Saved daily version · 2026-10-27
Daily evidence notes
OpenAI announced major scientific proof claims and released a cheaper computer-use model that it classifies as Critical in cybersecurity. Both developments require scrutiny beyond headline results: the scientific claims await open expert examination, while the model’s deployment safety remains a lab assessment. Independent mathematicians also called for immediate public release of the withheld physics proof artifacts.
Sources checked and limits of this review
Reviewed developments since the latest saved briefing, October 25, including the October 26 gap. Checked METR updates, ARC Prize’s blog, Epoch AI’s latest work, Google DeepMind news, OpenAI research/product announcements and safety documentation, Anthropic research, and relevant independent researchers. Reputable reporting supplied leads; selected claims and dates were checked against original announcements, the linked OpenAI proof repository and AGMAI’s dated statements. No comparably material new primary finding was verified from METR, ARC Prize, Epoch AI, Google DeepMind or Anthropic in this window. This is bounded coverage, not an exhaustive search. Source dates are publication dates, not dates of the underlying model outputs. SI relevance is editorial interpretation; no overall wisdom-trajectory judgment, proceed/stop recommendation or human approval is claimed. The source-opening helper failed; editorial instructions were read from the previously verified version-25 checkout. Persistent briefing storage is separate from source deployment.
Evidence note 1 · Lab-reported scientific claims; external examination pending
OpenAI claims new scientific proofs; public verification is only beginning
On October 26, OpenAI announced model-generated proofs concerning wave behavior, the Exponential Time Hypothesis and quantum field theory. It published mathematics and computing artifacts, including claimed Lean formalizations, while withholding the physics artifacts for an external group’s writeup. The computing claim challenges the belief that certain Boolean satisfiability problems necessarily require exponential time. OpenAI describes its outputs as raw research data, not the end of scientific understanding.
Our interpretation. Editorial interpretation: independently confirmed solutions to longstanding open problems would be stronger evidence of research capability than answering known benchmark questions. But a formal proof checks a statement as encoded; experts must still establish that the encoding matches the intended problem and explain its consequences. Wise research institutions need that distinction to remain visible rather than treating a lab announcement or a famous reviewer’s name as settled knowledge.
Main limitation. These are lab claims, not independently confirmed results in this review. We did not run the Lean code or audit the proofs. The physics materials are not publicly available, and theoretical advances do not by themselves establish practical algorithms or empirical physics predictions.
Next evidence. Watch for reproducible proof-checker builds, audits of theorem statements and assumptions, independent explanations and corrections, public physics artifacts, and disclosure of model assistance, attempts and compute. Practical computing implications require separate analysis.
Evidence note 2 · Lab-reported model evaluations and risk classification; independent replication pending
Cheaper computer-use capability also carries OpenAI’s highest cyber classification
OpenAI released GPT-6.1 Terra on October 26. Its launch tests report a 15-percentage-point gain over GPT-5.6 Terra on Agents’ Last Exam at medium reasoning effort, at about one-third the task cost. Its system card classifies Terra as Critical in cybersecurity and High in biological and chemical capability, using the Astra/Sol safeguards stack. The card also reports mixed safety outcomes, including lower performance than its predecessor on an adversarial dynamic mental-health evaluation.
Our interpretation. Editorial interpretation: progress toward powerful AI includes making existing capabilities cheaper and easier to deploy, not only raising the best available score. Lower costs could widen useful automation and the scale of potential misuse. Whether this is a wiser deployment depends on actual authorization boundaries, oversight and consequences—not on efficiency or an aggregate safety label alone.
Main limitation. The comparisons and safety judgments are lab-reported. Evaluation prompts, tools and reasoning settings can differ from production; some comparison models were rerun in later versions. A benchmark improvement is not a guarantee of reliable professional work, and “Critical” is a capability classification, not proof that safeguards fail or succeed.
Next evidence. Look for independent equal-budget replications, real-workflow error and unauthorized-action rates, prompt-injection tests across connected tools, and evidence that safeguards remain effective when lower prices increase usage. Track regressions alongside headline gains.
Evidence note 3 · Independent advisory recommendation; not a technical evaluation
Independent math advisers say proof access should not depend on lab-selected gatekeepers
The Advisory Group on Mathematics and Artificial Intelligence called on OpenAI on October 26 to release its Yang–Mills, QCD and Standard Model proof outputs immediately. Its October 25 statement opposed labs designating particular mathematicians to lead interpretation, citing unequal access, constrained community agency and borrowed reputation. It recommends independently hosted, versioned artifacts, production details and support for community-led understanding. The group says it operates independently and receives no payment for this advisory work.
Our interpretation. Editorial interpretation: as AI generates more research than people can immediately check, control over access and explanation can shape what becomes accepted knowledge. Open scrutiny and distributed expertise are conditions for correcting errors and preserving researchers’ agency. That is a governance question about the path toward wise SI, not a measurement of the model’s intelligence.
Main limitation. This is an independent advisory recommendation, not a proof evaluation or consensus of all mathematicians. Open access alone cannot guarantee correctness, fair credit or enough verification capacity. The tradeoff between rapid public release and a clearer initial explanation remains unresolved.
Next evidence. Watch for release of the withheld artifacts, independently chosen and funded review efforts, documented access and conflict-of-interest policies, and evidence that broader participation yields reproducible checks and clearer explanations.