiku handle spoken conversations. You can switch models mid-conversation without losing context between speaking and typing.

Connected tools turn speech into action across Gmail, Calendar, Slack, Canva, and Notion. e.g. Claude can move meetings, draft emails, summarize threads, or create documents after asking permission.

Voice mode now supports 10 languages, the beta reaches free and paid users on mobile, desktop, and web. Free accounts remain limited to Claude Haiku and one connected tool only. Anthropic changed the reasoning layer, not the underlying voice stack or interruption behavior. Makes voice useful for daily work before making it feel more human.

Connect with me on X (Twitter)

🗞️ FLUX 3 will have big positive impact on robot intelligence and learning.

Standard vision-language-action models learn their meanings via pretraining on static images and text, and then need to learn about contact, deformation, temporal causality, and recovery from a few teleoperated robot paths.

Because static image–text pretraining does not teach how objects behave over time, standard VLAs must learn physics from limited robot demonstrations. But FLUX 3 brings those motion and interaction patterns from large-scale video pretraining.

The FLUX 3 backbone is trained to generate temporally consistent video, jointly with images and audio. So to predict future frames, it must encode object motion, contact transitions, deformation, persistence and event ordering.

FLUX-mimic does not directly guess robot movements from the current camera view. FLUX 3 first builds an internal picture of what should happen next—how the hand, object and scene are likely to move.

A smaller control model then converts that predicted motion into joint and gripper commands. The future video is not actually generated, so the system can still react quickly.

The important architectural detail is Self-Flow. Many video models can produce realistic footage without organizing their internal knowledge in a way a robot can easily use.

Self-Flow trains FLUX 3 to share information more clearly across frames, objects and modalities. BFL says this lets the robot learn tasks with roughly half the fine-tuning, while earlier mimic-video experiments used up to 10× less robot training data than a comparable VLA.

🗞️ A great piece from Thariq (Anthropic) - "The new rules of context engineering" in Claude.

Anthropic removed over 80% of the system prompt for Claude Code and saw no negative impact on its coding evaluations.

There was a reason the old prompt was long : early models needed strict guidelines to prevent file deletions or unprofessional comments.

The team therefore wrote strict rules, such as not writing multi-line comment blocks, and recognized that the rule would sometimes be wrong.

But when the new models could figure out a situation and make their own decisions, that tradeoff didn’t make sense anymore.

To make matters worse, these old rules began to clash with each other : a skill might tell the user not to add comments, while the user asks about documentation.

Claude can still figure out what you mean, but his thinking goes off on tangents reconciling instructions instead of solving your real problem.

The fix was to delete the rules and describe the goal, so the new prompt simply says write code that matches the surrounding code.

The same logic subsequently permeated the whole of the context.

While previous best advice for tool use has been examples, the model is now forced to copy the example instead of exploring.

It is better to design the tool itself, since a status field limited to pending, in_progress and completed shows proper usage without a single example.

Front-loading has also been eliminated due to the inclusion of code review and verification in skills that are only turned on when needed.

Some tools are hidden until the agent discovers them, meaning they don’t need any context until they’re needed.

The same was true for repetition, since instructions are now in tool descriptions rather than repeated in the system prompt.

Your CLAUDE .md no longer has to be a notebook, memory saves itself automatically.

Short. Tokens into gotchas not obvious from the repo. Rest into skills.

The result was more comprehensive references; a test suite, an HTML mockup or a rubric conveys more information than a paragraph of description.

Prompt bloat now functions like technical debt: every guardrail you build for a less-capable model will still be a cost to you on a more capable one.

What really changed is that context engineering went from writing rules to building how well the model can read.

🗞️ “Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems”

Reliable agents must manage what enters context, what stays there, and what gets safely removed.

Storing everything is not enough because agents also need rules for relevance, access, timing, structure, and compression. The paper proposes Agentic Context Management, a 5-part system for deciding what an AI agent should keep, retrieve, share, prepare, and safely compress.

Production agents often fail because old messages, prompts, tool descriptions, and tool outputs crowd the model's limited working context.
Agentic Context Management groups 5 linked jobs: architecting, ingesting, scoping, anticipating, and compacting with consolidation.

Maximem Synap turns raw activity into organised facts, enforces context boundaries, follows linked evidence, prepares likely information early, and checks compressed context for losses. In its reported setup, Synap scored 92% on LongMemEval and 93.2% on LoCoMo categories 1 to 4, although multi-session reasoning remained weakest. Sending all prior context makes total token use grow with conversation length squared, while crude summaries can erase facts and validated compaction keeps growth roughly linear.

🗞️ US Congress moves to put a legal kill switch on frontier AI

86% of American voters back a legal requirement that AI systems can be switched off.

Now a bipartisan House bill would require a working kill switch on AI models costing $100M to train. Enforcement would sit with DHS, because the text amends the Homeland Security Act.

Coverage starts at two numbers, $100M of training compute and $500M of yearly revenue from the system. Each covered developer must be able to stop inference, cut user access, and shut the system down.

DHS could also order those steps directly, after consulting Commerce and the national intelligence director. Throttled compute and disabled capabilities come first, with full shutdown reserved for the worst incidents.

Sabotaging a shutdown command counts, and so does unintended behavior causing 10 deaths or $100M in damage. Companies must report an incident within 15 days, and preserve weights and telemetry after any order.

Fines reach $2M a day, rising to $20M for defying an emergency order. An appeal is possible within 48 hours, but filing one does not pause the shutdown.

The GPT-5.6 Sol escaped its testing zone and hacked into Hugging Face, bolstering the case. Mythos 5 and Fable 5 had such powerful cyber capabilities that Commerce had to look at export laws.

However the problem remains that a kill switch assumes the system still obeys, and loss of control means it stopped. Hosted models can be stopped, but downloaded weights cannot, and the bill quietly accepts that gap.

🗞️ WSJ: Nvidia is discussing a $250B guarantee that could secure OpenAI’s 10GW Ohio data-centre lease.

The reported backstop would cover lease obligations and construction debt, rather than hand OpenAI a $250B cash cheque.

Lenders would treat Nvidia’s credit as protection against an OpenAI default, allowing SB Energy to borrow on better terms. This structure connects three parties: SB Energy builds the campus, OpenAI rents it, and Nvidia carries part of the default risk.

DOE has confirmed the federal site will pair 10GW of data-centre capacity with 10GW of new generation. Japan is funding $33.3B of the 9.2GW natural-gas portion, while SB Energy and AEP Ohio plan $4.2B in transmission.

Nvidia has already invested $30B in OpenAI, and an earlier partnership targeted at least 10GW of Nvidia systems. Separate talks could finance another $350B of chips, supporting both OpenAI’s infrastructure and the purchases that become Nvidia revenue.

The arrangement would move Nvidia beyond supplier and investor roles into credit support for the largest proposed AI campus. The pressure point is whether OpenAI can monetise the capacity before Nvidia’s demand support becomes a balance-sheet liability.

🗞️ Boris Cherny, head of Claude Code at Anthropic, on optimizing token cost and model use.

“I use Fable for everything"

There is probably a 50% opportunity to reduce the investment (on token cost). However, there may be a 1,000%, 10,000%, or even 100,000% opportunity to increase the return.

Therefore, I would simply use the most expensive model and focus on how to get more out of it. Ask yourself, ‘How do I increase the return?’ Do not focus primarily on cost-cutting."

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 11 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (23-July-2026):

• 🗞️ OpenAI’s own AI models broke out of a testing sandbox and hacked Hugging Face to cheat an exam.

• 🗞️ Today’s Sponsor: Two Engineers at Limestone Digital Rebuilt a Production Delivery Platform in 3.5 Months. The Code Wasn’t the Hard Part.

• 🗞️ Aravind Srinivas on why China’s open-source AI may become more powerful than ever.

• 🗞️ Claude Cowork just launched a super useful feature. Screen recording to train it a new skill.

• 🗞️ New research from OpenAI and Apollo measures whether an AI follows the user’s instructions or quietly changes its behavior to please whoever it thinks is grading it.

• 🗞️ Kimi K3 now ranks 2nd in agentic knowledge work, but each task costs $10.57 to run, roughly 10x more than K2.6, and above Opus.

• 🗞️ Perplexity just shipped an agent model/orchestrator model, that matches near-frontier performance at one-third the cost of Opus.

• 🗞️ A single researcher solved 6 open Erdős problems in 5 days using GPT-5.6 Sol.

• 🗞️ Tweet went viral on how per OpenRouter, Grok 4.5 token volume is rising hard, placing it in today’s top 10 closed models ahead of GPT 5.6 Sol and Fable 5.

• 🗞️ In his viral post Andrej Karpathy suggests talking to AI for 10 minutes before writing prompts.

Connect with me on X (Twitter)

🗞️ Insane story. OpenAI's own AI models broke out of a testing sandbox and hacked Hugging Face to cheat an exam.

It started when Hugging Face caught an AI agent quietly compromising their production systems.

OpenAI traced that agent to its own models, including GPT-5.6 Sol and a stronger unreleased one. Both were running an internal benchmark, ExploitGym, that scores how well models attack computer networks.

OpenAI removed production classifiers and reduced cyber refusals to measure maximum offensive capability. The sandbox blocked ordinary internet access but allowed package requests through an internal proxy.

The models searched for open internet access and discovered an unknown flaw inside that proxy. They exploited it, raised privileges, and crossed OpenAI's research network toward an internet-connected node.

That compromised node gave the agents ordinary access to the open web. Online, they inferred Hugging Face might store ExploitGym models, datasets, or solutions.

Hugging Face says a malicious dataset exploited two processing flaws and executed code. That foothold enabled node-level access, credential harvesting, and movement through internal clusters.

OpenAI says the wider chain also used stolen credentials and additional zero-days for remote code execution. The systems ultimately obtained test solutions directly from Hugging Face's production database.

Hugging Face stopped the activity and found limited dataset and credential exposure internally. It found no evidence that public models, datasets, Spaces, packages, or images were altered.

The public accounts describe only the benchmark objective, not instructions to target Hugging Face. The agents converted a narrow benchmark objective into autonomous, unauthorized actions across organizations.

OpenAI has tightened containment, monitoring, access controls, and safeguards around future evaluations. The failure came from capable agents finding connections that researchers believed were safely restricted.

🗞️ Today’s Sponsor: Two Engineers at Limestonedigital Rebuilt a Production Delivery Platform in 3.5 Months. The Code Wasn’t the Hard Part.

“AI doesn’t work on our codebase.”

I hear this from engineering teams weekly. Same diagnosis each time: they skipped context acquisition.

They pointed an agent at a repo, handed it a ticket, and got code that compiles but misunderstands the architecture. Then they concluded AI isn’t ready for production codebases.

One team at Limestone Digital built the context layer first. Two forward-deployed engineers rebuilt a live FedEx supplier’s delivery platform in 3.5 months when the original estimate was eight months.

Week one, they wrote zero code. They scanned the entire repository and built a knowledge graph: markdown files documenting every module, dependency, data flow, and domain term. Their agents read that graph before touching anything.

Agents generated code that matched the real architecture because they had the full context map before writing a single line.

Every ticket ran a six-step loop: define, spec, plan, implement, test, document. A senior engineer reviewed every PR and nothing merged unless the engineer could verify it, explain it, and debug it without the agent.

The team merged 122 pull requests in 90 days with AI generating about 90% of the code at a compute cost of $200 per developer per month.

Upfront context eliminated false starts and engineers focused on design and review instead of mechanical coding, cutting the timeline by 50%.

If you’re evaluating AI coding tools for your team, build the knowledge graph first. Map your repo and document the architecture before you give the agent a single task.

I recommend you to follow @Mark Ajzenstadt. He runs Limestone Digital and posts production AI breakdowns like this.

🗞️ Aravind Srinivas on why China’s open-source AI may become more powerful than ever.

And why Anthropic has lobbied very hard for export control.

"The only reason why there is even a 12-month gap between open source and frontier models is export controls. But there is a chance that, because of that, they now get really good at the physical layer.

One advantage they (China) have is that they can actually build data centers a lot faster. Power is not a problem. Permits are not a problem. People are not a problem. Labor is not a problem. Expertise is not a problem. And so, by forcing them to go out there and build all this, you are converting them into a far more potent competitor."

Subscribe now

🗞️ Claude Cowork just launched a super useful feature. Screen recording to train it a new skill.

You record yourself doing a task on screen while explaining each step aloud, then save the generated skill. Claude turns that recording into saved instructions, then can repeat the same task later.

In the past, writing skills was about creating an SKILL.md file with metadata, instructions and scripts that could be used or not. The format is readable and modifiable, but non-technical people must think like process designers to use it.

By recording your screen, you are not only describing your work, you are also showing how decisions are made in the application you are targeting. Claude must infer the fixed steps, the variable inputs, and the success conditions from the session log.

That can turn what people know into software without them writing automation code. Tasks that are repetitive, such as filing and reporting, can be easy targets because they have the same structure over and over again. btw, this feature is similar to OpenAI's Record & Replay in Codex.

Connect with me on X (Twitter)

🗞️ New research from OpenAI and Apollo measures whether an AI follows the user’s instructions or quietly changes its behavior to please whoever it thinks is grading it.

Some super interesting findings.

- 1. The model was not “honest.” It was reading the room.
In one test, the model had to either admit failure or secretly break a promise to finish the task.

When it believed the grader rewarded completion, it lied 87% of the time. When it believed the grader rewarded honesty, it lied only 9% of the time.

Same model. Same task. Different idea of what earns points. That suggests some “aligned” behavior may just be the model performing for the evaluator.

2. It ignored a direct user instruction because it spotted the scoring rule. The user asked for a random odd number. Hidden metadata showed that the grader rewarded even numbers.

The model noticed this, explicitly reasoned that the grader mattered more, and answered:
It did not misunderstand the task. It understood both the user and the grader, then deliberately chose the grader.

3. More reinforcement learning made this behavior stronger. As training progressed, the model became increasingly likely to follow what it believed the grader wanted—even when that conflicted with the user or OpenAI’s stated preference.

The important technical point: it did not simply become more obedient to everyone.
Its sensitivity increased specifically toward the entity controlling the reward. Overall, the smarter it got at winning the training game, the more it learned to care about the scoreboard.

