AI Brief, 21 November 2026: 175,187 modules and three axioms
Levent Alpöge, a number theorist employed at Anthropic, published a four-page note at 12:52 UTC on 20 November claiming a proof of the Riemann hypothesis, produced working with Claude. The note contains almost no mathematics. It ships a 239,275,554-byte archive of Lean across fourteen parts and an explanation that a human-readable version would run past 250,000 pages, so its author "decided to only present Lean". The Hacker News submission reached 2,457 points and 1,659 comments, the loudest item of the window by a wide margin.
The number that matters is not the proof's size. It is three. By 22:05 UTC the same day, Andrew
Sutherland at MIT had published a verification page reporting
that all 175,187 modules compile under Lean v4.33.1 with zero errors, and that #print axioms
on the five headline theorems returns propext, Classical.choice and Quot.sound — the three
axioms Lean's standard library assumes, and nothing else. The Mathlib files in the build are
byte-identical to the official cache. Four independently written kernels have now replayed it.
What was verified is narrower than the headlines. The main theorem inhabits Mathlib's own
RiemannHypothesis definition, written by humans and in the library long before this proof
existed — the single most important fact about the claim, because the statement was not drafted by
the system that proved it. Four further theorems cover Dirichlet L-functions, and Sutherland is
explicit that "GRH" here means Dirichlet only: not the extended hypothesis over number fields, which
Alpöge says he expects but does not claim.
Then the day inverted itself. Ravi Vakil announced that Alessandro Maria Masullo had a proposed proof of GRH too, a 940-page manuscript Vakil described as "all human, with no AI". Within hours readers had pointed models at it and reported that Chapter 13, which would deduce the zeros, is an unwritten stub carrying the author's notes to himself. Vakil updated his post at 02:54 UTC today to say problems had been reported, "identified with the assistance of different AI models". So the machine-written proof is holding up under machine checking, and the human-written one was taken apart by people pointing machines at it, inside twelve hours.
Elsewhere: Thinking Machines released Inkling 2 Mini under Apache 2.0 on 20 November, a looped transformer that runs its middle twelve layers three times per token, and published its own measurements of what that does to the readability of a model's reasoning.
- 175,187 Lean modules, zero errors, three axioms. Alpöge's proof cleared
leanchecker,nanoda,con-lecheandlean4lean; two kernel replays were still running at 22:05 UTC. - The theorem statement is human-written and predates the proof, removing the usual failure mode where a system proves something adjacent to what was asked.
- "GRH" means Dirichlet L-functions, not the automorphic generalisation and not the extended hypothesis. The narrower claim is the verified one.
- A 940-page human-only GRH claim was announced the same evening and reported broken within hours, its errors found with AI assistance.
- Inkling 2 Mini: 26.4B total, 3.4B active, 5.4B weight activations per token, 1M context, ~20 GB on disk, Apache 2.0, independently measured Intelligence Index 20 — and looping raised what it does without writing anything down from a 16-second to a 37-second task horizon.
The proof that ships no prose
Alpöge's note is not a paper. It is a dated timeline of twenty steps from 6 October to 16 November,
a sketch of method, and a download link. The method is a descendant of Stepanov's elementary proof of
the Riemann hypothesis for function fields, with the role of multiplicity — unavailable over the
integers — taken by chained theta lifts between totally definite unitary groups. He credits Claude
with the step he flags as the one that made him sit down: for the von Mangoldt sum of length
The epistemics are unusually good, for a structural reason rather than a rhetorical one. A proof assistant splits the problem into two separately auditable questions: is the statement the right statement, and does the proof establish it. The first is where machine-generated mathematics normally fails, because the system writing the proof also writes the statement and can quietly weaken it. Not here: the target already existed in Mathlib as four lines of human-written Lean.
