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 continue our series about recursive self improvement. The learning loop dives into Reflection’s new model, EmbeddingGemma2 and Mistral4. Our opinion section dives into the debate whether value in AI is accumulating on the infrastructure vs. the application layer. We will discuss another robotics stack. Subscribe and don’t miss out: TheSequence is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. 📝 Editorial: Billions in Funding, Hundreds of Papers, One Big Leap? AI’s weekly news cycle increasingly resembles two competitions running in parallel: researchers compete to improve intelligence, while investors compete to price its future. This week, Reflection and Mistral supplied new models, TypeSafe supplied another spectacular financing, and Arena supplied a reminder that someone must measure what all this intelligence does. Together, they suggest a market developing interesting ambitions beyond a better chatbot. Reflection introduced Beam, a mixture-of-experts model with 501 billion parameters, of which 23 billion activate per token. Its technical story centers on reinforcement learning and inference efficiency: Reflection reports generating more than 100 million training rollouts and achieving competitive reasoning performance with substantially less estimated inference compute than some larger rivals. Those estimates are not measured deployment costs, and the weights remain promised for later this month. Still, the proposition matters: agents that execute lengthy workflows need intelligence with an operating budget. Mistral offered a European counterpart with Large 4, nicknamed Le Chonk. The trillion-parameter multimodal model entered public API preview, with weights scheduled for release by month’s end. Mistral emphasizes coding, cybersecurity, finance, and law, alongside training on infrastructure it operates in Europe. My reading is that sovereignty is becoming a product feature. For an enterprise, choosing a model also means choosing who controls deployment, customization, and continued access. An impressive benchmark cannot answer those questions. OpenAI added a more fundamental dimension: intelligence contributing to mathematical discovery. The company released more than 700 manuscripts from an unreleased model, reporting progress on open research problems and supplying Lean formalizations for a portion of the results. Verification remains unfinished, and several manuscripts have already been corrected or withdrawn. That distinction matters: generating a plausible proof and establishing mathematical knowledge are different achievements. Yet the possibility is extraordinary. AI could accelerate research enough that checking, interpreting, and connecting its results becomes a central bottleneck. Arena’s question—can we trust what the model did?—suddenly extends from enterprise workflows to the foundations of science. TypeSafe’s $870 million Series A, led by Andreessen Horowitz at a $7.5 billion valuation, put a price on a different architectural bet. Its Jev model produces typed decisions with confidence estimates that software can consume directly. Consider the thousands of small judgments inside a business process: classify this request, select a tool, escalate this exception. These decisions offer a plausible market for specialized intelligence. The valuation assumes that market will be enormous. The engineering opportunity is compelling; the commercial outcome still needs proving. Arena’s $200 million Series B at a $3.1 billion valuation rounded out the picture. Alongside the financing, it introduced an Alignment Index examining behaviors such as unauthorized actions and false claims of task completion. This is a consequential expansion of evaluation. A model can produce an excellent answer and still be an unreliable colleague. Once agents operate tools and modify business systems, measuring their behavior becomes part of the deployment stack. Investors are betting that the machinery for judging intelligence can become valuable infrastructure itself. The tempting interpretation is that AI has discovered four more ways to absorb capital. A more useful interpretation is that the industry is separating intelligence into distinct economic problems: producing capability, delivering it efficiently, retaining control, and verifying behavior. These problems reinforce one another. Cheaper inference makes more automation feasible. Greater autonomy makes evaluation more necessary. Open weights expand deployment options while shifting operational responsibility toward customers. 🔎 AI Research Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses AI Lab: Microsoft Research Asia Summary: This work introduces Harnessed Agentic RL, where the same agent harness used in deployment (such as OpenHands, Claude Code, or Codex) takes part directly in reinforcement learning through an OpenAI-compatible LLM proxy, so agents don’t need to be reimplemented inside the training framework. The open-source v1.0 fits a full RL control plane into about 3,500 lines, runs agents as Kubernetes jobs, and lifts Qwen3.5-9B from 41.8% to 56.4% Pass@1 on SWE-bench Verified with roughly 6,000 training samples. Long-WAM: Scaling the Context of World-Action Models AI Lab: NVIDIA, MIT, HKU, UCSD Summary: This paper presents Long-WAM, a framework for scaling the visual history that world-action models use for real-time robot control. Its key finding is that longer context only helps when the video foundation is pretrained autoregressively: on RoboCasa GR-1, going from no history to 19.2 seconds raises success from 63.3% to 78.7%, while a bidirectional start gains nothing. Streaming encoding and asynchronous execution keep each action chunk at 107.4 ms on an RTX 5090, and on real Unitree G1 and YAM robots it hits 95% on dynamic cup stacking where π0.5 and Fast-WAM failed all 20 trials. Recurrent