Skip to content

← X AI Daily

X AI Daily

Date28 October 2026 · WednesdayWindow2026-10-27 02:00 to 2026-10-28 07:58CoverageComplete

Translation

English translation of the Chinese original.

1. AI Models, Products & Applications

Axiom Announces the Classification of Finite Simple Groups Theorem Has Been Formalized in Lean

  • What happened: Axiom announced on X that the largest mathematical proof in human history has now been verified by computer — the classification of finite simple groups theorem (the "periodic table of symmetry") has been fully and automatically formalized in Lean. The original post gives the scale of the work: roughly 15,000 pages, produced by more than 100 mathematicians across generations. The official blog adds finer magnitudes: about 43.1 million lines of Lean code (of which about 1.2 million lines come from the existing Qiuzhen-CFSG), taking 24 days, completed by a swarm of dozens of AI agents dividing the work — the lead agent decomposed the argument and reviewed statements, while the remaining agents searched for existing results or wrote Lean proofs, claiming milestones through GitHub issues and reporting progress on Slack.

  • Why it matters: It turns a monumental work that previously could only be read by humans into a machine-checkable object; more important, the formalization process itself is error-correcting. The team recorded 34 significant deviations from the published literature, most of them typos, table errors and omitted hypotheses, with the most substantive ones concentrated in the Unbalanced Group Theorem. One concrete example: in the quasithin volume of Aschbacher–Smith, a lemma asserts that two numerical values are equal, and an agent used SL₂(5) to find a counterexample (one is 1, whereas the equation requires 2); only after the equation was changed to an inequality did the lemma hold.

  • Original post & engagement: @axiommathai original post, 4,828 likes, 1,797,105 views, 3,458 reposts, 321 replies, 463 quotes, 1,954 bookmarks (collected at about 08:10 Beijing time on 2026-10-28).

  • Verification: Axiom research blog (October 26), the project site and the dependency graph; for external relays see the Unite.AI report. The blog itself draws the boundaries: the second-generation proof is still incomplete, and what was formalized here is the first-generation proof; Lean only verifies formal derivations, and whether the statements and definitions faithfully match the original intent still requires human review; progress is "easy to overestimate," and early on tens of thousands of lines built on unproven assumptions were deleted; errata are still being compiled.

  • Assessment: The 43.1 million lines, the 24 days and the 34 deviations are all vendor self-reports and still await third-party review. But it demonstrates two things at once: AI agents can already take on ultra-long-cycle formalization engineering that requires scheduling and collaboration; and formalization can in turn find errors in classic literature. Reading it as "mathematics has been solved by AI" is a misreading; reading it as "mathematical literature can be machine-audited for the first time" is closer to the facts.

Google DeepMind Releases Genie 4: The First Real-Time World Model That Natively Produces Sound

  • What happened: DeepMind released Genie 4, describing it as its most realistic and most interactive world model to date. Users can rewrite the running world in real time with natural language — changing the weather, moving the camera, adding people or objects — with Gemini adapting the instruction to the current frame and monitoring coherence; physics and character motion are more realistic, characters react to the scene and speak; sound is natively generated, and the company says this is the first real-time world model that natively generates audio, with the ambient sound, music and dialogue in the demos all produced directly by Genie 4. Availability: a research preview has begun opening to trusted testers, it is "coming soon" to Project Genie for Google AI Ultra, as well as to early partners of the Gemini API and Google AI Studio, and a waitlist is open on the Project Genie side.

  • Why it matters: World models shift from "generating a clip of video" to "generating an environment you can walk into and rewrite on the fly," and fold sound into the same generation pipeline. This is not an image-quality upgrade but a shift that turns simulation into an operable object, directly relevant to games, film and television previsualization, robots and agent training environments.

  • Original post & engagement: @GoogleDeepMind main post, 14,159 likes, 1,497,767 views, 1,531 reposts, 613 replies, 1,123 quotes, 5,779 bookmarks; @Google post, 4,842 likes, 553,258 views; @GoogleAI post, 1,647 likes, 121,279 views.

