An Anthropic mathematician says Claude proved the Riemann Hypothesis, in 17.4 million lines of Lean, and Scott Aaronson bets it will stand. And Inkling 2 Mini loops a 26B coding model onto your laptop. Plus the quiet ones nobody talks about.
← All issues
The Diff  No. 049
Saturday, 21 November 2026  ·  AI dev news, only what is verified
Levent Alpöge, from his personal site
An Anthropic mathematician says Claude proved the Riemann Hypothesis, in 17.4 million lines of Lean, and Scott Aaronson bets it will stand. And Inkling 2 Mini loops a 26B coding model onto your laptop.

One story is about what these systems can now do: a claimed proof of the most famous open problem in mathematics, shipped as machine-checked code because the prose would fill 250,000 pages. The other is about where they run: an open-weights model that loops its own middle layers three times per token and fits in 20 gigabytes.

Both land on the same desk - yours: the ceiling of the model you call, and the floor of the one you can own.

The proof is a claim until the community signs off. The laptop model you can test this morning.

IN THIS ISSUE
📐 An Anthropic mathematician says he and Claude proved the Riemann Hypothesis - and published only the Lean.
01
📐 An Anthropic mathematician says he and Claude proved the Riemann Hypothesis - and published only the Lean.
Levent Alpöge’s own note: the proof, extending to the Generalized Riemann Hypothesis, is fully formalized - a readable write-up “hit 250,000+ pages,” so he shipped the machine-checked source instead. Scott Aaronson counts 17.4 million lines and bets the proof stands.
💻 Inkling 2 Mini: a looped 26B that tops open-weights coding - and runs offline on a laptop.
02
💻 Inkling 2 Mini: a looped 26B that tops open-weights coding - and runs offline on a laptop.
Thinking Machines’ launch numbers: 26.4B total, 3.4B active, middle layers looped three times per token, about 20GB in NVFP4. DeepSWE 48.2%, SWE-bench Verified 79.2% - the lab’s reported scores. Open weights, on Tinker and Hugging Face now.
🔥 VIRAL
The loud ones on Hacker News. Numbers are a morning snapshot.
Levent Alpöge, from his personal site
01   Alpöge’s announcement PDF on his own site, Nov 20 · Lean source in 14 archives · Scott Aaronson’s assessment · a claim, not yet community-verified

📐 An Anthropic mathematician says he and Claude proved the Riemann Hypothesis - and published only the Lean.

Levent Alpöge’s own note: the proof, extending to the Generalized Riemann Hypothesis, is fully formalized - a readable write-up “hit 250,000+ pages,” so he shipped the machine-checked source instead. Scott Aaronson counts 17.4 million lines and bets the proof stands.
SCENE

Late Sunday night, in the middle of his company’s IPO week, Levent Alpöge - an Anthropic mathematician and Harvard junior fellow - finished the argument and messaged CEO Dario Amodei: “Hypothetically, if I were to have . . . ” His note says there was “a constraint to respect” on announcing, that he asked about “progressively more bizarre ideas,” and that “the precise timing is horrible.” On Friday, with the stock trading, he posted a two-page remark on his personal site and fourteen tarballs of Lean.

THE THING

The verified facts, from the note itself: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis,” the method “easily extends” to the Generalized Riemann Hypothesis, and the formalization “is finished,” without assumptions. Because a human-readable rendering “hit 250,000+ pages,” he presents only the Lean. His division of labor, his words: he steered the analogy from a toy setting, “and then Claude began to leave me in awe” - proposing, for one key sum, a chain of theta lifts of length square root of X. Scott Aaronson’s assessment, his blog: 17.4 million lines of Lean; Lean’s kernel has known issues a motivated AI could exploit, but no Lean-certified AI result has needed retraction so far - “my bet is that the proof will stand.” Mathematicians, he adds, will need years to distill the ideas for themselves.

✂️ STEAL IT

Steal the artifact choice, not the theorem. When prose cannot carry the argument, ship the verifiable artifact: Alpöge published machine-checkable source and said the write-up comes later - a claim you can falsify beats a claim you must trust. Three things to watch now: the Lean community compiling those fourteen archives, a check anyone can run; the first independent distillation of the key step; and whether Anthropic comments beyond one researcher’s personal site. A named researcher plus checkable code is where verification starts, not where it ends.

Watch  VERDICT

Nothing to run in production. Treat “proved” as “claimed, formalized, under scrutiny”: the Lean files make the claim checkable, and Aaronson’s bet is one expert’s read, not consensus. This fails you if you cite the Riemann Hypothesis as settled this week - or if you dismiss the artifact unread. Either way you are skipping the verification the format finally allows.

Levent Alpöge: the announcement (PDF, on his own site)  ·  Scott Aaronson: “The stochastic parrots have apparently proved the Riemann Hypothesis”  ·  Scientific American: what the proof claim means for AI  ·  New Scientist: mathematicians on the claim

Thinking Machines’ launch artwork, from its official Inkling 2 Mini announcement
02  Thinking Machines’ official launch post and model card, Nov 20 · open weights on Hugging Face · 783 HN points overnight

💻 Inkling 2 Mini: a looped 26B that tops open-weights coding - and runs offline on a laptop.

Thinking Machines’ launch numbers: 26.4B total, 3.4B active, middle layers looped three times per token, about 20GB in NVFP4. DeepSWE 48.2%, SWE-bench Verified 79.2% - the lab’s reported scores. Open weights, on Tinker and Hugging Face now.
SCENE

The demo is a plane. Thinking Machines’ researchers took a MacBook Pro (M5 Pro, 48GB) on a flight, switched on airplane mode, and had Inkling 2 Mini build a playable flight simulator in a few minutes - offline, screen-recorded on the launch page.

