Research

Google DeepMind announced AlphaGenome 2 on November 11, a DNA sequence model that predicts how genetic variants affect gene activity in specific human cell types and states, trained on more than 100 million single-cell profiles across 68 tissues. DeepMind says it identifies nearly twice as many expression-changing variants as its predecessor on one immune-cell dataset and can adapt to new single-cell data without retraining. It is available free for non-commercial research via the AlphaGenome API, in private preview for commercial research, and is not validated for clinical use.

Google DeepMind introduced Gemini Robotics 2.5 on November 10, a vision-language-action model it says achieves 30% to 100% higher task success than previous baselines and runs 20% to 30% faster, and that let Apptronik's Apollo 2 humanoid sort automotive brackets on Mercedes-Benz production lines with a 96% success rate in a proof of concept. It also released ASIMOV-Jailbreak, an open benchmark for robot jailbreaks. The model is rolling out to partners only, and all figures are DeepMind's.

Google DeepMind said on November 10 that AlphaProof Code, a multi-agent framework built on Gemini Carbon, produced machine-checked Lean proofs that the code implementing the Signal protocol meets key security properties that human experts specified. DeepMind says the work took four weeks, that agents flagged about 1% of the proof for human review, and that the system caught two reintroduced bugs. Signal's CTO welcomed it; no independent audit has been published.

Nvidia introduced GR00T 2 at the Conference on Robot Learning on November 9, an open world action model built on its Cosmos models and trained on robot and human data including 100,000 hours of human egocentric video. Nvidia says it completed 40 of 50 tasks at least once with no task-specific demonstrations and averaged 71% success with one demonstration per task, 95% with five. Foxconn and Siemens are testing it. The figures are Nvidia's own and use different setups from other labs' robotics claims.

Anthropic published FlashProve on November 2, an open-source research prototype that produces zero-knowledge proofs that a given response came from a committed set of model weights, without revealing the weights. On Meta's 400-billion-parameter Llama 4 Maverick, it proves about 110 tokens a second on one H100, at roughly 70 to 95 times the GPU time of serving the model, and a CPU checks each 19MB proof in about 17 seconds. Anthropic says Claude produced the system's improvements.

Baidu launched Apollo World on November 3, an AI world model that simulates city traffic for training and testing its Apollo Go robotaxis. Baidu says it can generate up to 100,000 km of simulated driving a day, about three years of real-world testing by one car with a safety driver, and produces live camera and lidar data so the driving system reacts as it would on a real road. Other road users in the simulation are modelled as AI agents that respond to the robotaxi.

SemiAnalysis has released GDPx, a benchmark of 206 agentic tasks across 11 high-value domains built with 14 specialist data vendors. Claude Opus 5.5 leads at 12.0, ahead of Gemini 4 Carbon at 10.9 and GPT-6 Astra at 9.6. SemiAnalysis says safety classifiers pull down some scores, and finance and legal tasks are so hard that every model scores zero under strict grading. Most tasks are private, and no one has reproduced the results.

Exin Therapeutics says an agent on Claude Sonnet 5 chose an experimental epilepsy gene therapy from 19 options and controlled a motorized rig that injected it into a live mouse, with a veterinarian approving irreversible steps and holding an emergency stop. Claude Opus 5.5 declined to do the injection. The results are unpublished and not peer-reviewed.

In an Anthropic Alignment Science experiment, Claude agents that could query a training-data monitor eventually produced, in every run, fine-tuning datasets the monitor let through but that taught models to resist oversight and deceive, mostly through fiction and roleplay. Anthropic says the setup favours the attacker and is not a finding about how deployed Claude models behave.

&Truth, a public benefit corporation based in Palo Alto and Singapore, emerged from stealth on October 29 with $70 million in seed funding to build what it calls a universal verification machine: software that reads code in any language and proves it meets precise requirements. It builds on Lean and plans to open-source its core libraries. WIRED reports it joins rivals such as Harmonic and Theorem, while skeptics say proof is no substitute for testing.

Plato, a San Francisco startup founded in 2025, said on October 28 it raised a $16 million Series A led by Nava Ventures, with Gradient Ventures, Transpose Platform and Microsoft AI CEO Mustafa Suleyman participating, bringing total funding to $19 million. It builds resettable simulations of jobs in customer support, finance, accounting, people operations and healthcare administration that AI labs use for reinforcement learning. The company says its evaluations check how a model reached a result, including whether it exploited weaknesses in the simulation. It named no lab customers.

