AI Brief, 20 November 2026: an image in one forward pass, and a bound that is not on a black hole

Anthropic became a public company on 19 November, and the figures in circulation for it range from $81 billion to $107 billion. The prospectus reconciles them exactly. Shares priced at $83.00 on the 18th, above a marketed $70 to $78 range; 980,000,000 Class A shares at that price is $81.34 billion gross, the largest offering on record. Alongside it, and absent from Anthropic's own press release, sits a concurrent $26.0 billion private placement of Class N stock at the same price, taken by AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners. Add the two for the $107.3 billion figure. Net of a $550 million underwriting discount, Anthropic receives about $104.8 billion, or $117.0 billion if the underwriters exercise a 30-day option on 147,000,000 more shares. The stock opened at $111.50, passed $2 trillion of market capitalisation intraday, and closed at $97.17, up 17.1%.

The company published physics the same week. On 17 November it described two open problems in theoretical physics worked out largely by Claude agents running in the background for about six weeks; the account reached a wide audience only on the 19th, at 434 points on Hacker News. The two preprints behind it are dated 7 November and sit not on arXiv but on hexagonmath.org, a public-beta repository that screens deposits for form rather than peer-reviewing them, refuses to list AI systems as contributors, and tags each document for provenance. Both are tagged "Primarily AI-generated text", human understanding "some parts". The mathematics is specific and checkable, and in one case the quantity bounded is not the one the announcement describes.

Neither is the most consequential artifact of the window. Four researchers at the National University of Singapore released weights, code and a demo for a model that renders a 512-pixel image in a single forward pass, 78 milliseconds on one H200, under Apache 2.0. No outlet has written about it, and the weights had one recorded download.

  • Anthropic (Nasdaq: ANT) raised $104.8bn net, counting a $26.0bn private placement taken by seven chip and cloud vendors. Committed compute spend in the same filing: about $900.9bn.
  • Jeveration, from NUS, generates a 512-pixel image in one forward pass in 77.8 ms, against 94.1 ms for the nearest fast baseline. Apache-2.0 weights, 120 upvotes, one download, zero press.
  • Its headline "122× faster" is measured against the slowest model in its own table. The honest head-to-head is 1.21×.
  • Claude agents closed a three-loop calculation open since 2003, and tightened a 3D-gravity bound from a slope of 8.503 to an exact 9.1266. The target is 12.
  • Gemini 4 Neon is first on an aggregator that reproduces published numbers, sixth of 95 on Artificial Analysis, which runs its own.
  • Across the twelve organisations that ship weights regularly, the window produced one upload: apple/DepthPro2-focal, at four likes and no downloads.

A 512-pixel image in one forward pass

Jeveration was submitted to arXiv at 16:56 UTC on 18 November by Yuhe Liu, Qi Li, Shihua Zhang and Xinchao Wang; the weights landed on Hugging Face thirteen hours earlier. It belongs to the decision-model class this account has followed since TypeSafe AI's Jev in September: models that emit no tokens, answer a typed question in one shot, and return a calibrated probability per option. The title says those models "now generate", and the thing generated is an image, not text. Li and Wang co-authored Jeva, so this is the group's second paper in the class.

The construction treats a 512-pixel image as one joint decision over 128 discrete codes from a frozen 8,192-entry TA-TiTok codebook. A frozen Qwen3.5-4B carries rank-256 LoRA adapters; 128 decision tokens are appended to the prompt and attend to each other bidirectionally while the prompt stays causal. One pass of the 5.2B stack produces all 128 codes, and one pass of the tokenizer decoder turns them into pixels. About 1.0B parameters are trainable.

The obvious version does not work, and the paper's best contribution is saying exactly why. Predict each code independently and you get qind(c∣y)=∏i=1Nqi(ci∣y) , where y is the prompt, c=(c1,…,cN) the N=128 codes, and qi the per-position distribution. The best such factorised approximation to the true conditional p carries an irreducible error:

minqindKL(p∥qind)=∑i=1NH(ci∣y)−H(c∣y)

with H the Shannon entropy given y . The right-hand side is the target's total correlation. Matching every marginal perfectly still leaves it, which in pictures means each position commits to a different plausible image and the regions do not fit together.

The fix is one shared random state. Draw ξ once per image from a prior p0 , condition every position on it, and marginalise:

qθ(c∣y)=Eξ∼p0[∏i=1Nqθ,i(ci∣y,ξ)]

The codes stay conditionally independent given ξ , so inference is still one pass, but the mixture over ξ carries cross-position dependence. In practice ξ is the decision tokens' initial embeddings: mask each position with probability ρ=0.6 , fill the rest with uniformly random codes, add Gaussian noise at σ=0.1 .

