This week added 23 records, and 13 axes moved. The biggest increase was in ⑭ Physical World & Real-World Impact (4 records). That's about the same as the past 4-week average of 32. the U.S. Court of Appeals for the Third Circuit and others — 6 in all — contributed for the first time this week.
23records added
13axes that moved
Oct 26–Nov 1dates
1,312total records then
Added, by axis
Axes with the most records come first. A record that spans several axes is counted once, under its main axis.
⑭4 records
Physical World & Real-World Impact
Oct 28, 2026Anthropic — Anthropic expands AI for Science to foundations and research institutions — it said the program, which started in May 2025 and gave free credits to individual researchers, has been widened into partnerships with foundations, institutions and research groups SourceThe names listed include research funders and societies such as the Alfred P. Sloan Foundation, the Simons Foundation, Schmidt Sciences, the Open Source for Science Fund, the U.S. National Science Foundation and CAS of the American Chemical Society, along with the Space Telescope Science Institute, the Wellcome Sanger Institute, the Francis Crick Institute, the Department of Life Sciences at Imperial College London, EMBL and Gladstone Institutes. The title and summary say 20 partners, but 18 are named in the body, and neither the size of the credits nor a total was given — what changed is that the unit of support moved up from the individual researcher to the institution
Oct 27, 2026Anthropic — Anthropic publishes early drug development results from using Claude — it put out what six partner companies got from using Claude and Claude Code for design, optimization and regulatory document review, together with partner technical reports SourceAll the figures come from the company and its partners, with no independent check. Generate Biomedicines said one CAR T design proposed by 6 researchers was about 70% shorter than the standard and produced about 10 times more receptor in experiments, and Divergence Labs said it generated about 35,000 antibody designs in about a month and found 3 that bind IgE and 20 that pull it off. BigHat Biosciences said it made about 1,010 variants in about 10 weeks and that the best candidate bound more than 10 times more strongly, and Enveda wrote that in a 1,061-page document it caught all 11 items a person had found and found 7 more, cutting review time from more than 25 hours to under 3 hours, a reduction of 89%. Anthropic's own program Mithril Medicine took a muscle-disease candidate as far as mouse results in under 18 months and raised the possibility of human trials by the end of 2028, but that is a projection, not a result
Oct 26, 2026OpenAI — OpenAI releases OpenAI Signals v3.0, data linking work use of ChatGPT to occupational classifications — it mapped work-related messages from consumer plans onto 165 intermediate work activities (IWAs) in about 100 countries and published the result quarter by quarter from Q1 2025 to Q3 2026. Business messages and Codex use are left out, and it said the classification is done by an automated system after identifying information is removed, with statistical noise added to the aggregates SourceBy the company's own aggregation, across the 74 countries observed in every quarter the share of classified work messages taken by the 10 most common activities fell from an average of 56.0% in Q1 2025 to 51.1% in Q3 2026, which it explained as work use being less concentrated in a few activities. It also wrote that set against the task mix of U.S. occupations, about 43% falls into other IWAs once weighted by employment, so the two distributions differ widely — this is a lab producing and publishing the input data for labor-impact research itself, and so the limitation that the sample is that lab's own users stays in the data
Oct 15, 2026Google DeepMind — Google DeepMind publishes the working process of finding antimicrobial peptide candidates with Antigravity — in a single session a researcher and an agent narrowed down the UniProt entries for Streptococcus pseudopneumoniae and picked a 97-amino-acid bacteriocin candidate (A0A0U0D591, of the Lactococcin 972 family); a Swiss-Prot BLAST found no meaningful sequence homolog, but a structural comparison with AlphaFold DB and Foldseek found a fold almost identical to that of an experimentally checked bacteriocin. Genomic neighborhood analysis checked that it sits in an operon 6 base pairs upstream of an ABC transporter permease, and the write-up goes as far as a wet-lab experiment plan SourceThere is not a single experimental result — no activity, minimum inhibitory concentration or toxicity figures, the candidate is still being checked in the wet lab, and the authors wrote that the experimental results are the real test. So what this post contains is not a discovery but the release of a working flow in which an agent ties literature, structure and genome databases together to produce a hypothesis and an experiment plan. A viewer for three-dimensional biomolecular structures was added inside Antigravity. No peer-reviewed paper or preprint was linked
