AI News | October 28, 2026

Lead: October 27's AI news has one theme: AI is starting to do science itself. OpenAI dropped three manuscripts at once, claiming simultaneous breakthroughs in mathematics, computer science and theoretical physics — the most striking being that the Exponential Time Hypothesis is false. The same day, Axiom Math announced an AI agent swarm wrote 43.1 million lines of Lean code in 24 days, producing the first computer-verified proof of the finite simple group classification theorem. Models are turning from problem-solving tools into instruments that produce raw research data — but both results are still at the "claimed" stage, and the real test lies with peer reviewers.

1. OpenAI claims three scientific breakthroughs: local smoothing, the Exponential Time Hypothesis is false, and a QCD existence proof

On October 27, OpenAI announced that its AI models made new progress in mathematics, computer science and physics. In mathematics, the model proved the local smoothing conjecture in all dimensions, and with it the related Kakeya conjecture. In computer science, it claims the Exponential Time Hypothesis (ETH) — proposed by Impagliazzo and Paturi in 1999 and used as a premise by hundreds of papers — is false. In physics, it gave a rigorous mathematical existence proof for quantum chromodynamics (QCD) and Yang–Mills theory, touching one of the seven Millennium Prize problems. OpenAI says the mathematics and computer science results have been formally verified with the Lean proof tool, and the proof materials are public.

OpenAI's own wording deserves quoting verbatim: the manuscripts are "best understood as raw research data," and the company is "trying a new approach" — working with outside experts to understand the results. In other words, OpenAI itself stops short of a full "we proved it" claim. Independent breakdowns add a necessary cold shower: the ETH algorithm only beats the human record at astronomically large input sizes (on the order of 2^521,283 switches); it refutes the hypothesis at infinite scale, but it does not prove P equals NP and does not directly affect everyday encryption.

Summary: OpenAI on October 27 announced three results — local smoothing and Kakeya, ETH is false, QCD existence and Standard Model completion — with math and CS results said to be Lean-verified; the company itself frames the manuscripts as "raw research data" for outside experts to interpret.

Why it matters: If any one of these survives peer review, textbooks get rewritten — a false ETH rearranges the foundations of complexity theory, and a QCD existence proof is Millennium-Prize-level. But the more durable lesson for practitioners is the paradigm shift: AI moving from "helping humans solve problems" to "producing research data humans don't yet fully understand," with verification (Lean formalization, peer review) becoming the bottleneck. Until peers confirm, keep the word "claimed" attached.

Sources: Sohu / market news: OpenAI announces three-field progress (Oct 27) · Cailian Press via Sina: OpenAI's three new results (Oct 27) · YouTube breakdown: ETH paper algorithm details and applicable scale (Note: conclusions are OpenAI's claims, not yet peer-reviewed; the ETH algorithm beats the human record only at extremely large input sizes)

2. Axiom Math: an AI agent swarm wrote 43.1 million lines of Lean code in 24 days, completing the first computer-verified proof of the finite simple group classification

On October 27, Axiom Math announced the first computer-verified proof of the classification of finite simple groups (CFSG). Engineers and mathematicians directed a swarm of dozens of AI agents that completed the work in 24 days, writing about 43.1 million lines of code in the Lean proof assistant. The classification theorem identifies the basic building blocks of finite groups; its original proof spans more than 500 papers, about 15,000 pages, and roughly 100 mathematicians — earning it the nickname "the Enormous Theorem." A proof no single human can fully read now has a version a machine can check line by line.

Summary: Axiom Math on Oct 27 released the first computer-verified CFSG proof: dozens of AI agents, 24 days, ~43.1 million lines of Lean; the original proof ran ~15,000 pages across 500+ papers by ~100 mathematicians.

Why it matters: This pairs directly with item 1: OpenAI produces manuscripts "to be interpreted," Axiom produces a proof "to be checked" — and the latter is exactly what the former lacks. For math and formal-verification practitioners, AI agent swarms can now staff proof-engineering projects at enormous scale. For the wider AI field, the lesson is that "verifiability" is becoming the dividing line for AI research output: results checkable line by line in Lean are in a different value class from results that must simply be believed. Watch next: independent reproduction of the check, and the error rate in the AI-generated portions.

Sources: Unite.AI: Axiom Math releases Lean proof of group classification (Oct 27)

3. Liquid AI launches LFM3: six 2B-class open models — text, vision, audio, omni and decision — on one shared backbone

On October 27, Liquid AI announced LFM3, a new generation of open Liquid Foundation Models putting text, vision, audio, omni and decision models on one shared backbone. The first release covers six models in the 2B parameter class: LFM3-1.7B (text-only), LFM3-VL-2B (text, image and video in), LFM3-Audio-2B (text or audio in, text or speech out), LFM3-Omni-2B (the company's first omni model: four modalities in, text and speech out), plus two decision models, d1-2B and d1-omni-2B, which return calibrated answers rather than generated text. All branch from the same LFM3-1.7B-Base: 1.73B parameters, 256K-token context, 16 languages, under the LFM Open License v1.0, with weights on Hugging Face.

