The Diff

Archives
Log in
Subscribe
November 22, 2026

No. 050: The Lean community certifies the Claude Riemann proof. Someone shrunk Inkling 2 Mini to 5.4GB.

The Lean and Mathlib communities certify the Claude-assisted Riemann Hypothesis proof, independent kernels and all. And Inkling 2 Mini shrinks from 20GB to 5.4, overnight. Plus the quiet ones nobody talks about.
View in browser: The Diff No. 050
The Diff  No. 050
Sunday, 22 November 2026  ·  AI dev news, only what is verified
Levent Alpöge, from his personal site
The Lean and Mathlib communities certify the Claude-assisted Riemann Hypothesis proof, independent kernels and all. And Inkling 2 Mini shrinks from 20GB to 5.4, overnight.

Twenty-four hours ago the proof was a claim on a personal site. Now the people who maintain the proof checker itself have signed it: statement correct, proof correct, verified by kernels that come with their own proofs of honesty.

And the laptop model from yesterday’s issue already fits in a pocket: one person squeezed it to 5.4 gigabytes overnight, and it still beats quants twice its size.

“Certified by the community” is now the correct phrase for the proof - with a scope note. The 5.4GB model you can test with two commands.

IN THIS ISSUE
✅ Certified: the Lean and Mathlib communities sign Alpöge’s Claude-assisted Riemann Hypothesis proof.
01
✅ Certified: the Lean and Mathlib communities sign Alpöge’s Claude-assisted Riemann Hypothesis proof.
A joint statement, endorsed by the Lean FRO and the Mathlib Initiative: “We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct.” Independent kernels written in Rust agree. Two checks are still running.
🗜️ Overnight, one person squeezed Inkling 2 Mini into 5.4GB - and it still beats quants twice its size.
02
🗜️ Overnight, one person squeezed Inkling 2 Mini into 5.4GB - and it still beats quants twice its size.
The LookAtMyQuant build compresses the 18.89GB text weights to 5.38GB - 1.65 bits per parameter - fine-tuned across all 24 MoE layers for Apple Silicon. The author’s harness: GSM8K 88.0% versus the teacher’s 96.5%, HumanEval 70.7% versus 92.7%.
🔥 VIRAL
The loud ones on Hacker News. Numbers are a morning snapshot.
Levent Alpöge, from his personal site
01  The Lean community blog’s joint statement, Nov 21 · Andrew Sutherland’s MIT verification page, updated 04:35 UTC Nov 22 · 897 HN points overnight

✅ Certified: the Lean and Mathlib communities sign Alpöge’s Claude-assisted Riemann Hypothesis proof.

A joint statement, endorsed by the Lean FRO and the Mathlib Initiative: “We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct.” Independent kernels written in Rust agree. Two checks are still running.
SCENE

Friday morning this was a two-page remark on a personal site plus fourteen tarballs. By Saturday afternoon the Lean and Mathlib communities posted a joint statement, endorsed by the Lean FRO and the Mathlib Initiative: “We have full confidence that Alpöge’s proof of the Riemann Hypothesis is correct, and we congratulate him on this achievement.” Dozens of people had reproduced the compilation within a day.

THE THING

Certified means four specific things. The statement is Mathlib’s own vetted definition of the Riemann Hypothesis. All 175,187 modules compile with zero errors. Two independent comparison tools confirm the proved theorem is exactly Mathlib’s statement. And the 61.6-gigabyte proof export - byte-identical in an offline sandbox rebuild - was accepted by Lean’s official kernel and by independent checkers: nanoda, a Rust kernel, in 27 minutes on 80 threads; con-leche and con-ron, which carry their own formal proof that they cannot accept a false proof. That answers yesterday’s caveat, Scott Aaronson’s worry that Lean’s logic itself might be the hole. MIT’s Andrew Sutherland ran the checks - with Claude Code, so Claude checked Claude’s homework - and his page still lists Lean’s default and paranoid kernels as running, no verdict yet. The axioms audit found only the three standard ones. Scope, from the page: five theorems covering RH and GRH for Dirichlet L-functions; Dedekind, Hecke and automorphic L-functions are not among them.