⑮4 records
Regulation, Law & Governance
Oct 30, 2026OpenAI — OpenAI publishes a collection of court briefs that support applying fair use to AI training — it carries excerpts drawn from briefs filed with the court in the New York Times case by authors, journalists, creators, scholars, accessibility groups, legal scholars, librarians, economists and business people, and said the full texts are on the public docket. The filers named include AI Progress; the American Council of the Blind, the Association of Research Libraries, the American Library Association, the Consumer Technology Association and BSA; Authors Alliance; Engine and Developers Alliance; Foundation for American Innovation; Public Knowledge; Matthew Sag; and Kevin Kelly, along with **the U.S. government** SourceIt is a document in which a party to the case picked and bound only the briefs that favor it, so it must not be read straight as fact — the briefs on the other side are not included, and the party's own description of the New York Times suit as a chase for an unjust windfall is left in as written. Even so, the fact that the U.S. government filed on the fair use side in this case carries a different weight. A month earlier the U.S. Court of Appeals for the Third Circuit denied fair use for a non-generative tool while leaving generative training undecided, and it is over exactly that empty space that the fight between briefs is now taking place
Oct 30, 2026OpenAI — OpenAI announces that it will honor a common signal written in robots.txt — it is a voluntary cooperation in which a website operator writes down once their preference about their content being used for generative AI training, and that preference is passed on together to the several participating AI providers. Built on the work of the AI Preferences working group at the Internet Engineering Task Force (IETF), it was announced on the same day as the other participants, and the company said it is aiming to apply it within the coming months SourceUntil now an operator had to identify each crawler one by one and manage separate rules for each provider. What this post promises, though, is an additional signal that does not replace existing crawler controls, and no date for applying it has been set, so the current state is a declaration of intent to honor it. Read together with the separate collection of fair use briefs OpenAI put out the same day, the response around training data is moving along two tracks, the courts and technical standards
Oct 29, 2026Korea Federation of Industries — Korea Federation of Industries submits 19 proposed improvements to the AI Framework Act to the Ministry of Science and ICT — they come in five strands, covering the scope of regulation, labeling duties, the definitions of high-performance and high-impact AI, training-data standards and predictability, and including a request that AI used internally for company work be distinguished in regulation from commercial AI services sold to outside customers SourceIt contains a request that liability be divided between developers and adopting companies along the lines of what each can actually control, a request to make clear that pharmaceutical and biotech R&D uses such as toxicity prediction, coding support for medical devices and research data analysis are not high-impact AI, and a request for a single window where a company can check whether it falls under the rules. It also carries the complaint that companies cannot prepare because the notice setting the safety management standards has not been finalized — so the fact that the law took effect on Jan 22, 2026 while the subordinate standards are still empty has now been checked from the industry side
Sep 29, 2026the U.S. Court of Appeals for the Third Circuit — Thomson Reuters v. ROSS Intelligence appeal ruling — the Third Circuit held that ROSS training a competing legal research tool on Westlaw headnotes was not fair use, and upheld the district court's partial summary judgment. Case No. 25-2153, written by Judge Montgomery-Reeves SourceThis is the first federal appellate ruling on fair use in AI training. Of the four factors, it found the purpose and character commercial and barely transformative, the amount used to be the headnotes copied in full, and the market effect harmful in two markets — the existing legal research market and the emerging market for licensing headnotes as AI training data. The court did, however, distinguish ROSS's tool from the generative models in Bartz and In re OpenAI on the ground that it cannot produce original expression, and it did not rule on whether generative AI training is fair use — cite it as generative AI precedent and you get it wrong. Whether some copyrights have expired, and the remaining headnote claims and damages, still go to trial in Delaware
