Wednesday, 7 October 2026
Asked to Stop by Their Own Advisers, They Published 722 Proofs
OpenAI Published Its Largest Batch of Machine-Made Mathematics a Week After the Advisory Group It Consults Asked the Labs to Stop
On Tuesday OpenAI put 722 manuscripts, sorted into 372 “research families”, into a public GitHub repository — the output of putting roughly 4,000 open problems to an internal frontier model it does not name, at an average of about three hours of ChatGPT-Pro-equivalent compute per accepted result. The areas range from number theory and complexity to quantum Heisenberg ferromagnets and the relativistic Vlasov-Maxwell equations. The timing is the story. On 29 September the Advisory Group on Mathematics and Artificial Intelligence — nine mathematicians hosted at the Institute for Advanced Study, which OpenAI says it consulted on this very release — wrote: “At present, some frontier AI labs are testing advanced mathematical problems on proprietary models that remain inaccessible to the broader scientific community. We do not endorse this practice, and we ask them to stop.” OpenAI’s spokesperson, Lindsay McCallum Rémy, says the group’s “advice and public recommendations have informed how we’re sharing the results.” Many of the manuscripts carry no Lean formalisation yet; OpenAI says more will follow as verification work completes. And the reading has started: one headline claim, a proof of the full Birch–Swinnerton-Dyer leading-term formula, covers elliptic curves over the rationals under particular Selmer-group conditions — which is a real result, and is not the thing the title names.
Anthropic Will Now Verify You for Offensive Work, Not Just Defensive, and the Top Tier Is Reviewed With the US Government
Also Tuesday, Anthropic restructured who may use its models with the cyber safeguards turned down, and the shape of it matters more than the announcement. There are three tiers now. Defense Access covers SOC work, incident response, malware reverse-engineering and vulnerability analysis, is open to corporate teams, universities, nonprofits, critical-infrastructure operators, small firms, open-source maintainers and even individual researchers with a disclosure record, and is reviewed “within days”. Red Team Access adds authorised penetration testing and red-teaming — offensive work, explicitly — takes weeks, and is closed to individuals. Specialized Access carries the fewest blocks of all and is confined to organisations cleared to test things like flight operating systems and power grids; those applications are reviewed jointly with the US government. All three tiers get Opus 5.5, Sonnet 5.5 and Mythos 5.1, plus future models. Real-time blocks stay on ransomware deployment, physical damage and pen-testing high-risk safety systems. The number Anthropic leads with: partners in its Project Glasswing pilot found at least 129,000 verified software vulnerabilities between April and July, with its own open-source scanning adding 5,500 more and over 33,000 rated critical or high-severity.
Source: Anthropic, 6 October 2026
HackerRank’s AI Interviewer Is Now Generally Available, and What It Grades Is How Well You Handle an AI
Chakra went generally available today after roughly six months in beta, during which HackerRank says it ran more than 500,000 interviews, with Snowflake, Snorkel and Capgemini among the companies testing it. It does not set a coding puzzle. It watches a candidate work in a real repository with AI tools allowed, then asks them about their decisions, and scores problem-solving, judgement, communication and what the company calls “AI fluency” — framing a task for a model, assessing what comes back, steering it somewhere useful. Co-founder and CEO Vivek Ravisankar put the reasoning plainly: “The previous modality of evaluation was evaluating the output. Now, because of AI, anybody can produce an artifact.” Two numbers from the beta are worth sitting with: candidates rated the experience 4.8 out of 5, and flags for suspicious activity fell 70 to 80 per cent against traditional proctored tests — because when using the AI is the test, there is nothing left to smuggle in. Chakra scores; the employer still decides.
Source: TechAIWire, 7 October 2026
SAP Is Buying a Company Whose Product Is Knowing What Your Staff Actually Do All Day
At its Connect conference in Las Vegas on Tuesday, SAP announced a definitive agreement to acquire TechWolf, a Ghent company whose “context graph for work” infers what skills employees have from the tasks they actually perform inside the business software a company already runs — rather than from anyone’s self-assessment. Terms were not disclosed; the deal is expected to close in the fourth quarter subject to regulatory approval, TechWolf goes into the SuccessFactors portfolio, and co-founder Andreas De Neve stays on as its CEO. The agent angle is the whole point: the graph becomes the grounding layer under SAP’s Joule agents for skills-based hiring, workforce planning and role redesign. Manoj Swaminathan, president and chief product officer of SAP’s Autonomous Suite, named the second benefit without dressing it up — “This grounding layer perfectly matches our strategy. It makes token usage more efficient.” A map of your employees’ real work, bought partly because it is cheaper to feed a model.
Google Put Text, Code, Images, Video and Audio Into One Searchable Space and Made It Small Enough for a Phone
EmbeddingGemma 2 landed on Tuesday under Apache 2.0: 740 million parameters total, but modular — a 270M text model with a 170M vision encoder and a 300M audio encoder, loadable in whichever combination you need, all of them writing into the same 768-dimensional space. Which is the actual feature: a voice memo and a video frame and a line of code become comparable objects, so you can search one with another. The context window is 8,192 tokens across every modality, four times its predecessor’s, and it covers over 100 languages. The clearest gain is on code, where its MTEB score goes from 68.76 to 78.68. And it runs where the data is: roughly 191MB of active RAM for text alone on a Pixel 11 Pro, about 567MB for the full multimodal set, with Matryoshka truncation holding around 95 per cent of retrieval quality at a third of the dimensions. Weights are on Hugging Face and Kaggle now. Entirely local retrieval, nothing leaving the handset — unglamorous, and the most immediately usable thing released this week.
Source: Unite.AI, 6 October 2026