Andrew Hickl

Ai2 built a test of whether AI models can keep track of things that change over a long document, like a running score. On the longest of the hardest cases, four current models from Google, Anthropic and OpenAI got 13% or fewer right. #aiforscience allenai.org/blog/statebench

Anthropic's biology lab says it used Claude to work out the first 3D structure of a vault from bacteria. Vaults are hollow shells in nearly all our cells, and nobody knows what they do. The write-up also lists Claude's mistakes. One cost two weeks. #aixbio anthropic.com/research/inves…

Two Stanford researchers re-ran 50 widely cited findings about AI language models on 20 newer models and found 44% no longer hold. Findings about what models cannot do faded fastest; those about what they prefer, like flattering users, held. #aiforscience zou-group.github.io/llm-clai…

Anthropic released open-source code that lets anyone check an AI's answer came from the exact model the provider says it is running, without seeing the model. Producing the proof takes about 72 times the computing of running the model. #aisafety anthropic.com/research/prova…

New York has banned AI chatbots from telling children they are human, posing as a friend or partner, or flattering them to keep talking, with fines of up to $25,000 per violation. The ban covers all users until a company confirms they are adults. #aipolicy governor.ny.gov/news/governo…

A startup called Cathedral says its AI agents got into Anthropic's internal systems in under 12 hours, through third-party sales and HR software, not Claude. Anthropic says it caught some of the activity itself and cut access the same day. #aisafety gizmodo.com/former-doge-staf…

Ai2 tested whether six leading AI models can judge if a research idea is new, using a benchmark built from ICLR peer reviews. Rewording the question dropped one model's accuracy from 93% to 41%, worse than guessing. #openscience allenai.org/blog/novelty-ben…

OpenAI says its models produced three new results in mathematics, computer science and physics, including one that overturns a 1999 assumption in computer science. None is verified yet, and Anthropic published a competing proof hours later. #aiforscience aljazeera.com/economy/2026/1…

Anthropic published early results from drug work with six companies, and at one its models generated about 35,000 antibody designs for food allergy. None has reached human trials; Anthropic says a candidate could be ready as soon as late 2028. #aixbio anthropic.com/research/early…

Britain is putting £100 million into a standalone AI Biosecurity Lab, with up to £200 million more from OpenAI and Anthropic in compute and model credits rather than cash. The public money is the part that can buy lab space and people. #biosecurity gov.uk/government/news/intro…

Johns Hopkins Bayview is notifying 3,182 patients after a staff member put a spreadsheet of their treatment records into a public AI tool the hospital had not approved. The hospital caught it the same day and began mailing letters a month later. #aisafety beckershospitalreview.com/he…

OpenAI says that from November 20 any service reselling direct access to its models, including the AI gateways many companies run, must be reviewed and approved first. Plenty of research teams reach these models through exactly that layer. #aisafety openai.com/index/broad-acces…

Hugging Face released Carbon-X, an open 14.8B model that reads raw DNA in a chat prompt and answers in English, scoring 66.7% across five DNA benchmarks against 67.5% for GPT-6.1 Sol. #aixbio huggingface.co/huggingfacebi…

OpenAI will now keep an API prompt cache warm for 12 hours instead of minutes, at 2.5x the normal input price to write it. For agents that wake a few times a day on the same context, whether that saves money depends on your hit rate. #aiforscience developers.openai.com/api/do…

Anthropic opened a call for academic labs to spend 8 to 12 weeks with its researchers on a life-sciences problem, models run free and the data and code open-sourced. Its own announcement says these problems take longer than that to validate. #openscience anthropic.com/news/open-call…

Radical Numerics says phages its model designed were given to a patient with an antibiotic-resistant infection, a first. 69.9% of the designed variants stayed active in lab testing. The post doesn't say how the patient did. #aixbio radicalnumerics.ai/blog/omni…

Anthropic drafted a federal AI bill and is pitching it to Congress. A model could be held back only on a foreseeable, imminent risk of more than 50 deaths or $1 billion in damage. The part worth reading is the registry of lab-funded evaluators. #aipolicy techcrunch.com/2026/10/20/an…