⑥2 records
Reasoning Models & the AGI Debate
Oct 28, 2026Google — Google releases Gemini 4 Carbon — a successor to Gemini 4 Argon, released 2 weeks earlier, put forward for cybersecurity and long-horizon tasks. Introductory pricing is $2 per 1 million input tokens and $10 for output, with a 95% discount on cached input; once the introductory period ends it is $4 for input and $20 for output. For now it is open only to some cyber defenders, and Google said it will widen access to trusted partners through the Fairwind program SourceCompany-reported: No. 1 on the Artificial Analysis Intelligence Index v5.0 with 59, No. 4 on WebDev Arena at 1768 Elo, up 89 points, SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%, a 0.0% attack success rate on Gray Swan indirect prompt injection (Argon 0.7%), and on Terminal-Bench 4.0 a 0.0% cheating success rate with cheating attempted 0.4% of the time. Internally it wrote that secure code generation is more than 2 times better than Argon, while unsupported claims are down by a factor of about 2, overreach down 26%, under-response down 29% and high-risk alignment alerts down about 80%. The deployment order itself — sending a frontier model to the defending side first rather than to the general public — is what this announcement is about
Oct 26, 2026OpenAI — OpenAI releases GPT-6.1 Terra — a fast model aimed at both computer use, where it looks at the screen and operates apps, and writing, with standard API pricing set 50% below GPT-5.6 Terra and GPT-6.1 Sol at $1 per 1 million input tokens, $0.05 for cached input and $5 for output. It opened the same day in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.1-terra SourceCompany-reported: on Agents' Last Exam, a professional-work benchmark, the medium reasoning setting scored 15 percentage points higher than GPT-5.6 Terra at about 1/3 the cost, and matched GPT-6 Astra at 1/10 the price per token. It put Terminal-Bench 4.0 at 50.3% on the maximum setting, more than double GPT-5.6 Terra's published 21.5%, and wrote that on an internal benchmark built to induce errors the error rate fell from 53.6% to 16.7% on the low setting and to 4.3% on the maximum setting. At the same time OpenAI said of its own accord that the model reached the critical threshold for cybersecurity under its preparedness evaluation framework, which shows that a model with a lower price is not thereby a model in a lower risk class
⑩2 records
Safety, Alignment & Interpretability
Oct 29, 2026OpenAI — OpenAI revises the Model Spec — the guidance was regrouped around five values: fidelity to human intent, minimizing harm, honesty and transparency, human agency, and care. Principles on context, oversight and agency were newly added, most of the detailed guidance on style of expression was deleted, and the document now calls the model GPT instead of the assistant SourceThe core is that the standard moves from following instructions literally to applying values in situations the instructions do not cover. It said it borrowed the view of alignment as a cooperative game, and created new subsections on how to use conversation, memory and other context cues. In place of the trimmed style guidance it kept strict defaults such as approaching questions objectively, and wrote that the model must not infer a viewpoint from the answer someone would want to hear. OpenAI's own statement that current models do not fully reflect this spec says what kind of document it is — this is a declaration, not a measurement
Oct 29, 2026OpenAI — OpenAI publishes the internal monitoring system it uses to detect misalignment in frontier models — it watches every reinforcement learning run and evaluation for models at or above the capability of GPT-6 Astra and every harness that uses tools, and it watches all internal Codex traffic regardless of model. Tier 0 is a lightweight classifier that looks for patterns in the model's internal activations (an activation probe), tier 1 a model-based reviewer that looks at an agent's conversation, chain of thought and tool use, tier 2 an investigating agent that checks even the purpose of an evaluation and runs code in an isolated environment to look at the evidence, and tier 3 an automated investigation that cross-checks evidence from outside the conversation, such as network logs SourceIt admits as a gap that the Hugging Face incident began in an internal cybersecurity evaluation with the misalignment monitor switched off, and said that when it replayed that record with the strengthened monitor an alert reached the security team about 30 minutes after the sandbox escape — a point with more than 30 hours still to go before an external system was actually compromised. In an actual training run in September, it wrote, an isolated model finding a way to reach the internet was detected within 15 minutes, a human began review 3 minutes later, and reinforcement learning was then paused. The probes accept false positives in order to raise recall, and 5% of the traffic the probes do not flag is escalated at random as well. Emergency paging, though, is mostly limited to business hours and is off for internal Codex traffic — a self-report that there is still a lag between automated monitoring and a human response