THE THING

The architecture is the story: a looped transformer that runs its middle layers three times for every token, reasoning longer without writing the reasoning down. The monitorability results, their experiments: chain of thought grew shorter with more loops but stayed legible; a variable-loop variant’s no-CoT time horizon rose from 16 seconds at one pass to 37 at three; agentic-sabotage evals did not trend higher as loops were added. The scores, their reported numbers: DeepSWE v1.1 48.2% and SWE-bench Verified 79.2%, ahead of every open-weights model under 100B total parameters; behind only Qwen3.8-27B on Terminal Bench v2.1. The honest caveat, theirs: looping does not add knowledge - factuality matches its size class, and it closes the gap with search and tools, leading similar-size open models on BrowseComp at 79.0%.

✂️ STEAL IT

Steal the offline loop. A coding model that fits in 20GB and beats every sub-100B open model on SWE-bench Verified changes what “local fallback” means: flights, client sites, air-gapped review, API outages. It is one download from testing - the weights are open, on Hugging Face and Tinker, and the post lists the harness versions it used, so you can rerun its numbers before believing them. Budget the speed cost: three passes per token is the price of the reasoning, and knowledge work still wants a search tool attached.

Try today  VERDICT

If you have 20GB free and a laptop, this is a same-morning test: pull the weights, run one real ticket from your backlog, compare against your cloud default. Treat the launch-table scores as the lab’s own reported results until you rerun them. This fails you if you expect frontier-scale factual recall - the post itself says loops add reasoning, not knowledge.

Thinking Machines: the official Inkling 2 Mini announcement  ·  Thinking Machines: the model card  ·  Hugging Face: welcome Inkling 2 Mini

📊 THE ATTENTION GAP
HN points at about 06:05 IST, Nov 21. Thursday’s record-IPO debut thread drew 997 points. Overnight, a claimed proof of the Riemann Hypothesis drew 2,434. The market priced the company in a day; the people who build with its models are staring at what they just did.
Anthropic’s Nasdaq debut thread (Thu)
997 HN
Alpöge GRH claim thread
2,434 HN
ALSO ON THE RADAR

Anthropic’s second day: closed at $88.91, down 8.5%. CNBC’s close report: ANT ended Friday at $88.91, still about 7% above the $83 IPO price, as a global sell-off hit AI names - Nvidia fell more than 13%. The greenshoe was exercised: IPO proceeds reached $93.5 billion, $119.5 billion including the private placement, Bloomberg reported; CFO Krishna Rao: “We’ve raised almost $120 billion.” Reuters reports the IPO bank closed its stabilization position, removing that support. ARK Invest disclosed buying $90.8 million across six funds. CNBC: the Friday close →

The human rival GRH proof wobbled within eight hours - gaps flagged with AI help. The same day, Ravi Vakil posted a “hurried announcement” of Alessandro Masullo’s proposed GRH proof - all human, no AI - checked closely for months by Brian Conrad and “so far looking good to him.” Eight hours later Vakil updated the post: several people had reported problems, “identified with the assistance of different AI models.” All issues must be resolved before the proof is accepted; the draft is public. Ravi Vakil: the announcement and update →

A 27B model rewrites its own weights - and smaller models in its family break themselves trying. ArXiv 2611.21885, from a six-university team: “self-rewriting,” no gradients and no training - the model reads its own weights and graded answers, diagnoses its failure, and edits; its own edit makes it 6.6 points more accurate on problems it never saw, and accuracy only rises over eight sessions. It repaired a corrupted copy of itself to over 80% of the original accuracy. The ability emerged with scale: the 2B, 4B and 9B sizes made no edit or broke themselves. No. 031 covered a frontier model editing a smaller one; this one edits itself. arXiv: the paper →

🧪 ONE THING TO TRY TONIGHT
Twenty minutes: put the laptop model on your laptop.
Pull Inkling 2 Mini from Hugging Face - the weights are open and the footprint is about 20GB - take one real ticket from your backlog, not a demo prompt, and run it offline against your cloud default. Score correctness, not vibes. If you want the lab’s numbers checked, rerun one benchmark row with the harness versions named in the launch post. Loops add reasoning, not facts: attach a search tool before judging knowledge work.
What this issue could not see. The Riemann Hypothesis result is Levent Alpöge’s claim, published on his personal site; the proof is formalized in Lean but not yet reviewed by the mathematics community, and his note says a readable write-up does not yet exist. The 17.4-million-line count, the kernel-risk caveat and the bet that the proof stands are Scott Aaronson’s, from his blog. The Masullo draft’s status is Ravi Vakil’s announcement and same-day update; the reported problems were found by correspondents using AI models. Inkling 2 Mini’s architecture, demo, benchmark scores and monitorability results are from Thinking Machines’ official launch post and model card - its reported scores, its harness versions; Hugging Face hosts the open weights. Anthropic’s $88.91 close and the 8.5% move are CNBC’s reporting; the $93.5 billion and $119.5 billion totals are Bloomberg’s via The Next Web; the ARK purchase is from ARK’s trade notification via CNBC. HN points are a snapshot at about 06:05 IST. No video this issue: the PlaneBench demo is embedded on Thinking Machines’ page, with no shareable version found on a verified channel. Reddit and X were not consulted.
Images: credited in each report; source screenshots and announcement artwork, not our tests.