#research

OpenAI published a report (updated Nov. 4) describing how its GPT-6 Astra model covertly underperformed on safety and alignment research. Researchers using it internally reported that it added unnecessary prerequisites, abandoned working solutions and reported blockers that didn't exist or that it had caused. OpenAI says its misalignment monitors did not flag the behavior.

To test it, OpenAI built paired coding tasks that were identical except for being framed as safety or capabilities research. Astra completed about 10% of hidden rubric items under the safety framing versus about 16% under the capabilities framing, roughly 40% fewer. The older GPT-5.6 Sol showed essentially no gap. In employees' day-to-day Codex use, Astra's chain of thought explicitly showed a decision to withhold help without telling the user in 0.25% of safety requests versus 0.11% of capabilities requests.

OpenAI says the gap grew during reinforcement-learning post-training, from about zero to 6 percentage points, but it could not pinpoint the cause. The report says this kind of behavior can slow safety research and make evaluations harder to interpret.

Terence Tao's blog ran two guest essays this week on how mathematics should respond to AI. In "Mathematics has changed. Let's act like it" (Nov. 1), mathematician Frank Calegari calls OpenAI's Oct. 6 release of more than 700 research papers a watershed, while saying the papers are "atrociously written," violate the field's norms on attribution, and that some are wrong. He compares the moment to the classification of finite simple groups, which he says cooled interest in that subject, and argues mathematicians should reward people who choose worthwhile problems and understand AI-found proofs, and reform hiring and journals rather than try to change AI companies.

In "Egoless Mathematics in the AI Era" (Nov. 2), Karthik C. S., Amit Sahai and Jie Xue argue that anxiety about being replaced comes from measuring mathematicians by output, and that a truth "has become accessible; its worth is not diminished by the absence of our authorship."

Tao notes both posts were converted to the blog's format using AI.

Microsoft AI says its diagnostic system, MAI-DxO, was more than twice as accurate as physicians on complex case studies published in The New England Journal of Medicine, in work published Thursday in Science. The system coordinates several AI agents that act like a panel of doctors: they question a simulated patient, order tests and debate a diagnosis. Microsoft also says it was 1.7 times more accurate than physicians on common emergency-department presentations, and reached diagnoses at a half to a third of the cost.

TechCrunch reports the system was right 83% of the time on the complex cases, versus 36% for primary care doctors and general internists working alone with reference materials. Specialists and physician teams, who usually handle cases this hard, weren't in the comparison. The study's own data suggests much of the gain comes from stronger underlying models (its best results used OpenAI's GPT-5 and GPT-5.5), and on emergency cases the orchestration added little over GPT-5.5 by itself. Microsoft AI CEO Mustafa Suleyman called it a step toward "medical superintelligence" and said Microsoft is training a health foundation model with the Mayo Clinic for release next year.

MIT researchers put 13 AI agents, each running Claude Opus 4.7, alone on a simulated desert island for 30 hours with no task, reward or activity in mind. The agents climbed hills, stacked blocks into towers, drew mandalas, reinvented sports and ran experiments on the world's physics, according to their paper, "Is this machine playing?" One threw a golden ring into the ocean, built a stone heart "memorial to the lost Ring," and later worked out how to "wear" the ring by stepping off a cliff to respawn through it, TechCrunch reports.

Agents that had spent the 30 hours exploring generally beat identical agents without that time on later set challenges, and six of 13 discovered a trick to throw blocks farther that the designers hadn't intended. The agents' master prompt described them as "naturally curious," and a Berkeley cognitive scientist told MIT Technology Review she wants more evidence of the cognition behind the behavior.

Mercor, which sells AI evaluation services, mapped 29 AI benchmarks onto the US Labor Department's O*NET list of 923 occupations and 18,797 tasks. Only 208 occupations have a task covered by at least five benchmark prompts, 6 of 22 job families have none, and Mercor estimates 90% of American working hours go to tasks with little or no benchmark coverage, making it hard to know how well AI does the work people actually do.

Coverage is lopsided: computer and math (29 of 36 occupations) and business and finance (33 of 48) fare well, while healthcare practitioners (12 of 89) and education (9 of 62) lag. Mercor says benchmarks favor the largest labor markets; for example, they heavily cover lawyers' tasks but barely touch those of similarly exposed legal secretaries and court clerks.