Claude Now Leads 26% of Anthropic R&D — AI Digest

Claude Now Leads 26% of Anthropic R&D — AI Digest

Anthropic put its own development automation on the record: the share of model R&D tasks led by Claude went from 1% to 26% in six months. Mozilla puts China's open-weight lag at 4.4 months, and OpenAI's internal monorepo fell in under 72 hours.

Today's highlights

Anthropic: the share of model R&D tasks led by Claude went from 1% to 26% in six months

Anthropic published three measurements of its own development process — how much AI R&D is done by AI, how well agents are overseen, and how compute is allocated. Labs rarely put internal automation on the record as a number rather than a feeling. A breakdown of the release carries the sharper figures: the share of model R&D tasks led by Claude rose from 1% to 26% in roughly six months, Claude collaborates on or leads more than 90% of model R&D work, and about 30,000 agents run inside the company. Those three numbers come from a third-party summary rather than the publication itself, so they deserve more caution than the three headline metrics. The specific values matter less than the gesture: lab automation has been made an object of measurement.

Mozilla: the China–US open-weight gap has narrowed to 4.4 months

Mozilla's "State of Open Source AI" report puts China's leading open-weight models 4.4 months behind frontier US systems — the figure comes from the report itself, published at stateofopensource.ai. Open weights are models whose parameters are published, which is what lets you run and fine-tune them on your own hardware instead of calling someone else's API. Tom's Hardware frames the finding differently: there is no parity on the harder benchmarks, but inference on the Chinese models is drastically cheaper. In the discussion on 17 September 2026 readers pushed back on the methodology, disputing the model ordering in the rankings and the design of the survey section. Open tools you can deploy yourself are collected in the Open Source catalogue on AI SKILLS, and what open weights actually mean is covered in our guide to open-source models.

Three researchers reached OpenAI's internal monorepo in under 72 hours

Three researchers used Claude Opus 5 to chain an image-upload bug into a ChatGPT and Codex account takeover and then into OpenAI's connected services, proving the access with a pull request in the company's internal monorepo — this is the researchers' own account. An exploit chain is a sequence of small vulnerabilities, each harmless on its own, that together produce full access. According to a walkthrough of the work, the whole thing took under 72 hours and a few thousand dollars in tokens; the WSJ reported on the incident. The technical lesson is not that AI cyber is frightening — it is that chain automation is now practical against the ordinary integration surface: single sign-on, forums, email, connected productivity tools.

Stellar Colosseum and Agora: multi-agent harnesses with explicit memory instead of "swarms"

Stellar Colosseum is a model-agnostic many-agent harness for mathematics and theoretical computer science that splits strategy, decomposition, subproblem solving and verification into separate roles; the claimed results are a Codeforces rating of 4263 and 71.0% on TCS-Bench. A harness is the scaffolding around a model: which tools it gets, how the task is cut up, and how many turns it is allowed. Alongside it came Agora, where Git commits serve as shared memory for thirteen workers: over twelve days the system made reproducible progress on model initialisation without a single gradient update. The direction is away from vague agent swarms and toward explicit memory structures, decomposition patterns and reproducibility. Ready-made scaffolding for working tasks is collected in the automation templates on AI SKILLS.

Numbers and facts

Different views: is a 10% chance of extinction this decade an estimate or a rhetorical device?

A group of mathematicians led by Timothy Gowers sent an open letter to the Royal Society calling the situation an emergency and insisting that an estimated 10% chance of human extinction this decade must not be dismissed as hype — the letter was published on 18 September 2026. The signatories argue that recent OpenAI and Anthropic models are approaching the level of the strongest human mathematicians, and they call for government attention to risks in cybersecurity, weapons, biology and disinformation.

For. A sharp jump in mathematics is not a niche event but a sign of rising general capability; if it arrived ahead of expectations, other thresholds may be crossed before anyone is ready for them.

Neutral. Superhuman mathematics is useful as a tool: part of the letter's audience sees it accelerating work on hard problems in physics and biomedicine, provided domain experts stay inside the decision loop.

Against. Critics aim straight at the number: the 10% estimate arrives without a methodology, a base rate or uncertainty bounds, which makes it read as elicited expert opinion rather than a calculation. A separate argument holds that the likelier danger is AI in the hands of careless people, not autonomous extinction.

Tools and techniques

1. A Gaussian splatting scene from an AI-generated orbit video. The recipe: ask the video model to keep the subject rigid while the camera completes a full 360° orbit, extract the frames, run COLMAP with the SIMPLE_PINHOLE camera model, then hand the reconstruction to Postshot or Brush. The author of the walkthrough flags the main source of failure: the first and last frame must be the same image, or the orbit will not close. Commenters suggest GLOMAP as a faster replacement for COLMAP.

2. A character from a reference image in two steps. The 12-billion-parameter vision model Gemma 4 reads the reference and writes a detailed prompt from it, which Krea 2 then renders; the finished workflow is posted on Pastebin and the style is held by a LoRA on Civitai. The strong part of the pipeline is Krea 2's ability to hold a long, complex prompt — and that prompt can also be assembled with the AI SKILLS prompt generator.

3. Activation probes against reward hacking. Prime Intellect showed that activation probes catch reward hacking about as well as an LLM judge while costing less, after Goodfire argued that reward hacking is pervasive among open models on agentic benchmarks. If you run agents on your own tasks, this is a cheap way to tell whether the model is solving the problem or gaming the metric.

In brief