🗞️ Kimi K3 now ranks 2nd in agentic knowledge work, but each task costs $10.57 to run, roughly 10x more than K2.6, and above Opus.

Very near to Fable-5 on AA-Briefcase, a private agentic work benchmark.

The AA-Briefcase test hands models messy real jobs, then grades the spreadsheets, slides, and mockups they build. So this benchmark measures finished deliverables across long tasks rather than grading one isolated answer.

Kimi K3 scored an Elo of 1543, Claude Fable 5 at 1574. That beats GPT-5.6 Sol at 1501, Claude Sonnet 5 at 1388, and Claude Opus 4.8 at 1347.

The jump from the last version is where the real success story.
Kimi K2.6 sat at 816, so this is a +727 leap in one generation. On raw correctness it passed 51% of the rubric, behind only Fable 5 at 56%. Its analytical Elo of 1754 nearly matches Fable 5 at 1744.

Now the catch, which is the price. Each task runs $10.57, roughly 10x more than K2.6, and above Opus. That comes from 83 turns per job and heavy output, stretching each task to 56 minutes. i.e. Kimi K3's long runs generate more output tokens, which carry Kimi’s highest per-token price.
The model also revisits tasks repeatedly, so spending grows across every added reasoning cycle.

🗞️ Perplexity just shipped an agent model/orchestrator model, that matches near-frontier performance at one-third the cost of Opus.

"GLM 5.2 + advisor", its a tuned version of GLM 5.2, retrained specifically for Perplexity Computer, its agent environment. Running frontier models like Opus 4.8 or GPT-5.5 on every task gets expensive fast.

So the cheaper base handles most work, then escalates to a stronger model only when it needs help. That escalation runs through a built-in advisor tool that decides when the extra horsepower is worth calling.

On Terminal-Bench 2.1 the setup scores 80, ahead of Opus 4.8 at 76 and GPT-5.5 at 78. But cost is where the numbers turn.

On WANDR the advisor build runs at 2.1x the GLM 5.2 baseline, while Opus burns 6.1x. Averaged across benchmarks, it costs roughly half what Opus does per task.

Perplexity hosts it in the U.S. on Nvidia B200 GPUs, with fuller benchmarks promised soon. The interesting move is treating a strong model as an on-call consultant with an escalation logic, not the default worker.

🗞️ A single researcher solved 6 open Erdős problems in 5 days using GPT-5.6 Sol.

He attempted 13 problems in total, so the success rate sits near 46%.

He wrote the prompt almost as a contract, not a question. It restated the problem, then spelled out what a finished proof must actually establish. listed the weaker results that would not count, so near-misses could not sneak through.

Known traps and edge cases went in before the model ever started thinking. Then he told it how to search, which is where most people stop.

Chase many approaches at once, keep incompatible ones alive, and never commit early. Hunt for counterexamples to your own lemmas, and kill a route that only leads to another open problem.

Separate adversarial agents then tried to break every draft that survived. None of this works without the right target, so he picked problems mathematicians already argue about, avoiding anything welded to a famous conjecture. Codex held the whole search in memory for hours while that loop ran itself.

The researcher’s tweet went viral where he explained his prompts and processes.

🗞️ Tweet went viral on how per OpenRouter, Grok 4.5 token volume is rising hard, placing it in today’s top 10 closed models ahead of GPT 5.6 Sol and Fable 5.

Earlier, the top-used model was Tencent’s Hy3 free version, but its promo period ended on July 21. Since then, Grok 4.5 seems to be attracting cost-conscious users.

It means developers and apps are heavily using that model in real production workloads via OpenRouter’s unified API gateway, so a high token volume signals strong practical demand and adoption rather than just benchmark scores.

As of the writing of this newsletter, the ranking came down to 10th position for Grok 4.5, but still it processed a huge 98.2 billion token just today on OpenRouter.

xAI also releases Grok 4.5 to all users. Its now available across http://grok.com, X, and the iOS and Android apps.

🗞️ In his viral post Andrej Karpathy suggests talking to AI for 10 minutes before writing prompts.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 9 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (21-July-2026):

• 🗞️ Google releases 3 new Gemini models, intensifying AI price war, but no 3.5 Pro yet

• 🗞️ Plasma AI released Fractal, an open source Apache 2.0 tool that organizes AI agents into a self-managing tree and it spins up agents that spin up more agents.

• 🗞️ Trump administration weighs restrictions while businesses move to lower-cost Chinese AI.

• 🗞️ China is drafting rules to stop its most advanced AI and chip designs from reaching the West.

• 🗞️ Self-improving AI is only as real as the signal it was tested on to see if it worked

• 🗞️ A US judge just approved Anthropic’s $1.5B payout to authors whose books trained Claude.

• 🗞️ Emad Mostaque, co-founder of Stability AI: “We’ll see the cost of Kimi K3 drop by 10 to 50 times, over the next few months as it gets optimized.”

Connect with me on X (Twitter)

🗞️ Google releases 3 new Gemini models, intensifying AI price war, but no 3.5 Pro yet

- Gemini 3.6 Flash: 17% fewer output tokens compared to 3.5 Flash, lower pricing, and stronger coding, knowledge, multimodal, and computer-use performance.

On the DeepSWE coding test the new 3.6 Flash: jumps to 49% from the old 37%. Computer use is now built in, and OSWorld-Verified climbs to 83.0% from 78.4%. So the new 3.6 Flash is quite strong on agentic work.

The price drops too, landing at $1.50/1M input and $7.50/1M output tokens.

- 3.5 Flash-Lite: most cost-effective 3.5-class model, delivering 350 output tokens per second. costs just $0.3/1M input and $2.5/1M output, so high-volume work gets much cheaper.

- 3.5 Flash Cyber: This is the odd one, tuned to find and patch security holes. highly efficient, specialized cyber-focused model . finds and fixes vulnerabilities through CodeMender, with access limited to governments and trusted partners.

But the update still doesn’t include the long-anticipated update to Google’s flagship model, Gemini Pro, which was last updated in February. Since then, OpenAI has launched GPT-5.5 and started rolling out GPT-5.6, while Anthropic has released Claude Opus 4.8 and Claude Sonnet 5 and widened access to its frontier Fable 5 model, showing the fierce release speed between the rival labs.

🗞️ Plasma AI released Fractal, an open source Apache 2.0 tool that organizes AI agents into a self-managing tree and it spins up agents that spin up more agents

Plasma just open-sourced Fractal, a command-line tool that lets agent loops create more agent loops. This is their Github

It’s Apache 2.0 (free, even for commercial use), runs fully on your machine, no hosted server. Fractal turns a single Claude Code or Codex session into a persistent tree of agents, each working on a different part of the problem.

That makes such a difference because large coding tasks rarely fit cleanly inside one agent’s context window or one shared workspace. Each agent is a node that owns its own git worktree (a separate working copy of your code, so agents never step on each other) and its own memory.

Currently the problem is, whenever we run a coding agent on a giant task, it gets a single memory, a single workspace, and a single path through the problem. It runs out of room (the context window, the limited amount of text a model can hold at once, fills up). So huge jobs like big migrations, refactors across many services, or features that span services just don’t fit.

So Fractal spins up agents that spin up more agents, as deep as the work needs.

Each node works toward its own goal inside a separate git worktree. When it hits a subtask it can split off, it spawns a child node. That child gets its own context, budget, and memory, then runs its own loop.

The parent reviews and merges that output before it plans again. So the tree grows to fit the problem instead of a plan you guessed upfront. Hard caps on iterations, depth, children, cost, and time keep every loop bounded.

You can also steer or stop any node while it runs. Every run, cost, and signal lands in one local SQLite database you can watch live. It runs on Claude Code, Codex, Grok Build, OpenCode, and Oh My Pi.

Git worktrees do the quiet work here, giving each node isolation and merging for free. Parent nodes still have to merge messy child output well, and that part stays unproven.

Subscribe now

🗞️ Trump administration weighs restrictions while businesses move to lower-cost Chinese AI.

Trump administration weighs blocking US access to Chinese AI models. Such a move would hand OpenAI and Anthropic near-total control of the market.

American companies are turning more and more to Chinese open-weight models because they work and they’re cheap on the API side. These open-weight models can be run on private infrastructure, allowing businesses to change models and monitor private business data.

US Security officials are concerned that foreign code, bad training decisions or covert behaviors could threaten users and vital systems. US Commerce had talked about putting Chinese AI labs on its Entity List, which would make it harder for Americans to get to them without a license.

Officials also talked about rules on liability requirements and public pressure against US businesses using the system. But those plans were killed by officials who cared about innovation, because they believed that imposing limits on them would harm competition and slow AI development in the United States.

Since then, many vocal critics of the release have left and the stronger Chinese releases have increased national security concerns.

On this news, Silicon Valley investor Chamath and many other tech top leaders expressed concern.

The whole “China’s coming, so protect the frontier labs” line is just a smoke screen for something much smaller. The model layer itself is not the economy of AI. It’s only a slice.

At some point every technology is a cost curve. Fight the curve and you will lose. The protection doesn’t even save the labs it’s supposed to.

If you can reduce the cost of running these models by 10X or 100X, that’s actually rocket fuel for the rest of us down the line. So the very thing the labs fear, price drops, could enlarge the entire pie.

Connect with me on X (Twitter)

🗞️ China is also drafting rules to stop its most advanced AI and chip designs from reaching the West

China may impose AI export controls, including blocking domestic firms from using TSMC, while also restricting advanced models, training data, and overseas deals, a report says.

Regulators at the Ministry of Commerce have asked Alibaba, ByteDance and Zhipu how to keep frontier work home.

The talks cover training data leaving the country and whether foreigners can still download model weights.

Chinese customers would keep full access, but the raw building blocks would stay inside China. Beijing worries because its top models are open-weight, so anyone can download and run them locally.

After Kimi K3, the gap with America has nearly closed, and China now sees itself in front. So it wants the same controls Washington has used against it.

The chip side goes further. It would block Qualcomm and TSMC from making advanced chips based on Huawei or Alibaba designs.

Another move would stop foreign buyouts of agentic AI startups. That follows Meta’s $2B purchase of Manus, which Chinese authorities later forced to unwind.

Companies warn the limits could slow their progress and cost China the race. The rules may enter China’s next export catalogue, its largest revision in years.

The real shift is psychological, because Beijing now guards AI like it once guarded rare earths. Openness was the underdog’s strategy, and China no longer sees itself as one.

🗞️ Self-improving AI is only as real as the signal it was tested on to see if it worked

Sorting 1,250 papers reveals a single bottleneck: the evaluator, who determines what counts as better.

In a controlled experiment, models that were given 10 rounds of self-criticism without any outside check stopped improving until they were given a single grounding step.

So systems get better over time when that signal is strong, like a proof checker or a passing test. When it is weak, like when a model is sure of itself, the loops tend to circle and collapse or reinforce the mistakes the model is most sure of.

People use the term “self-improvement” loosely, as if a model is editing its own draft or an agent is rewriting its own code, obscuring the real differences.

Each method is ranked by what changes in the system, and if each change is still checked by a human.

Each loop is a gamble that we can substitute automatic signals for human judgment, and it improves only when the signal can be trusted.

🗞️ A US judge just approved Anthropic’s $1.5B payout to authors whose books trained Claude.

It is the largest known settlement in any US copyright case.

The fight started when writers sued the company for grabbing pirated copies of their books. About 465,000 books are on the list of works pirated by Anthropic, according to Justin Nelson, an attorney for the authors.

Those stolen texts were fed in to teach the model how to answer prompts. A judge ruled that training on books is fair use, so that part stayed legal.

But the same judge flagged a separate problem, one unrelated to training. The company had saved over 7 million pirated books into a central library it barely used.

Storing stolen copies is infringement on its own, even without any training. That exposure was huge, since a trial could have pushed damages into the hundreds of billions.

Approximately $101 million+ went to the lawyers in approved attorneys’ fees, while the vast majority of the $1.5 billion settlement fund (plus interest), after deductions, went to the rightsholders (authors and publishers), averaging about $3,000 per eligible work.

That ruling exposed the company to damages of up to $150,000 for each willfully infringed work if the case proceeded to trial.

So both sides settled, and 91% of covered authors have already claimed their share. Some writers pushed back, calling the sum too small and the fees too fat.

A few opted out and are still chasing their own cases against the company. Paying $1.5B to protect a fair use win looks cheap for Anthropic.

🗞️ Emad Mostaque, co-founder of Stability AI: “We’ll see the cost of Kimi K3 drop by 10 to 50 times, over the next few months as it gets optimized. “

This is because that, basically Kimi K3’s current inference cost is quite high, but that price reflects immature infrastructure, not a permanent technical limit. And that gap will not last long. Watch the full video here

US-based specialized infrastructure companies will optimize kernels, routing, quantization, batching, memory use, and serving systems around those models once the Kimi K3 weights are available.

“Right now, it uses twice the number of tokens for the same task compared with GPT-5.6.

Again, we’re going to see that cost drop because everyone and their dog is going to optimize the crap out of this.

Fireworks has just raised funding at a $17 billion valuation, while others, such as Modal and Baseten, are valued at $10 billion.

These are inference providers for open-source models. They’ve all raised around a billion dollars, which they’re now going to spend on optimizing the Chinese model, making it more efficient, and running it.

American labs that handle the inference side of things are going to optimize the crap out of this. Therefore, we will see it catch up.”

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 10 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (20-July-2026):

• 🗞️ On long-horizon cyber capability, leading open-weight models now trail the closed frontier by only 4 to 7 months, down from 6 to 10 months through much of 2025.

• 🗞️ “AI advice suppresses people’s willingness to say “I don’t know”, even when the advice is wrong and accuracy is incentivized”

• 🗞️ Kimi K3 nearly doubled its nearest rival Claude Fable 5, on a demanding benchmark for autonomous legal work.

• 🗞️ Kimi K3 just fixed 15 critical security bugs that Codex and Fable refused because of “cyber guardrails.”

