The Polylog AI Intelligence Brief

Morning Edition · Saturday, November 21, 2026

Tech

Anthropic mathematician releases a Lean-checked claimed proof of the Generalized Riemann Hypothesis made with Claude

Early independent checks report that the Lean code compiles on the standard axioms and encodes the Generalized Riemann Hypothesis. On the same day, the American Mathematical Society's president posted a separate, human-written proof draft.

5 sources
Plausible

Markets

Anthropic closes its first Nasdaq day at a $1.76 trillion valuation, and SoftBank shares fall more than 15% the next day

The offering raised $81.3 billion, and Masayoshi Son called SoftBank's choice to back OpenAI instead of Anthropic one of his mistakes.

4 sources

Tech

OpenAI prices GPT-6.2 Astra at $8 and $40 per million tokens, the same list rate as Anthropic's Claude Fable 5.2

GPT-6.2 Sol keeps GPT-6.1 Sol's $2 and $10 rates and scores within half a point of Astra on the FrontierCode 2.0 coding benchmark, according to third-party compilations of OpenAI's figures.

3 sources

Tech

OpenAI says GPT-6.2 Astra leaks protected reasoning in 0.2% of adaptive extraction attacks, down from 54% on GPT-6 Astra

The company says it now acts against thousands to tens of thousands of accounts a week and shares indicators of compromise with other model providers.

3 sources
Plausible

Tech

Anthropic says Claude Fable 5.2 agents solved two open black-hole physics problems in six weeks of largely autonomous work

An Anthropic researcher gave each of 22 problems that the model proposed to its own agent, and one reported result uses holographic duality to show that abrupt black-hole formation also occurs in very small universes.

3 sources
Plausible

Tech

Epoch AI finds agentic inference capacity of AI chips sold grew 4.6 times in a year, against 2.6 times for peak compute

Nvidia's Blackwell generations accounted for 74% of the added inference capacity, according to Epoch's new Inference Performance Index, which accounts for memory bandwidth.

3 sources

Tech

Nvidia and Microsoft open a limited Azure preview of Vera Rubin NVL72 systems, with public preview planned for early 2027

The companies also extended confidential computing to rack-scale AI and added Nemotron 3.5 open models to Microsoft Foundry, but they have not published pricing for Vera Rubin virtual machines.

3 sources

Tech

OpenAI's MedChemBench finds models rebuild molecular structures reliably but score about 50% on compound selection

The internal benchmark draws 99 problems from 117 papers and patents, and OpenAI warns that those sources may be in the models' training data.

3 sources

Tech

Perplexity's Multiagent mode lets Claude Fable 5.2 direct up to 16 agents running rival labs' models in parallel

The agents exchange findings and cross-check each other before a single answer returns, and the mode is open to Perplexity Max subscribers on the web.

3 sources

Tech

Dario Amodei says Anthropic could already move far faster with models building themselves but has chosen a slower pace

Speaking to CNBC on Anthropic's first trading day, the chief executive compared the company to a driver who could reach 200 miles an hour and holds at 90 to leave time for safety work.

3 sources
Plausible

Tech

A public GitHub repository publishes extracted system prompts for Claude Fable 5.2 and other production models

The asgeirtj/system_prompts_leaks collection covers ChatGPT, Claude and Gemini, and its Claude files contain behavioral rules such as a ban on cursing unless the user asks for it or swears heavily.

2 sources

Tech

New arXiv papers aim to predict model-merging interference and control perturbation density in sparse adversarial attacks

One paper targets the cost of evaluating candidate merges to tune task-vector coefficients, and the other replaces the fixed sparsity setting used by generator-based black-box attacks with feedback control of perturbation density.

2 sources