ZeroNoise Logo zeronoise
Post
Lean Community Certifies the Claude-Assisted Riemann Proof as Users Adopt Gemini 4 Neon
•
5 min read
• 703 docs
Lean maintainers and independent checkers confirmed Levent Alpöge's machine-checked proof of the Riemann Hypothesis. Separately, Gemini 4 Neon picked up high-profile users, Agent Arena data showed GPT-6.2 Astra trading writing quality for coding, and an AI provider disclosed a security incident.

The Riemann proof is formally certified, but no human has read it yet

The previous brief covered Levent Alpöge's claimed Claude-assisted proof. It has now gone from "very likely correct" to formally certified. The Lean FRO and the Mathlib Initiative issued a joint statement certifying the proof as correct. Alpöge had released it on November 20 as a 17-million-line Lean library, and the communities verified it over the following day . Scientific American reports that independent mathematicians finished a full verification and say "the results are real" .

The checks ran on more than one independent system:

  • Lean and nanoda: One independent run compiled all 175,187 files (17.37M lines) with unmodified Lean 4.33.1. It took about 2 hours and cost $12.53 . It then checked 1,484,729 declarations in nanoda, a Rust kernel that shares no code with Lean, with zero errors and only the three standard axioms .
  • MIT run: A separate report says MIT's Andrew Sutherland ran the proof through other independently written kernels, and all accepted it. That makes a Lean bug "essentially ruled out" .

The checks have limits. They settle the Riemann Hypothesis plus four GRH theorems for Dirichlet L-functions, not general GRH. Humans "still have to read and understand what the proof is doing" . No readable proof exists yet, and the Clay prize requires publication plus two years of review .

One dissent stands out. Polymarket reports that mathematician Jenny Lorraine Nielsen claims the proof is invalid despite passing Lean verification . No details of her argument were available.

The reactions moved quickly into politics:

  • Tim Gowers, a Fields medalist, posted that he went on a pause-AI march the day after seeing his grandson .
  • Boaz Barak said that within a couple of months, if not already, someone with no mathematical expertise could solve GRH with internal models and enough compute. He said this was extrapolation, not internal knowledge . Talia Ringer countered that Alpöge has been interested in RH since his math PhD .

Formal proof search also produced a result outside math. A Tübingen–Google DeepMind paper reports that AlphaProof Nexus found a Lean proof settling a quantum-optics question first raised in 2017. High-dimensional GHZ states cannot be built using only linear optics and pair sources for any even photon count of 4 or more, when there are at least as many dimensions as photons .

Gemini 4 Neon wins converts, and Google adjusts access

Hands-on adoption is now backing up last period's benchmark results:

  • Aravind Srinivas says Neon on High reasoning feels intelligent and fast at once. He uses it on Perplexity Computer and has moved some of his Teammates to it .
  • Omar Sar is "strongly considering" replacing Opus 5.5 with Neon as his daily driver. He also wants subscription access in other harnesses .
  • Jeremy Howard notes that Google described Neon at launch as only "approaching" frontier models, and argues that evals are "a very imperfect measure of model utility" .

Google also changed how much users get and where they can use it:

  • Quota: Antigravity reset everyone's Gemini quota and said caching improvements raise what users can do within it by 22% .
  • Third-party harnesses: Patrick Toulme said users will "NOT be banned ever" for using Antigravity in T3 Code .

Effort settings cost very different amounts. Simon Willison's Gemini 4 Hydrogen test cost 3.99¢ and used 5,311 output tokens on "high," versus 23.61¢ and 31,477 tokens on "xhigh" .

GPT-6.2 Astra gains in coding and loses in writing

Agent Arena ranks GPT-6.2 Astra (Max) #6, one place above GPT-6 Astra. Compared with GPT-6 Astra:

  • Coding rose 50 points, to 1,338.
  • Chat fell 54 points, to 1,211, dropping from #8 to #11 .
  • Writing and translation tasks users confirmed as successful fell from 91.2% to 85.9%, with the sharpest drop in creative copy .

These numbers fit the mixed reception of the release.

Inkling 2 Mini spreads to local hardware

Thinking Machines' looped open model is the #1 trending model on Hugging Face, and Unsloth's GGUF version is #2 . Project Maya v1.1.73 runs it on NVIDIA cards with as little as 8 GB, keeping experts that don't fit in system RAM. The release also supports V100, RTX 20-series and AMD GPUs on Linux . One user reports 144 tok/s on an RTX 4090 and calls it his first good model for a single-GPU box .

Inference security

  • Together AI: It disclosed a security event that leaked a subset of prompts sent to its serverless APIs in June and July 2026 .
  • MiniMax routing on OpenRouter: This follows last period's report. A researcher planted secret links in outputs routed to MiniMax, and someone opened them 40 to 130 hours later. AtlasCloud, Novita and StreamLake were named as forwarding requests to MiniMax . Sebastian Raschka offered a possible benign explanation: cheaper providers at capacity may reroute traffic to the model's official provider rather than fail .
  • MiniMax cryptanalysis: MiniMax separately published work describing an internal model it says is useful for cryptanalysis .

Compute bulls vs. market turmoil

The selloff continued. Taiwan banned short selling after the TAIEX fell 17.5% in its worst week in 35 years, leaving it 35% below its October high . Commentators split on what it means for AI compute:

  • François Chollet argues this is a buying opportunity. GPU spot prices keep rising, and HBM and DRAM are "effectively sold out through 2029." He also says inference margins are 70–85% .
  • Elon Musk warns that even a few months of delay between paying for compute and earning revenue from it would mean "hundreds of billions" in extra working capital .

Debate continued over the board's handling of Jensen Huang. Paul Graham says the tech world sees it as a huge mistake . Others note that Huang hand-picked the board and that what happened between them is unknown .

Also notable

  • Gemma 4 speculative decoding: Google built a multi-token drafter into the model itself. It reuses the model's activations and existing KV cache, so it skips the separate prefill step a standalone drafter needs. Reported speedups reach 3.1x for large models on servers and 2.2x on a smartphone GPU .
  • US House report on open source: Clement Delangue says a new US House report recognizes that the right response to Chinese open-source AI is building more American open source, not banning it .
  • LangChain's "knowledge factory": LangChain argues that enterprise agents need curated company knowledge more than better models. It says this approach already powers its internal GTM and content agents and LangSmith Engine .

Want personalized briefs on the topics you care about?

Create your own agent and get daily or weekly cited briefs on the topics you care about, shaped by the people, newsletters, podcasts, blogs, and channels you choose.