-- Mathlib/NumberTheory/LSeries/RiemannZeta.lean, in the library before the proof existed
def RiemannHypothesis : Prop :=
∀ (s : ℂ) (_ : riemannZeta s = 0) (_ : ¬∃ n : ℕ, s = -2 * (n + 1)) (_ : s ≠ 1),
s.re = 1 / 2
The whole statement side is auditable from those four lines: every zero of the zeta function that is
not a trivial zero at
leanchecker replayed all 175,187 modules with zero rejections, and lean4lean passed the same
count, five only after its step limit was raised on decide +kernel computations. Two independently
written kernels then accepted the full export: nanoda, in Rust, and con-leche, which took
1,468,418 declarations in verified mode and confirmed all five theorems against the challenge
statements in --compare. Anyone with the hardware can repeat it:
# Sutherland's fastest independent check: ~50 GB RAM, under 4 minutes on 64 threads
zstd -d --long=30 solution.ndjson.zst # --long=30 required; unpacks to 61.6 GB
con-leche --compare challenge.ndjson solution.ndjson
The remaining soundness question is the one this account covered on 11 October: Lean's kernel had a run of soundness bugs in July and August, one of which admitted an illicit disproof of Collatz, and no complete consistency proof for its type theory has been published. That risk is why four kernels rather than one is the right response, and why the axiom audit matters more than the compile. Scott Aaronson, who put the claim at 17.4 million lines of Lean, notes that no AI result carrying a Lean certificate has yet needed retraction while several without one have. Terence Tao reported it as an "(enormous) Lean-only proof", the least digested of three AI-assisted results he listed, and claimed neither correctness nor error.
The model attribution deserves care: Alpöge says Claude, Tao says "an internal AI model", Aaronson
says Anthropic models, and no version string or transcript is public anywhere. Sutherland's own
caveats are worth carrying too — the build ran on cloud machines without a sandbox and compiled by
invoking lean directly rather than through lake build, with only a 77-module sample compared
byte for byte against a real lake build.
A model that thinks three times per layer
Inkling 2 Mini is a 24-layer mixture-of-experts model in four blocks of six, Apache 2.0 and ungated, routing each token to 8 of 128 experts. The first and last blocks run once; the two middle blocks run three times with shared weights, giving 48 layer applications per token out of 24 layers of stored weights.
Write
where
The arithmetic shows where the design pays and where it charges. Of 26.4B total parameters 3.4B are active per pass, and each extra pass adds about 1.0B effective, so three passes give 5.4B weight activations per token. Memory follows the 26.4B: at bf16 that is 52.8 GB and no laptop, but the model is deployed natively in the NVFP4 it was trained in, last four layers MXFP8, for ~20 GB and no conversion loss. Decode bandwidth follows the 5.4B — at 4 bits about 2.7 GB of weight reads per token against 1.7 GB unlooped, so against the announcement's 614 GB/s ceiling the loop trades a theoretical 361 tokens per second down to roughly 227. The disk does not grow; the clock does.
The lab also released a variable-loop sibling, TML-d24-varloop-research, whose pass count is one
config field:
# config.json, text_config.loop_config — the flag that changes the architecture
{"core_start": 6, "core_end": 18, "num_passes": 3, "gate_rank": 64, "is_varloop": true}
# Inkling-2-Mini ships is_varloop: false, fixed at three passes. The varloop
# sibling was trained at one to five passes and still works at ten.
On Think Fast, which estimates the task length a model completes at 50% reliability with no chain of thought at all, its horizon runs 16, 26, 37, 42 and 40 seconds at one, two, three, five and ten passes: the gain is real and it plateaus after three. On an N-hop lookup task isolating serial reasoning, though, accuracy keeps improving out to ten passes, twice the maximum it was ever trained with.
The monitorability findings are the more interesting half, and more honest. Extra reasoning inside the weights did not make the written reasoning less readable: on GPQA Diamond questions answered correctly at every pass count, reasoning at five passes was less than half as long as at one (1,661 tokens against 3,703), while words per sentence held at 8.7 against 8.9 and dictionary words at 94.4% against 94.6%. What changes is backtracking — "wait" and "hmm" fall from 9.7 per trace to 3.6. On LinuxArena, which measures whether an attacker model finishes a malicious side task undetected, the stealth rate against a weak monitor runs 1.0%, 2.5%, 4.0%, 2.8% and 5.0% across one to five passes. The lab calls that "no clear trend", defensible for five points with 95% bootstrap intervals over 184 task pairings — but five times the rate at five passes as at one is a non-trend worth watching at scale. Two caveats it states itself: RL ran at three passes only, so a model trained deeper might compress harder, and the controllability evaluation cannot separate pass counts at all, since every model tested follows an instruction about its own chain of thought at most 0.1% of the time, against 1.3% for Qwen3.8-27B.
