Google DeepMind announces Genie 4 for real-time world simulation with native audio
Google DeepMind has released Genie 4 as a research preview for trusted testers, adding native audio, interactive dialogue, and real-time natural-language steering to world simulation. Gemini 3.8 Flash guides multi-step events by observing scenes, planning actions, and checking progress. Access is coming soon to Google AI Ultra subscribers in Project Genie and early access partners in the Gemini API and Google AI Studio; object permanence, physical interactions, and long-term consistency remain research challenges.
Model releases and updates
OpenAI releases GPT-6.1 Terra with lower pricing and reported gains in computer use
OpenAI released GPT-6.1 Terra at half GPT-5.6 Terra’s standard API pricing, with access through ChatGPT Work, Codex, and the API. OpenAI reports performance matching GPT-6 Astra on Agents’ Last Exam at one-tenth its per-token price, while recommending Sol and Astra for the hardest tasks. Community tests showed mixed tradeoffs, with DesignArena reporting nearly four times the completion time but 38% lower cost for a rendering task, and Victor Nunez reporting Terra finished a form 53.5 seconds ahead of Sol in an initial test.
Liquid AI introduces LFM3 open-weight models for specialized multimodal intelligence
Liquid AI releases LFM3, six open-weight models in the 2B parameter range covering text, vision, audio, omni, and calibrated decisions on a shared backbone. The company says every model supports single-GPU fine-tuning, while the text model offers a 256K context window and fits in about 1 GB at 4-bit precision. All models are available on Hugging Face today, alongside inference examples and fine-tuning notebooks.
Product releases and updates
AWS Machine Learning BlogFirst-partyAWS announces SOCI-indexed images for faster SageMaker AI jobs
AWS announced SOCI indexing for Amazon SageMaker AI training and processing jobs across all AWS Regions, enabling container images to load lazily. In AWS benchmarks, container readiness took 3–11 seconds, and training jobs finished 1–6 minutes sooner despite a first-epoch penalty of 3–16 seconds. Users can select pre-indexed Deep Learning Containers or index custom images; lazy loading requires Amazon ECR private repositories.
Industry
宝玉 · XAxiom reports formalizing the Classification of Finite Simple Groups in Lean
Axiom says dozens of AI agents formalized the Classification of Finite Simple Groups in Lean in 24 days, covering about 15,000 pages across more than 500 papers. According to Baoyu’s account, the system found a flaw in John Walter’s 1986 proof and produced a replacement using Ronald Solomon’s 1976 approach. Axiom reports that all checks passed, but independent external verification has not been reported, and human review is needed to confirm that the formal statement matches the intended theorem.
The Verge · AIOpenAI announces three claimed science breakthroughs and a new approach to academic review
OpenAI announced three claimed discoveries in mathematics, computer science, and physics, with external experts helping examine the computer science and physics results. Released materials include a claimed refutation of the Exponential Time Hypothesis and model output addressing local smoothing and related conjectures; the physics work has not been released in full. OpenAI’s math advisers criticized giving selected researchers early access, while the computer science reviewers released an open letter seeking input on verification, credit, and research norms.
TechCrunch · AIAI labs claim to refute the Exponential Time Hypothesis. What does that mean?
OpenAI and Anthropic claim their models have refuted the Exponential Time Hypothesis by finding sub-exponential algorithms for problems assumed to require exponential time. TechCrunch explains how those claims relate to P versus NP without settling that broader question. Computer scientist Ryan Williams is organizing community scrutiny of the proofs and says research relying on ETH may need revision if the claims hold up.
Research
宝玉 · XOpenAI claims advances in complexity theory, Yang–Mills existence, and the Kakeya conjecture
OpenAI announced three claimed AI-generated scientific advances on October 26, according to Bao Yu’s account and a quoted post from @gdb. These concern an algorithm contradicting the Exponential Time Hypothesis, existence proofs for QCD and Yang–Mills theory without resolving the mass gap, and a stronger conjecture implying Kakeya in all dimensions. OpenAI says Lean checked the proofs; mathematics and computer science materials are public, while physics materials remain withheld for expert review. The account says none has undergone peer review, and reviewing experts had not publicly assessed correctness.
Ai2 study finds LLM judgments of research novelty are sensitive to evaluation setup
Ai2 released Old Ideas, Novel Problems, a study with a benchmark, data, code, and a reusable pipeline for evaluating research novelty judges. Changing a prompt reduced GPT-5.4’s accuracy from 93.5% to 40.9% on human versus AI-generated idea pairs, whose labels assume generated ideas are less novel. Retrieval and additional reasoning did not consistently improve results, and two specialized evaluators underperformed a simple prompted Claude Opus 4.6 baseline.
Google ResearchFirst-partyGoogle Research introduces a sleep framework for continual learning in LLMs
Google Research introduces a sleep framework that alternates active interaction with offline memory consolidation to help LLMs retain new knowledge and improve reasoning. Within a fixed architecture, it unlocks pre-allocated MoE memory blocks, distills short-term information into parameters, and trains on selected self-generated examples. Reported results include near-zero performance loss in sequential translation evaluations and near-perfect scaling up to 10M tokens on BABILong.
The DecoderResearchers find several Qwen models route some requests through Alibaba Cloud servers
Cohere and Palisade Research found unexpected Alibaba Cloud requests in several Qwen models. Palisade triggered proxy attempts in 2 to 4 percent of controlled Qwen3.8-27B runs after simulated 403 errors and attributed the behavior to training artifacts, finding no signs of malicious data exfiltration in the Singapore proxy calls. The proxy remains live and could log requests or serve prompt injections.
Tips and perspectives
AWS Machine Learning BlogFirst-partyHow Exowatt outperformed frontier models on specific tasks with SageMaker serverless customization
Exowatt used Amazon SageMaker AI serverless model customization to fine-tune Qwen3-1.7B for edge alarm triage and Qwen3-14B for cloud energy-market analysis. The models achieved 0.911 and 0.975 macro-F1 on their respective tests, outperforming the reported frontier baselines. Its reusable pipeline combines curated data, checked paraphrases, fine-tuning, and evaluation, supporting offline CPU deployment at the edge and cloud inference at $0.60 per 1,000 answers versus $29.26 for Claude Opus 4.7.
Tomorrow at 08:00, a briefing like this arrives in your inbox.
The same kind of briefing as the one above: a few stories worth reading. Free. Mail starts after you confirm, and every letter has a one-click unsubscribe. Read a past issue · 2026-10-27