hiw3c 日报精选
Latent Space RSS 全文

[AINews] 黎曼猜想被证明……两次独立

原文标题 · [AINews] Riemann Hypothesis proved...twice independently

2026年11月20日 —

来源:Latent Space · latent.space

正文为英文,可一键机器翻译(仅首次需要等待)

Within a few hours of each other (timeline), two mathematicians claimed independent proofs of the Riemann Hypothesis, the most famous open problem in mathematics (Scientific American).

  • Levent Alpöge of Anthropic first made his Claude-driven proof in a note released with Lean code (1.3GB worth, ~250,000 page equivalent) rather than a full paper. Elliot Glazer confirms that the formalized statement is GRH and puts the chance of a Lean bug well under 1%. So if it finishes compiling with no extra axioms, it is probably correct, and no human will ever have read all of it…

  • …prompting Alessandro Masullo to release his 940 page, 100% human, work in progress proof he had been working on for the last 12 years (and had been undergoing secret manual review and revision per Kevin Hartnett).

This is the third in the trilogy of frontier math breakthroughs, following from the Navier Stokes solutions and the Quasi-Riemann bounding, but it doesn’t feel over. OpenAI's September Navier-Stokes claim is still the cautionary tale: critics argued that its Lean proof did not match the natural-language result. Both AI-claimed Millennium Prize results will take years to verify.

Apart from the incredible accomplishment of the most famous problem in mathematics, one cannot miss the contrast that the AI-written solution is probably correct but incomprehensible, and the other solution is probably technically incorrect but understandable.

Sponsored by Agentfield.AI

Open models are ready. Most coding harnesses aren’t. On DeepSWE, CodeAF finished #1 of ten on the same open model, ahead of Claude Code, at the lowest cost per fix.

Try CodeAF!

As Levent eloquently put it: “Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line.”

This also ties into other AI themes we’ve covered on the podcast and on this newsletter:

  • The nature of Slop: Levent says “I am presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics. It being obviously inappropriate to attempt to post some massive unreadable travesty of "slop", I decided to only present Lean.” One of the key tenets of the No Slop Manifesto is that you should understand what you ship (“Slop is when we put less effort into producing something than we ask the consumer to spend understanding it.”). And yet, this giant pile of Lean is verifiably correct and is incredibly significant in human history. Should we have delayed its release by a decade simply because we didn’t understand every single bit of it?

  • Weak to Strong alignment: the typical experience of using agents is Strong to Weak - we are general intelligences guiding not-as-smart AI to do things we know how to do well. We also often expect AI to automate most work “within distribution”, and expect it to run into more trouble at the limits of capability where the “centaur” pattern of human-AI-collaboration is required. Levent’s lived experience is the opposite - collaborating with Claude only in the early stages and then leaving him far behind with the critical parts involving automorphic forms on higher-rank groups. We are starting to need Weak to Strong alignment — the open ended domain of controlling AI smarter than us.

Sponsored by Backblaze B2

If your data is growing faster than planned, you need storage that grows with your business, feeds flash and GPUs with terabit-scale throughput, and doesn’t charge egress.

Backblaze B2 has delivered capacity tier storage for twenty years, manages more than six exabytes, and is trusted by leading neoclouds, labs, and gen AI companies.

Growth shouldn’t be something you sweat. We’ve got room for whatever you’re building.

Come and try it out!


AI Twitter Recap

Claimed Proof of the Riemann Hypothesis with Claude

  • The claim: Anthropic mathematician Levent Alpöge says that, "working with Claude," he has found a proof of the Riemann Hypothesis. He says the method also covers the Generalized Riemann Hypothesis (GRH) (Scientific American, statement summary, remark.pdf link).

    • Artifact: Only Lean code was released. A written-out prose version reportedly ran past 250,000 pages (timeline).

    • Timeline: The work began on Oct 6. Eight of the 20 steps landed on Nov 15, and the final entry on Nov 16 reads "GRH." Alpöge credits Claude for the parts involving automorphic forms on higher-rank groups, an area where he says he has little mastery (breakdown).

    • Unknowns: The Claude model used and the degree of human steering have not been disclosed (open questions).

  • Verification status: Elliot Glazer confirms that the formalized statement is GRH. The Lean code is "still compiling, will take a few hours." He puts the chance of a Lean bug at <<1% (analysis).

    • Provisional status: Artificial Analysis notes that both AI-claimed Millennium results are unverified: this one and OpenAI's September Navier-Stokes claim, which used an unreleased model (tally).

  • Competing human proof: Alessandro Masullo has announced a separate, purely human proof of RH. It is reported as a roughly 940-page GRH draft (announcement, page count).

  • Interpretation debate: Some argue the result shows we have little idea how hard math problems objectively are, rather than showing broad superhuman ability (Theo Jaffee, Ege Erdil).

    • Brute-force framing: Others compare the approach to olympiad-style "bashing" at a scale humans lack the willpower for (cloneofsimo).

