翻訳待ち：The Sequence Radar-Issue #923: Last Week in AI: AI’s Industrial Turn

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

Next Week in The Sequence: More distillation coming to you. We break down GLM and Qwen new models. We discuss some ideas about moats in the era of AI. We will discuss some of the new platforms in AI for science. Subscribe and don’t miss out: 📝 Editorial: Last Week in AI: AI’s Industrial Turn For the past three years, we have watched AI through a microscope pointed at the model. Which system reasons better? Which benchmark moved? Which lab discovered the next scaling trick? This week, the camera pulled back. The most important developments were not new models at all. They were about ownership, power, and capital—the machinery required to turn intelligence from a research breakthrough into an industrial system. The reported $12.9 billion agreement for NVIDIA to acquire Hugging Face is the clearest example. Hugging Face is not simply another AI software company. It is the town square of the open-model ecosystem: the place where developers discover models, exchange datasets, publish evaluations, and assemble applications. For NVIDIA, this would connect two extraordinarily powerful control points. The company already owns much of the computational substrate on which AI runs. Hugging Face would give it a distribution layer through which AI is discovered and adopted. NVIDIA would no longer be selling only the engines. It would own part of the highway system directing traffic toward them. That logic is powerful, but it introduces a tension. Hugging Face became important because developers viewed it as relatively neutral infrastructure. Under the industry’s dominant chip supplier, every recommendation, integration, and technical default will receive more scrutiny. Vertical integration can accelerate an ecosystem. It can also make that ecosystem feel less open. Anthropic’s reported $45 billion, six-year agreement with NScale shows the same industrial transition from another angle. The eye-catching figure is the price, but the more revealing number may be 460 megawatts. Frontier AI companies are no longer purchasing cloud capacity like ordinary software startups. They are reserving power-plant-scale infrastructure years in advance. This is less like buying cloud credits and more like negotiating an energy treaty. A frontier lab must increasingly behave like a hybrid of a software company, a utility, and an infrastructure-finance operation. The scaling laws of AI now extend beyond parameters and tokens into electricity, cooling, networking, real estate, debt, and depreciation. Then there is a16z’s new $1.1 billion Machine Age fund, focused on chips, memory, networking, data centers, robotics, and energy. The symbolism is difficult to miss. The firm that popularized “software is eating the world” is now funding the physical systems needed to feed software’s enormous appetite. Software is still eating the world. It has simply started consuming steel, copper, concrete, and electricity. Taken together, these developments form a coherent picture. NVIDIA is moving toward developer distribution. Anthropic is locking in industrial-scale compute. a16z is financing the physical stack beneath both. The model still matters. But the model is becoming one component inside a much larger machine. AI began as a race to build intelligence. It is becoming a race to build, finance, and control the industrial system around it. 🔎 AI Research WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution AI Lab: Google Research, Virginia Tech Summary: This paper introduces a framework that allows AI agents to co-evolve skills alongside a persistent knowledge base that continually organizes raw execution traces into reusable patterns. As detailed in the referenced file “2608.27454v1.pdf”, this approach significantly outperforms existing skill-evolution methods, demonstrating that larger models particularly benefit from these transferable skills. SWE Refactor Bench: Can Coding Agents Complete a Long-Horizon, Whole-Repository Stack Migration? AI Lab: Navers Lab, Einsia.AI, Tsinghua University Summary: This benchmark evaluates whether coding agents can autonomously complete whole-repository stack migrations while preserving observable system behavior. Using a strict three-stage evaluation protocol, the study reveals that only 5.4% of tested models successfully complete migrations without breaking functionality. A Programming Paradigm for Spatiotemporal Composability AI Lab: Peking University, DeepSeek-AI Summary: The authors propose a unified formal foundation for dynamic composition in modern software by lifting classical effect and coeffect concepts to runtime mechanisms. This paradigm enables the complete reversal of a component’s side effects upon removal (temporal composability) and the reactive management of inter-component dependencies (spatial composability). Stream4D: 4D-Consistency for Streaming Autoregressive Diffusion Video Models AI Lab: UCLA, Tsinghua University Summary: To fix geometric drift and static motion collapse in streaming autoregressive video models, this paper introduces a reinforcement-learning framework utilizing a feed-forward 4D Gaussian Splatting reconstruction reward. By explicitly modeling scene dynamics and applying a motion prior, Stream4D improves 4D consistency and preserves realistic motion across extended video generation horizons. ONE SUCCESS ISN’T RELIABILITY: THINKINGBOX, A SANDBOX AND BENCHMARK FOR AGENTS IN STATEFUL BUSINESS WORKFLOWS AI Lab: University of Pittsburgh, Northwestern University, University of California, Irvine, Microsoft Summary: This paper presents an isolated sandbox and a 507-task benchmark to evaluate LLM agents on complex business workflows that require strict policy adherence and backend state transitions. Extensive testing reveals a significant discovery-reliability gap, showing that while models occasionally find successful trajectories, they consistently struggle to execute stateful tasks reliably across repeated attempts. 🤖 AI Tech Releases GLM-5.3-Flash Z.ai introduced GLM-5.3-Flash, its first multimodal model. Qwen3.8-Flash-Next Qwen released Qwen3.8-Flash-Next, a multimodal MoE model that uses new architecture ideas. Pipette Liquid AI open sourced Pipette, a benchmarking suite for on-device intelligence. 📡10 AI News You Need to Know About Nvidia has reportedly agreed to acquire Hugging Face for $12.9 billion, a nearly 3x jump from its $4.5 billion 2023 valuation, in a bet that a thriving open-model ecosystem keeps more of the market dependent on Nvidia hardware. Lambda raised about $1 billion of short-dated private debt, arranged by JPMorgan, to buy Nvidia GPUs it will lease to Microsoft, its third major debt raise since May and part of over $400 billion in AI-related debt issued globally this year. Hugging Face and Pollen Robotics launched Microduck, a $399 open-source bipedal duck robot with a camera, lidar, and a full RL training stack on GitHub, shipping before Christmas. Generalist raised a roughly $200 million extension led by 8VC at a $3 billion valuation, bringing its Series B to $600 million just months after Radical Ventures led the first $400 million at $2 billion. Instinct, the year-old AI assistant startup led by 23-year-old Noah Shinn, raised a $250 million Series B co-led by Index and Benchmark at a $2.5 billion valuation while still in private beta and under fire for aggressive app permissions. a16z closed a $1.1 billion Machine Age Fund, its first dedicated hardware vehicle, to back chips, memory, networking, data centers, and robotics, arguing that 20-30% annual hardware supply growth cannot keep up with triple-digit AI compute demand. Salesforce posted Q2 FY27 revenue of $11.3 billion, up 11%, with cRPO up 14% to $33.5 billion, and raised full-year revenue guidance to $46.1 to $46.4 billion. Anthropic agreed to a six-year, roughly $45 billion deal with Nscale to rent about 460 MW of Nvidia Vera Rubin capacity from its Monarch campus in West Virginia starting late 2027, its latest in a run of compute deals with Volta, AMD, SpaceX, Amazon, Google, and Broadcom. Bengaluru-based Runable raised a $21 million Series A co-led by Susquehanna and Nexus at a $65 million valuation to extend its general-purpose agent from building websites and apps into running ads, SEO, and social for small businesses, despite negative gross margins from subsidizing over 1 trillion tokens of usage in 90 days. Presentation startup Gamma acquired Accel-backed Lica, whose co-founders will lead a new design research lab exploring multimodal and personalized presentation formats, as the AI presentation category consolidates following OpenAI’s purchase of NextSlide earlier this month.