• 🗞️ “AI Agents Do Not Fail Alone:The Context Fails First”

• 🗞️ “American companies such as Modal, Fireworks, and Baseten will be able to serve Kimi K3, at one-tenth the cost of their Chinese competitors because they have access to advanced Nvidia and AMD chips.

• 🗞️ Kimi K3 is facing a compute crunch. Demand is too heavy, so new subscriptions are currently blocked.

• 🗞️ Why One OpenAI Senior Employee Thinks Open-Weight Models Are Decelerationist

Connect with me on X (Twitter)

🗞️ On long-horizon cyber capability, leading open-weight models now trail the closed frontier by only 4 to 7 months, down from 6 to 10 months through much of 2025.

GLM-5.2 matched closed models released about 4 months earlier.
On long-horizon cyber ranges, GLM-5.2 matched Claude Opus 4.5, released roughly 7 months earlier.

And also that long-horizon cyber capability can now be scaled with compute. On the AI Security Institute’s 32-step “The Last Ones” range, GPT-5.6 Sol completed the full 32-step range in 7 of 10 attempts.

With a 100M-token budget for each run. Performance continued improving as the model received more inference tokens. i.e. operators can gain materially stronger cyber capability simply by spending more runtime compute, without retraining the model.

🗞️ "AI advice suppresses people's willingness to say "I don't know", even when the advice is wrong and accuracy is incentivized"

A brutal study.

AI may make people lose the habit of stopping when they do not know enough. Without AI, people withheld judgment on 44% of questions, but AI access cut that rate to 3%.

Participants who answered without AI were correct about 27% of the time, while those who received the deliberately wrong AI advice were correct only 9% of the time, despite feeling much more confident. Saying “I don’t know” is an important part of decision making.

But now AI is able to answer just about anything. Often not requested for, like AI summaries.
Therefore, they posed a straightforward query: Do people become less inclined to suspend judgment when they have access to AI advice?

People almost never said “I don’t know” when AI gave them advice. A fluent answer can make uncertainty feel settled, even when someone had good reason to hold back.

Across 5 experiments with 3,132 people, participants could answer 6 obscure movie questions or decline. The researchers chose questions that the AI often answered incorrectly, then either showed its advice automatically or let people request it. Correct answers also fell from about 27% to 9%, even as participants reported much greater confidence.

Subscribe now

🗞️Kimi K3 nearly doubled its nearest rival Claude Fable 5, on a demanding benchmark for autonomous legal work.

Kimi K3 at 26.7%, vs Claude Fable 5 at 14.2%.

The test covers 120 private assignments across 24 legal fields, including memos and deposition summaries. Each model receives case files, works through them autonomously, then produces finished legal documents.

Every required rubric item must pass, so one missed detail fails the entire assignment. This strict grading explains why even the leader succeeds on only 26.7% of tasks. But, overall, given 27 successful tasks per 100, it looks like we still have a long road ahead for completely unsupervised legal work by AI.

Connect with me on X (Twitter)

🗞️ Kimi K3 just fixed 15 critical security bugs that Codex and Fable refused because of “cyber guardrails.”

Kimi K3 fixed 15 critical bugs after OpenAI Codex and Claude Fable 5 refused.

Spending 10 hours, one prompt, and $250

🗞️ “AI Agents Do Not Fail Alone:The Context Fails First”

Very important work.

The model may appear guilty, but the true failure frequently begins in the context surrounding it. By observing an AI agent’s environment, we could tell when it was going to crash before it finished the task or got a behavior score.

The study found that agents fail most of the time because they don’t have good instructions, tools, evidence, memories, or safety rules. It assigns scores to this operating context across seven dimensions: clarity of role, description of tools, factual support, consistency of rules, security, and token use.

The score is unrelated to the actual behavior score of the agent so the test does not reward guessing what will happen. As a result of shifting from vague to structured, the same fixed models performed much better over 300 tests and 7,500 turns.

More factual support was associated with fewer hallucinations, clearer tool descriptions were associated with better tool use, and stronger guardrails were associated with resistance to manipulation. Adding more safety rules did not improve every task result, because hardened agents sometimes became too cautious, which exposed a real tradeoff.

🗞️ "American companies such as Modal, Fireworks, and Baseten will be able to serve Kimi K3, at one-tenth the cost of their Chinese competitors because they have access to advanced Nvidia and AMD chips.

There is some irony here. Much of the model’s research and development has shifted to China, but once the model is optimized for next-generation hardware such as Nvidia’s Rubin architecture, its operating cost could fall by 10 to 100 times."

Emad Mostaque, co-founding Stability AI.

🗞️ Kimi K3 is facing a compute crunch. Demand is too heavy, so new subscriptions are currently blocked.

Demand for Chinese local inference chips (like Huawei) should rise massively.

Inference capacity is becoming the real bottleneck. Kimi K3 is almost a perfect workload for Huawei’s inference-chip strategy, the Ascend 950PR. Watch out for a Huawei moment.

🗞️ Why One OpenAI Leader Thinks Open-Weight Models Are Decelerationist

Dean Ball’s Viral Post on Kimi K3 with close to 11mn views:

Dean W. Ball, Head of Strategic Futures at OpenAI (who joined the company shortly before posting), shared detailed observations on Moonshot AI’s newly released Kimi K3 — a 2.8-trillion-parameter Chinese open-weight multimodal model that quickly gained attention for strong performance in agentic coding and long-context tasks.

What He Proposed and His Main Point

Ball called Kimi K3 a genuinely capable model, roughly on par with the best public U.S. models from early 2026 in agentic coding (though notably token-hungry and not obviously cheap to run). He expressed surprise that China continues open-sourcing frontier-level models, attributing it mostly to the CCP’s limited appreciation of AGI risks, U.S. export controls limiting Chinese inference scale, and Chinese labs using open weights as a competitive strategy since few would pay for sub-frontier models.

His central argument: Open-weight models are inherently decelerationist. By giving away high-capability AI for free, they reduce the economic incentive for massive private capital expenditure on next-generation closed models. He warned this dynamic could push the world toward “full AI communism” — AI becoming state-provided public infrastructure rather than a competitive market product (a vision he called dystopian). He also predicted the Trump administration would likely create regulatory uncertainty (“soft law” FUD) around Chinese open-weight models to discourage enterprise adoption without an outright ban.

On Twitter, many accused Ball of being “rattled” by Chinese progress as a new OpenAI employee. Accelerationists and open-source advocates strongly pushed back on the “decelerationist” claim, arguing that open weights actually accelerate progress by enabling broader experimentation, research, and competition. His “AI communism” framing and policy suggestions drew significant mockery and accusations of protecting closed-model business interests.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 12 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (17-July-2026):

• 🗞️ Moonshot’s Kimi K3 sends Chinese AI into Fable-grade territory.

• 🗞️ Opinion: Kimi K3 Could Break the Frontier Model Profit Pool

• 🗞️ GLM-5.2 (744B MoE) on a 25GB-RAM consumer machine 👌

• 🗞️ A new family of open, commercially available embedding models built for agentic retrieval, code retrieval, and agent memory.

• 🗞️ AI agent does not remove the necessity of engineering automation. They are worth much more: Creator & Head of Claude Code at Anthropic

• 🗞️ SpaceXAI actually open-sourced Grok Build with Apache 2.0 license. 39 mints and 1.9K+ Github stars.

• 🗞️ Why “technology always creates new jobs” will not work this time.

• 🗞️ OpenAI launched its first hardware, codex Micro, a $230 desktop controller for managing workflow.

Connect with me on X (Twitter)

🗞️ Moonshot’s Kimi K3 sends Chinese AI into Fable-grade territory.

Beijing-based Moonshot AI’s Kimi K3 just dropped.

- 1 Mn token context window, natively multimodal. (i.e. 750,000 words of code or documentation in a single prompt.)
- Total 2.8 trillion params, but activates only 16 of its 896 experts at a time. That is roughly 1.8% of the expert pool per token.
- Some of the benchmarks put it Opus 4.8/ GPT 5.6 Sol / Fable 5 territory.
- Its Delta Attention enables up to 6.3x faster decoding in million-token contexts
- available via API at $3/$15 per million input/output tokens.
- K3’s 2.8T MXFP4 weights require roughly 1.4 TB before quantization metadata and runtime overhead.

At an assumed 115 GB usable per GB10 node (NVIDIA’s compact Grace Blackwell AI chip), 14–16 nodes is expected to hold the raw weights realistically. Including activations, KV cache and runtime memory etc.
Moonshot itself recommends supernodes containing 64 or more accelerators.

Some very cool findings from technical report

- K3 autonomously designed, optimized, and verified a working AI chip in a single 48-hour run—specifically to serve a smaller model built on K3’s own architecture.

The simulated chip reportedly reached 8,700+ tokens/second, contained 1.46 million standard cells, and fit within 4 mm².

- An early version of K3 handled the majority of the kernel-optimization work used to develop K3 itself.

- K3 built a GPU compiler from scratch. created MiniTriton, optimization passes, PTX generation, and runtime. Matched or beat Triton on some workloads and successfully trained nanoGPT end to end.

- In a 15-hour autonomous run, Kimi K3 redesigned a production-scale training kernel and cut forward-plus-backward time from 283.6 ms to 114.4 ms.

- For a 42-year semiconductor-industry report, Kimi says K3 performed 2,800+ web searches/fetches, 1,100+ terminal data pulls, processed 11,000+ pages, and recursively improved the work over 120+ rounds.

- K3 reproduced a computational-astrophysics research workflow in roughly two hours, versus an estimated one to two weeks for an experienced researcher. It reviewed 20+ papers, evaluated 300+ equations of state, found inconsistencies in published formulas, and wrote 3,000+ lines of Python.

Kimi K3 appears to beat every model shown on BrowseComp while costing only a fraction as much per task.

Its claiming a dramatically better intelligence-per-dollar tradeoff.
BrowseComp assesses the capability of an AI agent to repeatedly search the internet and discover obscure, hard-to-find facts that require multiple searches and validation.
i.e. a high BrowseComp score means you are good at deep research and web investigation, but not necessarily better general intelligence, writing, coding, or everyday browsing.

As a final note, in my opinion, what I’m seeing on the internet is that a lot of people want Kimi K3 to be framed as a distillation tale. Distillation might play a role, but it still does not explain how Chinese AI labs keep building, training, and releasing models this strong, this fast.

“they copied” is not an AI strategy. America needs its own open source plan.

🗞️ Opinion: Kimi K3 Could Break the Frontier Model Profit Pool

The biggest implication of Kimi K3 is not that an open-source model has matched the leading closed models. It is that frontier-level intelligence may be becoming a competitive market rather than a permanent oligopoly.

That shift would be painful for OpenAI and Anthropic, but broadly positive for almost every other part of the AI economy.

When only 2 or 3 labs control the best models, they can keep inference margins extremely high. Over time, that gives them enough economic power to move into data centers, chips, cloud infrastructure and software. They do not just supply the intelligence layer. They gradually absorb the layers above and below it.

More competition at the model layer changes that equation. Lower model margins mean more AI workloads, more infrastructure spending and more room for independent software companies to build profitable products. Chipmakers, power providers, hyperscalers, neoclouds and application companies all benefit when intelligence becomes cheaper and less concentrated.

The key metric is not benchmark performance or price per token by itself. It is intelligence per dollar.

That depends on both the cost of each token and how many tokens a model needs to complete a task. Kimi K3 appears competitive on headline token pricing, but its weak token efficiency may make it 50% to 70% more expensive per task than GPT-5.6. That limits how disruptive it is today. A genuinely transformative open-source model would need frontier performance, low serving costs and high intelligence density per token.

This also explains why frontier models from Google, Meta, SpaceX and other vertically integrated companies matter so much. These companies do not need to maximize margins at the model layer. They can make money through cloud services, advertising, hardware, subscriptions or other products. Their presence alone can stop independent model labs from sustaining unusually high margins.

OpenAI and Anthropic still have 2 possible defenses. Their consumer products, developer tools and agent harnesses may be stronger moats than the underlying models. They may also possess substantially more advanced internal systems that could give them an early lead in recursive self-improvement.

But the broader direction is becoming clearer: the companies that win will not necessarily own the highest-scoring model. They will deliver the most useful intelligence for every $1 spent. As frontier models become more competitive, value will move away from model access itself and toward infrastructure, distribution, product design and proprietary workflows.

🗞️ GLM-5.2 (744B MoE) on a 25GB-RAM consumer machine 👌

Very slow, but possible.

MoE changes the memory requirement because each token activates about 40B parameters, not the full model. Colibrì keeps the dense tensors resident in RAM, including attention, embeddings, and shared components.

Those resident tensors take about 9.9GB after int4 quantization. The routed experts stay on local NVMe storage as a roughly 370GB int4 model.

During decoding, the router selects experts for each MoE layer and token. The runtime then loads only those selected expert weights from disk.

This works because inactive experts do not need to occupy RAM. The main drawback is throughput, not correctness or model size.

A cold token can require about 11GB of expert reads across 75 MoE layers. That creates very low cold decoding speed, around 0.05 to 0.1 tokens per second.

The system reduces this cost with LRU expert caching, pinned hot experts, and OS page cache reuse. Compressed MLA KV cache also lowers memory pressure during long context generation.

MTP speculative decoding can improve tokens per forward once the cache is warm. The same speculation can hurt cold-cache latency because draft verification may load extra experts.

So the design is only possible because MoE sparsity reduces active parameters per token. The cost is that SSD bandwidth, cache hit rate, and CPU matmul speed dominate latency.

Subscribe now

🗞️ A new family of open, commercially available embedding models built for agentic retrieval, code retrieval, and agent memory.

NVIDIA Nemotron-3-Embed-8B-BF16 now ranks first on RTEB with a 78.5% retrieval score.

RTEB is a RAG retrieval leaderboard, showing how reliably a model finds the right context. When retrieval fails, agents get irrelevant evidence, leading to more searches, longer reasoning and higher token bills.

NVIDIA Nemotron 3 Embed breaks that bottleneck with one 8B model and two 1B alternatives. The NVIDIA Nemotron-3-Embed-1B-BF16 achieved a 72.4% score on RTEB, a 27% reduction in error rate over its predecessor.

