AI Digest Nov 20, 2026

AI progress — 20 November 2026

Anthropic’s Nasdaq debut is the day’s biggest event: a record-breaking IPO that raised nearly $105 billion at $83 a share, according to TechCrunch. It makes a frontier lab a public company. Elsewhere, the technical news is incremental: a Codex release that changes the default model, and a Gemini API preview for large tool catalogs.

New models and papers

  • NarrowSpec: draft hidden width need not match the target — A speculative-decoding paper showing the draft model’s residual stream can run at half or a quarter of the target’s hidden width. Why it matters: if the result holds, draft models for EAGLE-3-style decoding can be made smaller without being tied to the target’s width (the authors’ claim; not independently checked).
  • On-Policy Morphing into Hybrid Attention Models — A method for converting pretrained Transformers into hybrid full/linear-attention models by choosing which layers keep full attention under a budget; the authors address two off-policy mismatches in FlashMorph’s layer selection. Why it matters: it targets the cost of retrofitting existing models for efficiency instead of training hybrids from scratch.
  • Tool Retrieval in the Gemini API (preview) — Google lets agents draw on hundreds of tools while using fewer tokens. Why it matters: anyone with a large tool catalog can test retrieval instead of loading every tool definition into the prompt.

New tools and software

  • openai/codex rust-v0.171.0 — The release adds GPT-6.2 Sol and Astra to the model catalog (Sol is the default), enables the TUI workspace by default, and adds codex exec --remote. Why it matters: the default model in the CLI changes on upgrade, so Codex users should check what they now run.
  • AgentCore payments supports AWS Marketplace x402 products — Agents can find and buy x402-enabled APIs, data and content and pay in USD on the AWS bill, with no stablecoin wallets or API keys. Why it matters: per-request agent purchases can go through existing AWS billing rather than separate payment infrastructure.
  • PostgreSQL 19 released — The release came two months later than usual, mainly because of a flurry of AI-reported vulnerabilities; Simon Willison highlights a new feature for apps that send reads to standbys. Why it matters: a major database release slipped because of AI-found security issues, which is a concrete maintenance cost for maintainers and operators.

Overall digest

  • Anthropic’s Nasdaq debut — Reuters covers Anthropic’s record-breaking IPO; TechCrunch reports the price was raised to $83 a share, raising nearly $105 billion. Why it matters: the lab behind Claude is now publicly traded and answers to public-market scrutiny.
  • Teens sue school and police after an AI gun alert flagged a Doritos bag — A lawsuit says over 99% of the school system’s AI gun alerts were false positives. Why it matters: it is a legal test of what happens when a detection system’s alerts trigger a police response; the figure is the plaintiffs’ allegation.
  • Token fraud is here and it’s bad — Claire Vo, who runs ChatPRD, describes in a video the flood of token fraud from groups abusing her product. Why it matters: anyone offering free or trial access to an LLM-backed product should expect this abuse and plan for it.
Share
Subscribe

Get the AI digest by email

One email per weekday morning: new models and papers, new tools, and the overall picture. No ads, unsubscribe in one click. See the privacy policy.