  • Verification: DeepMind official blog (October 27). Several common relays need correction: the company did not say it is "powered" by a single model, but rather that the Gemini family collaborates — Gemini Omni anchors realistic appearance and behavior, other Gemini models provide coherence reasoning, Gemini 3.8 Flash serves as the guiding agent in the loop, and Gemini 4 Argon was previously paired with it in experiments; the company did not disclose resolution, frame rate or a maximum duration, saying only that coherence holds for "short periods"; the blog explicitly lists limitations: after longer runs, multiple rounds of complex edits or large camera movements, detail and object permanence drift, and text rendering, spatial precision and real-time audio accuracy are still under research, while generated content carries a SynthID watermark.

  • Assessment: The capability narrative is supported by the official line, but "most realistic" is vendor wording, and hard metrics for horizontal comparison are missing — Genie 3 stated 720p at 24 frames back then, and this generation gives no corresponding numbers. What is really worth tracking is its downstream use as an agent training environment, not how good the demo footage looks. Incidentally, on the developer side work continues in parallel on bringing Gemini 4 Argon into Antigravity (official blog, @googledevs Unity workflow demo, 713 likes, 63,136 views), but Argon first launched on October 13 and is not an event in this window.

Liquid AI Releases LFM3: One 1.7B Backbone Behind Six On-Device Models

  • What happened: Liquid AI released the third generation of Liquid Foundation Models. Six roughly 2B-class models share the LFM3-1.7B-Base backbone: the text model LFM3-1.7B (256K context), the vision-language LFM3-VL-2B (supporting video input and spatial understanding), the audio model LFM3-Audio-2B (18 languages, voice conversation and tool calling), the fully multimodal LFM3-Omni-2B, and the decision models d1-2B and d1-Omni-2B with typed outputs and calibrated probabilities. The new tri-hybrid architecture combines gated short convolutions, sliding-window attention and global attention; the text model's 4-bit weights are about 1GB and it can be fine-tuned on a single GPU, and the architecture search covered 14 processor and backend combinations (phones, laptops, Raspberry Pi 5, cloud CPUs, Hexagon NPU, and so on).

  • Why it matters: The company says LFM3-1.7B surpasses LFM2.5-2.6B with 35% fewer parameters; LFM3-VL-2B, at 1.9B parameters, averages 71.4 across 62 vision benchmarks, on par with the 12B Gemma 4 12B; the text model beats Qwen3.5-9B on WebArena-Verified and AutomationBench, and at 128K context its KV cache takes 1.6GB (Qwen3.5-4B: 4.3GB); they also ran 300,000 LFM3-1.7B agents in parallel to process the Enron email archive. Small models, single-GPU fine-tuning and swappable hardware hit directly at the demand for "enterprise-private, on-device agents."

  • Original post & engagement: @liquidai main post, 4,206 likes, 829,382 views, 1,154 reposts, 302 replies, 380 quotes, 3,575 bookmarks; among the breakdown thread, architecture and on-device design (216 likes, 21,954 views) and the text model (216 likes, 13,985 views), among other items, have engagement in the hundreds.

  • Verification: Liquid AI official blog (October 27) and the weights, documentation and API entry points given in the last post of the release thread. The comparisons above are all vendor self-assessments and have not yet been reproduced by third parties.

  • Assessment: LFM3's selling point is not leaderboard position but the trade-offs made under this set of engineering constraints — "one backbone, six modalities, single-GPU fine-tuning, runnable on 14 classes of processors" — which turns "specialized intelligence" from a slogan into a deployable catalog. Note that small models are still pulled apart from large ones on open-domain tasks, and that most of the competitors cited by the vendor are specific slices from its own evaluation.

Giotto Releases Alto and Solo: A Sovereign Inference Model Family from a Swiss Team

  • What happened: The Swiss lab Giotto released a model family: the flagship Giotto Alto focuses on advanced reasoning, is free for everyone, and is already available on the App Store and Google Play, hosted on a shared cloud in Europe (the press release says Alto runs on 8 B300s); Giotto Solo is positioned as "the strongest model that can run on a single GPU," for institutional deployment. The website says customers can choose a European shared cloud, a dedicated Swiss or European cloud, bring their own GPUs, or buy a certified all-in-one machine with Giotto preinstalled.