Two official numbers are worth noting: the base model was pretrained on 22 trillion tokens, and LFM3-1.7B outperforms the previous LFM2.5-2.6B with 35% fewer parameters; at 4-bit precision the text model fits in about 1GB. The architecture keeps gated short convolutions for more than half its layers plus sliding-window attention, and was tuned on 14 device and backend combinations including Samsung and OnePlus phones, Raspberry Pi 5, Apple and AMD laptops and Qualcomm's Hexagon NPU. Day-0 partners include Ambarella, AMD, Arm, Intel via OpenVINO, NVIDIA, OpenRouter, Qualcomm and Vercel; Mercedes-Benz and Shopify joined pre-release testing.

Summary: Liquid AI on Oct 27 released LFM3: six 2B-class open models sharing one 1.7B base, covering text, vision, audio, omni and decision; 256K context, 16 languages, Hugging Face weights; ~1GB at 4-bit, built for on-device deployment.

Why it matters: On-device AI competition is shifting from "can it run" to "is one backbone enough": one backbone growing six capabilities means developers can cover chat through decision-making with a single fine-tuning and deployment pipeline instead of maintaining a model per modality. For phone, vehicle and edge teams, LFM3 is worth benchmarking on-device — but first verify multilingual coverage (which 16 languages, per the model card) and whether the decision models' calibration lives up to their name.

Sources: Unite.AI: Liquid AI launches LFM3 family of six open-weight models (Oct 27) · AI Brief: LFM3 on-device release details (22T tokens, 14 device tunings)

4. Google DeepMind rolls out Genie 4: a real-time interactive world model in research preview, with sound and dialogue generated in sync

On October 27, Google DeepMind announced Genie 4, its new world model for real-time interactive simulation, released as a research preview to trusted testers. Genie 4 is natively multimodal: music, ambient noise and interactive dialogue are generated simultaneously alongside the simulated world, not dubbed in afterwards. DeepMind says it "unlocks unprecedented realism," with more photorealistic visuals, fluid movement and audio produced directly by the model. The technical base comes from the Gemini family: Gemini Omni anchors simulations in real-world appearance and behavior, while other Gemini models supply the reasoning that keeps a simulation coherent — a user sets an event in motion, and the model simulates what happens next, how it sounds, and how the people in it move, react and speak.

Summary: DeepMind on Oct 27 released Genie 4 world-model research preview: real-time interactive simulation with music, ambient sound and dialogue generated natively in sync; Gemini Omni grounds appearance and behavior, Gemini reasoning maintains coherence.

Why it matters: World models are a key link toward robot training and agent simulation environments: training robots has meant either expensive real data or rule-based simulation; a model that generates infinitely variable, enterable, audible worlds changes that trade-off. It remains a trusted-tester preview, so ordinary developers can't try it yet — note the direction, not the immediate hands-on: interactive simulation is upgrading from "generating video" to "generating a world you can enter," with implications for games, robotics simulation and interactive content.

Sources: Unite.AI: DeepMind rolls out Genie 4 research preview (Oct 27)

5. Baidu releases and open-sources ERNIE Unlimited OCR 2: document parsing goes from tens of pages to "unlimited" — 3,600 pages in one pass

On October 27, Baidu released and open-sourced ERNIE Unlimited OCR 2, the first system to parse documents of any page count — pushing OCR's limit from tens of pages to "unlimited." In testing, the model parsed all seven Harry Potter books — over 3,600 pages — in a single pass while maintaining high accuracy and efficiency. On public benchmarks, it scored 84.6 on MPDocBench (multi-page document parsing) and 77.1 on MarathonDocBench (hundred-page-scale long-document parsing), reaching end-to-end state of the art and leading Qwen 3.8-Max, Gemini-3.8-Flash, GLM-5.3-Flash and DeepSeek-4.1-Flash.

The key technique is Sparse Reference Sliding Window Attention (SR-SWA), modeled on how a human copies a book: focus on the current page, and only look back at relevant pages when cross-page checking is needed, rather than re-reading the whole book. A page gate narrows each query's visual reference range from the entire document to relevant pages while modeling cross-page positions; top-scoring visual token buckets then capture long-range visual dependencies. Baidu reports single-step decoding in hundred-page scenarios 7× faster than full attention and 5.6× faster than the previous generation R-SWA.

Summary: Baidu on Oct 27 released and open-sourced ERNIE Unlimited OCR 2: any-page-count parsing, 3,600-page Harry Potter series in one pass; MPDocBench 84.6, MarathonDocBench 77.1, end-to-end SOTA; SR-SWA decoding 7× faster than full attention on hundred-page documents.

Why it matters: For Chinese document digitization, earnings-report parsing and legal archives — where long documents are the norm — this is a direct productivity change: the old split-into-chapters-and-stitch approach loses cross-page information and costs more. Open weights mean teams can deploy locally and verify accuracy on real Chinese long documents; benchmark scores are vendor-reported, so performance on real business documents (blurry scans, cross-page tables, mixed layouts) still needs testing.

