Skip to content

← X AI Daily

X AI Daily

Date21 November 2026 · SaturdayWindow2026-11-20 02:00 to 2026-11-21 07:50CoveragePartial

Translation

English translation of the Chinese original.

1. AI Models, Products and Applications

Two claimed Riemann Hypothesis proofs on the same day: Alpöge's Lean code independently compiled, Masullo's human-written manuscript still being revised

On November 20 Beijing time, Anthropic mathematician Levent Alpöge published a four-page memo, "Remark," on his personal website, saying that he and Claude had found a proof of the Riemann Hypothesis and that the same method extends to the Generalized Riemann Hypothesis. He published no paper, only Lean formalization code (14 parts); the memo says that expanding the argument into a readable document "would hit 250,000+ pages," so it was "obviously inappropriate to attempt to post some massive unreadable travesty," and he presented only Lean. The memo also notes that the parts where Claude is the expert and he is not are his paraphrase of a summary, which "might make zero sense," and that the Lean should be treated as authoritative.

This was the most widely spread AI-related item in the window: kimmonismus's relay drew 571,470 views, 5,924 likes, 551 reposts, 234 replies, 229 quotes and 1,045 bookmarks; DeItaone 176,721 views and 805 likes; AiBattle_ 44,672 views and 330 likes; clashreport 27,366 views and 511 likes; Cointelegraph 23,919 views and 64 likes, with the text stating the proof has yet to be independently verified. The Japanese account itnavi2022 (35,059 views, 101 likes) translated the memo paragraph by paragraph. No original X post from Alpöge himself was found in this window's retrieval, nor any announcement from Anthropic's official account; all of the attention came through relay accounts.

A second claim appeared the same day. Ravi Vakil, president of the American Mathematical Society and a Stanford mathematician, published on his blog a draft proof of the Generalized Riemann Hypothesis by Alessandro Masullo, a young mathematician also at Stanford, writing that it is "all human, with no AI," that Brian Conrad had been studying it closely for months and giving extensive feedback, that it was "so far looking good," and asking people to check and formalize it quickly. Conrad told Scientific American that judging whether the method succeeds requires a version closer to a final draft, and that the paper is "undergoing major revision." On X, posts placing the two claims side by side drew little engagement: zhouruijin had only 149 views and 1 like.

Verification has to be separated into two layers. On formalization there has been progress: according to Scientific American, researchers including Andrew Sutherland of MIT have independently compiled Alpöge's Lean code, and Sutherland said "the Lean proof has been compiled and independently rechecked by the Lean kernel with only the three standard axioms." But Kevin Buzzard of Imperial College noted that a comparator run is still missing: confirming that the theorem statement in the code really corresponds to the Riemann Hypothesis, and that the code is not exploiting an unknown bug in the Lean kernel. On the mathematics there has been no progress: nobody has verified the argument itself, and Masullo's manuscript has not even been formalized. Both can only be recorded as claimed and pending verification.

Assessment: the real change here is not "the Riemann Hypothesis has been proved" but that formal verification has been pushed this far forward — that the code compiles is solid, while whether the theorem statement matches and whether the argument is understood by humans remain open. Writing "compiles" as "the proof holds," or framing the two same-day claims as "AI versus humanity," goes beyond the available evidence.

Thinking Machines releases Inkling 2 Mini: a 26.4B-parameter looped transformer that runs offline on a 32GB Mac

At 04:10 Beijing time on November 21, Thinking Machines' official account @thinkymachines released Inkling 2 Mini: 517,496 views, 5,646 likes, 562 reposts, 245 replies, 369 quotes and 3,066 bookmarks, the highest-engagement product launch in the window.

According to the official blog and the Hugging Face model card, it is a looped transformer: its 24 layers of weights are split into four blocks, the two middle blocks run three times per token and the first and last blocks once each, so 24 layers of weights produce 48 layer applications; the MoE has 26.4B total parameters and 3.4B active, with about 5.4B effective active parameters per token because of looping, and each layer routes each token to 8 of 128 experts. The checkpoint is about 20GB, with weights mostly in NVFP4 and the last four layers in MXFP8, matching training precision; Macs with 32GB of memory or more can run it through Ollama, and DGX Spark is also supported; the context window is 1 million tokens; the license is Apache 2.0.