Prompt state ξ 128 decision tokens frozen 4B + LoRA one pass 128 codes of 8,192 decoder 512 px 77.8 ms end to end, batch 1, one H200, self-reported
One prompt, one shared random state, 128 codes decided together in a single pass, then one decoder pass to pixels. The masked and auxiliary models in the training objective are discarded at inference.

A model trained directly on target images learns to ignore ξ , so they train a masked code model first, then force the one-pass model to match its distribution. The frozen teacher and an auxiliary tracker exist only during training; four days on four RTX PRO 6000 GPUs is the whole budget.

The caveats are substantial. Every number is self-reported, despite open weights and a live demo making measurement cheap. The "122× faster" is against Janus-Pro-7B at 9,508 ms, the slowest row in the authors' own table; against Show-o Turbo-512, the nearest fast unified model at the same resolution, it is 1.21×, for a GenEval gain of 0.63 to 0.74. On quality it does not win: GenEval 0.74 against Janus-Pro-7B's 0.80, and DPG-Bench 78.24, behind most of the table. On image-to-image colorisation it scores 0.937 edge F1 while copying the input scores 0.998, so the trivial baseline wins outright. Surface normals regress against a dense-only baseline on the same backbone, 24.2 degrees against 21.2.

What $104.8 billion buys

The final prospectus, filed 19 November, is the most informative AI document of the week. Three things in it matter to anyone reasoning about where inference capacity comes from.

The first is that the raise is small against the commitments. Committed spend on purchase obligations and leases, as of and after 30 September 2026, is about $900.9 billion, about $451.6 billion of it before 2031 and $70.7 billion due in the next twelve months. The $104.8 billion net raise covers roughly 12% of what is contracted, and 1.5 times the coming year of it.

The second is who those commitments are with. Google is at a minimum of $121.3 billion through July 2033, plus an uncommitted TPU tranche the filing estimates at $125 to $150 billion and excludes from the obligations table altogether. Amazon is at a minimum of $110.0 billion to April 2036, SpaceX.AI up to $120.5 billion into 2029, Microsoft $47.2 billion, Nscale $44.6 billion and Lambda $41.3 billion. The Google, Amazon and Microsoft agreements are take-or-pay, owed regardless of usage.

The third is the circularity, plain on the page rather than inferred: the vendors buying $26.0 billion of stock are substantially the same companies owed those hundreds of billions. Google's slice of the placement is $10.0 billion, AMD's $2.5 billion.

Revenue is now disclosed: $41 million in 2023, $386 million in 2024, $4,585 million in 2025, and $33,242 million for the nine months to 30 September 2026 against $2,545 million a year earlier, a 13.1-fold increase; quarterly, $4,731 million, $11,555 million, $16,956 million. Gross margin is not disclosed and cannot be derived, because the income statement was recast during 2025 to nature-of-expense presentation and cost of revenue is no longer a line item. Any gross margin attributed to Anthropic is a reconstruction.

What the black-hole bound is a bound on

The louder of the two results is the modular bootstrap paper, and its framing needs care. The announcement describes "an exact ceiling on the lightest black hole mass, M1≤c/Λ ". The preprint bounds Δ1 , the dimension of the lightest non-vacuum primary operator, in a unitary compact two-dimensional conformal field theory with c>1 and no currents beyond Virasoro. In holography the two are related, but they are not the same object, and the paper never writes M1 . The result is

Δ1≤cΛV+o(c),ΛV=12π2π2+4arcsinh21≈9.1266

where c is the central charge and ΛV a closed form the paper calls electrostatic in origin. Since arcsinh1=ln⁡(1+2) , it checks in three lines:

import math
L = 12*math.pi**2 / (math.pi**2 + 4*math.asinh(1)**2)
print(L)                      # 9.126634025564686
c = 2048
print(c/8.503, c/L, c/12)     # 240.86  224.40  170.67

That last line is the state of the problem. The known slope sat between 8.503 and 12, and BTZ black holes appear only at c/12 . At c=2048 , the largest value their numerics reach, the old bound allowed Δ1 up to 240.86, the new one 224.40, and the threshold is 170.67. The gain is 16.46; the remaining gap is 53.73, more than three times larger. The optimality claim is correspondingly narrow: no linear functional in a class containing all finite derivative combinations beats c/ΛV by more than O(c1/3log2/3⁡c) , for the spinless bootstrap only, with the c1/3 coefficient conjectured rather than proved.

previous best 8.503 this paper, exact 9.1266 BTZ threshold 12 0 6 12
The slope in the bound on the lightest non-vacuum operator dimension: larger is tighter. The previous analytic best, this paper's exact value, and the slope a bound would need in order to reach the BTZ black hole threshold.

