18 Jun 2026

Digest: OpenAI's autonomous AI chemist, new life-science benchmark

A near-autonomous AI chemist headlines the day, alongside a new expert-built life-science evaluation, methods for auditing model deception, and chip-control politics in Congress.

1. A near-autonomous AI chemist improves a challenging reaction in medicinal chemistry

   OpenAI and Molecule.one report that a near-autonomous system built on GPT-5.4 reviewed literature, proposed and ranked experiments, and improved a hard-to-run medicinal-chemistry reaction, with human chemists steering and validating the result. It is an early demonstration of a frontier model driving much of the scientific research loop end-to-end, with clear dual-use relevance for chemistry.

   OpenAI
2. Introducing LifeSciBench

   OpenAI released LifeSciBench, an expert-authored and expert-reviewed benchmark for how AI systems handle real-world life-science research tasks — the kind of evaluation that bears directly on tracking bio-relevant capabilities in frontier models.

   OpenAI
3. “Did you lie?” Evaluating Lie Detectors across Model Scale and Belief-Verified Model Organisms

   New work evaluates LLM 'lie detectors' by building model organisms that verifiably believe the opposite of what they say, finding most existing testbeds fail to meet that bar — a methodological step toward auditing and monitoring deception in models.

   Alan Cooney via LessWrong
4. Will the MATCH Act Change Chip Controls?

   An analysis of the proposed MATCH Act examines Congress's push to take a more direct role in semiconductor export controls toward China, a key lever over the compute underpinning frontier AI development.

   Aqib Zakaria via ChinaTalk