Subscribe

The Day in AI

Saturday 21 November 2026

←

In short

  • NVIDIA released Nemotron-Labs-Sigma on 20 November, 14 open checkpoints of a diffusion language model that denoises text in a 16-dimensional continuous space.
  • Levent Alpöge, an Anthropic mathematician, says he proved the Riemann Hypothesis with Claude and posted a Lean formalization that nobody has independently verified yet.
  • Thinking Machines released Inkling 2 Mini, an open-weights looped model with 26.4B total parameters that it says scores 48.2% on DeepSWE v1.1.
  • Tristan Buckmaster and Levent Alpöge published an AI-aided Navier-Stokes blowup proof on 21 November that they say closes a loophole in OpenAI's September solution.
  • Tencent opened a public beta of Computer Use in its WorkBuddy agent, which operates Windows and macOS apps that have no API.

The day in brief

NVIDIA released Nemotron-Labs-Sigma on 20 November, a family of 14 open-weights checkpoints at 3B and 8B parameters that form a continuous diffusion language model.

A diffusion language model starts from noise and refines a whole passage over several steps, where a standard model writes one token after another. LLaDA, Dream and most other diffusion language models do that refining on discrete tokens, and Sigma does it in a smooth 16-dimensional embedding space. NVIDIA says the continuous space makes few-step distillation and trajectory steering practical for text. The checkpoints have a 4,096-token context, or 8,192 tokens for Sigma-8B-instruct, and the release record gives no benchmark scores, so how Sigma compares with token-based models of the same size is unknown.

On 20 November Levent Alpöge, a mathematician at Anthropic, posted a document on his own website saying he proved the Riemann Hypothesis with Claude, with a method that extends to the Generalized Riemann Hypothesis. He says the argument is formalized in Lean without assumptions, in about 17.4 million lines by Scott Aaronson's count. Mathematicians on Zulip are examining it and most appear confident the Lean result will hold, but it has not been independently verified and a human-readable explanation is still to come. The same day Ravi Vakil posted an all-human proof of the Generalized Riemann Hypothesis by Alessandro Masullo, a Stanford postdoc.

Thinking Machines released Inkling 2 Mini on 20 November as open weights. It's a looped mixture-of-experts model with 26.4B total and 3.4B active parameters, and it runs its middle layers three times for every token, so it does more computation per token and reasons with fewer output tokens. Most weights were trained and deployed in NVFP4, which gives a footprint of about 20GB and lets it run on a laptop. SenseTime's SenseNova-Loop-9B, released on 17 November, used the same looping idea. Thinking Machines also released a research model that loops one to five times so outside researchers can study monitorability.

Thinking Machines reports 48.2% on DeepSWE v1.1, against 42.2% for Qwen3.8-27B. Artificial Analysis scores it 20 on its Intelligence Index, the same as Qwen3.8 27B, at $0.59 and 9 minutes per task where Qwen takes $3.53 and 38.2 minutes. On OpenRouter it costs $0.075 per million input tokens and $0.30 per million output tokens.

Also in the news

  • Tristan Buckmaster of NYU and Levent Alpöge published an AI-aided proof of finite-time blowup for the Navier-Stokes equations on 21 November, and they say it closes a loophole in OpenAI's September solution.
  • Tencent opened the WorkBuddy Computer Use beta on 20 November, which needs client 5.11.0 or later, and Tencent advises users to stay present for sign-in, payment and deletion steps.
  • RadixArk released its first model, DaisyRose-27B, on 18 November, a reinforcement learning fine-tune of Qwen3.8-27B that it says raises Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%.
  • OpenAI released GPT-Live-1.1 in the API on 19 November with better multilingual pronunciation and transcription, plus a standby setting that pauses inference when nobody is speaking.
  • Google added a collapsible sidebar listing AI Mode chats to the Search homepage on 19 November, rolling out globally on desktop.
  • Anthropic shares slid after its IPO banks exercised the greenshoe option to sell extra shares, according to The Information.
  • Oracle says it has completed the funding for its infrastructure contract with OpenAI, according to The Information.
  • Ars Technica reports that researchers who put canary credentials in requests to AI routers such as OpenRouter and Vercel saw keys stolen, even under zero data retention policies.

Everything from this day

Research