Sources: Tencent News: ERNIE Unlimited OCR removes the document-parsing page limit, 3,600 pages in one pass (Oct 27) (Note: benchmark scores and speed multiples are vendor-reported)

6. Apple launches M6 MacBook Pro and iMac: on-device AI up to 5× faster, on sale November 6

On October 27, Apple unveiled a new 14-inch MacBook Pro and iMac powered by its latest M6 chip, pitching faster on-device AI. The 14-inch MacBook Pro starts at $1,999, the iMac at $1,499; pre-orders open October 28 in 35 countries and regions, with availability from November 6. Apple's figures: the M6 MacBook Pro delivers up to 30% faster graphics and up to 25% faster AI performance than the M5, with up to 24 hours of battery life — the longest ever in a Mac; the new iMac delivers up to 5× faster on-device AI than the previous M4 model. The launch comes a day before an Apple event in Los Angeles on October 28, which Bloomberg reports will feature Apple's first touch-screen MacBook Pro.

Summary: Apple on Oct 27 launched the M6 MacBook Pro ($1,999+) and iMac ($1,499+): M6 AI 25% faster than M5, 24-hour battery; iMac on-device AI 5× faster than M4; on sale Nov 6.

Why it matters: On-device AI is becoming a reason to upgrade: when 2B models like LFM3 run on phones and OCR can read 3,600 pages locally, a chip's local inference capability directly caps the user experience of those models. Apple putting AI performance in the launch headline signals that "running models locally" has moved from enthusiast demand to mainstream selling point. For developers, the M-series local inference ecosystem (MLX, Core ML) deserves a fresh look as an alternative to cloud API cost and latency.

Sources: Reuters: Apple launches M6 MacBook Pro and iMac, pitches faster on-device AI (Oct 27)

7. Switzerland's Giotto launches a sovereign AI family: Solo and Alto, pitched as "AI public infrastructure" with no US dependency

On October 27, Lausanne-based AI lab Giotto.ai announced Giotto Solo and Giotto Alto, a family of Swiss-developed reasoning models for public cloud, dedicated cloud and on-premises deployment, alongside free anonymous chat and iOS and Android apps. Alto is free to everyone and runs on eight B300 GPUs; Solo is billed as the smartest model to run on a single GPU. Deployment routes deliberately avoid US dependency: European shared cloud, dedicated cloud in Switzerland or Europe, or on-premises. Giotto's stated mission: "to make advanced AI a public good that people and organisations can use on their own terms."

Summary: Giotto on Oct 27 launched sovereign AI family Solo and Alto: efficient reasoning, free anonymous chat; Alto free, 8× B300; Solo single-GPU; European cloud and on-premises deployment, no US dependency.

Why it matters: "Sovereign AI" is turning from slogan into product form: after government statements, here is a lab writing deployment location, data ownership and portability into the product spec. For European institutions and companies operating abroad, the value is not benchmark scores but compliance and control: data stays in jurisdiction, models can run locally, the vendor sits outside US jurisdiction. For Chinese readers, Giotto's model is also a reference point: as AI becomes infrastructure, "who operates it, where the data sits, and whether you can move it" is becoming a selection criterion alongside performance.

Sources: TradingView / Reuters release: Giotto launches sovereign AI family (Oct 27)

8. Policy signal: US think tank proposes disclosure of "AI building AI" — quarterly reports, major changes within 4 business days

On October 27, the US nonprofit Americans for Responsible Innovation (ARI) proposed a framework: major AI developers would have to disclose whether and how extensively they use their own technology to develop new models, filing quarterly reports to a new Commerce Department office, with significant changes in research practices notified within four business days. The premise: AI increasingly performs the research needed to build more capable successors, shortening development timelines and leaving government and safety researchers less time to evaluate emerging risks — "Washington's visibility into that shift should not depend on what companies choose to announce."

Summary: ARI on Oct 27 proposed a disclosure framework: major developers report the extent of AI use in developing new models, quarterly to a new Commerce office, major changes within 4 business days; the backdrop is AI-built-AI potentially accelerating development and compressing risk-evaluation windows.

Why it matters: This proposal connects directly with items 1 and 2: if AI is producing math proofs and designing successor models, then "how much AI participated in building the next model" stops being gossip and becomes baseline data for regulation and safety evaluation. It is currently only a think-tank framework, not law — but it marks a regulatory shift from governing "what models output" to governing "how models are made." For practitioners, R&D auditability (data, compute, AI involvement) will sooner or later enter compliance checklists; keeping records now is cheaper than reconstructing them later.

Sources: PYMNTS: AI disclosure proposal targets how models build successors (Oct 27)

Closing: In one sentence: AI is shifting from a tool that does work for people to an instrument that produces new knowledge — and humanity's next homework is learning how to verify what it produces.

Chinese version: AI 大事件 | 2026年10月28日