Blog

Jensen Huang Admits Affair as Nvidia Opens Probe, METR on Anthropic, ChatGPT Images 2.6 and More

A surprised robot looking at a soaring chart on a laptop — Nvidia and Jensen Huang, METR report, ChatGPT Images 2.6, weight editing, GPT-6.2 Astra, Claude Fable 5.5

Nvidia CEO Jensen Huang admits affair; board opens independent probe

  • In a November 8 email to employees, Nvidia CEO Jensen Huang admitted to an extramarital relationship and said he takes "full responsibility."
  • The New York Times had reported that the woman sought close ties with Huang to learn about Nvidia's technology, and that her employer, Megaspeed, had become one of Nvidia's largest buyers.
  • Nvidia's board formed an independent committee to look into reports that Huang was the target of a foreign intelligence effort. It says it has no reason to believe national security or company secrets were compromised, and Huang will keep leading the company during the review.
  • Beijing denied any spy plot, and the reports drew strong bipartisan backlash in the U.S. Congress.

CNN: Huang admits affair as Nvidia opens investigation · Axios: Nvidia's board to launch investigation

METR: AI research at Anthropic is moving "about twice as fast"

  • On November 5, the AI evaluation group METR published a report on how much AI is speeding up AI research at Anthropic, based on high-access review of Anthropic's internal material.
  • It estimates that Anthropic is improving capabilities on well-specified tasks about twice as fast as it did from March 2024 to October 2025, with a plausible range of 1.3x to 3x.
  • A separate check using benchmark scores found growth about 1.9 times faster since October 2025 (95% confidence interval 1.5–2.8x).
  • METR says its reviewers were not confident in the assessment, that Anthropic redacted much of the evidence it found most informative, and that a planned check on compute allocation was canceled partway through.

METR report: AI R&D acceleration at Anthropic (October 2026)

OpenAI's Brockman: death "could become a choice"

  • Speaking at a WSJ tech event, OpenAI president Greg Brockman said AI should make real progress on curing diseases within the next 12 to 18 months.
  • He also floated the idea that advances in biology could make death a matter of choice, without giving any timeline.
  • Critics pushed back that biology and clinical trials do not move at the speed of software.

Summary of Brockman's remarks on curing disease in 12–18 months

ChatGPT Images 2.6 adds GIFs and reference images

  • Released November 4 for every ChatGPT and Codex plan on web, desktop and mobile.
  • You can now create animated GIFs and save reference images of yourself, pets or characters to reuse later.
  • Developers get two API models: GPT-Image-2.6 Flare for speed (as little as about six seconds) and Sunburst for higher quality and more consistent edits.
  • Image markup for targeted edits and SVG export were added as well.

OpenAI: ChatGPT Images 2.6 · OpenAI: Flare and Sunburst API models

MIT's "weight-space programming": an AI edits another AI's weights directly

  • Early research shared by MIT CSAIL researchers in a November 2 blog post (not yet peer reviewed).
  • With no training loop, GPT-6 Astra read the raw weights of a small model (Qwen3.5-0.8B) and rewrote values directly.
  • Changing just 0.016% of the weights reportedly lifted the MATH-500 score from 7.9% to 48.1%, in the most generous test setting.
  • It failed, however, to tell an honest fine-tune from a sycophantic one by looking at the weights alone.

VentureBeat: MIT researchers used GPT-6 Astra to edit Qwen weights directly

GPT-6.2 Astra spotted in Codex usage

  • Some Codex users saw gpt-6.2-astra listed next to GPT-6 Astra in their usage analytics, suggesting some requests are being routed to a new model in testing.
  • OpenAI has not announced it, and it does not appear in the model documentation.
  • Separately, on November 5 OpenAI added GPT-6-Cyber, a security-focused variant of GPT-6 Astra, to its API changelog.

TestingCatalog: GPT-6.2 Astra appears in Codex usage

Claude Fable 5.5 said to be close

  • Reports keep pointing to an imminent release of Anthropic's next model, Claude Fable 5.5.
  • An entry briefly appeared in Microsoft Azure's model catalog before being removed, and some partners reportedly already have access.
  • Anthropic has not announced it; the newest released Fable model is still Fable 5.1.

TestingCatalog: Claude Fable 5.5 reaches partners

Blog

← Back to the list Back to top ↑