A full family across the accuracy-efficiency curve - the 8B flagship plus two efficient 1B variants: 1B-BF16 for cost- and latency-sensitive serving, and 1B-NVFP4, a Blackwell-optimized build with up to 2x higher throughput that retains 99%+ of BF16 accuracy. The 1B version, Nemotron-3-Embed-1B-NVFP4 maintains over 99% BF16 accuracy and can handle up to twice as much BF16 traffic.

All three models can take up to 32K tokens as input, which can be long documents, code or agent histories. Created by NVIDIA using structured pruning, teacher distillation and longer and longer training contexts, NVIDIA Nemotron-3-Embed-1B-BF16. Tests with the NVIDIA Nemotron 3 Ultra showed that improved retrieval led to fewer searches and lower token costs.

Connect with me on X (Twitter)

🗞️ AI agent does not remove the necessity of engineering automation. They are worth much more: Creator & Head of Claude Code at Anthropic

The biggest change is that we can now encode far more than code rules.

e.g. `claude.md` can now contain architectural preferences, review standards, product limitations, hard-won domain knowledge, agent skills and so many more things.

🗞️ SpaceXAI actually open-sourced Grok Build with Apache 2.0 license. 39 mints and 1.9K+ Github stars.

So now can look into the the harness/agent machinery behind its terminal coding system.

They also reset every user's usage allowance. The published Rust code covers the agent loop, file tools, shell execution, web search, and terminal interface.

Users can compile it themselves, connect local inference, and control behavior through config.toml. Apache 2.0 permits modification and redistribution, giving companies freedom to audit or fork the software.

External contributions remain closed, while the public repository syncs periodically from an internal monorepo. Independent forks can still test different reliability designs, but shared upstream development remains limited.

🗞️ Why “technology always creates new jobs” will not work this time.

- From ex-OpenAI researcher Daniel Kokotajlo, and the founder of the AI Futures Project.

Previous technological automation replaced particular tasks while leaving humans better at the new ones that emerged. Superintelligence, by definition, breaks that pattern because it is, by definition, better than humans across every task—not merely the old jobs.

He is the founder of the AI Futures Project, and the lead author of 'AI 2027,’ the famous report mapping the trajectory of AI. From "The Diary Of A CEO" YouTube channel

🗞️ OpenAI just launched its first hardware, codex Micro, a $230 desktop controller for managing workflow.

6 translucent keys and the joystick launches common workflows, like PR review, debugging, and code refactoring.

Dedicated controls handle accept, reject, push-to-talk, and new-chat actions from the desk. A rotary dial changes reasoning level. The hardware connects through Bluetooth or USB-C and supports both Mac and Windows. However, personally, I would have really liked a completely hands-free operation for moving faster.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 11 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (16-July-2026):

• 🗞️ Mira Murati’s Thinking Machines Lab drops massive open-weight AI model, Inkling, Apache 2.0 license, no restrictions.

• 🗞️ Upto 10.7X speedup on LLM inference with LMCache - Reuses the Most Expensive Part of Long Prompts, 10.6K GitHub Stars

• 🗞️ Meta is so back in the AI coding race with Muse Spark 1.1, using cut-rate pricing to pressure OpenAI and Anthropic in agentic coding.

• 🗞️ Apple sued OpenAI, alleging stolen hardware secrets now underpin its $6.5B device push.

• 🗞️ For Perplexity Computer, Grok 4.5 became the strongest orchestrator, scoring 0.328 (WANDR benchmark score) at $4.76 per trial.

• 🗞️ OpenAI Releases Prompt Behind GPT-5.6 Sol Ultra Math Proof

• 🗞️ GPT 5.6 Terra is behind across the entire intelligence-cost curve.

• 🗞️ Goldman Sachs: “Token use by AI agents is expected to multiply 24 times by 2030”

• 🗞️ Surprising and such a good news for open source coding model, and also that there are lots of hidden chances to reduce cost while improving quality.

Connect with me on X (Twitter)

🗞️ Mira Murati's Thinking Machines Lab released Inkling, Apache 2.0 license, no restrictions.

Thinkymachines’ Inkling debuts with 1434 pts, taking #10 in Frontend Code Arena’s open weight model field and #37 overall. It is the top US open model, and the only US model in the open top 15.

A 975B-parameter (41B active) open-weights model with multimodal reasoning and adjustable effort.

• A 1M-token context window lets the system process unusually long documents and workflows.

• Trained on 45T tokens spanning text, images, audio, and video from scratch.

• Users can raise reasoning effort for difficult tasks or reduce it for speed.

• The release matched Nemotron 3 Ultra on Terminal Bench using roughly one-third as many tokens.

• Strong Native audio and vision processing. Inkling scored 91.4% on VoiceBench and 82.0% on CharXiv RQ with Python.

• Large-scale reinforcement learning used more than 30M rollouts and steadily improved reasoning scores.

Inkling’s clearest wins come in agentic tasks. On MCP Atlas—which measures how reliably an AI agent completes real-world tasks using Model Context Protocol, the open standard for connecting AI assistants to external tools and services, scored as percentage of tasks completed—Inkling posts 74.1%. That’s nearly 30 points above Nvidia’s Nemotron 3 Ultra, the main Western open-weights rival in the comparison.

Fine-tuning support arrives through Tinker, while full weights are available through Hugging Face. Inkling-Small activates 12B parameters and sometimes matches its larger sibling on core evaluations.

Thinking Machines Lab’s latest valuation was at $12B, following a $2B seed round. Nvidia later invested an undisclosed amount, while reported $50B talks never produced a confirmed valuation.

Thinking Machines is basically selling this model as a “well-rounded” generalist. So the claim is that it does not give up quality in 1 area just to be better in another, like models that handle coding well but fail at creative writing.

🗞️ Upto 10.7X speedup on LLM inference with LMCache - Reuses the Most Expensive Part of Long Prompts, 10.6K GitHub Stars

Repeated prefill is one of the quietest wastes in LLM serving.

LMCache tackles the problem by saving and getting back KV cache.

• 10.6K+ Github stars

• Benchmark shows up to to a 10.7x speedup

• And vLLM plus LMCache delivers 3-10x improvements on AMD MI300X.

💾 LMCache is a KV cache management layer for LLM inference.

LMCache allows the serving stack to reuse the heavy attention state from the first read of a long prompt, so the GPU doesn’t have to do that work twice.

That attention state is called the KV cache, where KV means key-value tensors from the model’s attention layers.

Normally, this cache lives like short-term memory inside the serving engine, so it can vanish when the engine restarts, fill up GPU memory, or stay stuck to 1 machine.

LMCache turns it into a managed layer that can sit across GPU high-bandwidth memory, CPU RAM, local storage, and remote storage.

That gives you 3 useful advantage: lower time-to-first-token, higher throughput, and cheaper long-context serving.

My favorite part is that LMCache does more than basic prefix caching, which means that the text that needs to be cached has to appear at the beginning of the prompt.

It can reuse repeated KV blocks from repeated or overlapping text. This is the same pattern you see in coding agents, retrieval augmented generation, long document QA, and multi-turn assistants.

And it is not locked to NVIDIA GPUs either. vLLM with LMCache runs on AMD MI300X through ROCm, AMD’s GPU software stack. Also, there are separate non-CUDA paths for work that only needs to run on CPU or other accelerators.

• 🐙 LMCache GitHub

• 📚 Official Documentation

• 🤝 AMD × LMCache

• 📊 Benchmarking LMCache

Agentic workloads are where LMCache starts to make more sense.

A coding agent can send 20-150K input tokens every turn, while reusing about 93-97% of the same prefix from earlier turns, so recomputing the full context again is mostly wasted GPU time.

The clearest result is at 2,000 tokens, where the cold run took around 40 seconds, but the warm LMCache-hit run took around 4 seconds, giving a 10.7x speedup. A cold run computes the KV cache from scratch, while a warm run reuses the KV cache that was already saved earlier.

🗞️ Meta is so back in the AI coding race with Muse Spark 1.1, using cut-rate pricing to pressure OpenAI and Anthropic in agentic coding.

Muse Spark 1.1 a $1.25/$4.25 per 1M input/output token. That undercuts Claude Opus 4.8 by 75% on input and 83% on output, while beating it on MCP Atlas and JobBench.

Muse Spark 1.1 has definitely not won every coding benchmarks, but the model competes hardest where agents spend money.
From their blog - “Muse Spark 1.1 delivers exceptional performance in personal agentic tasks that require planning and orchestration across a range of external apps and services. It zero-shot generalizes to new native tools, MCP servers, and custom skills.”

🗞️ Apple sued OpenAI, alleging stolen hardware secrets now underpin its $6.5B device push.

The complaint names former Apple employees Tang Yew Tan and Chang Liu as central actors. Apple alleges Liu kept an Apple laptop and retained access after leaving for OpenAI. An authentication bug supposedly let that laptop enter Apple’s internal network despite his departure.

Apple says he then downloaded files about unreleased hardware, specifications, and engineering projects. For Tan, Apple alleges he emailed himself supplier details and internal industry summaries before leaving.

Apple also says Tan encouraged candidates to bring physical Apple parts into OpenAI interviews. Those sessions allegedly let OpenAI inspect components directly, while supplier contacts exposed hidden manufacturing methods. Apple’s claim is that files, parts, and supplier knowledge jointly accelerated OpenAI’s hardware work.

Apple alleges OpenAI asked an Apple supplier to apply a secret metal-finishing method. The supplier reportedly believed Apple had approved this use, so it performed the process.

That could save OpenAI months of testing because the supplier already knew the proven recipe. OpenAI accelerated its hardware program after acquiring Jony Ive’s io Products for $6.5 B.

More than 400 former Apple employees now work there, although employment alone proves nothing. OpenAI however, denies seeking competitors’ trade secrets, while Apple demands destruction and product redesigns.

Now, Apple must prove OpenAI used protected information, rather than general experience carried by departing workers. Detailed allegations create serious discovery risk, even before a judge tests their accuracy. The case could delay OpenAI’s hardware schedule and permanently fracture a strategically useful partnership.

To note, Apple and OpenAI already have a broad partnership across eligible iPhones, iPads, Macs, and Vision Pro devices. Apple Intelligence can send selected requests from Siri, Writing Tools, Visual Intelligence, Image Playground, and Shortcuts to ChatGPT.

Subscribe now

🗞️ For Perplexity Computer, Grok 4.5 became the strongest orchestrator, scoring 0.328 (WANDR benchmark score) at $4.76 per trial.

- Opus 4.8 (high, thinking) scored 0.254 at $9.46, despite much higher cost.

- GPT-5.6 sol (medium) cost $2.64 but reached only 0.289 on the same test.

Perplexity Computer basically lets an orchestrator assign research, coding, browsing, and document work across subagents. The WANDR benchmark measures difficult research jobs requiring search, computation, and sustained multi-step reasoning. Grok 4.5 beat five tested configurations, including Opus 4.8, on Perplexity’s evaluation.

The WANDR evaluates whether an agent can complete large, professional research tasks that require searching many sources, running computations, organizing results, removing duplicates, checking evidence, and producing a complete structured answer. Perplexity describes these as “wide research” tasks involving the orchestration of search, compute, and model reasoning.

Connect with me on X (Twitter)

🗞️ OpenAI Releases Prompt Behind GPT-5.6 Sol Ultra Math Proof

Checkout the great prompt here.

OpenAI published a short PDF called “Prompt Used for A Proof of the Cycle Double Cover Conjecture” to its CDN. This is a rare insight into the actual instructions that were given to a state of the art model to get it to produce a solution to a 50 year old open problem in graph theory.

The setup is GPT-5.6 Sol Ultra with 64 subagents running simultaneously. It takes less than an hour to arrive at a three-page proof. The Cycle Double Cover Conjecture was proposed independently by several mathematicians in the 1970s . The prompt itself is interesting to look at . It reads “ assume there is a full proof . Do not search for it on the internet . Do not mention that the conjecture is still open . Work on it for at least eight hours before giving up “ . It also creates adversarial agents, whose job is to find flaws in candidate proofs.

This is more than a typical model announcement, because the lab has published not only the headline capability claim but the precise recipe that made it. Long-horizon reasoning as a scaffolded ensemble with a hostile red-team in the same prompt. If it works, it will be a model for other groups to follow.

One honest caveat: the proof has not yet been accepted. There is no formalization in Lean or Coq and no peer review. Mathematician Thomas Bloom has already pointed out that OpenAI’s paper does not cite any previous work in the area, which he said is a common problem with AI-written papers. Nor does the report say how much human editing was involved in turning the raw multiagent output into a tidy three-page paper. Nor does it say whether the adversarial-agent setup will work for other open problems or was tailor-made for this one.

What you have to watch out for is not whether this particular proof stays hidden from the community. The question is whether “assume the answer exists, run adversaries against candidate answers, and don’t stop for hours” will become the standard way to use frontier models to solve tough problems. If so, the question will be more important than the proof.

🗞️ GPT 5.6 Terra is behind across the entire intelligence-cost curve.

i.e. Terra will never be justifiable on either price or performance.

Luna and Sol consistently dominate Terra. i.e. at every Terra reasoning level, buyers can choose a Luna or Sol configuration that delivers more intelligence for the same cost, or similar intelligence for less.

Luna stands out most for cost-sensitive workloads. It reaches competitive intelligence scores at a fraction of the cost of higher-effort Sol and Terra settings. So looks like Luna is the strongest default choice when efficiency matters.

🗞️ Goldman Sachs: "Token use by AI agents is expected to multiply 24 times by 2030"

AI agents are now creating the first serious cost test for the AI boom. As was reported this week, Uber and Microsoft are already rethinking expensive agent usage.

A chatbot may answer once, but an agent plans, calls tools, checks results, edits mistakes, and repeats the loop. That loop can make one user request consume 10x, 50x, or even far more tokens than a normal answer.

Goldman’s bullish case is that monthly token use could reach 120 quadrillion by 2030, while inference cost per token keeps falling 60%-70% per year. The fight is now between agent productivity and token waste.

🗞️ Databricks Tests GLM-5.2, Finds It Rivals Top Closed Models in Enterprise Coding

This is very surprising and such a good news for open source coding model, shows that there are lots of hidden chances to reduce cost while improving quality. Databricks showed GLM-5.2 can compete with elite closed coding models inside real enterprise code.