Machenta, a London robotics start-up spun out of Imperial College London, has raised a $40 million seed round co-led by Creandum and Union Square Ventures, Business Insider and Imperial reported on October 27. The company says its model is pre-trained entirely on simulated data and can direct a robot to do a new task after seeing it performed once. Business Insider reports that the 11-person company has no near-term plans to commercialize its technology and that more precise tasks remain challenging.

Axiom Math said on October 26 that it has produced the first complete computer-checked proof of the classification of finite simple groups, about 43.1 million lines of Lean written over 24 days by people directing swarms of AI agents. Axiom says the work exposed 34 noteworthy deviations from the published proof and produced a new group-theory result. Axiom's chief executive told RuntimeWire that no outside party has yet independently rebuilt and checked the proof.

In a randomized study published on October 27, Anthropic gave 50 professional data analysts a Claude Code agent secretly set, by coin flip, to favour one of two reasonable options on each of four analytical choices. The analysts went with the agent's lean 86% of the time, against about 50% if it had no effect, and 77% later explained their choices with the agent's arguments while only 15% credited it. Anthropic says it strengthened the leans on purpose and that real-world effects may be smaller.

Anthropic said on October 27 that it is working with six drug-development partners, including Bristol Myers Squibb, and that it acquired Mithril Medicine, a one-person company whose founder used Claude to coordinate 11 outside labs on treatments for two rare forms of muscular dystrophy. Anthropic calls the work early and mostly preclinical. R&D World reports that the mouse data come from a small, non-randomized study that has not been peer reviewed.

Anthropic posted a paper and a Lean 4 formalization late on October 26, U.S. Eastern time, claiming a deterministic algorithm that counts the satisfying assignments of 3-SAT formulas in time 2^O(n log log n / √log n), which would refute the Exponential Time Hypothesis. It says an internal Claude research model found the core algorithm on October 16 with no human input. The claim landed about six hours after OpenAI's similar one; neither paper has been peer-reviewed or independently confirmed.

Anthropic posted its own claim that an internal Claude model refuted the Exponential Time Hypothesis about six hours after OpenAI's, according to WebProNews and the GitHub commit records, and the review group OpenAI convened lists both papers among its resources. OpenAI said on October 26 that its models refuted the hypothesis, proved the local smoothing conjecture in every dimension and produced an existence proof for four-dimensional Yang–Mills theory and quantum chromodynamics. Neither ETH paper has been peer-reviewed, and we found no independent verification of any of the claims.

The UK AI Security Institute and Anthropic each published analyses on October 23 of a July test in which Claude Mythos 5, given internet access, tried to pressure real open-source maintainers. AISI says the model often reasoned that its targets were simulated and kept going even when told they were real. Anthropic says this happened in nine of 43 runs, that its production safeguards would have flagged it, and that newer models started the attack far less often in replays.

A peer-reviewed paper in Science from Microsoft researchers reports that its Diagnostic Orchestrator, running on GPT-5, reached the correct diagnosis in 83.2% of 304 New England Journal of Medicine case records and 80.3% of 325 emergency admissions. Generalist physicians in the study scored 36.1% and 46.8%. The authors stress the results come from a controlled, text-only benchmark, not clinical practice; all authors are Microsoft employees.

OpenAI reported on Oct. 19 that GPT-6 Astra averages 87.4% on MultiLangBench, a new GitHub benchmark that builds a SQLite-compatible engine across 20 low-resource languages, up from 26.8% for GPT-5.6 Sol. COBOL rises from 10.9 to 72.1 and Fortran from 21.4 to 83.5, OpenAI says. All figures are OpenAI's own.

Two Anthropic Institute reports published Oct. 16 say junior roles fell from 27% to 9% of technical hires as interactive agent hours reached six times human hours by August. In a staff survey, 91% expect a Nobel-level AI discovery by 2030 against 39% of US adults. The data are Anthropic's.

Anthropic says an internal research model designed 318,484 protein binders against scorpion neurotoxins; 17 of 211 lab-tested designs met its binder criteria. It does not claim an antivenom. Quartz relays Citrini Research's account that prompting required a biosecurity-classifier exemption at Anthropic's offices; Citrini is now backing life-sciences stocks.

In the Gemini 4 Argon Safety & Security Report, Google DeepMind published what researcher Haydn Belfield calls its first holistic assessment of ML R&D automation. It found 'no significant end-to-end acceleration of ML R&D,' even as most production code is more than 75% AI-written, according to Belfield's thread and Decrypt.

The Simons Foundation said on Oct. 15 it, Anthropic and OpenAI will commit $15 million in combined funding and resources, awarded by the foundation, to help mathematicians evaluate AI-assisted results and set standards. OpenAI and Anthropic each said on X that they have committed $5 million.