📊 Today's Overview
The AI hardware trade cracked on both sides of the Pacific. Cambricon fell 10.81% and lost its ¥500B market cap, Zhongji Innolight dropped 12.69% on a Goldman downgrade, and the STAR 50 slid 3.81% — while CoreWeave jumped 7.16% as Oracle's $20B raise calmed AI-debt nerves. Meanwhile OpenAI disclosed that GPT-6 (Astra) designed a working Codex device end-to-end in under a week, and Meta open-sourced F3 Lightning to ship ad-ranking features independently. On the research side, Ai2's StateBench shows frontier models collapsing below 13% on hard state tracking.
🔥 Trend Insights
- Compute bottleneck moves to power: Musk says AI chip output will outrun electricity supply by end-2026; SemiAnalysis pegs SpaceX at 13GW of compute within 18 months.
- World models as data engines: 3D World Simulator and NVIDIA's Cosmos recipes both push simulation-trained policies toward parity with real-robot data.
- Small models, engineered hard: IBM's 20B Palette beats frontier models on deck generation via role-splitting, synthetic data, and a repair harness.
🐦 X/Twitter Highlights
📈 Trends & Buzz
- Musk: AI chip output will exceed power supply by end of 2026 - He says the bottleneck has shifted from chips to electricity, with generation capacity outside mainland China barely growing. Cooling, networking gear, and batteries are also constrained @elonmusk
- SemiAnalysis: SpaceX will build 13GW of compute in 1.5 years, becoming the world's largest AI cloud - The article title literally reads "Elon Is Doing It Again," putting the 13GW buildout at eighteen months @SemiAnalysis_ (semiconductor and AI analysis firm)
- Fireworks serves over 40T tokens per day of "specialized models" - Nathan Lambert defines a specialized model as "any model that has been post-trained for a customer." The 40T figure is single-day inference volume @natolambert (Ai2 researcher / author of Interconnects)
- Kimi K3.1 Pro drops weights tomorrow - Together AI hosts an architecture deep-dive the same day, led by Moonshot's Feihu Tang (Kimi architecture lead), covering long-running agents and production inference @togethercompute (Together AI, open-model infrastructure)
🔧 Tools & Products
- vLLM v0.33.0 ships: 927 commits, 389 contributors - New support for Reflection Beam-1.0, Step-5 Preview, Nemotron Parse 3.5, and MiniCPM-OCR 1.0. Speculative decoding jumps from 1373 to 3390 tok/s on GB300; GB300 decode latency drops 1.5–2x and time-to-first-token falls 39–48%. PyTorch moves to 2.14 with CUDA 13.0-only wheels, and ROCm 10.0 becomes the default @vllm_project (vLLM open-source inference engine / UC Berkeley)
- Mistral Large 4 lands on Open Math OS: 1.05T open weights, 1M context - Open Math OS is a research workbench shared by mathematicians and AI agents, running only open models, initially for approved teams. The same model independently built a 3D app, Orbital Trace, tracking ~19,000 satellites and debris in real time — acting as its own user, testing, speeding things up, and validating @MistralAI｜@MistralAI
- Pruna open-sources P-H3-Image: 8-step generation, 4MP in just 2.6 seconds - Adapted from MiniMax-H3 (a video generation model); the original pipeline took 50 steps and 12.4 seconds. The model is 7x smaller and runs on smaller GPUs. Supports 2–16MP rendering and editing with up to 8 input images, already integrated with Diffusers and ComfyUI (node-based image workflows) @PrunaAI (model compression startup)
- DR Pixel: a deep research agent that retrieves and reasons directly over page images without OCR - A collaboration between Berkeley NLP and Rental Dynamics; the trained 27B model matches GPT-5.4 across 1.06M pages @Hughie_Hu43 (Berkeley NLP researcher / first author)
⚙️ Engineering Practice
- Raschka rounds up 10 open LLM architecture releases in October, calling it "Open October" - Meta's first open-weight release since Llama, Muse Spark Open 816B, uses GQA+SWA plus LatentMoE; Ling 3.1 Flash 560B uses Kimi Delta Attention and gated MLA; Nemotron 3.5 Super VL 120B keeps only 8 attention layers and swaps the rest for Mamba-2; Mistral Large 4 1T combines MLA with sparse attention @rasbt (author of "Build a Large Language Model from Scratch")
- Ai2's new StateBench: accuracy collapses below 13% on the hardest tier - It tests state changes in long documents, like objects changing hands or variables being reassigned. GPT-6 Sol and Claude Opus 5.5 both fail on the hardest, longest problems @allen_ai (Allen Institute for AI)
- LlamaIndex speeds up OCRia-1 document parsing 10x on a single B200 - From 285 tok/s to 2800 tok/s. The approach: quantization plus dflash2 speculative decoding, along with 17 small optimizations (replaySSM, kernel fusion, vLLM changes), most of which add only 2–5% individually. Some optimizations will flow back to vLLM @llama_index (LlamaIndex, document parsing and RAG framework)｜@jerryjliu0 (LlamaIndex CEO)
- Krea 3's diffusion Transformer drops from 22.9s to 3.3s on AMD MI355X - Hedra used an autonomous research system for optimization, hitting 6.9x speedup while preserving 98.3% SSIM quality. MXFP6 GEMM is 2.8x faster than BF16, and the share of time spent waiting on GPU transfers falls from 18% to 2% @krea_ai (image generation platform)｜@hedra_labs (Hedra, generative media company)