Performance figures are a mix of vendor self-tests and third-party leaderboards: an Artificial Analysis Intelligence Index of 20.1, tied with Qwen3.8-27B; DeepSWE v1.1 48.2%, SWE-bench Pro v2 Hard 77.9%, Terminal Bench v2.1 71.6% (the last below Qwen3.8-27B's 73.0%); ARC-AGI-2 49.3% and ARC-AGI-1 84.5%, with ARC Prize giving a cost of $0.14 and $0.06 per task. The vendor also states the weak spots: knowledge depends mostly on total parameter count, AA-Omniscience scored -8.6 with a 63.4% non-hallucination rate, and it trails Qwen3.8-27B on multi-turn chat and many-constraint instructions.

Why it matters: looping lets the model "think more and write less," which goes straight at monitorability — reasoning that never enters the chain of thought cannot be seen by a chain-of-thought monitor. The vendor therefore also released a variable-loop research model, TML-d24-varloop-research, and reports no meaningful reduction in monitorability at this scale, but very low CoT controllability: among chains of 750 to 1,250 tokens, at most 0.1% followed instructions to reshape their own reasoning. On spread, Artificial Analysis had 31,165 views and 540 likes; Goodfire 22,092 views and 562 likes (early access, studying information flow between layers); Ollama 14,858 views and 174 likes; and author ziqiao_ma drew 5,340 views and 168 likes on the variable-loop model.

Assessment: this is not "a frontier model on a laptop," but it turns "trade looped layers for serial depth" into a downloadable engineering artifact, at the cost of knowledge and multi-turn ability; for people building local agents, a 20GB checkpoint plus 32GB of memory is a genuinely usable starting point, and for people working on safety, the "reasoning that is not written down" enabled by looping is a new monitoring blind spot.

OpenAI publishes its defenses against adversarial distillation: reasoning-trace extraction success falls from 54% to 0.2%

At 03:33 Beijing time on November 21, @OpenAI published a research post: 317,363 views, 2,119 likes, 148 reposts, 341 replies, 161 quotes and 586 bookmarks.

According to the official article, the defenses have several layers: adversarial training against extraction attempts, input and output classifiers, API-level limits on reusing encrypted reasoning across conversations, and investigations and enforcement against coordinated extraction. The hardest numbers are the adaptive black-box red-team results (aggregated across six attacker models): the average attack success rate for "extracting protected reasoning from an earlier request" fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra with updated safeguards; "extracting from the current request" fell from 22.9% to 0.9%. A separate cyber-capability transfer experiment found that a student model trained on the teacher's exact reasoning traces generally outperformed one trained on reasoning summaries after supervised fine-tuning, showing that reasoning traces themselves are high-value training data, not just final answers. The article also says weekly enforcement actions range from thousands to tens of thousands of accounts.

Why it matters: it reframes distillation from an intellectual-property issue into a safety-boundary issue — a student model can inherit capabilities without inheriting safeguards. It should be clear that these numbers are OpenAI's internal red-team self-tests with no third-party replication; the later sections on national security, export controls and legislation are political content and are not used in this report, and the "aimed specifically at Chinese labs" reading circulating on X is likewise not used.

Assessment: the technical point is "protecting the chain of thought," not the slogan "banning distillation"; for people working on data and evaluations, it means usable reasoning traces will become scarcer and more expensive.

Tencent Hunyuan opens the Hy4 flagship weights: 770B total, 49B active, with a MaaS clause in the license (follow-up)

At 11:52 Beijing time on November 20, @TencentHunyuan announced the open weights for the Hy4 flagship: 113,602 views, 1,683 likes, 186 reposts, 89 replies, 118 quotes and 289 bookmarks. This is a follow-up in the Hy4 series: Hy4 flash weights were opened on November 18 (reported in the November 19 daily), and today the flagship, the FP8 quantized version and the pretrained base were opened; the flagship model itself was released earlier, and this item covers only the weight release.

The verified specifications come from the GitHub repository and the Hugging Face model page: 770B total parameters (excluding the vision encoder and other components), 49B text parameters plus 1.23B vision parameters activated per token; a 1 million (1,048,576) token context; 78 layers, one dense plus 77 MoE, each MoE layer with 256 routed experts and one shared expert and top-8 routing per token; plus a 1.23B vision encoder and a 3.3B DFly2 draft model for speculative decoding. The most concrete set of self-reported numbers is on financial tasks: 579 real tasks scored by 213 finance professionals (153 from leading institutions), with Hy4 averaging 3.04, DeepSeek-V4.1-Flash-1017 at 2.97, and Kimi K3.1 Pro and GLM-5.3 both at 2.68; the vendor also says internal evaluations cut the hallucination rate from 15.76% to 5.45%. These are vendor self-tests.