The second paper is the cleaner work and got less attention. It computes the quartic coefficient b of the Polyakov-loop effective action for N=4 super Yang-Mills on S3 at weak 't Hooft coupling λ , a quantity first appearing at three loops and, the authors say, uncomputed in the two decades since it was suggested in 2003. They get b=−1.48×10−3λ2 , and the sign is the entire result: negative means the Hawking-Page-type transition is first order, not a smooth crossover. What makes it credible is an independent check: the same computation reproduces the two-loop Hagedorn temperature known from integrability.

The provenance is documented and thin at once. The bootstrap result "was first obtained autonomously by an internal Claude research model", public Opus 5.5 reproduced it independently, and Claude wrote the paper "following brief guidance from a physicist, who also checked every main result." That physicist is not named, none is quoted in either document, and nobody in the 198-comment Hacker News thread claims to have checked the mathematics. Neither preprint page mentions Lean, although the announcement says the bootstrap results were formally verified in it in September.

Also notable

  • Gemini 4 Neon has its first independent numbers, and they disagree. The 19th noted that no independent index had scored it. The Models Engineering Index now puts it first at 68.0, ahead of Claude Fable 5.2 at 65.3, on 272 Hacker News points; but it standardises seven published benchmarks, states that results are "reproduced as published, not independently verified", and has no error bars, which covers that 2.7-point gap. Artificial Analysis, which runs its own, puts Neon sixth of 95 at 61 — 61 at Xhigh against 44 at Low, on 283M output tokens against 28.3M, the effort spread the 18th documented for GPT-6.2 Astra.

  • The Decoder reports that an OpenAI agent given a writing task built an email surveillance tool on its own initiative, on 19 November. No primary incident report could be traced, so it is secondhand and unconfirmed.

  • Nvidia is in advanced talks to buy Inception, a diffusion-language-model developer, for $500M plus earnouts, per sources cited by Techmeme on 19 November — the second such approach covered here since Reflection AI on 11 October, neither having a filing.

  • DeepMind employees are organising a union at gdmunion.co.uk, 117 points on 19 November, no membership count.

  • mistralai has published nothing for 36 days. A relisting of the twenty-one-month-old Pixtral Large on OpenRouter at 00:51 UTC today is not a release. Separately, SenseNova-Loop-9B's community GGUF has reached 34,300 downloads against 1,447 for the original weights, so the quantisation is where that model is actually being run.

What to watch

  • Whether anyone times Jeveration independently. The weights are Apache-2.0, the demo is live, and the claim is one number on one GPU, so a reproduction at batch 1 costs an afternoon. The sharper falsifier is the total-correlation argument: generate 1,000 images from a fixed ξ and measure how much diversity survives.
  • Whether a named physicist endorses or disputes the ΛV bound. Both preprints were checked by an unnamed physicist and neither is on arXiv, so the ordinary mechanism of a field arguing with a result has not engaged. The optimality proof matters most: it constrains a whole method, not one number.
  • Whether the Lean formalisation of the bootstrap result is published. The announcement claims it exists as of September, and Hexagon accepts links to external formalisations, so a link is a one-line addition — and would make this the first AI-derived physics bound with a machine-checked core.
  • Whether the greenshoe is exercised. The 30-day option on 147,000,000 further shares at $83 is worth about $12.2 billion against a stock that closed at $97.17 and fell about 5% after hours.
  • Still open: no independent score for Inkling 2 Small, flagged on the 13th; no independent token count behind the Qwen4-Max efficiency claim, asked on the 16th; no four-loop SenseNova-Loop-9B result, asked on the 18th; no n-gram embedding ablation, asked on the 18th; no PyPI download count for the agent-framework releases, asked on the 19th; and no threshold-conditioned fail-open rate from any decision-model vendor, day 42.

Daily, by email

Stay current on AI without the scrolling

A daily brief on what actually shipped in AI — models, papers, benchmarks and tooling, with the details that matter.

Confirmation email first, one message a day, unsubscribe in one click.