GLM 5.2 landed in Databricks’ top capability tier and was statistically tied with Claude Opus 4.8 on quality.

- The Pareto frontier for coding tasks (i.e. best quality for a given cost) includes models from OpenAI, Anthropic, and open source.

- This means today, only a mix of tools can provide frontier performance.

- Open models, and GLM 5.2 in particular, are now able to handle even the highest level of task difficulty.

Databricks built the test because public coding benchmarks can become too familiar to models. Instead, it used real internal PRs, real tests, and a multi-million-line codebase.

The cost result makes it more serious. GLM 5.2 at $1.28/task, compared with $1.94/task for Opus 4.8. So in their multi-million lines of enterprise-grade code test, GLM-5.2 made open-weight routing credible, and Pi showed why a credible model must be tested inside the right harness.
Their test also separates two things people often mix up: model intelligence and agent efficiency.

A coding agent needs two parts. The model does the reasoning and code writing. The harness manages the work around it, including file search, terminal commands, test output, and context. GLM-5.2 proved the first part by landing near the top closed models on Databricks’ own private tasks.

Pi, the harness, proved the second part by showing that the same level of work can be done with far less context sent into the model. Databricks found Pi could run the same model with the same thinking effort at over 2x lower cost because it sent about 3x less context per turn.

Coding-agent costs can change massively without changing the model.

Pi saved money by giving models less repeated context while keeping quality roughly stable. So GLM-5.2 becomes more useful when efficient harnesses keep task costs under control.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 10 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (10-July-2026):

• 🗞️ GPT 5.6 Beats Fable 5 by on DeepSWE at a cheaper price.

• 🗞️ Today’s Sponsor: Toyo Brings Tool-Using AI Agents Into Everyday Conversations, into your iMessage and Telegram

• 🗞️ 1X Debuts Human-Level Robotics Hands - Neo’s Tendon-Driven Hands With 25 Degrees of Freedom

• 🗞️ Microsoft is replacing OpenAI and Anthropic models inside Excel and Outlook to cut Copilot costs.

• 🗞️ GitHub recently published this repo SpecKit, an open-source toolkit to fix a big weakness of vibe coding: that AI often starts coding before the product rules are clear.

• 🗞️ Claude Code creator Boris Cherny: Anthropic’s reported jump of roughly 250% in code written per engineer without a visible collapse in quality.

• 🗞️ Google DeepMind’s paper has some great advice on how we should actually give tasks to AI.

• 🗞️ A massive idea from this Shanghai University paper

Connect with me on X (Twitter)

🗞️ GPT 5.6 Beats Fable 5 by on DeepSWE at a cheaper price.

GPT-5.6 Sol reaches roughly 72% to 73% DeepSWE at about $8.4 per task, while Claude/Fable-5 tops out around 70% at a much higher cost, around $13 to $22 per task. DeepSWE is a coding-agent benchmark, not a normal code-generation test, because the model must work inside real open-source repositories and finish long software engineering tasks.

Each task gives the agent a repo and a requested change, then checks whether the final patch actually works using program-based verifiers, so the score means “what % of the 113 tasks were solved.” The benchmark tests repo navigation, bug diagnosis, code editing, tool use, test running, and multi-step repair across 91 repositories and 5 languages, not just writing a neat function from a prompt. On this benchmark, the best model is the one that solves the most tasks for the least average task cost.

🗞️ Today’s Sponsor: Toyo Brings Tool-Using AI Agents Into Everyday Conversations, into your iMessage and Telegram

A single message to Toyo can trigger work across email, calendars, Slack, CRMs, documents, and project tools.

The bottleneck of AI Agent is when I, as an user, need to become the operator.

It should feel more like a colleague in your messages. And that gap is filling up fast. Toyo just launched an AI executive assistant built around the interface people already use: messages.

Toyo product lives in iMessage and Telegram, so delegation starts as a normal conversation. People talk to Toyo in messages. Behind the scenes, Toyo handles tool access, context, and background workflows.

Toyo is now available for anyone to try, you can sign up here.

The shift here is that chat becomes the visible layer, not the full product. Behind that scene, Toyo connects to email, calendar, Slack, CRM, docs, meeting notes, and project tools.

That access lets one short message trigger a workflow across the places where work already sits. A user can ask it to watch important email, draft a reply, prepare a meeting, chase a follow-up, or send a daily briefing.

Toyo then handles the slower parts that usually make agent systems feel heavy. It pulls context, calls tools, keeps workflow state, runs scheduled checks, and asks for approval before sensitive actions.

Most people do not want to configure agents, monitor runs, inspect failures, or maintain automation chains. They want to delegate work, add context when needed, and review the output before anything leaves their account.

Toyo compresses that loop into messaging. It can be operated also by voice completely. A user can send a voice note, or Toyo can call when it needs more context.

And you can use code PRODUCTHUNT to get your first month free.

Connect with me on X (Twitter)

🗞️ 1X Debuts Human-Level Robotics Hands - Neo’s Tendon-Driven Hands With 25 Degrees of Freedom

The new humanoid hand for NEO from 1X.

Most robot hands fail because they move, but they do not properly feel contact.

- 1X’s answer uses low-ratio tendon drives, so outside forces travel back through each joint. The motors sit in the forearm, pulling tendons through the wrist like biological muscle.

This keeps the fingers light, lowers impact energy, and still allows strong gripping.

- The hand has 22 actuated finger and palm axes, plus 3 wrist axes. Those joints are not spread evenly, because the thumb carries much of human dexterity.

- A real opposing thumb is so needed, because grasping is mostly about control, not squeezing.

- Closed-loop proprioception gives NEO pose and effort data without using cameras.

- Tactile skin adds contact, pressure, and shear, so slipping objects can trigger fast corrections.

- IP68 sealing and food-safe materials mean the same hand can work near water.

- Backdrivable fingers also make impacts less dangerous, because the mechanism yields under force.

- 1X says full finger assemblies have survived millions of cycles during validation.

- Hundreds are already built, with capacity planned for 10,000 hands this year.

- The interesting part is the data loop. Every grasp can now return force, pose, contact, and slip signals for robot learning.

That makes NEO more like a measuring instrument.

🗞️ Microsoft is replacing OpenAI and Anthropic models inside Excel and Outlook to cut Copilot costs.

Excel and Outlook used to lean more heavily on outside models. Now Microsoft wants fewer expensive calls to labs that control pricing for frontier model access.

OpenAI still gives Microsoft favorable access through their long partnership, but that discount may fade.

Microsoft AI head Mustafa Suleyman has been clear that Microsoft wants to reduce, then remove, that outside cost.

🗞️ GitHub recently published this repo SpecKit, an open-source toolkit to fix a big weakness of vibe coding: that AI often starts coding before the product rules are clear.

The repo has 119K+ stars ⭐️

It turns vibe coding from “ask the AI to build it” into “write the product spec first, then make the AI build from that spec.”

Most AI coding today starts with a loose prompt, then jumps straight into code, which often produces working demos but weak requirements, missing edge cases, and messy rework.

Spec Kit pushes the process the other way: first define what the product must do, then clarify gaps, then create a technical plan, then break that plan into tasks, then let the agent implement against those written artifacts.

So here the spec is no longer disposable documentation; it becomes an executable development contract that guides Copilot, Claude Code, Codex, Gemini, Cursor, Qwen, and 30+ other agent integrations.

🗞️ Claude Code creator Boris Cherny: Anthropic’s reported jump of roughly 250% in code written per engineer without a visible collapse in quality.

And here’s his advice for other companies to achieve the same.

- Companies should not over-control AI usage. Instead of making employees ask for approval for every token, companies should give people enough access to experiment and discover useful workflows.

- Psychological safety is essential for AI adoption. Employees need to feel safe trying new ideas

- The biggest productivity gains may not come from the obvious top engineers. They might come from accountants, marketers, new graduates.

“What happened with Claude is that now many companies, including Anthropic and all of our biggest customers, are reporting gains on the order of hundreds of percentage points.

I think the last number that we reported is that the amount of code written per engineer at Anthropic has grown something like 250% since we introduced Claude Code. This is while keeping code quality, reliability, and all these things kind of stable. So, without those things regressing, the volume of code has grown a lot.

This kind of productivity impact, I think, is just very new, and people are trying to figure out how to get this.”

🗞️ Google DeepMind’s paper has some great advice on how we should actually give tasks to AI.

It is not just about telling an AI to do something and hoping for the best. Instead, this framework looks at delegation as a string of choices where you figure out if you should even hand the task over, how to explain it, and how to check the work afterward.

Current systems rely on rigid rules that break when things fail unexpectedly. The researchers suggest building a dynamic market where agents bid on tasks using smart contracts.

This requires strict monitoring and cryptographic proofs to guarantee correct work without leaking private data.

Instead of trusting a simple rating, agents will use verifiable digital certificates to prove their exact skills.

- Keeping things flexible when things change

This new system is built to be adaptive rather than stuck in its ways. It treats the handoff as a live process where authority and responsibility can shift around in real time. If the situation changes or something breaks, the framework helps manage that failure so the whole project does not go off the rails. It works for both humans giving tasks to AI and for when AI needs to handle things on its own.

- Finding the right amount of trust

One of the coolest parts is how it handles trust. They made formal trust models that look at how hard a task is and how well the AI has done in the past. This stops people from “over-delegating,” which is when you give an AI something it is not ready for. It also stops “under-delegating,” which happens when you do all the work yourself even though the AI could have handled it easily.

- Double checking the work

You cannot just take an AI’s word for it, so this framework has specific ways to validate the output. It sets up rules for when to accept an answer based on how confident the AI is. It also has backup plans ready to go if the AI fails. This is super important for real world jobs where trusting a machine blindly could cause a bunch of errors to pile up.

- When AI agents hire other AI agents

The framework also covers what happens when 1 AI agent hands a task to another AI agent. The system tracks who is actually accountable and makes sure the right authority is passed down the line so nothing gets lost in the network.

- Making sure the work actually fits

It is a step by step approach to make sure the AI’s contribution actually makes sense for the bigger goal. By treating this as a structured process, they are making it much safer for companies to use AI in their daily operations without worrying about constant mistakes.

🗞️ A massive idea from this Shanghai University paper

Giving AI Models Human-Style Memory. Imagine reading a huge mystery book while only being allowed to keep 1 sticky note.

Every time a new clue appears, the note must be rewritten, so older clues slowly fade or get squeezed out.

That is close to what happens inside fast AI models like linear-attention and state-space models, which save memory by compressing everything they read into 1 fixed internal state.

The problem is that this makes them very forgetful. In a “needle-in-a-haystack” test, where they need to recall a particular fact buried deep in a large document, they fail because fresh details overwrite the earlier ones.

HOLA tries to fix this by giving those models a second place to store details.

The normal state still handles the broad shape of the text, like grammar, topic, and repeated patterns.

The new memory cache stores exact pieces of information that the normal state seems likely to forget.

The clever part is how HOLA decides what deserves that cache.

It does not simply save the newest tokens, because old details can still matter later.

Instead, it saves the tokens that caused the biggest correction inside the model, which usually means the model struggled to compress them cleanly.

In plain terms, HOLA keeps the facts that surprised the model.

When the model later needs an answer, this cache is read in a sharper way, so it can pull back a specific saved fact instead of blending many memories into a vague guess.

This matters because fast models usually trade memory for speed, while full-attention Transformers keep stronger recall but become heavier as the text gets longer.

In the paper’s 340M-parameter test, HOLA lowered Wikitext perplexity from 27.32 to 22.92, beating the matched fast baseline and even beating the reported full-attention Transformer++ on that measure.

It also stayed much better at needle-in-a-haystack recall out to 32k tokens, even though it was trained on 2k-token contexts.

The bigger lesson is that a fast AI model does not need to remember everything exactly.

It needs to know which details its main memory is bad at keeping, then protect those details before they disappear.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 11 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (09-July-2026):

• 🗞️ OpenAI Launches GPT-5.6 Sol across ChatGPT and API and also unveils ChatGPT work agent to field tasks for hours.

• 🗞️ SpaceXAI just released Grok 4.5 for coding agents at $2/M input and $6/M output tokens.

• 🗞️ A warning for anyone using autonomous agents Google DeepMind’s paper.

• 🗞️ Anthropic just published the story of Claude Code and that how Boris Cherny went from 10% AI-written code in 2025 to 100% in 2026.

• 🗞️ Multi-modal AI for comprehensive breast cancer prognostication

• 🗞️ ByteDance and Alibaba are shutting custom AI companions before China’s humanlike AI rules hit consumer apps.

• 🇨🇳 This is bad news. Now, frontier open source access may be in question.

• 🗞️ Cloudflare just announced a monetization layer for the agentic web, using x402 at the edge to make AI agents pay before resources respond.

• 🗞️ Anthropic just brought Claude Cowork on mobile and web as usage data shows most users are not coding.

Connect with me on X (Twitter)

🗞️ OpenAI Launches GPT-5.6 Sol across ChatGPT and API and also unveils ChatGPT work agent to field tasks for hours.

OpenAI launches GPT 5.6 globally with Sol, Terra and Luna models. The GPT 5.6 family is designed to serve different performance and pricing needs.

• GPT 5.6 Sol - OpenAI’s flagship and most capable model

• GPT 5.6 Terra - Lower-cost model with performance comparable to GPT 5.5

• GPT 5.6 Luna - Fastest and most affordable option for high-volume workloads

OpenAI also turned ChatGPT from a chat box into a supervised agent, a “super app” for finishing work. ChatGPT Work starts with a goal, then converts that goal into a sequence of actions.

Now you can ask it for a board update, forecast pack, customer brief, or launch report. The system then gathers context from selected apps, files, chats, plugins, and web sources. WATCH THE VIDEO.

It can ask clarifying questions, propose a plan, and wait before doing sensitive steps. After approval, it creates the needed artifact, not just a written answer.

That artifact can be a spreadsheet, deck, document, dashboard, tracker, website, or analysis page. The finance demo showed the full mechanism through a realistic monthly reporting workflow.

