Saved daily version · 2026-11-16
Daily evidence notes
OpenAI reports progress on healthcare computer-use tasks and describes GPT-6 Astra’s contribution to a candidate physics construction. The healthcare ratings come from selected evaluations, not measured patient outcomes; the physics paper was submitted October 6, while the new account appeared November 15. A newly reviewed November 12 study shows why checking AI-generated research requires more than plausible prose: the benefit of verification tools varied by scientific field. These are bounded findings about models, tools and human oversight, not a demonstrated superintelligence threshold.
Sources checked and limits of this review
Review since the November 15 briefing. Opened primary watchlists from METR, ARC Prize, Epoch AI, Google DeepMind, OpenAI and Anthropic; searched current reporting and independent-research leads, then checked original evidence and dates. Selected OpenAI’s November 16 healthcare evaluation report and November 15 scientist account, cross-checking HealthAdminBench’s original project page and the physics paper’s October 6 arXiv record. The physics item is a new account of earlier research, not a newly dated proof. Independently examined the November 12 AI peer-review benchmark and its full methods and limitations; this is newly reviewed earlier evidence. Its authors also developed one tested reviewer, so it is an academic comparison rather than outside replication of their system. Google’s Hydrogen announcement is dated November 12, updated November 13; it was not treated as a November 16 launch. No further material new finding was verified on the other named watchlists within this bounded review. Coverage depends on accessible pages and indexing. Lab reports, researcher evaluation and SI interpretation are distinguished. No human editorial approval, overall wisdom verdict, stop/proceed recommendation or pivotal designation is claimed.
Evidence note 1 · Lab-reported evaluations and deployment accounts; independently developed benchmark; SI relevance is editorial interpretation
Healthcare agents: strong selected-task ratings still need outcome evidence
OpenAI’s November 16 report says 99% of 325 reviewed GPT-6 Astra healthcare tasks were rated clinically safe. Separately, it reports 94.3% of available rubric points on HealthAdminBench. That benchmark was developed by Stanford and Kinetic Systems researchers; this model result is reported by OpenAI.
Our interpretation. Editorial interpretation: moving between clinical records and insurer portals extends AI from answering questions into institutional workflows. Whether patients benefit depends on the surrounding approval process and error correction.
Main limitation. A safety rating on selected tasks is not a measured patient-harm rate. Rubric points are not the fraction of entire workflows completed correctly. The MRI demonstration uses synthetic data and prepares work for human review. Independent replication of these model results was not verified.
Next evidence. Prospective independent studies measuring errors, delays, clinician workload and patient outcomes, including whether review gates catch consequential mistakes.
Evidence note 2 · Lab-hosted scientist account of AI assistance; earlier research preprint; SI relevance is editorial interpretation
A new account of AI-assisted physics, based on an October paper
In a November 15 OpenAI guest post, physicist Nathaniel Craig says GPT-6 Astra helped turn his strategy into a candidate lattice definition of the Standard Model: a formulation on a grid representing spacetime. The linked paper was submitted October 6, not this week. Craig describes repeated model collaboration and input from human experts.
Our interpretation. Editorial interpretation: this is evidence worth examining for AI’s contribution to specialist research. The useful unit of analysis is the collaboration and its verified output, rather than the model’s account alone.
Main limitation. The construction remains a candidate. The arXiv record shows no journal reference, and this review does not independently validate the proof. Craig identifies unresolved questions about a condition connecting the lattice construction to probability-conserving quantum theory and about practical simulations. No controlled human-only comparison establishes acceleration.
Next evidence. Independent specialist scrutiny, checks of the stated mathematical conditions, and reproducible practical uses.
Evidence note 3 · Academic benchmark and developer evaluation, newly reviewed earlier paper; SI relevance is editorial interpretation
Newly reviewed: verification tools help AI peer reviewers unevenly
A November 12 academic paper tests twelve reviewer systems using 357 errors inserted into 30 astrophysics and machine-learning papers. Its Skepthical reviewer retrieves cited papers and uses algebra and code to check claims. Citation retrieval produced a clear lead in astrophysics, but no statistically distinguishable lead over the best single model in machine learning.
Our interpretation. Editorial interpretation: producing research faster makes checking it more consequential. Verification should be judged by the errors it actually catches across fields, not by the fluency of a review.
Main limitation. The authors developed Skepthical and evaluated it themselves. Each field contains only 15 papers; injected errors omit novelty and overall scientific merit. Judgments rely largely on models, with a human check covering one paper. Comparisons do not match computational budgets. Funding includes an unrestricted Google gift.
Next evidence. Independent replication with broader human adjudication, naturally occurring errors, additional fields and equal-cost comparisons.