Looped Transformer AI Lab: Princeton University, University of Pennsylvania Summary: This paper introduces the Recurrent Looped Transformer (RLT), which splits its layers into a parallel causal encoder and a recurrent decoder that feeds each token’s final state into the next, so computation grows with sequence length at a fixed per-token cost. Trained on at most 40 bits, it generalizes parity to 256 bits with 100% accuracy where a matched Transformer stays at chance, and it reaches 97% on S5 permutation tracking at 8x the training length versus under 1%. Ablations show the gains vanish without the recurrent feedback. Structuring MoE Expert Selection for Agentic Reinforcement Learning AI Lab: Apple, Purdue University Summary: This paper finds that off-the-shelf MoE routers already group experts by agentic operation, such as READ or UPDATE, but standard RL ignores that structure and lets routing drift. The authors add hierarchical routing control, aligning turn-level expert selection with operation labels and keeping token-level routing consistent, plus an entropy gate that switches it off when training destabilizes. It needs no architecture changes and improves success rates by more than 10 points on AppWorld and AutomationBench. Frozen Models, Evolving Expertise: Model-Agnostic Learning from Deployment Experience for Multimodal Medical AI AI Lab: University of Maryland, College Park, NVIDIA Summary: This paper proposes a model-agnostic framework that lets frozen LLMs and VLMs learn from deployment experience without changing weights, using a Skill for reasoning and tool use, a Knowledge Memory of verified facts, and a Multimodal Knowledge Base of visual cases. Updates are kept only if they help on new cases without hurting earlier ones. Across six benchmarks and four base models, it improves medical task performance by up to 34.2%, transfers to other models, and also works outside medicine. A GPU-Parallel Framework for Heterogeneous Multi-Task Reinforcement Learning AI Lab: NVIDIA, HKUST (Guangzhou), Tsinghua University Summary: This paper introduces Hebero, a GPU-parallel Isaac Lab benchmark for training one policy across 40 heterogeneous manipulation tasks, along with Demonstration-Guided Policy Optimization (DGPO). Its IW-ABC method uses per-task learning progress to relax behavior cloning and upweight lagging tasks in PPO. With 50 demonstrations per task it reaches 90.1% mean success from state inputs, 7.8 points above the best baseline, and 93.5% with visual inputs, and one simulated policy performs four tasks on a physical Piper robot. 🤖 AI Tech Releases Mistral Large 4 (le Chonk) Mistral launched a preview of Mistral Large 4, a 1-trillion-parameter natively multimodal open-weight model with 49B active parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs. The preview API is live now, and weights follow by the end of the month. Mistral reports strong coding and agentic results (61.7% on DeepSWE v1.1) and the top open-weight score outside China on the Artificial Analysis Cyber Index. Claude Haiku 5.5 Anthropic released Claude Haiku 5.5, its cheapest and fastest model, at $0.10 per million input tokens and $0.50 per million output tokens, 90% below Haiku 4.5. It’s the first Haiku with an adjustable effort setting and scores 72.4% on OSWorld 2.1 and 39.2% on Terminal-Bench 4.0, ahead of OpenAI’s GPT-6 Luna on both. Reflection AI Beam Reflection AI introduced Beam, its first open-weight model: a text-only mixture-of-experts with 501B total and 23B active parameters and up to 1M tokens of context. Reflection says it matches GLM-5.2 on reasoning with 3–4x less inference compute, which hasn’t been independently verified. It’s in early access now, with Apache 2.0 weights due later this month. EmbeddingGemma 2 Google DeepMind released EmbeddingGemma 2, an Apache 2.0 embedding model that maps text, code, images, video, and audio into one shared 768-dimensional space. At 740M parameters with modular vision and audio encoders, it’s built for on-device search and retrieval, and it’s available on Hugging Face. 📡 10 AI News You Need to Know About GPU cloud provider Lambda is raising up to $4 billion at a $14.5 billion pre-money valuation in what could be its last private round before a planned 2027 IPO, led by Coatue Management and Blackstone—after its backlog jumped from $15 billion in June to $50 billion in September, largely on a ~$35 billion Anthropic commitment, and days after Lambda raised an additional $1 billion in debt for data-center buildouts. Sierra and Meta, with partners Genesys, Instinct, Rocket, Shopify, Stripe, and Walmart, announced Personal Agent Protocol—an open standard for how consumer personal agents authenticate with and act on company websites, APIs, or company agents via OAuth sessions (guest or signed-in, read or write), with a v0.1 spec, design workshops, and reference implementation planned later this month. Chinese AI lab DeepSeek is nearing a second external funding round of at least ~80 billion yuan (~$12B)—and considering expanding it to as much as 100 billion yuan (~$15B), double its original target—at a ~500 billion yuan (~$75B) valuation ahead of a possible early-2027 Shanghai STAR Market IPO, with Tencent and CATL among the largest backers alongside Geely Auto, Monolith Management, and Loyal Valley Capital (after a ~50 billion yuan / ~$7.4B first round closed in June). Robot-data startup Mecka AI raised a $60 million Series B led by Sequoia, with participation from Nvidia and Microsoft’s M12, after TechCrunch reported it was nearing a round at a $500 million valuation. The 2024-founded company aims to be a Scale AI or Mercor for robotics, paying people to record everyday tasks like making coffee or fixing cars while wearing body sensors and using smartphones to generate human-motion training data for humanoids and other robots. AI leaderboard Arena, wh [truncated for AI cost control]