✂️ STEAL IT

Steal the verification ladder, not the theorem. One person with a good model produced more machine-checked mathematics in weeks than a field could review in years - and the field responded by building the review machine: a public challenge file, independent kernels, byte-identical rebuilds. That ladder now exists for any claim you ship: a checkable artifact, an isolated statement, a second implementation. What still has no shortcut: understanding. Aaronson’s point stands - mathematicians will spend years distilling why the proof works. Certified is not yet understood.

Watch  VERDICT

Update your language, not your roadmap: “claimed” is now stale - the correct phrase is “certified by the Lean and Mathlib communities, at the machine level, for RH and Dirichlet GRH.” The proof export, the challenge file and the logs are all public if your work touches formal methods. This fails you if you read certification as comprehension - the community certified the proof, not a human explanation of it - or if you extend the claim past its stated scope.

Lean and Mathlib communities: the certification statement  ·  Andrew Sutherland (MIT): the verification files and live check results  ·  Xena Project: the four criteria for a checked proof  ·  Scott Aaronson: the original assessment and kernel caveat

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperAlready knew it
The quant’s Hugging Face repository card
02  The quant’s Hugging Face repository, posted overnight · 385 HN points this morning · weights, runtime and drafters all public

🗜️ Overnight, one person squeezed Inkling 2 Mini into 5.4GB - and it still beats quants twice its size.

The LookAtMyQuant build compresses the 18.89GB text weights to 5.38GB - 1.65 bits per parameter - fine-tuned across all 24 MoE layers for Apple Silicon. The author’s harness: GSM8K 88.0% versus the teacher’s 96.5%, HumanEval 70.7% versus 92.7%.
SCENE

Yesterday’s issue shipped a 20GB laptop model. Overnight, a quantizer going by shinmegamiboson cut it to 5.38GB - small enough to run in about 6GB of active memory with a pair of 179MB drafting sidekicks - and posted the weights, the runtime and the eval tables before breakfast in California.

THE THING

The numbers, all from the author’s own harness: 1.65 bits per parameter across the text weights, packed for the MLX runtime on Apple Silicon. On an M5 Max he reports 1,807 tokens per second of prefill and 75 per second serial, rising to 124-143 with lossless drafted decoding. Quality, measured against the original BF16 teacher on identical harnesses: GSM8K 176 of 200 versus 193, executable HumanEval 116 of 164 versus 152, vision untouched at 4 of 4 - an 89.2% three-task mean at 3.5x compression. The comparison that matters: it outscores Unsloth’s 7.76GB and 9.14GB GGUF quants on both math and code while being 41% smaller. The honest limitation, his wording: calibration weighted English, chat, code and math heavily - for multilingual fidelity he points you to his 8.9GB to 12.8GB builds.

✂️ STEAL IT

Steal the size class. “Local fallback” stopped meaning “the small dumb model” this weekend: a 5.4GB file that keeps 89% of a coding model’s bench scores fits on the free space of a phone backup. If yesterday’s 20GB was a download you had to plan, today’s is an impulse. The harness is public and the eval tables name their splits - GSM8K test rows, all 164 HumanEval problems - so disbelief is runnable.

Try today  VERDICT

On Apple Silicon this is two commands and a 5.38GB download: install the LookAtMyQuant runtime, generate against one real ticket, compare with your cloud default. Treat every number here as the author’s until you rerun them - the harness is published for exactly that. This fails you if you need strong non-English fidelity - use the larger builds - or if you expect teacher-level scores at a third of the size.