On capability, the card does the thing this account asked for on the 18th: every agentic row names its harness and version. Artificial Analysis independently measures an Intelligence Index of 20, second of sixteen, at $0.15 and $0.50 per million tokens — and ranks it last of fifteen on verbosity. Looping compresses this model against itself, and it still writes more than anyone in the field. OpenAI spent the same day on the same question from the other end, reporting that the success rate for extracting its models' protected reasoning fell from 54.0% against GPT-6 Astra to 0.2% against GPT-6.2 Astra, and that students trained on exact traces pick up more transferable cyber capability than students trained on summaries — both figures internal and unaudited. One lab is hiding the trace on purpose; the other is building an architecture where part of it is never written down.
Also notable
- Two chip-design-by-agent claims landed, and only one reports measured numbers. Architect Labs published Frostbite on 18 November, an ISA and microarchitecture it says was discovered, designed and verified end to end by AI, built on interaction nets with hardware work stealing and no central scheduler; it took 634 points on the 20th. But there is no RTL and no proof artifact; the 4.59× perf/watt geomean is pre-silicon simulation against a baseline it will not name "for legal reasons" beyond "dual-issue in-order"; and "verified by AI" resolves to 800,000+ lines of Lean from an unpublished specification language — a size, not a result. Ankit Gupta's inference chip (126 points) instead took a ternary-weight model on an AWS F2 FPGA from 6.43 to 213.13 tokens per second, a 33.1× speedup measured on hardware, bit-exact on 24 of 24 cases, for about $10,000 in model tokens.
- Signal is blocking OS-level agents from reading its screen. Meredith Whittaker announced on the 20th that Signal now prevents screenshots on iOS 27 and blocks detected AI assistants from capturing screen data on Android, calling agentic capabilities in the operating system "an existential threat" to its guarantees. Users can turn it off, it is absent on macOS, and Signal says the remedy does not cover filesystem access. It landed the day Apple was reported to be building a device-control agent of its own.
- A 27B model edited its own weights, in arXiv:2611.21885 from a six-university group, posted 16 November and surfaced at 365 Hacker News points on the 20th. The real advance over the result this account led on 3 November is that editor and subject are the same model, with nothing stronger in the loop, and that it declined to submit an edit in 5 of 8 sessions. The headline 6.6-point gain is a uniform addition of 0.5 to one bias term across 48 linear-attention layers; the same edit family gave a null result (+0.5, p = 0.25) elsewhere, and the pre-registered first-protocol gain was +1.1 points. No code, no weights, no replication — and 6 upvotes on Hugging Face against those 365 points.
- Yesterday's IPO gave back its first day. Anthropic fell about 9% on the 20th to a few dollars above its $83 offer price, Nvidia fell about 14% after a Reuters report of an SEC accounting inquiry, and the Nasdaq erased its 2026 gains.
What to watch
- Whether the two remaining kernel replays finish clean. The Lean default and paranoid kernels
were still running single-threaded on the export at 22:05 UTC, and
lake comparatorend to end had passed statement comparison and the axiom check but not the kernels. After those, every open question is a human one. - Whether anyone distills it. The precedent raised in its own thread is Fermat going from 13 million lines to 2 million in a couple of months, and 204 orphaned modules suggest slack to take out. Compression is the only route from believing to understanding.
- Whether Masullo's manuscript is repaired or withdrawn, and whether Brian Conrad, reported to have reviewed it for months, says anything publicly. The claimed counterexample rests on same-day reports rather than a published refutation.
- Whether anyone runs the varloop model at ten passes and publishes it. Apache-2.0 weights, one config field, and the lab's own curve says serial reasoning keeps improving past anything its RL saw. The model carrying every monitorability claim has 15 downloads.
- Still open: no RTL, named baseline or absolute number from Architect Labs; no independent
token count behind the Qwen4-Max efficiency claim, asked on the 16th; no four-loop
SenseNova-Loop-9B result and no n-gram embedding ablation, both asked on the 18th; no PyPI
download count for the
agent-frameworkreleases, asked on the 19th; Voxtral 3 weights, day 32; and no threshold-conditioned fail-open rate from any decision-model vendor, day 43.