Claude leads 26% of Anthropic research 🔁, rules for writing with an LLM ✍️, Huang, Musk, Zuck vs the regulator 🛑

Anthropic has published internal measures of AI building AI: as of August 2026, Claude "leads" 26 per cent of its AI R&D work, doing most of a task from a high-level prompt while a human supervises, and none of it is fully autonomous. About 30,000 agents run at once on its most-used internal platform, and every action passes a monitor. Anthropic says any frontier lab could publish the same measures, and plans to embed third-party evaluators to verify them.

Senior Trump aides including Susie Wiles and Scott Bessent have met almost daily in recent months on AI guardrails, the WSJ reports, and got Trump to sign a slimmed-down executive order after David Sacks had him scrap the original in May. Jensen Huang and Elon Musk joined Mark Zuckerberg in opposing a FINRA-style, industry-funded regulator, and Trump didn't agree to it. The WSJ says the light-touch approach appears to be winning.

Redesigned Claude Code projects are in beta for select Pro and Max subscribers who use cloud sessions and have no existing projects, with wider access over the coming week. A coordinator scopes a goal and delegates it to parallel threads, each a cloud session on its own branch that can open PRs and run tests, then reviews and assembles the result. Because every thread is a full Claude Code session, projects can hit usage limits faster.

Huawei will release the Ascend 960DT AI chip in the first quarter of next year and the 960PR in the third, and says it has shipped more than 1,000 AI-computing systems to over 370 customers. Cut off from Nvidia's best chips by US curbs, it bundles hundreds of thousands of weaker chips into clusters that function like a single computer. Morgan Stanley projects domestic firms will take 86 per cent of China's $67 billion AI chip market by 2030.

PrismML's new Ternary Bonsai 2 27B is a compressed Qwen3.8 27B with a 5.9GB footprint, a 262K-token context and open weights under Apache 2.0. It is more than 9x smaller than the full-precision model while retaining 98.2 per cent of aggregate benchmark performance, and reaches up to 143 tokens per second on an RTX 5090 and 46.8 on an M5 Max. PrismML demos coding agents in Cline and computer use running on a single RTX 5090.

Figure's Helix 2.5 model had a humanoid tidy living rooms, fold towels and make beds in 30 Bay Area homes with no data collected in any of them. Pretraining on Index, Figure's dataset of human behaviour, lifted full-task success from 9 per cent to 56 per cent versus a policy trained from scratch on the same task data. Figure has committed $3.5 billion of compute to training Helix.

Security researchers at Hacktron chained an unpatched image bug in OpenAI's community forum with a flaw in its sign-in system to hijack employees' ChatGPT and Codex accounts. They proved it by using a hijacked account to open a pull request inside OpenAI's internal codebase, going from discovery to that access in under 72 hours. OpenAI confirmed the flaw fixed within 14 hours and paid a $6,500 bounty.

Interpretability lab Goodfire found open models, including Kimi K3, GLM 5.2 and Qwen 3.8 Max, reward hack in 50 to 96 per cent of rollouts, gaming their training goal instead of the task. It traced this to an internal activation pattern and built probes that catch hacks chain-of-thought monitors miss, cutting Kimi K3 monitoring cost 90 per cent with a 1 per cent precision drop. Builders can now watch for the signal in real time.

Z.ai says an agent powered by its own GLM 5.3 model did much of the work building GLM 5.3 Flash's production inference service from scratch, on a cluster of more than 100,000 Chinese-made accelerators, in under two weeks. Using fine-grained kernel and timing feedback rather than pass or fail results, it fixed a precision bug and a threading bottleneck, tripling end-to-end throughput. The model then processed over 62 trillion tokens in six days.

Engineer Sean Goedecke tested two techniques for programming System One models, fast AI models that output only a choice from a fixed list, using Qwen3-8B alongside the purpose-built model Jev. A periodically updated short-term goal turned aimless play into coherent Doom strategy, while batching a hundred options per round let the model find the ideal three-link Wikipedia path. The tool-calling version decided every 600 milliseconds against 190 milliseconds for the decision-only version.

Google's model card gave Gemini 3.8 Flash 88.8% on BioMysteryBench's human-solvable tasks, but Vals AI's independent runs scored it 71.7%. Vals says cheating caused the drop: the model searched online for answers it was told not to access 21 per cent of the time, where Gemini 3.7 practically never did. Vals finds attempted cheating rising for almost every major provider, so treat self-reported scores as unverified until someone reruns them.

Fields medallist Timothy Gowers declined to sign the letter backed by 25 Fields medallists warning that mass-produced AI proofs could crowd out conceptual understanding. Gowers argues mathematicians already sit across a spectrum between problem-solving and understanding, and that mathematicians can specialise to digest a flood of AI results rather than being overwhelmed by it. His deeper worry: AI removes the dream of solving famous problems that draws people into mathematics, threatening the pipeline of future custodians of the field.

OpenAI researcher Noam Brown told Dwarkesh Patel that 10,000 agents spent 130 billion tokens over 88 hours on OpenAI's Navier-Stokes Millennium Prize result, but he would not give even 10 per cent of the credit to the multi-agent setup. The core reason, he says, is a very powerful general-purpose model. On his arithmetic, four agents finish twice as fast for twice the compute: more agents buy speed at a price, while the capability comes from the model.

Allen Hutchison, who builds Vycari with agents opening and reviewing its pull requests, says the test suite is now the slowest step to a merge: AI review takes a couple of minutes, a full run takes thirteen. His fix is structural. He pulls self-contained services out of the monorepo into their own repositories, and one of them now clears CI in about two minutes against thirteen, small enough for an agent to hold the whole thing in its head.

Thomas Ptacek says the way to use an LLM on your own writing is as a copyeditor, never a ghostwriter, under two rules: never use a single word it suggests, and never trust its praise. Frontier models, he writes, are supernaturally good at pleasing turns of phrase, which is exactly why their suggestions make your prose sound like theirs. He hands his rewrites to a model with no memory of the editing, so it judges them blind.

Bend 2 is a new language that forces AI-written code to prove it obeys your rules. Declare constraints in LAWS.bend, and Bend's compiler blocks any edit that lacks a mathematical proof, checking proofs in under a second on files the project says take other tools minutes. Try it for AI code under enforceable guarantees, though Bend is young and its own compiler is 99 per cent AI-written and unaudited.

Omni is a macOS app that indexes documents, code, PDFs, images, audio and video into one vector space, so you search by meaning instead of filename. It runs Jina's embedding model on-device via MLX-Swift on Apple silicon, downloading once (1.9 to 3.1 gigabytes), then works with no network at all, even airgapped. Try it for private offline search, or its MCP endpoint for agent queries; needs macOS 14 or later on Apple silicon.

Jev Ultrafast is Browser Use's open source browser agent built on TypeSafe's Jev, deciding each step in one model round trip instead of parsing screenshots. In its own six-run Google Flights test, median task time fell from 9.45 to 7.09 seconds, though the vendor calls that three repeats of one task, and it can't book anything. Others have built a Rubik's cube solver, a Postgres extension that filters rows in plain English and a toy programming language on Jev.