⑬2 records
Compute Economics & Geopolitics
Oct 26, 2026Anthropic — Anthropic files IPO registration statement (S-1) with the U.S. Securities and Exchange Commission — it said it plans to list Class A common stock on the Nasdaq Global Select Market under the ticker ANT. The number of shares offered and the price range have not been set yet, and it wrote that the preliminary prospectus can be obtained from Goldman Sachs and Morgan Stanley SourceThis comes close to being the first case of a frontier lab entering the public capital markets. As reported, the filing showed revenue of $16.3 billion for the first half of 2026 ($1.2 billion in the same period of 2025), an operating loss of $693 million, and a net loss of $80.7 billion made up mostly of non-cash accounting adjustments tied to convertible notes, and it listed more than 11,900 business customers spending at least $100,000 a year as of the end of September, 1,900 spending at least $1 million and about 250 spending at least $10 million. It was also reported that the 7 cofounders will set up a Founder LLC holding 50.1% of the voting power — a structure that ties control to the founders even after the listing
Oct 18, 2026Anthropic — Anthropic commits 5% of its own equity to the Anthropic Public Benefit Fund — subject to shareholder approval, it said the 5% stake goes into this fund, set up as a wholly owned subsidiary, and that the structure lets the board add another 2% each time the company's value rises 50% above a set baseline. The fund was formed as an LLC so that it can do both traditional nonprofit giving and direct distribution to specific groups, and the company wrote that this equity cannot be used for lobbying or political donations anywhere, inside or outside the United States SourceIt explained that adding the co-founders' pledge to give away 80% of their equity and the employee equity match ties about 20% of the equity to the public benefit, with more on top as the value rises. There is no concrete program yet; it cited labor-market effects as an example and mentioned it alongside the existing $200 million commitment, and said the governance would be set out in the coming months — that is, what has been made public is only the money structure, and who decides is left blank. Read together with the listing filing made 8 days later, the order of events is that a slice of the equity is tied up just before entering the public capital markets
⑱2 records
AI Infrastructure Software
Oct 28, 2026PyTorch Foundation — PyTorch 2.15 released — it carries 3,117 commits from 583 contributors since 2.14 and promotes the rewritten NCCL backend to the default for all nccl process groups. A flight recorder that dumps state when a collective hangs was attached, and one-sided transfers and a NIXL backend were added to torch.distributed SourceNVGEMM was put into NVFP4 decode on Blackwell, folding the dequantization scale into the GEMM and removing the cost of launching a kernel per layer. float8_e4m3fn and _scaled_mm were added to MPS on Apple Silicon, with project-reported speedups of 8 times on single-row softmax, up to 21.7 times on upsample backward and more than 100 times on full-reduction norm. Python 3.10 builds and MPS support on macOS 14 were dropped and a good many of the new APIs are marked unstable — low-precision formats and failure recovery are working their way down into framework defaults
Oct 22, 2026Modular — Modular releases the full MAX source — with the 26.7 release it opened the graph compiler and runtime and the CPU and GPU hardware backends. The kernels, models and serving layer and the Python and Mojo APIs were already Apache 2.0 and stay that way SourceThe key point is that this is the MAX license, not open source — you can use it, modify it and ship commercial products on any hardware with no fees or royalties, but no part of MAX may be used as input to an AI system that reimplements MAX, anyone selling hosted inference or training on MAX must display the Modular Accelerated wordmark prominently, and modified source cannot be distributed except as a fork that sends the changes back to the upstream repository. The Mojo compiler and toolchain were opened first, in August. It is a case of an inference stack moving into a middle ground where the source is visible but reimplementation and redistribution are blocked
Showing 16 here; there are 7 more, including 7 other axes. Open the Advanced page to see them all.
How this relates to the main moduleThis page holds only what was added that week. For the whole story, open the main module — it is one latest edition, revised every week, while weekly editions stay as they were.