The license deserves separate attention: the Tencent Hy License v1.0 is derived from Apache 2.0, and Section 10 requires a commercial party running a MaaS business (offering inference or fine-tuning interfaces to third parties) to obtain a separate written license if its aggregate worldwide gross revenue exceeds $50 million in any consecutive 12-month period, or if it has more than 100 million monthly active users. So these are open weights, not open source in the OSI sense. The system card link exists (PDF), but it could not be parsed in this fetch, so the report does not cite its internal conclusions; the vendor said in a separate post that the system card includes evaluations by Tencent's Xuanwu, Zhuque and Keen Labs plus outside organizations (38,646 views, 469 likes).

Chinese-language relays mostly restate the specifications: NFT_Chen (6,197 views, 41 likes), xiax0603 (1,459 views, 6 likes) and huagechen (1,248 views, 11 likes); wellwue (260 views, 2 likes) noted that the license had switched to Tencent's own and that commercial use by large companies is restricted, which matches the license text.

Assessment: the real barrier is not the parameter count but deployment cost and licensing terms; the vendor's own statement that a clear gap remains to top proprietary models is closer to the material than the relayed "Chinese open source wins again." Neither the system card nor independent third-party benchmarks have been checked, so vendor self-tests should not be treated as conclusions.

2. AI Compute and Industry

Arcfront comes out of stealth: turning expert judgment into training data, with the company saying four of the five frontier labs already work with it

At 02:40 Beijing time on November 21, @arcfront posted its coming-out-of-stealth announcement: 5,240 views, 77 likes, 17 reposts and 12 bookmarks — newly released, with high engagement not yet formed. The company describes itself as working on "the record of judgment": turning how experts trade off options and when they change course in real work into training data, reinforcement-learning environments and evaluations, sold to frontier model companies; the official post says "four of the five frontier labs already work with us, 300+ enterprise partners, 1,000+ RL environments." The website describes the business the same way, and the jobs page lists roles in San Francisco and Bengaluru; the site carries no financial figures.

The attention came from another thread: gokulr (13,701 views, 120 likes) confirmed this is the company he had previously described anonymously as the fastest to a $1 billion annualized revenue rate, and accounts such as NoContextHumans (19,090 views, 5 likes) relayed "less than a year old, over $1B in annualized revenue, under $2M raised." Tracing back, the original claim is gokulr's October 7 post: zero revenue in November 2025, $80 million in in-month revenue in September 2026, and "they will hit $1 billion annualized." That is an investor's projection and relay, not realized or audited revenue, and no first-hand disclosure of the funding amount was found.

Assessment: data and evaluations are becoming an external supply chain for frontier labs, and that structural change is what is worth recording here, not the "fastest to $1 billion" headline; the annualized run rate, the monthly revenue and the funding amount all need confirmation from the company or authoritative media, and for now can only be written as claimed and pending verification.

3. Chinese AI Ecosystem

Chinese-language circles attribute the slower cadence of domestic models to distillation defenses

One of the more widely spread Chinese-language AI discussions in the window was Wavequester relaying AI blogger Da7em's claim that "Anthropic closed the door on distillation, and Chinese models have not moved an inch since," attributing the recent cadence of Kimi, DeepSeek, Zhipu and Qwen4-Max to that cause. The post drew 25,034 views, 31 likes, 67 replies and 7 bookmarks, with replies clearly outnumbering likes — a question-driven discussion.

A second observation, on pricing, came from xiaowuxiaowu66 (622 views, 5 likes): Chinese open models keep cutting pay-as-you-go prices, with the post saying Kimi K3.1 Pro costs $2 and DeepSeek-V4.1-Flash $0.3, while complaining that "prices came down, expectations went up." Those price figures are relays and were not checked against official pricing pages for this report.

Verification: no first-hand evidence links the release cadence of Chinese models to any provider's distillation defenses; the research OpenAI published in this window describes only its own defenses and enforcement and does not explain other countries' model cadence, and this round's retrieval found no new Anthropic announcement on the subject. The specific release intervals of the models named were also not verified one by one.

Assessment: the value of this item is that it records one explanatory frame in Chinese-language circles: attributing a change in cadence to outside defenses travels better than admitting the iteration itself has slowed; the pricing observation is more concrete: competition has moved from "can they catch up" to "who is cheaper to use." The causal claim is currently only a blogger's opinion and cannot be cited as fact.

Coverage gap for this issue: the Hy4 system card PDF could not be parsed, Alpöge's Lean code was not reviewed line by line, and the Arena Alignment Index was not given its own item because the candidate text was truncated — all three need follow-up checks next round.