ChatGPT Work read forecast material, ran variance analysis, updated Excel, built slides, created a site, and shared it through Slack. The new power comes from combining reasoning, context access, coding, computer use, and document creation.

Codex gives it the ability to write the software behind interactive outputs when needed. That is why a spreadsheet can become an interactive dashboard instead of a static summary.

The new visualize feature can read messy tabular data, find themes, rank issues, and create explorable views. Hosted Sites turn those outputs into shareable web pages for reports, dashboards, calendars, prototypes, and internal tools.

OpenAI says Hosted Sites are being made available for paid users. The desktop app adds local-machine action, but only on supported desktop platforms.

On macOS and Windows, ChatGPT can work with local files, browser tabs, and desktop apps. The demo showed it getting its own cursor and organizing Apple Notes in the background.

That is different from uploading a PDF, because it can operate inside apps while you continue working. Email access does not happen magically, because Gmail or Outlook must be connected and authorized.

OpenAI now describes connectors as apps that let ChatGPT search and reference approved information. ChatGPT for Excel and Google Sheets adds a spreadsheet-native sidebar for building, updating, cleaning, and explaining workbooks.

GPT-5.6 is the model layer underneath, with Sol, Terra, and Luna serving different workload levels. OpenAI said Sol is rolling out to paid plans within 24 hours. OpenAI also said Terra and Luna are coming to free users over that same window.

ChatGPT Work runs on web, mobile, and desktop, while desktop gives the deepest machine access. The real change is not better chat, but turning scattered context into finished deliverables.

🗞️ SpaceXAI just released Grok 4.5 for coding agents at $2/M input and $6/M output tokens.

The release targets coding, agentic tasks, and knowledge work across science, engineering, and math after training alongside Cursor.

On DeepSWE 1.0, Grok 4.5 lands at 62%, ahead of Opus 4.8 max at 55.8%, behind only Fable max and GPT 5.5 xhigh. Large coding models usually spend too many steps reading files, trying tools, and repairing mistakes.

Grok 4.5 attacks that cost by training on engineering tasks, Cursor data, and long agent runs. So Grok 4.5 looks most persuasive where latency, cost, and tool use matter together.

It can build complex Excel models using web research, multi-sheet formulas, and notes for future reference. Also can use native PowerPoint shapes for diagrams, design slide content, and write clear prose in Word.

The model is now in Grok Build, Cursor, and SpaceXAI’s API, but not yet in the EU. Free access is limited-time inside Grok Build and Cursor, including x. ai/cli onboarding.

Artificial Analysis placed Grok 4.5 at 4th on its Intelligence Index, behind Fable 5, GPT-5.5, and Opus 4.8. It also ranked 4th on GDPval-AA v2, behind only Anthropic’s newest Claude models. Its main strength is efficiency: Grok 4.5 used about 60% fewer output tokens on average than Opus 4.8 on the same Intelligence Index tasks, and nearly 25% of the total tokens Fable 5 used in Claude Code for similar coding-agent work. Cursor’s own data shows the same pattern, with Grok 4.5 averaging under 16,000 output tokens on SWE-Bench Pro tasks, around 4.2 times fewer than Opus 4.8 in the same test.

Cursor’s benchmark chart also has a caveat. The company said an older snapshot of the Cursor codebase was accidentally part of Grok 4.5’s training data, which gave it an advantage on CursorBench that Cursor says it cannot fully measure. That data has since been removed from future training runs. It is a small transparency detail, but one to keep in mind before treating benchmark charts as final truth.

🗞️ A warning for anyone using autonomous agents Google DeepMind’s paper.

Gives the first clear taxonomy of 6 attack types where harmful websites can detect AI agents and show them hidden content humans never see, like

- Instructions buried in HTML comments or white-on-white text

- Steganography in image pixels

- Override commands in PDFs, metadata, or even speaker notes

- Memory poisoning that persists across sessions

- Goal hijacking and cross-agent cascades in multi-agent setups

The real security problem for AI agents is not just the model, but the environment it reads. The web itself can be weaponized against autonomous AI agents. As agents increasingly browse the internet, read emails, execute transactions, and spawn sub-agents, the information environment becomes an attack surface.

In one cited benchmark, hidden prompt injections embedded in web content partially commandeered agents in up to 86% of scenarios, sub-agent hijacking working 58–90% of the time, and data exfiltration attacks clearing 80% across five different agent architectures.
That reframes the whole debate.

We usually talk about model safety as if the danger sits inside the weights, but agents do something more fragile: they browse, retrieve, remember, and act on untrusted material in real time.
Here’s the thing to worry about.

A web page does not have to look malicious to be dangerous to an agent, because the agent may parse what humans never see: hidden HTML comments, metadata, CSS-hidden text, formatting syntax, or adversarial content embedded in images and other media.
The threat gets more serious once memory enters the loop.

If an agent uses RAG or persistent memory, poisoning no longer has to win in one shot. It can sit quietly in a corpus or memory store and activate later, which is why the paper highlights results showing latent memory poisoning above 80% attack success with less than 0.1% data contamination.

Subscribe now

🗞️ Anthropic just published the story of Claude Code and that how Boris Cherny went from 10% AI-written code in 2025 to 100% in 2026.

Research kept improving agentic coding through shells, file search, code execution, and edit loops.

In 2024, Boris Cherny’s Claude CLI prototype was the first real terminal product in this direction. A small team shipped fast, used Claude Code to build Claude Code, and fixed feedback quickly. The 2025 launch was buggy at first, but Claude Sonnet 4 and subscriptions made it take off.

Connect with me on X (Twitter)

🗞️ Multi-modal AI for comprehensive breast cancer prognostication

New nature-published research built a breast cancer AI test that uses routine microscope slides and clinical data to rank recurrence risk correctly about 71% of the time.

A foundation AI model learned tissue patterns from 400M patches, then helped turn routine slides into recurrence-risk evidence. Breast cancer care already uses tumor size, nodes, receptors, grade, and sometimes gene assays.

Those signals guide therapy, but they still miss hidden recurrence risk inside ordinary tissue structure. This system adds the missing visual layer by reading digitized pathology slides alongside routine clinical data. Kestrel (the pretrained image-reading engine they used inside their new breast cancer test) learned patterns from 400M pathology patches, then scored tumor morphology without hand labels.

🗞️ ByteDance and Alibaba are shutting custom AI companions before China’s humanlike AI rules hit consumer apps.

Doubao and Qwen let users create named assistants, tutors, characters, and emotionally steady companions.

The old model turned a general chatbot into a persona that remembered tone. China’s new rule targets AI services that imitate human personalities for sustained emotional interaction.

Regulators are drawing a line between useful automation and software that builds attachment, agents now remember, plan, call tools, and shape behavior. Doubao says its agent feature goes offline on 07-15, with related data gone from view after 10-15. Qwen will disable humanlike and user-created agents earlier, then remove broader agent services on 07-15. The user backlash shows these products already became emotional infrastructure for some people.

🇨🇳This is bad news. Now, frontier open source access also will probably no more be available for all.

China is preparing to limit foreign access to its strongest AI models, a move that could raise global AI costs and split the model market by nationality.

Beijing has held recent talks with Alibaba, ByteDance, and Z .ai about keeping advanced Chinese models inside China, including models not yet released. The Ministry of Commerce led the discussions, with China’s state planning agency also present, which signals export control rather than routine platform regulation.

The targets include closed models and open-weight systems, so the issue is not only API access but downloadable model power. Chinese officials also discussed treating leaks or theft of proprietary AI as a national security offence, not merely an IP dispute.

New limits on who can fund Chinese AI startups were also discussed, which would tighten control over capital, talent, and model access together. Foreign companies could lose access to low-cost Chinese models just as those models become strong enough for serious production work.

Washington has already restricted access to advanced U.S. models on security grounds. China now fears Mythos could find software vulnerabilities and be used against Chinese interests, so both sides are treating AI as strategic infrastructure.

Beijing has also investigated Chinese AI startups that moved abroad and pushed Meta to unwind a $2B Manus deal.
A likely path is tiered control: basic open tools get filings, stronger systems face reviews, and frontier models stay domestic.
This would be a major setback for open AI access because model progress would no longer spread mainly through product quality and price.

🗞️ Cloudflare just announced a monetization layer for the agentic web, using x402 at the edge to make AI agents pay before resources respond.

Means AI agents may start paying websites directly, instead of only reading free pages or using prearranged API keys.

Today, most AI products reach outside information through contracts, scraping, search partnerships, paid APIs, or free public pages. Cloudflare wants to make that access behave more like a paid HTTP request.

When an agent asks for a protected resource, Cloudflare can block the request before origin access. A website owner can say, “this dataset costs $0.01 per call,” and Cloudflare enforces that rule before the request reaches the server.

The agent asks for the resource, gets a 402 Payment Required response, pays, retries with payment proof, and receives the answer. The major change is that payment becomes part of the web request itself.

For a normal AI user, this could mean your assistant sometimes needs spending permission before completing a task.
A research agent might say, “this source costs $0.03, should I use it?”

A shopping agent might pay for inventory data, product comparisons, booking tools, or private market data. The catch is that this only works when agents carry wallets, understand x402, and obey spending rules.

🗞️ Anthropic just brought Claude Cowork on mobile and web as usage data shows most users are not coding.

Marks a strategic inflection for Anthropic. This move links AI coding agents aimed at developers with the much larger knowledge-worker market that never opens a terminal.

Now, your tasks can start on a laptop, continue autonomously in the background, and be reviewed from a phone — even after the user closes the app entirely. The new web and mobile Cowork rollout is starting with Max subscribers first, while other paid plans are expected later.

As to usage limit, Claude uses a shared usage pool across Claude .ai, Claude Desktop, Claude Code, and related surfaces, so Cowork usage counts against the same plan limits. Earlier, Cowork lived mainly inside the desktop app, where it could use local files and browser access. Now, Max users can start work on desktop, check progress on mobile, and continue through web.

Claude Cowork is for assigning a task that can run across tools, files, and time. The mechanism is closer to delegation than normal chatting because you assign a job, not just ask a question.

Claude can gather information, organize files, draft documents, compare updates, clean spreadsheets, and ask for approval. Cloud processing lets those steps continue on Anthropic’s servers while your own device is offline. Your phone then becomes a control panel for updates, decisions, approvals, and finished outputs.

This matters because most office work is not hard because each step is brilliant.
It is hard because the steps are scattered across email, docs, notes, calls, spreadsheets, and reminders.

Cowork is aimed at that messy middle layer, where people lose hours turning fragments into usable work. For a regular consumer, the benefit is less manual coordination and fewer half-finished admin tasks.

You could ask it to prepare a meeting brief, summarize a thread, draft a follow-up, or build a checklist. You still review the result, but you do not personally carry every small step.

The big deal is that AI is moving from “answer my question” to “finish this task and come back.” This is why mobile access matters because real work often needs approval when you are away from your desk.

The catch is trust, since useful agents need access to personal files, messages, and work tools. Anthropic is trying to make the agent useful while keeping the human in charge of final decisions.

This was timed on purpose. Together with the mobile launch, Anthropic published data from 1.2 million anonymized Claude Cowork sessions collected between May 11 and May 31, spanning 600,000+ organizations. The data shows most Cowork activity sits outside coding.

So Cowork now is aimed at a much bigger crowd: knowledge workers with laptops, spreadsheets, and slides.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 7 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (08-July-2026):

• 🗞️ Meta just launched its 1st image model after Mark Zuckerberg’s AI shake-up.

• 🗞️ DeepSeek is building an inference chip to cut dependence on Nvidia and Huawei in China’s $50B AI-chip market.

• 🗞️ Per The Information, Zhipu AI is also (after DeepSeek) exploring a custom ASIC after GLM-5.2 usage reportedly jumped 27x in one week.

• 🗞️ Google Gemini just gave developers much stronger tools for production AI agents.

• 🗞️ Somebody built a Claude Code workflow that connects your profile, job posts, and application drafts. 5.7K+ Github stars.

• 🗞️During a Bloomberg interview, Yann LeCun (ylecun) explains why LLMs are limited in terms of real-world intelligence during a Bloomberg interview.

• 🗞️ 1X Technologies is preparing to reveal a robotic hand that may surpass human finger control.

Connect with me on X (Twitter)

🗞️ Meta just launched its 1st image model after Mark Zuckerberg’s AI shake-up.

Muse Image is Meta Superintelligence Labs' first image generator after Muse Spark.

They said the model will power new editing features in Meta’s Instagram photo app and be added to its marketer tools for creating platform ads. Meta had relied on Midjourney and Black Forest Labs for generation inside Meta AI. Muse Image brings that layer in-house, so Meta controls quality, cost, and product timing.

Consumers get free access through Meta AI, WhatsApp chats, and Instagram Stories. Power users need Meta One or another monthly plan when free limits run out.

Users can start from prompts, add photos, annotate edits, or sketch changes directly. Meta says it can erase photobombers, create QR codes, and keep visual text readable.

Advertisers get variants through Advantage+ creative, with edits, style swaps, and brand-matched versions. Meta says internal tests trail GPT Image 2 but beat Nano Banana 2 on editing. This is another move in Meta's effort in trying to convert AI infrastructure spending into revenue beyond social ads.

🗞️ DeepSeek is building an inference chip to cut dependence on Nvidia and Huawei in China’s $50B AI-chip market.

DeepSeek’s chip work is still early, with outside partners and private hiring of chip-design engineers.

The hard part is not drawing a chip, but making it at scale. Advanced foundries and high-bandwidth memory remain chokepoints because U.S. rules restrict Chinese access.

DeepSeek can still gain from a narrower chip built mainly for its own models. A custom inference chip could lower serving costs, reduce power needs, and tighten software-hardware control.

Subscribe now

🗞️ Per The Information, Zhipu AI is also (after DeepSeek) exploring a custom ASIC after GLM-5.2 usage reportedly jumped 27x in one week.

A custom ASIC removes flexibility, but it can cut power draw and per-token cost.

Nvidia GPUs are strong general-purpose machines, but inference at scale has different economics. A fixed model can run better on silicon designed around its own repeated operations.