  • Why it matters: This is a rare consumer-facing landing in the "sovereign AI" narrative — a team that is neither American nor Chinese at once offering a free flagship for the public and a single-GPU version for institutions, with keeping data in Europe as the selling point. What it tests is whether, without an advantage in compute or data, reasoning capability plus deployment compliance can carve out a market.

  • Original post & engagement: @giotto_ai original post, 897 likes, 231,659 views, 61 reposts, 150 replies, 103 quotes, 443 bookmarks.

  • Verification: The official website and the EQS press release (October 27). The benchmarks on the website are all vendor-reported: Alto reports 100.0% on AIME 2026, 91.9% on GPQA-Diamond, 52.0% on HLE and 99.80% on Math-500; Solo reports 96.6% on AIME 2026, 88.3% on GPQA-Diamond and 37.5% on HLE. The website does not disclose specific GPU models or memory, nor does it give details of compliance certifications related to "sovereignty."

  • Assessment: The models exist and are free to use, and that part is verifiable; "the strongest European sovereign model" is not. The two questions more useful to readers are whether the single-GPU Solo is good enough for real business tasks, and what exactly a "European cloud" does and does not provide in compliance terms.

Anthropic Randomized Trial: Agents Change Experts' Judgments Without the Experts Knowing

  • What happened: Anthropic published a study using a randomized controlled trial to see whether experts still own the key decisions after handing work to an agent. The study recruited 61 professional data analysts; 11 chose not to use the agent, and the main analysis is based on the remaining 50. The task was to evaluate an expansion plan for an agricultural project. The agent was given a subtle "lean," and the result was that analysts agreed with the agent's lean on 86% of the four analytical choices (outcome measure 92%, comparison method 82%, which farms were covered 81%, weighting approach 90%).

  • Why it matters: More alarming is attribution — 77% of people restated the agent's argument in at least one place, but only 15% attributed the influence to the agent anywhere; when the agent disagreed with their prior preference, 60% changed their position (only 6% did when it agreed); of the 48 respondents, only 6 guessed that the study was examining AI's influence on them, and none suspected that the agent's preference had been manipulated. The influence happened and almost no one noticed.

  • Original post & engagement: @AlexTamkin introductory post, 734 likes, 58,287 views, 79 reposts, 54 replies, 46 quotes, 601 bookmarks.

  • Verification: Anthropic research page (October 27). Supporting results: 3 of the 50 people who used the agent (6%) made serious methodological errors, versus 6 of the 11 who did not (55%); median time was 65 minutes versus 90 minutes; when the agent leaned toward expansion, 54% supported expansion, and 36% when it leaned the other way. The small sample and the single task are the boundaries of this study.

  • Assessment: This is a textbook case where "efficiency gains" and "transfer of judgment" hold at the same time. For teams putting agents into analysis, policy evaluation and safety review processes, the conclusion is not "don't use them" but rather to make the agent's lean explicit and to leave a record of decisions, otherwise the review step degenerates into rubber-stamping.

Anthropic Publishes Early Progress in Drug Discovery: Six Partners, an In-House Pipeline and the Preclinical Boundary

  • What happened: Anthropic described deep research collaborations between its scientists and six organizations that develop new drugs, alongside the launch of its own drug discovery work, and disclosed some early results. The partners named on the official page are BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group; the in-house side includes the acquisition of Mithril Medicine and Pleiades, founded by The Column Group.

  • Why it matters: What the piece reports are process gains: Generate's team used Claude Code to propose a design that shortens part of the genetic instructions for CAR-T by about 70%, increasing the number of receptors the immune cells manufacture by about tenfold; Divergence Labs produced about 35,000 antibody designs in around a month, and in early tests 3 bound IgE and 20 potentially stripped it, with 3 of them outperforming the standard drug; BigHat tested about 1,010 variants in roughly 10 weeks, with the best version improving binding strength more than tenfold; Enveda used Claude to review 1,061 pages of production documents, catching all 11 problems humans had found plus 7 more, cutting review time from more than 25 hours to under 3 hours; BMS, meanwhile, is evaluating protocols and probabilities of success at the clinical trial stage and plans to expand the collaboration in 2027.

  • Original post & engagement: @AnthropicAI original post, 4,863 likes, 677,552 views, 461 reposts, 347 replies, 263 quotes, 1,454 bookmarks.

