AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。
Next Week in The Sequence: We start a new series about recursive self-improvement. Can’t miss it. In the learning loop, we dive into DeepSeek’s new release, Meta’s Muse and DeepMind’s amazing AlphaGenome. We will cover another AI robotics startup you need to know about. The opinion section explores the culture clash between massive scaling in the West vs. algorithm improvements from China Subscribe and don’t miss out: 📝 Editorial: Last Week in AI: DeepSeek V4.1-Flash, AlphaGenome Atlas, Meta Muse, and OpenAI’s Proposed Math Breakthrough This week’s AI news looked like four different conferences accidentally sharing a venue. DeepSeek introduced a faster model. Google DeepMind mapped billions of possible genetic changes. Meta launched a personal agent. OpenAI announced a proposed solution to a Millennium Prize problem. Somewhere between your grocery list and the mathematics of fluid motion, a useful pattern emerged: the industry is getting increasingly creative about turning computation into work. Start with DeepSeek’s V4.1-Flash, released September 10. It brings native visual understanding and, according to DeepSeek’s own evaluations, surpasses the previous V4 Pro across capability, cost, and speed. The company considers the improvement substantial enough to begin retiring the older Pro model. These are vendor results, but the proposition deserves attention: yesterday’s premium capability is becoming today’s smaller, faster workhorse. For developers, that changes the budget for intelligence. An agent might inspect a screen, propose an action, execute it, check the result, and retry. Every additional step consumes time and tokens. Better inference efficiency makes more of these feedback loops practical. Think of an engine becoming efficient enough that you can finally afford to drive the vehicle somewhere interesting. AlphaGenome Atlas applies a different computational strategy: do an enormous amount of work upfront and make the results reusable. DeepMind’s new resource contains predicted molecular effects for roughly nine billion possible single-letter substitutions across the human genome. Its Variant Impact score combines information from AlphaGenome and AlphaMissense to help researchers prioritize changes for investigation, including those outside protein-coding regions. Imagine inheriting a vast codebase without documentation. You can read every character, but figuring out which edits break which functions is another problem entirely. Atlas offers a predictive index of those edits. The predictions require biological validation; their immediate value is helping scientists choose better experiments. Precomputed inference becomes shared scientific infrastructure. Meta’s Muse brings the question closer to everyday life. The personal agent runs in a dedicated virtual machine with a browser, can continue working after the app closes, and uses connected services to pursue tasks. Meta also describes a separate Sentinel agent that checks outbound activity, alongside user approvals for sensitive actions. The interesting engineering unit here is the entire system: model, memory, computer, permissions, and execution history. A useful personal agent needs all of them. Planning dinner sounds trivial until software must reconcile calendars, dietary restrictions, reservations, and somebody changing their mind. Everyday competence has an impressively large test suite. OpenAI’s mathematics announcement explores the opposite end of the difficulty spectrum. The company reports that an internal model, deployed through roughly 10,000 concurrent agents, produced a proposed Navier–Stokes solution in 88 hours, followed by 17 hours of Lean formalization and verification. The claimed result constructs finite-time breakdown from smooth initial conditions with smooth external forcing. That extraordinary claim deserves independent mathematical scrutiny, including examination of the formal statement and assumptions. The computational approach is itself revealing: researchers coordinated parallel searches, shared intermediate discoveries, and redirected effort toward promising results. Here, inference starts to resemble a research organization. My reading of the week is that AI progress increasingly depends on how intelligence is deployed. DeepSeek expands the computation developers can afford. AlphaGenome makes predictions reusable. Muse connects reasoning to persistent action. OpenAI explores coordinated search at extraordinary scale. Each approach creates a different verification problem: did the task succeed, does the biological prediction hold, is the proof correct? The next phase of AI will reward systems that can turn all those tokens into outcomes we can actually check. 🔎 AI Research On the Navier–Stokes Millennium Prize Problem AI Lab: OpenAI Summary: OpenAI reports an analytical proof—and a Lean formalization—that smooth three-dimensional incompressible Navier–Stokes flow can develop a finite-time singularity under a smooth external force with finite energy, resolving Millennium Prize statements C and D via a self-similar inward-spiraling vortex. The writeup was produced by a large multi-agent system powered by an internal model beyond GPT-6 Astra, after agents first resolved an unforced Euler blowup question. Procedural Graphs: Self-Evolving Execution Structures for LLM Agents AI Lab: Google, Georgia Tech, Peking University Summary: This paper introduces Procedural Graphs, editable (procedure, relation, procedure) structures that answer *what-to-do* the way knowledge graphs answer *what-is*, with online generative guidance that soft-biases ReAct without hard constraints and offline self-evolution that