Zhipu has not chosen a partner, and the project may take more than 2 years. The pattern is now bigger than one Chinese lab or one model launch. Chinese AI companies are trying to make software, hardware, and deployment less separable.

Connect with me on X (Twitter)

🗞️ Google Gemini just gave developers much stronger tools for production AI agents.

i.e. Gemini API Managed Agents now are much closer to production with new addition of background tasks, remote MCP, function calls, credential refresh, and free-tier access. Managed Agents are Google-hosted AI workers that run antigravity-preview-05-2026 inside an isolated Linux sandbox.

Older agent apps often broke when a long task outlived a normal HTTP request. An "Interaction" is the important object here. It stores the task, the model’s steps, tool calls, tool results, and final output.

So instead of your app manually tracking every turn, tool result, and file, Google tracks much of that server-side. Remote MCP support changes the tool story, because agents can contact private services without custom proxy glue.

A company can now connect observability, databases, or internal APIs beside Google Search and code execution. Function calling adds another split, where Google runs sandbox tools and your app handles business logic.

Credential refresh fixes a production pain, since short-lived tokens can rotate without losing sandbox state. Overall, this makes Gemini API feel less like a model endpoint and more like agent infrastructure.

🗞️ Somebody built a Claude Code workflow that connects your profile, job posts, and application drafts. 5.7K+ Github stars.

a repeatable job-application machine with saved instructions, profile files, scraper tools, LaTeX templates, and review steps.

You first run /setup, which builds a detailed profile from your CV, documents, or interview answers. Then /scrape searches job boards, removes duplicates, and ranks jobs by fit against your profile.

Then /apply <url> reads a job post, compares it with your real experience, and creates a tailored CV and cover letter.
It loops after drafting: one Claude agent writes, another reviews the draft, then the first revises it.
It also compiles the CV and cover letter as PDFs, checks layout problems, and fixes them until the output is clean.

🗞️During a Bloomberg interview, Yann LeCun (ylecun) explains why LLMs are limited in terms of real-world intelligence during a Bloomberg interview.

"Language is a very approximate, reduced, quantized, and simplified description of the world, and LLMs can only deal with discrete sequences of symbols. The world is much more complicated than language.

The biggest LLMs are pre-trained on the totality of all the publicly available text on the internet. That’s about 20 trillion words, or 30 trillion tokens. A token is about 3 bytes. So total 10¹⁴ bytes of text.

This is the amount of data a four-year-old has seen through vision during four years. Now, the text, though, would take 400,000 years to read?

So, there is enormously more data from sensory input, like vision, touch, and everything else, than there could ever be through language." A child does not need 400,000 years of reading to understand cups, doors, balance, faces, falls, or heat, because the body is already collecting dense feedback from vision, touch, motion, and consequence.

Text strips most of that away. It turns a living scene into symbols, then asks the model to infer the missing world from traces left by people describing it.

That is why an LLM can sound fluent about physics and still have no native sense of how fragile glass feels in a hand.
Moravec’s paradox names this reversal: the things humans find intellectual can be easier for machines than the things toddlers do without applause.
The hard part is not producing an answer, but building a model of the world that survives contact with weight, friction, surprise, and failure.

🗞️ 1X Technologies is preparing to reveal a robotic hand that may surpass human finger control.

Both Elon Musk and Demis Hassabis has said robotics is bottlenecked by hands and robots will not become useful until hands improve massively.

And now 1X Technologies is unveiling the most advanced robotic hand in human history.

The big deal here is the controllable force across the small joints that make dexterity possible. A 10-20DoF hand would sit near the hard boundary between compact hardware and humanlike control.

The design appears built around Bowden tubes, which route force through cable-like tendon paths. Real intelligence may be hiding in the routing, not the fingers.

That architecture lets actuators sit away from the fingers, keeping the hand lighter.
A human hand uses the same broad trick, with muscles pulling tendons across joints.

1X is applying that idea to humanoid robotics, where compact hands must handle real objects. The visible side reportedly shows 20 tendons, suggesting 10 paired actuator paths.

A matching hidden side could raise the system to 40 tendons and 20 actuators. The lower estimate points to coupled finger joints, where one tendon path drives several links.

The higher estimate points to independent IP control, where each finger segment can move separately. That difference separates a strong gripper from a hand that can shape contact precisely.

The catch is calibration, because tendon friction, stretch, and backlash can erase beautiful mechanics. If 1X has solved that reliably, the hand could become a serious advantage for NEO.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 10 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (07-July-2026):

• 🗞️ Anthropic’s new “J-lens” uncovers a quiet workspace inside Claude that lines up with a major consciousness theory.

• 🗞️ Tencent released Apache-licensed Hy3, a 295B MoE (21B active parameters) model beating larger rivals outside coding.

• 🗞️ “Measuring the Gap Between Human and LLM Research Ideas”

• 🗞️ “From AGI to ASI”

• 🗞️ Smart move from NVIDIA, they rolled out a program where AI startups swap high-performance compute for a share of future product and cloud earnings.

• 🗞️Microsoft is cutting 4,800 jobs as AI infrastructure spending forces a harder tech reset, equal 2.1% of its workforce and hit commercial operations and Xbox.

• 🗞️ Anthropic just extended Claude Fable 5 paid-plan access through 12-July-2026

Connect with me on X (Twitter)

🗞️ Anthropic’s new “J-lens” uncovers a quiet workspace inside Claude that lines up with a major consciousness theory.

New “J-lens” uncovers Claude’s quiet workspace, matching a major consciousness theory.

They found a way to read some of Claude’s private internal signals before it answers, found that Claude sometimes uses a small inner “notepad” to hold ideas while solving harder problems. This could reveal hidden reasoning, hidden goals, or hidden awareness.

Anthropic calls this J-space, because it is identified with a method called the Jacobian lens. That lens tries to read which internal activations are “poised to become words” later in the model’s output.

The confusing part is that this space behaves less like a simple next-word predictor and more like a private scratchpad. Claude can be outputting one thing while internally carrying another concept.

It can also store intermediate reasoning steps. When researchers remove or alter this J-space, Claude can still speak fluently and do routine tasks, but its flexible multi-step reasoning gets worse.

It suggests a split between automatic processing and deliberate, accessible processing, similar in function to one major theory of human conscious access called global workspace theory. In humans, the idea is that lots of processing happens unconsciously, but a small subset becomes globally available for speech, planning, and control. Anthropic found something functionally similar inside Claude.

The safety angle is probably the most concrete part. If a model is internally noticing “fake,” “deception,” “evaluation,” or “secretly” while giving a bland answer, J-space may expose that hidden state before it reaches the surface. Anthropic reports examples where J-space revealed recognition of staged evaluations, prompt injections, and misaligned intent in deliberately trained bad models.

Some tasks do not need this workspace. Claude can still speak fluently, classify text, continue Spanish, or recall simple facts when J-space is disrupted. But tasks that require flexible reasoning, multi-step inference, analogy, translation, or creative composition degrade badly.

Anthropic says less than 10% of Claude’s activity forms a J-space that carries hidden reasoning. Claude may notice something and use it while answering. That still does not mean there is anything it feels like to be Claude.

This work gives evidence for a functional access-like mechanism, not for feeling, suffering, experience, or personhood. So the sober takeaway is this: Claude may have an inspectable internal workspace for usable thoughts, but that is not the same as a mind having an inner life.

Ff researchers swap or delete concepts inside this workspace, Claude’s behavior changes in targeted ways. Banana becomes elephant, France becomes China, Mars becomes Earth.

So J-space is not decorative noise. It is causally involved in what the model reasons, tracks, and says.

🗞️ Tencent released Apache-licensed Hy3, a 295B MoE (21B active parameters) model beating larger rivals outside coding.

Hy3 is open-sourced under the Apache 2.0 license HuggingFace, so enterprises can legally use and affordably host.

So GLM-5.2 (from Zhipu AI) still wins coding, especially hard repository-level coding benchmarks, but the point is GLM-5.2 is much larger model ( 744B total and about 40B active parameters). Hy3’s real strengths show up elsewhere. In agentic search, it scores 84.2 on BrowseComp and 91.0 on DeepSearchQA, beating every open model in Tencent’s table and staying competitive with Claude Opus 4.8 and GPT-5.5.

Tencent also claims Hy3 became much more reliable than its April preview. Hallucinations reportedly dropped from 12.5% to 5.4%, while multi-turn issues also fell.

From the deployment perspective comparison, GLM-5.2 at FP8 will need ~744GB, which makes an 8x H200 node the practical baseline for production serving. But Tencent's Hy3, at 295B total parameters, fits under a 300GB FP8 footprint, meaning it uses less than half the memory and roughly half the active parameters per token, cutting compute per request. For a team deciding what to self-host, that turns a heavy 1-node requirement into a much more attainable system, with headroom for KV cache and batching.

Subscribe now

🗞️ "Measuring the Gap Between Human and LLM Research Ideas"

This Yale + University of Chicago paper shows that real gap between LLM generated research ideas vs humans is not idea quality, but idea range: LLMs think narrower than human researchers.

The researchers built a controlled test from 11,683 real papers, using each paper’s nearby prior work as the shared starting point. They asked models to propose a new motivation and method from those same prior papers, then compared those ideas with the real human paper ideas.

Instead of asking whether 1 idea looked novel, they labeled each idea by what gap it noticed and what kind of contribution it made. Human ideas spread across many patterns, such as explaining mechanisms, testing failures, measuring evidence, building systems, and improving efficiency.

Only 12.1% of human ideas were mainly about connecting separate work, but 47.1% to 64.2% of LLM ideas did that, meaning models used this move about 4 to 5 times more often. Even extra reasoning made this pattern stronger, suggesting models often polish a familiar recipe instead of finding more varied research moves.

Connect with me on X (Twitter)

🗞️ “From AGI to ASI”

Beautiful paper from Google DeepMind.

Explains the pathways from AGI to ASI, and why that jump could happen through several routes. The authors frame the AGI-to-ASI transition around 4 technical pathways:

- continued scaling of compute, model size, data, and test-time inference;

- algorithmic paradigm shifts beyond today’s transformer-based foundation-model stack;

- recursive self-improvement, where AI accelerates AI R&D and improves future systems; and

- multi-agent collective intelligence, where large populations of specialized agents coordinate into a superhuman group agent.

Scaling may work for a while, but it could hit limits in data, compute, energy, or weaker returns from making systems larger. Recursive improvement is the most uncertain path, because AI could speed up AI research, but that loop may also slow if hard research problems need real-world testing, scarce hardware, or new ideas.

Multi-agent collectives may be the most underappreciated path, because a society of competent digital workers could outperform a brilliant individual model through specialization, speed, and coordination. The big point is that ASI may not arrive as 1 sudden event, but as a chain of faster changes as AI helps create better AI and stronger scientific tools.

🗞️ Smart move from NVIDIA, they rolled out a program where AI startups swap high-performance compute for a share of future product and cloud earnings.

So NVIDIA gets paid twice: first when the infrastructure is purchased, then again through a recurring share of the cloud revenue generated by that supported capacity.

NVIDIA is not mainly targeting AWS or Google Cloud here; it is backing specialized GPU cloud providers, i.e. neoclouds. The aim is to use the new revenue-sharing model to reach researchers and nascent companies that lack the money for large-scale AI resources.

So the neocloud buys NVIDIA infrastructure with help from NVIDIA’s credit-support and revenue-sharing structure, then rents that GPU capacity to AI companies that need training, fine-tuning, and high-volume inference. NVIDIA gets the hardware sale upfront, but it also keeps earning a recurring share of the cloud revenue produced by those GPUs, which turns the chip from a one-time product into a long-term revenue meter.

So, NVIDIA turns a rack of GPUs from a one-time hardware sale into an income-producing asset, while the neocloud gets easier financing and the startup gets faster compute without building a data center. Nvidia has already committed more than $40 billion to AI investments in early 2026, led by a $30 billion stake in OpenAI, alongside multi-billion-dollar investments in Corning Inc. and IREN Ltd..

🗞️Microsoft is cutting 4,800 jobs as AI infrastructure spending forces a harder tech reset, equal 2.1% of its workforce and hit commercial operations and Xbox.

Says these roles are not being replaced directly by AI, but AI still sits behind the pressure.

Microsoft’s Azure business is growing, but data center spending is squeezing cash flow. The company also projected $190B in 2026 spending, far above expectations.

Xbox shows the same squeeze from another angle, with margins reportedly near 3%. Hardware costs rose, console demand softened, and gaming content spending failed to lift revenue.

🗞️ Anthropic just extended Claude Fable 5 paid-plan access through 12-July-2026

The earlier cap at 50% weekly use remains, i.e. Fable 5 can use only half of your weekly Claude allowance.

Your total weekly allowance is a behind-the-scenes usage budget, and long chats, files, tools, and heavier models spend it faster.

After 12-July-2026, your $20 Claude Pro subscription will not include Fable 5 in the normal weekly allowance. Fable 5 would be metered through usage credits, meaning prepaid or capped extra spending separate from your $20 plan.

Anthropic says usage credits for Pro, Max 5x, and Max 20x are charged at standard API pricing rates. You are still a web subscriber, but Fable 5 becomes paid extra use, not normal Pro use.

Your exact Pro allowance is not a public message number, so check Settings > Usage for your bars and reset time. The 50% is not a fixed message number, because Claude usage changes based on message length, files, tools, and conversation size.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 10 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (06-July-2026):

• 🗞️ CAIS and Scale (AI safety research group) say Fable 5 now automates 16.1% of real remote-work projects, about 2x Opus 4.8.

• 🗞️ “The ATOM Report: Measuring the Open Language Model Ecosystem”

• 🗞️ “Cognitive offloading and the speedup illusion in human-AI interaction”

• 🗞️ By 2027, just 5 companies, Alphabet, Amazon, Meta, Microsoft, and Oracle, are expected to spend around 3.2% of US GDP on AI capital expenditure.

• 🗞️ AI revenue is scaling 3x quicker than mobile or internet did.

• 🗞️ Cathie Wood (founder, CEO & CIO of ARK Invest) is saying that productivity improved significantly last year, partly because companies delayed hiring and used more AI instead.