Hugging Face: the 5.4GB weights, runtime and full eval tables  ·  Hacker News: the discussion  ·  Thinking Machines: the original Inkling 2 Mini launch

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperTest it for me
📊 THE ATTENTION GAP
HN points, snapshots a day apart. The claim that Claude helped prove the Riemann Hypothesis drew 2,434 points on Friday. Saturday’s certification - the moment the claim became machine-verified fact - drew 897. Being first still outdraws being certain, nearly three to one.
The claim thread (Fri)
2,434 HN
The certification thread (Sat)
897 HN
ALSO ON THE RADAR

The rival human GRH proof now has a titled rebuttal - treat it as unverified too. A manuscript called “Invalidity of Masullo’s Proof of the Riemann Hypothesis” appeared on PhilArchive late Friday, and physicist Manlio De Domenico flagged it as “a counterexample to a key step.” Caution cuts both ways: the same author has a separate PhilArchive paper claiming RH is independent of ZFC, which the mainstream does not accept. Ravi Vakil’s update stands as the reliable status: problems reported, all of them must be resolved before the proof is accepted. PhilPapers: the rebuttal manuscript record →

OpenRouter off-boarded three inference providers and pulled 125 endpoints from zero-data-retention. After security researchers reported evidence that some providers were not upholding their data policies, OpenRouter removed the endpoints from ZDR eligibility as a precaution, notified affected customers, and says OpenRouter itself was not breached. Provider write-ups so far: Together disclosed prompt leakage affecting requests from June and July 2026; Novita posted a privacy notice about MiniMax requests; Phala posted a preliminary report. OpenRouter credits a Dragos report and is making the testing ongoing. OpenRouter: the announcement →

OpenAI shipped GPT-6.1 Terra: a mid-tier computer-use model at half of 6.1 Sol’s price. Their numbers: $1 per million input tokens, $5 output; Terminal-Bench 4.0 at 50.3% at maximum effort, more than double GPT-5.6 Terra; an internal error-inducing benchmark fell from 53.6% to 16.7% at low effort. OpenAI claims ~9x lower cost per task than Claude Opus 5.5 at max effort, and over three times faster task completion than GPT-6 Astra. Available now in ChatGPT Work, Codex and the API as gpt-6.1-terra. OpenAI: the announcement →

🧪 ONE THING TO TRY TONIGHT
Two commands: run the 5.4GB coding model on your Mac.
Install the LookAtMyQuant MLX runtime from the author’s GitHub, then "lamq generate shinmegamiboson/Inkling-2-Mini-LookAtMyQuant-5.4GB" with a real ticket from your backlog as the prompt - not a demo question. Score the answer against your cloud default for correctness, then try the drafter flag for the speed he claims. If the numbers hold for you, you have an offline coding model that fits in 6GB of memory.
What this issue could not see. The certification is the Lean and Mathlib communities’ joint statement, endorsed by the Lean FRO and the Mathlib Initiative; the module counts, export sizes, check timings and running-status items are from Andrew Sutherland’s MIT page as of 04:35 UTC Nov 22, where two Lean kernel replays were still in progress, and the page notes the checks were run with Claude Code. The quant’s compression figures, benchmark scores and speeds are the author’s own reported numbers from his Hugging Face repository. GPT-6.1 Terra’s benchmark, error-rate and cost-comparison figures are OpenAI’s. OpenRouter’s actions and provider list are from its own announcement. The rebuttal manuscript’s existence and upload time are confirmed on PhilPapers; its claims are unverified, and its author separately claims RH is independent of ZFC. HN points are snapshots around 06:47 IST Nov 22 (certification) and 06:05 IST Nov 21 (claim). No video this issue: no relevant video was found on a verified channel. Reddit and X were not consulted.
Talk to the editor 💬

Reply to this email with anything: what to chase, what to drop, what you want more of. Every reply shapes the next issue.

More of this kind of issueLess of this
Images: credited in each report; source screenshots and announcement artwork, not our tests.
Don't miss what's next. Subscribe to The Diff:
Older → No. 049: An Anthropic mathematician says Claude proved the Riemann Hypothesis. Inkling 2 Mini loops a 26B onto your laptop.
Powered by Buttondown, the easiest way to start and grow your newsletter.