Add/Delete/Updates topology under a validation gate. Across HotpotQA, MultiChallenge, GDPval, ALFWorld, τ-bench, BFCL, and EnterpriseArena, PG-guided Claude, Gemini, and Grok solvers often set or match the best score versus memory and workflow baselines—including large survival lifts on EnterpriseArena. Online Draft Co-Training for Speculative Decoding in RL Post-Training AI Lab: NVIDIA Summary: The authors make online draft co-training practical for large-scale, long-context RL by fixing two systems bottlenecks: branch-aware packed zigzag ring attention under context parallelism (EAGLE-3, DFlash, DSpark) and TapChannel side-path transport of target features under pipeline parallelism, integrated in NeMo-RL. Co-trained drafts keep acceptance high as the policy evolves, delivering up to ~1.88× end-to-end RL speedups and scaling through 122B-class targets and 256K-token contexts. Kalman Delta Networks: Uncertainty-aware Associative Memory AI Lab: Yale University Summary: This work casts delta-rule associative memory as a linear–Gaussian SSM and derives Kalman Delta Networks that propagate both memory state and uncertainty so write gains track evidence; Diagonal and Isotropic variants stay scan-compatible at low cost. At 750M/50B and 1.3B/100B FineWeb-Edu scales, KDNs beat strong linear-attention baselines (including Mamba-3 and gated delta variants) on perplexity, zero-shot averages, and RULER retrieval. Recognition–Refusal Misalignment in LLMs AI Lab: University of Southern California, Arizona State University Summary: The paper shows that models linearly encode structural unanswerability (math/code) with mean probe AUC 0.939 across 11 models, yet that recognition direction is nearly orthogonal to safety-refusal directions (mean cos ≈ 0.087)—so confident answers to impossible questions look like a routing failure, not missing knowledge. Steering the recognition axis flips abstention behavior by +33–52 percentage points, and the geometry largely appears before instruction tuning. 🤖 AI Tech Releases Agents API OpenAI introduced the Agents API, a managed Codex harness for durable cloud agents—sessions, orchestration, context compaction, recovery, tools/MCP, and parallel subagents—with OpenAI-hosted or self-hosted environments. Muse Meta introduced Muse, a personal AI agent that runs on Muse Secure VM, acts across everyday apps with user-gated access, and is powered by Muse Spark—rolling out free in the US on iOS, Android, and muse.ai, with AI glasses coming later. AlphaGenome Atlas Google DeepMind released AlphaGenome Atlas, a 1-petabyte catalogue of predicted molecular effects for all ~9 billion single-nucleotide variants in the human genome, with AVI impact scores, free academic web/API access today and commercial access on Google Cloud coming soon. DeepSeek-V4.1-Flash DeepSeek released DeepSeek-V4.1-Flash, a 552B MoE with an asymmetric Causal Encoder–Decoder (8B active on input, 16B on output), native multimodal support via deepseek-flash, and a much smaller KV cache—now live on the API as older Flash/Pro endpoints begin routing over. 📡10 AI News You Need to Know About Cognition raised over $2 billion at a $48 billion valuation in a Series E led by Andreessen Horowitz and Accel, with Founders Fund, General Catalyst, and Avenir returning, as Devin’s annualized run-rate revenue climbed from $492 million to nearly $900 million since May. Mistral closed a €3 billion Series D at a post-money valuation of more than €21 billion—what it calls the largest equity raise ever by a European tech company—led by Samsung Electronics with Scaleup Europe Fund and PSG Equity as co-leads. Harvey raised $550 million at a $15.6 billion valuation in a round co-led by Lightspeed Venture Partners and Diffusion, aimed at funding the legal AI startup’s push to build its own models. Bloomberg reported that Alibaba is set to lead a $300 million investment in UniPat AI, an AI training and benchmarking startup founded by a former Alibaba staffer, at a $2.5 billion valuation, with Tencent and existing backer HSG also participating (talks still open). Listen Labs walked away from a signed $125 million Series C term sheet at a $1.5 billion valuation (Menlo Ventures to lead) amid Salesforce talks to buy the AI customer-research startup for around $2 billion—discussions that are not final. Cymphony launched with $30 million in funding ($25 million Series A co-led by Sequoia Capital and SMBC Fin Atlas Beyond Fund) for an AI-agent governance and security platform that maps how employees and agents access enterprise data and systems. TechCrunch reported that XDOF, a robotics teleoperation-data startup less than three months out of stealth, is in late-stage talks for a Series B at about a $1.2 billion valuation led by 8VC, with annualized revenue approaching $50 million (terms not final). Dario Amodei called for companies to “pace the frontier,” proposing embedded third-party evaluators (such as METR), industry coordination on safety standards and rate limits among democratic-country labs, and limited global coordination—with Anthropic unilaterally committing to embedded evaluators now. TechCrunch reported that Mecka AI, which collects egocentric human-motion data for robot training, is nearing a Sequoia-led round at about a $500 million valuation, just three months after a $60 million Framework-led raise (size and terms not final). Bloomberg reported that ByteDance founder Zhang Yiming is personally overseeing a Seedance-based real-time spatial-video “world model” aimed at interactive 3D environments for livestreams, dramas, games, and Pico headsets—possibly as early as October, with cloud rendering cited around 20 fps and ~0.05s latency (timing not final; no company announcement).