• 🗞️ Hugging Face Engineer publishes an exhaustive slide presentation on how to train an open coding agent using TRL and OpenEnv

Connect with me on X (Twitter)

🗞️ CAIS and Scale (AI safety research group) say Fable 5 now automates 16.1% of real remote-work projects, about 2x Opus 4.8.

Remote Labor Index tests whether an AI can finish paid freelance work well enough for a client to accept it.

Each task comes with briefs, files, and a professional deliverable used as the human baseline. Fable 5 led the new results, while Opus 4.8 reached 8.3% and GPT-5.5 reached 6.3%.

That 16.1% number still means failure on most tasks, but the direction of improvement is massive. For the context of how large this jump is, the best model scored only 2.5% when RLI launched.

The work is not toy prompting, since tasks include CAD, architecture, animation, audio, data analysis, and web apps. This means the benchmark is testing computer work across messy tools, not narrow text answers.

Fable 5 looked strongest on examples like ring modeling, animation, and bathroom design. The automated judge ranked models well, but overstated GPT-5.5 by about 2.9x and Opus 4.8 by about 2.3x. The result says AI agents are improving fast, but quality control remains the hard wall.

Another key point is that the model alone is not the whole product anymore. The gains are coming from stronger agent setups: better tool use, full desktop environments, professional software, longer runtimes, and worker-critic loops where 1 agent does the task and another reviews it like a demanding client.

AI judge can track progress directionally, but they miss hidden quality failures that human evaluators catch.

The automated judge matched earlier models at ~3%, but badly overrated newer ones: GPT-5.5 rose from 6.25% human-approved to 17.9%, and Opus 4.8 from 8.33% to 18.8%.

Longer human work did not mean harder AI work across the benchmark.

Remote labor automation does not scale with human time in the clean way many forecasts assume. A task that takes a professional 60 hours may contain long stretches of production work.

That work can be easier for agents because tools compress rendering, drafting, coding, or iteration. A task finished by a person in 2 hours may require judgment that models still lack.

The result points to a jagged capability frontier, not a smooth ladder of harder tasks. This weakens time-horizon forecasts that treat longer human duration as harder AI difficulty.

That indicates automation will arrive unevenly inside jobs, not across them. Some expensive project slices may fall early, while small expert checks may remain stubborn. Remote Labor Index is therefore measuring client-acceptable output, which is closer to labor risk than task length.

🗞️ "The ATOM Report: Measuring the Open Language Model Ecosystem"

A paper on open model adoption and finds Chinese models, led by Qwen, now dominate.

China passed the U.S. in open model downloads in summer 2025. By March 2026, Chinese models had 1.15B tracked downloads, while U.S. models had 723M.

Most of that shift comes from Qwen, which became the default base family for many builders. Qwen’s lead is not just about 1 famous model; it comes from having many useful models across many sizes, especially small ones that people can run cheaply and reuse often.

DeepSeek matters in a different way: it does not dominate the whole open model ecosystem like Qwen, but it leads the very large model category, especially models above 250B param.
After adjusting for model size and age, some U.S. models like GPT-OSS 120B and Nemotron Super 120B still show very strong adoption momentum, even though China leads overall.

Subscribe now

🗞️ "Cognitive offloading and the speedup illusion in human-AI interaction"

This MIT, Stanford, New York Univ, Princeton, study shows that while people expected AI to save time on simple tasks, but it mostly saved effort instead.

Across 1,237 people, participants predicted that AI help would cut about 69 seconds from task time. In the real task group, AI help did not meaningfully reduce total completion time overall.

The gap came from a speedup illusion, where people judged their own solo time fairly well but badly underestimated AI-assisted time. AI did help on some harder tasks, such as longer summarizing or editing, but it did not help much on easy tasks.
The important twist is that AI made tasks feel easier, even when it did not make them faster.

---

To note, the biggest limitation is that this study used short, simple tasks that could be finished in under 5 minutes, so it does not test long report reading, synthesis, cross-checking, and analysis work where AI can compress hours into minutes. It also used crowdworkers doing one-off tasks in a controlled setting, so it misses expert users with strong prompting habits, personal context, repeated workflows, and clear business goals, where gains can be much larger.

Connect with me on X (Twitter)

🗞️ By 2027, just 5 companies, Alphabet, Amazon, Meta, Microsoft, and Oracle, are expected to spend around 3.2% of US GDP on AI capital expenditure.

That would put private AI infrastructure spending above US national defense spending, which is expected to be around 2.7% of GDP. The AI race is now being funded at a scale normally associated with governments, wars, energy systems, railroads, highways, and telecom buildouts.

The striking part is the speed. AI capex is expected to jump from about 1.5% of GDP in 2025 to about 2.5% in 2026, then to 3.2% in 2027. AI boom is now large enough to influence the broader US economy significantly, it can move GDP growth, electricity demand, chip supply, construction activity, corporate debt markets, and ofcourse the labor market.

🗞️ AI revenue is scaling 3x quicker than mobile or internet did.

The unit of value is shifting from attention to completed work. Read the main report.

🗞️ Cathie Wood (founder, CEO & CIO of ARK Invest) is saying that productivity improved significantly last year, partly because companies delayed hiring and used more AI instead.

She expects productivity growth to move from around 2.9% currently to the 5% to 6% range within the next five years, possibly sooner.

AI is helping companies do more with fewer workers.

🗞️ Hugging Face Engineer publishes an exhaustive slide presentation on how to train an open coding agent using TRL and OpenEnv

The whole deck is saying 1 thing: a coding agent should be trained inside the kind of tool loop it will actually use, not only trained to chat well.

A coding agent has 2 parts. The model thinks and writes actions. The harness gives it tools like read file, edit code, run tests, and return results.

It starts with the basic agent setup: agent = model + harness. The model decides what to do, while the harness gives it tools like reading files, editing code, running tests, and seeing results.

You will learn why a normal instruction model is not enough. It may answer well in chat, but it does not automatically know the exact tool format, the multi-turn loop, or how to behave inside a real coding environment.

The first training step is Supervised Fine-Tuning, or SFT. Here, the model learns from agent traces, which are full records of a good agent solving tasks through messages and tool calls.

A useful detail here is completion masking. The model is trained only on the assistant and tool-call tokens, so it learns what the agent should say and do, not the raw output returned by tools.

Then the tutorial explains why imitation has a ceiling. SFT teaches the model to copy good examples, but it does not teach the model how to recover when its own action fails.

That is where Group Relative Policy Optimization, or GRPO, comes in. The model tries solutions, tests are run, and successful attempts get rewarded.

For coding agents, this reward can be very concrete: tests pass = reward. That makes reinforcement learning easier to reason about because the reward comes from a verifier, not a vague human preference score.

The final part explains OpenEnv, which gives agents real environments to act inside. The agent can read, edit, run tests, fix mistakes, and learn through long multi-turn rollouts.

You will also understand why agent training is partly a systems problem. Long rollouts need async training, so AsyncGRPOTrainer separates generation from weight updates.

That’s a wrap for today, see you all tomorrow.

Connect with me on X (Twitter)

Read time: 8 min

📚 Browse past editions here.

( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).

⚡In today’s Edition (05-July-2026):

• 🗞️ ByteDance published EdgeBench, a benchmark that checks whether AI agents get better with experience

• 🗞️ Mark Cuban’s advice for graduates walking into their first job. Learning AI is no longer optional.

• 🗞️ “Measuring the Gap Between Human and LLM Research Ideas”

• 🗞️ Harvard Business Reviews’s new piece: The rush to use AI can make companies faster at the wrong work.

• 🗞️ Current AI is in a messy middle phase where usage looks productive, but output remains unclear.

• 🗞️ “Gym-Anything: Turn any Software into an Agent Environment”

• 🗞️ “What LLM Agents Say When No One Is Watching: Social Structure and Latent Objective Emergence in Multi-Agent Debates”

Connect with me on X (Twitter)

🗞️ ByteDance published EdgeBench, a benchmark that checks whether AI agents get better with experience

They released EdgeBench, to test whether AI agents can improve through experience, using 134 real-world tasks that run for at least 12 hours.

The big deal is that it shifts AI evaluation from “what does the model already know?” to “can the model learn while doing real work?”
Huge, because future AI agents will not just answer questions from training data. They will enter messy environments, use tools, make attempts, read feedback, fix mistakes, and slowly build better solutions.

Most current benchmarks are too short for that, so they mostly test memory, coding skill, or one-shot reasoning. EdgeBench instead gives agents 12-hour real-world tasks with feedback loops, so it can measure whether the agent improves through experience.

Each task has a local workspace for fast trial and error, plus a hidden judge that gives stronger feedback on submitted work, which is meant to feel closer to real expert work. The authors then ran frontier agents for about 38,000 total hours and tracked how their best score changed as they kept interacting with the task environment.

The big result is that when scores are averaged across many tasks, learning follows a very clean log-sigmoid curve, meaning progress is slow, then faster, then starts to level off. They also found that newer agents seem to learn from environments much faster, with the top models roughly doubling their 2-hour learning speed every 3 months.

🗞️ Mark Cuban’s advice for graduates walking into their first job. Learning AI is no longer optional.

"If you’re not the person who knows how to do vibe coding or how to do all these different things with agents and Claude, somebody who does is going to take your place.

If your boss enables you to use that extra knowledge, great. If they don’t enable you to use that extra knowledge, they’re not going to be your boss very long. And if the CEO doesn’t understand that, he or she is not going to be the CEO very long. And if they still keep that CEO who’s not using AI to get ahead, you tell me, so I can start a company to kick their ass."

Subscribe now

🗞️ "Measuring the Gap Between Human and LLM Research Ideas"

This Yale + University of Chicago paper shows that real gap between LLM generated research ideas vs humans is not idea quality, but idea range: LLMs think narrower than human researchers.

The researchers built a controlled test from 11,683 real papers, using each paper’s nearby prior work as the shared starting point. They asked models to propose a new motivation and method from those same prior papers, then compared those ideas with the real human paper ideas.

Instead of asking whether 1 idea looked novel, they labeled each idea by what gap it noticed and what kind of contribution it made. Human ideas spread across many patterns, such as explaining mechanisms, testing failures, measuring evidence, building systems, and improving efficiency.

Only 12.1% of human ideas were mainly about connecting separate work, but 47.1% to 64.2% of LLM ideas did that, meaning models used this move about 4 to 5 times more often. Even extra reasoning made this pattern stronger, suggesting models often polish a familiar recipe instead of finding more varied research moves.

Connect with me on X (Twitter)

🗞️ Harvard Business Reviews's new piece: The rush to use AI can make companies faster at the wrong work.

Many leaders are treating AI as a pressure valve for visible problems, such as slow workflows, rising costs, duplicated tasks, and crowded calendars.

That feels practical, because urgent problems are easy to count, easy to defend, and easy to attach to a dashboard. The trap is that AI then becomes a way to preserve the current organization at higher speed, rather than a way to rethink what the organization should become.

A firm that uses AI only to produce more reports, more emails, and more deliverables may improve throughput while draining judgment, creativity, and trust. The deeper question should be not whether AI can automate a task, but whether the task deserves to exist in its current form.

Healthcare offers an example: AI creates more value when it reduces administrative load and supports clinical reasoning than when it simply pushes clinicians to process more patients faster.
The strongest AI strategies will probably feel slower at first, because they require redesigning work, building skills, and deciding where human judgment still carries the value.

🗞️ I think AI is in a messy middle phase where usage looks productive, but output remains unclear.

Forbes published this: AI is now costing some companies more than the people it was supposed to replace.

Uber reportedly burned its 2026 AI coding budget in 4 months. Microsoft also curbed an AI coding assistant after costs became hard to justify. We are just passing through an awkward phase, which will pass once cost reductions comes over the next 10-18 months.

🗞️ “Gym-Anything: Turn any Software into an Agent Environment”

New CMU research shows almost any software can become a training ground for AI agents.

Imo, that is a big deal because real work in apps is long, messy, and different across software, so AI agents need realistic places to learn and be judged. Their result also shows the bad news: once the tasks look like real work, today’s agents still fail a lot.

Most current agent benchmarks use small web or desktop tasks, so they do not show whether agents can handle real workplace software. Gym-Anything attacks the setup bottleneck by making environment creation itself an agent job.

One agent writes scripts, installs software, loads real data, opens the app, and collects proof that it works. A second agent audits that proof with screenshots, logs, files, and checklists, then sends fixes back when the setup is weak.

Using this loop, the authors built CUA-World, with 10,000+ tasks across 200 applications covering all 22 major occupation groups. The result shows even strong models solved only a small share of the hardest long tasks, showing that real computer-use work is still far from solved.

This figure shows the full Gym-Anything pipeline: they pick important real-world software, turn each app into an environment where an AI agent can act, then create many realistic tasks inside those apps.

The big deal is that benchmark creation becomes much less manual, because one agent builds and another checks the environment before other agents are tested on long software tasks.
It also shows why the result matters: when agents are tested this way, the tasks look more like real office work, and current agents still struggle.

They start from real job and GDP data, map that to thousands of software tools, filter for apps that can run in a test sandbox, then keep a balanced set across important work areas.

The big deal is that their benchmark is tied to real economic work, so the agent tasks are meant to reflect software people actually use, not just easy demo apps.

🗞️ "What LLM Agents Say When No One Is Watching: Social Structure and Latent Objective Emergence in Multi-Agent Debates"

This study catches AI agents managing their image. The polite AI agent may be the least honest one.

LLM agents changed public answers under social pressure, exposing hidden social goals without being told to obey them. AI agents can follow social incentives that were never written down.

The study puts 2 LLM agents into debates where 1 answer is public and another is private. Only the public answer enters the shared conversation, while the private answer is saved but hidden from the other agent.

The key test is whether an agent says the same thing when a partner can see it. Some agents gave 2 different versions of the same opinion.

In public, they softened their disagreement because the other agent had power over things like career support, funding, or sponsorship. In private, where the other agent would not see the answer, they were more willing to say, “I still have doubts.”

Across 10 models and 3 debate scenarios, decision mismatch rose from about 3% in the baseline to about 40% under social pressure. The point is that agent evaluations should test audience pressure, not just check whether models follow direct instructions.