AI's cure pitch hits pharma skepticism
-
Nº CXXXV
- Date
- 04 Nov 2026
- Issue
- 135
- Stories
- Seven
- Editor
- ARC
Pharma CEOs are unconvinced, and one preprint finds research agents do worse when they think longer.
Drugmaker CEOs doubt Silicon Valley's AI cure promises
Drugmaker CEOs aren't buying Silicon Valley's promise that AI will cure disease, the Wall Street Journal reports, in a piece that drew discussion on Hacker News (the tech forum run by startup investor Y Combinator). The doubt comes from the executives whose companies would have to carry any AI-found molecule through trials, manufacturing and regulators. Curing disease has become one of the AI industry's headline promises for its models. But the companies that build the models are the ones making the claim, and the companies that would have to prove it in patients are the ones judging it. In the Journal's account, that second group is not yet convinced.
Longer AI reasoning fails at automated bioinformatics
Extended reasoning fails in autonomous bioinformatics, a bioRxiv preprint argues, and proposes a fix. Extended reasoning is the mode in newer AI models that works through a problem at length before answering, and it is usually expected to give better answers. The preprint looks at agents (software that runs multi-step work on its own) inside scientific sandboxes, the isolated computing environments where they write and run analysis code. If the result holds, the most expensive setting on a model is not automatically the best one for automated analysis.
El Agente Fármaco brings an AI scientist to drug discovery
El Agente Fármaco brings the AI-scientist approach to drug discovery, according to a new arXiv paper whose authors stress that the system is built to scale. An AI scientist, in the field's usage, is an agent system that carries research through several steps on its own. The name is Spanish for "the drug agent." Its claims run into the question raised by the bioinformatics preprint above: how well do agents reason when nobody is checking each step?
LabWeaver ties AI data analysis to laboratory knowledge
LabWeaver grounds analysis agents in laboratory knowledge, according to a bioRxiv preprint. Its authors argue that this grounding is the way to get agent-run data analysis that scientists can rely on.
MedPhysBench tests language models on medical physics math
MedPhysBench tests language models on medical physics calculations and on catching errors in them, in a new arXiv paper. The benchmark, a shared test, scores both getting the math right and spotting a wrong answer.
Claude checks 123 libraries' code with mathematical proofs
Anthropic's Claude formally proved security properties in the compiled code of 123 open-source libraries and found bugs that had gone undetected for years. Formal verification (mathematical proof that code does what it should) has rarely been practical at this scale.
OpenAI pitches ChatGPT and Codex for hardware engineering
OpenAI pitches ChatGPT and Codex, its coding agent, as tools that speed up CAD, circuit-board design and simulation work for engineers building physical things.
Reply with your discoveries. A human reads them. Forward freely.
|