    • Jaggedness: Omar Khattab notes the contrast between solving Millennium problems and being "this dumb on 1000 other things" (post).

  • Market reaction: ANT stock fell about 10% on the day (scaling01, Nabeel Qureshi).

    • Unconfirmed rumor: A DeepMind "math horizon" repo appearing the same day prompted speculation about Gemini math results (unconfirmed).

Thinking Machines Releases Inkling 2 Mini, an Open-Weights Looped Transformer

  • Inkling 2 Mini: A 26.4B-parameter looped MoE under Apache 2.0, pitched as small enough for a laptop. It scores 20.1 on the AA Index (launch, Mira Murati).

    • Architecture: There are 12 shared core layers run for 3 passes, with gated messages carried between passes (vLLM).

      • Active parameters: 3.4B unique active parameters, or 5.4B effective counting the loops. It has a 1M context window and accepts text and image input (AA).

    • Training recipe: Two-thirds of the experts were pruned during mid-training, followed by logit distillation from a larger Inkling (eliebakouch).

    • Claimed results: It leads open models under 100B on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%) (via teortaxes).

      • Comparison to Inkling 1.0: It reportedly beats Inkling 1.0 at 2.7% of the size (Ziqiao Ma).

    • Independent evals:

      • Vals: It ties Claude Opus 5.5 on CyberBench at under a quarter of the cost and keeps 96–97% of Inkling 2 Small's Vibe Code and SAGE scores (Vals, details).

      • Artificial Analysis: AA places it on the Pareto frontier for intelligence vs. active parameters (AA).

      • ARC-AGI-2: It reportedly scores 49.3% at $0.14/task, well below Inkling 2 Small (teortaxes).

  • Safety and interpretability rationale: Thinking Machines says the open release is meant to let researchers study how loops affect chain-of-thought monitorability (Alex Robey).

    • Variable-loop model: A companion research model, TML-d24-varloop-research, shows long-context scores (GraphWalks, MRCR v2) improving with more passes. Knowledge benchmarks such as SimpleQA and AA-Omniscience do not improve (Ziqiao Ma).

    • Goodfire analysis: Goodfire mapped information flow across layers, blocks and loops (Goodfire).

    • NDIF hosting: NDIF hosts both models with tooling for activation patching across loops (NDIF).

  • Serving and local availability: vLLM had day-0 support.

    • GB200 decode: TP1 decode on GB200 drops from 10.29 to 3.84 ms at batch 1, with kernels per step falling from 4,278 to 927 (numbers).

    • Local runtimes: Ollama 0.40.17-rc3 (Ollama), Unsloth GGUF quants targeting 16GB RAM (Unsloth), transformers, llama.cpp and MLX (HF).

    • Hands-on reports: Soumith Chintala reports about 75 tok/s decode on a DGX Spark and evals "roughly as good as Qwen 3.8 27B" (post).

Gemini 4 Neon: Developer Reception, Evals and Access

  • Developer reception: Strong practitioner praise for Neon's usability (Theo, systematicls).

    • Lucas Beyer comparison: In a side-by-side against Opus 5.5, Muse Wonder Full and Sol 6.2, all four models built a working dataset viewer. Neon's was the most data-aware (Lucas Beyer).

    • Cyber evals: Thoughtful Lab says Neon found subtle vulnerabilities other models missed while using a fraction of the tokens (Karina Nguyen).

  • Benchmarks:

    • Company Agent Bench: 87.4%, ahead of Fable 5.2 (82.2%) and GPT-6.2 Astra (79.7%) (results).

    • Agent Arena Index: Debuted at #2 with 1336 points (Arena).

    • FutureSim: Highest score, using 500+ subagents (post).

    • FrontierCode 2.0: 46.2% vs. Opus 5.5's 47.3%, at roughly a third of the cost ($4.38 vs. $12.82 per task) (roundup).

  • Pricing and access concerns: One tester paid $4.60 for 15M tokens on a simple demo and called it non-competitive against Anthropic's subscription limits (teortaxes).

    • Antigravity terms: Developers raised concerns about account bans and headless-use rules in Antigravity (linaqruf).

    • Third-party harnesses (attributed): Gemini subscriptions will reportedly work in any harness starting Dec 2 (kimmonismus).

  • Naming confusion: Users are unsure how Neon, Carbon, Argon and Hydrogen relate (Simon Willison).

    • Gemini 4 Hydrogen: The speed-focused model is live in the Gemini app and Search (Google).

  • API additions:

    • Dynamic Tool Retrieval: Supports deferred functions and remote MCP servers with server-side tool search (Phil Schmid).

    • Gemini 3.5 Transcribe: Voice editing is now in public preview in the Gemini API (Google).

    • GIF generation: Available in the Gemini app, powered by Gemini Omni 1.2 Flash (Gemini App).

Evals, Alignment and Measurement

  • METR research update: METR says AI R&D-relevant capabilities are improving rapidly and are getting harder to measure (METR,