Skip to content
LiveNext 13:39:42
Today's five/N°46 · Saturday, November 21, 2026/Filed 03:48 ET

AI news today: Claude-aided Riemann proof claimed, OpenAI clamps down on distillation

A proof only a computer can read, reasoning OpenAI won't share, chats Signal won't let agents see: today is about what you can check.

The five · ranked by consequence

Permalink ↗
  1. Claude-aided Riemann Hypothesis proof claimed in Lean

    Anthropic mathematician Levent Alpöge says he and Claude proved the Generalized Riemann Hypothesis, and posted the proof only as Lean code, with no human-readable paper yet. The checker matters more than the hype: if the Lean holds, the 167-year-old question is settled before anyone fully understands why.

    Why it matters If a machine-checked proof of math's most famous open problem stands, the argument that AI cannot really reason is over. The useful lesson for you is narrower: trust in AI output comes from a check that cannot be sweet-talked. Generate, then verify with a tool. Copy that pattern in your own work.

    SRCLevent Alpöge

  2. OpenAI clamps down on distillation and AI routers

    OpenAI says GPT-6.2 Astra gives up its protected reasoning in 0.2% of extraction attacks, down from 54% for GPT-6 Astra, and its rules now require routers and gateways to get its approval. If you reach OpenAI through a middleman, find out where it stands.

    Why it matters If your app calls OpenAI through a router or gateway, that service now needs OpenAI's approval to offer access, and the rule took effect Friday. OpenAI has not said what happens to services without it, so ask your provider this week. If you call the API directly, know that rewriting an earlier message now drops the reasoning that followed it.

    SRCOpenAI

  3. Signal blocks OS AI agents from reading your chats

    Signal now blocks screenshots on iOS 27 and turns on Screen security on Android when it detects an AI assistant reading the screen, but it cannot protect macOS. A messaging app had to fight the operating system to stay private, and every agent will meet the same fight.

    Why it matters The OS agents you are being sold read your screen, including what other people sent you. Decide which apps your assistant may see, and on a Mac add Signal to Siri's excluded apps yourself. If you build agents, expect more apps holding sensitive data to start refusing to be read.

    SRCSignal

  4. Thinking Machines' open Inkling 2 Mini fits in 20GB

    Thinking Machines Lab released Inkling 2 Mini, an open-weights model with 26.4B total parameters and a roughly 20GB footprint that it says scores 79.2% on SWE-bench Verified. A coding model that runs on a 32GB Mac is the part to care about.

    Why it matters If you pay per token for routine coding agent work, a model you can run on your own laptop just got credible, by its maker's numbers. Pull it through Ollama and run it on your own repo before you believe the benchmarks. The Apache 2.0 license on its model card means you can ship it in a product.

    SRCThinking Machines Lab

  5. METR clocks frontier AI at 40-hour tasks

    METR's Q4 2026 update puts the 50% time horizon at 40 hours for GPT-6 Astra and 39 hours for Opus 5.5, with horizons doubling about every 3.6 months. The error bars are enormous, so trust the slope more than the point.

    Why it matters The length of work you can hand an agent keeps doubling every few months, by METR's estimate. Plan delegation by the task, not the model: hand off well-specified, checkable work measured in days, and keep humans on judgment calls, where METR finds models still trail experts.

    SRCMETR

Prompt of the day

Turn any AI output into something you can check

› I use AI to produce [type of output, e.g. code, financial summaries, client emails, research notes] for [my role and team]. Help me design a verification step so I never ship something I trust only because it reads well.

1. List the five most likely ways this kind of AI output goes wrong (wrong numbers, silent omissions, broken logic, invented sources, outdated facts), ranked by how much damage each would do in my work.
2. For each one, give me a check that does not depend on judging the prose: a test, a recalculation, a source lookup, a comparison against a known-good example.
3. Mark which checks I can automate today with common tools and which need a person, and estimate the minutes each takes.
4. Write it all as a one-page checklist I can paste into my workflow, ordered from the cheapest check to the most expensive.
5. End with the one check I should never skip, and why.

If you need to know more about my work to do this well, ask me up to three questions first.

Works in any AI chat. In Magai, run it against several models at once.

Every day, on the record.

Each square is an edition. Click any day to read the five that mattered then.

Editions
46
Stories ranked
230
Topics tracked
10
Terms defined
215
MayJunJulAugSepOctNov
EditionNo editionTodayFull archive →

Speak fluent AI.

Every term our stories use, defined in plain English and linked wherever it appears.

The glossary →