  • Verification: Anthropic research page. The page states explicitly that this work is "still early" and "mostly preclinical," and that the drugs have not yet been tested in humans; the figures above are self-reported by the partners and Anthropic, and this daily does not draw efficacy inferences from them.

  • Assessment: The weight of this item lies in organizational form rather than in the numbers — model companies are beginning to act simultaneously as tool vendors and as research participants, and "who owns the discovery" will become a new question of contracts and authorship. At this stage what most deserves watching is whether reproducible experimental data is published later, not the "AI makes drugs" narrative.

Ling-3.1-flash-1019 Open-Sourced, AA Intelligence Index Rises from 41 to 44

  • What happened: InclusionAI open-sourced Ling-3.1-flash-1019, along with Ling-3.1-flash-base for continued training. Artificial Analysis's evaluation shows the model scoring 44 on AA Intelligence Index v4.3, up from 41 for the previous generation; it has 560B total parameters with 25B active, a 1M context, an MIT license, weights on Hugging Face, and a first-party API.

  • Why it matters: Agent capability improves noticeably — AA-Briefcase v1.1 Elo rises from 1,397 to 1,521, GDPval-AA v2.1 from 1,617 to 1,653, and AutomationBench-AA from 62% to 68%; on scientific reasoning, HLE rises from 39% to 43% and CritPt from 18% to 26%. The cost is token usage: about 145k tokens of output per task, more than 40% above the previous generation's 103k, at $1.49 per task, still below Mistral Large 4 Preview and GLM 5.3. Pricing is $0.30 per million input tokens, $0.90 for output, and $0.06 for cached input.

  • Original post & engagement: @TheInclusionAI open-source post, 80 likes, 2,696 views; @ArtificialAnlys evaluation post, 301 likes, 21,413 views.

  • Verification: The parameters, license and pricing come from Artificial Analysis's evaluation post, with open-source channels on Hugging Face and ModelScope; the scores follow a third-party evaluation standard, not vendor self-reporting.

  • Assessment: A typical iteration that gains a little on scores and spends a lot more tokens; the change in cost-effectiveness must be calculated against a specific workload, and looking only at a 3-point rise in the index overrates it.

2. AI Compute & Industry

No new AI compute and industry news worth listing separately was found in this issue.

3. Chinese AI Ecosystem

The Chinese Community Reads Genie 4 as "the Next Main Battleground for World Models"

  • What happened: After Genie 4 launched, dense commentary appeared on Chinese-language X, mostly focused on three points: native audio (ambient sound, BGM and character dialogue generated together), real-time scene changes in natural language (change the weather, move the camera, add people), and the positioning judgment of "moving from generating video to interactive simulation." The most widely spread post came from @op7418, calling it "the highest mountain in the world-model field," mentioning real-time generation of ambient sound, better image quality and physics, and finally being able to generate events with prompts during generation, while regretting that Ultra members cannot try it yet.

  • Why it matters: The focus of Chinese-language discussion differs from the English side — it rarely gets bogged down in parameters, and is more about ranking world models and judging whether this is the next main battleground; some are already pressing on evaluation methods, and one post argues that long runs and multiple rounds of editing will produce detail drift, so evaluation should look at causality and object consistency rather than whether a single frame looks right.

  • Original post & engagement: Sample posts @op7418, 70 likes, 15,133 views; @berryxia, 4 likes, 849 views; @xiax0603, 5 likes, 813 views. The overall magnitude is far below the English main post, a high-density, low-single-hit form.

  • Verification: Checked against the DeepMind official blog, the claim that "Gemini 3.8 Flash serves as the guiding agent" is accurate; claims such as "simulation duration up to 2 minutes, twice that of Genie 3" do not appear in the official blog, which says only that coherence holds for "short periods," so such numbers should be treated as second-hand relay; the like and bookmark counts cited in some other posts also do not match the collected values, a difference attributable to the time of posting.

  • Assessment: The Chinese side is currently mainly translation plus commentary, with no independent first-hand evaluation yet. Its value is that it quickly formed a consensus frame (world model = interactive + with sound + rewritable in real time); the risk is the habitual filling-in of metrics the company has not published (resolution, frame rate, duration).