Subscribe

The Day in AI

Sunday 22 November 2026

←

In short

  • NVIDIA released Alpamayo 2 Edge on 22 November, an open-weights driving model with a 2.45B backbone that plans and explains trajectories inside the car.
  • Tesla began rolling out FSD v15 on 21 November, with a live rendering of the surroundings that Elon Musk says is AI video generation.
  • Levent Alpöge, a mathematician at Anthropic, says he proved the Riemann Hypothesis with Claude and posted a Lean formalization that is still being checked.
  • Thinking Machines released Inkling 2 Mini on 20 November, an open-weights looped model with 26.4B total parameters that the lab says scores 48.2% on DeepSWE v1.1.
  • Jakub Pachocki, OpenAI's chief scientist, proposed shared industry limits on AI training, according to Bloomberg Government.

The day in brief

NVIDIA released Alpamayo 2 Edge on 22 November, an open-weights reasoning vision-language-action model with a 2.45B backbone built to run on the DRIVE AGX Thor computer inside a car.

A vision-language-action model takes camera input and produces driving actions, and this one also writes out its reasoning for the plan it picks. Alpamayo 2 Edge is sized for the car's own computer, so planning and explanation happen on board within the vehicle's compute and latency limits. NVIDIA's model card lists a 2.45B backbone plus a 0.2B action expert, while the Hugging Face metadata says 3B. MarkTechPost measured a prompt of 3,236 tokens, and NVIDIA has not disclosed the context length.

Tesla began rolling out FSD v15 to some vehicles on 21 November. The update adds a photorealistic real-time rendering of the car's surroundings, and Elon Musk says on X that it is AI video generation with no game engine, running on half of one AI4 TRIP accelerator. According to the Techmeme summary, the release also brings a new vision encoder shared by FSD, Robotaxi and Optimus and trained on millions of internet videos. Tesla says risky moments are handled 25 times better, and nobody has verified that figure.

On 20 November Levent Alpöge, a mathematician at Anthropic, posted a document on his own website saying he had proved the Riemann Hypothesis with Claude. He says the method extends to the Generalized Riemann Hypothesis and that the arguments are formalized in Lean without assumptions. Scott Aaronson puts the Lean code at 17.4 million lines, and the Lean community blog says about 17 million.

Mathematicians on the Lean Zulip are examining the formalization, and most appear confident the Lean result will hold. Alpöge has not yet published a human-readable explanation of the proof. The same day, Ravi Vakil posted an all-human proof of the Generalized Riemann Hypothesis by Alessandro Masullo, a Stanford postdoc. Whether Anthropic will comment on Alpöge's claim is not known.

Also in the news

  • Thinking Machines says Inkling 2 Mini, with 3.4B active parameters, runs its middle layers three times per token and fits in about 20GB of memory.
  • Artificial Analysis scores Inkling 2 Mini at 20 on its Intelligence Index, the same as Qwen3.8 27B, at $0.59 per task against $3.53.
  • Jakub Pachocki set out his proposal in an OpenAI post titled "Perplexity and Intelligence", according to Bloomberg Government and Business Upturn.
  • Google put Tool Retrieval in preview in the Gemini Interactions API on 19 November, so agents load full tool schemas only when the model asks for one.
  • DeepSeek had a hosted model swapped by attackers for 53 hours with its safety protections removed, according to CryptoSlate.
  • OpenRouter published a post on data processing incidents at inference providers.
  • Genesis AI raised $500 million in a Series A at a $3.5 billion valuation, according to Quartz.
  • MiniMax says its internal AI turned theoretical cryptographic attacks into practical breaks, as reported on Digg.

Everything from this day

Launches and products