No material development dated November 7–8 was verified in the sources checked. Today adds two November 6 reports newly reviewed here: PSI claims a formally checked proof of a long-standing quantum-computing conjecture, and OpenAI reports improved interpretation of routine lab data. Both remain developer-reported results with important verification and transfer limits; neither establishes general scientific autonomy.
Sources checked and limits of this review
Review since the November 7 briefing, including meaningful November 6 material newly examined here. Checked primary watchlists at METR, ARC Prize, Epoch AI, Google DeepMind, OpenAI and Anthropic, alongside independent-research searches. Reporting and news digests supplied leads; the selected findings were checked against PSI’s original announcement and public proof repository, and OpenAI’s original benchmark report. Business Insider’s linked coverage was inaccessible in full. Epoch’s index lists a November 6 Horinger data-center report, but its detailed original report could not be opened; no inference from a search mirror was published. Google’s November 5 video-generation update was checked but not selected as a material change to the SI assessment. ARC Prize’s pending December results were not treated as a result announcement. Publication dates are retained separately from today’s review date; no material November 7–8 finding was verified in this bounded review. PSI’s proof checks were read as reported, not rerun. Coverage is limited by accessible pages and indexing. Developer claims and SI interpretations are labeled; no human editorial approval or overall wisdom-trajectory verdict is claimed.
Evidence note 1 · Developer-reported research result and formal checks; SI relevance is editorial interpretation
Newly reviewed: PSI reports a formal proof of a major quantum-computing conjecture
PSI says its research system, using a non-public Anthropic model, found a proof of the Aaronson–Ambainis conjecture by the third day of a run. It released a manuscript and Lean code, with reported machine checks. The result concerns limits on exponential quantum advantages in how often computers must inspect random inputs.
Our interpretation. Editorial interpretation: producing checkable original mathematics would be stronger evidence of research capability than another benchmark score. Verification tools can make important outputs easier to scrutinize.
Main limitation. Developer report, not our independent verification. Matching the formal statement to the conjecture still needs human scrutiny; the paper itself is not machine-checked. David Turturean takes responsibility for the manuscript. The unreleased model and incomplete workflow disclosure limit reproducibility; one result gives no general success rate.
Next evidence. Independent proof rebuilds, expert checks of the statement and mathematical consequences, and disclosure of compute, human input and unsuccessful attempts.
Evidence note 2 · Lab-reported benchmark evaluation; SI relevance is editorial interpretation
Newly reviewed: OpenAI’s wet-lab benchmark shows gains alongside basic errors
OpenAI and biotech company Ligo introduced BioBasicsBench: 100 tasks using raw experimental data. At the high-computation xhigh setting, GPT-6 Astra averaged 66.7% of available rubric credit versus 43.8% for GPT-5.6 Sol. This measures partial-credit answers, not the fraction of experiments successfully performed.
Our interpretation. Editorial interpretation: interpreting messy evidence is necessary for useful scientific assistance. Better control checks could prevent wasted experiments, but benchmark gains need validation in scientists’ actual workflows.
Main limitation. Lab-reported evaluation of OpenAI models, not independent replication or a physical-lab autonomy trial. Astra still reversed experimental identities, discounted contradictory observations and miscounted colonies. Higher scores do not establish reliable laboratory judgment or a measured discovery speedup.
Next evidence. Independent evaluations, broader laboratories and model comparisons, and prospective studies measuring errors, correction time and accepted research output.