- Extracting task-specific robot policies from a general language model: success rate doubles on three tasks - The author's view is that the language model itself is a general robot policy, but deploying it directly is too slow and too unstable. The extracted policies complete tasks 4–8x faster, and that multiple doesn't even count LLM inference latency @jkraut_ (paper author)
- Toyota TRI: pretrained models match single-task training with only 15% of the demonstrations - Russ Tedrake gave an 82-minute talk at Stanford. Data normalization alone contributes ~40 points of improvement — more than switching from diffusion to flow policy. Each task is evaluated 50 times, because at 10 runs the error bars are ~60 points wide. TRI is working with Boston Dynamics on Electric Atlas @0x_meden (speaker Russ Tedrake: MIT professor / head of TRI's large-scale behavior model team)
⭐ Featured Content
Meta F3 Lightning: decoupling publish and execution boundaries for 20K+ ad-ranking features — 3.6x faster independent feature launches at no extra cost ｜ The feature platform moves from whole-package releases to per-feature canary rollouts
F3 compiles 20K+ ad-ranking features into a single merged DAG with shared computation, but the whole graph is the release unit — any bad feature definition or failed model-related test blocks all features from shipping. F3 Lightning's core idea is to decouple the "publish boundary" from the "execution boundary," letting each feature compile, validate, deploy, and roll back independently. The hard part: once decoupled, shared computation disappears and CPU usage rises. Four optimizations claw it back — stable computation identifiers, a single shared execution plan, one memory buffer shared by all features, and dense index mapping for direct array access. For anyone building recommendation-system feature platforms, this is a complete engineering case study on "how to do per-feature canary rollouts without sacrificing shared computation."
Sources: engineering.fb.com
OpenAI officially discloses GPT-6 (Astra) doing end-to-end hardware engineering design: two people finish the form factor, PCB layout, and firmware for a conversational Codex device in under a week; a contract manufacturer produces a prototype in a month ｜ A first-hand sample of agents moving from software into the physical world
OpenAI's blog systematically discloses, for the first time, progress on models doing physical-world engineering design: starting from functional requirements, designing the form factor, enclosure, internal structure, circuit component selection and PCB layout, and UI plus firmware integration for a device used to talk with Codex — all without direct human editing, completed by two members as a side project in under a week, with a contract manufacturer producing a small batch of prototypes within a month. The post also candidly admits shortcomings: Astra can't reliably account for factory capacity limits, tolerances, and assembly details, still requiring multiple rounds of human review and iteration, and it offers an evaluation approach for engineering tasks. Reading it gives you the capability boundaries and failure list for agents entering hardware engineering.
Sources: openai.com
3D World Simulator: jointly predicting appearance/geometry/tactile in a unified explicit 3D representation space, 50+ FPS interactive long-horizon simulation on an RTX 5090, policies trained purely on simulated data rival real-robot data ｜ A real-time and modality breakthrough for embodied world models
The research team (Yunzhu Li / Berkeley / Jimmy Wang et al.) released 3D World Simulator: within a unified 3D representation space of point clouds/voxels/3D Gaussians, it simultaneously generates novel-view appearance, multi-view-consistent RGB-D, 3D geometry, and spatially aligned tactile signals, covering rigid, articulated, and deformable objects (folding clothes, part assembly, threading rope). It focuses on scenarios that are hard for pose estimation and point tracking, like occlusion during in-hand manipulation. The biggest highlight is the controllability that 3D brings — you can pick any camera viewpoint and combine learned interaction dynamics with Gaussian backgrounds and kinematic robot models into different simulation backbones, so policies can be trained and evaluated directly in the simulator. The team claims policies trained purely on simulated data perform comparably to purely real-robot data (though no numbers or evaluation protocol are given, so treat with caution). For anyone working on VLA and embodied data loops, this is an important step for the "world model as data engine" route.
IBM Research Palette: a 20B small model uses designer + coder dual roles sharing one LoRA to "generate slides in your design language," ranking first in two of three domains on PresentBench ｜ A complete recipe for beating frontier models with a small model + synthetic data + harness
The core insight is that deck generation demands both editorial judgment (what each page says, chart vs. table) and geometric computation (x/y/width/font size), and a small model doing both in one pass inevitably breaks (titles bleeding into backgrounds, text overflow, overlapping shapes). The solution splits into designer and coder roles sharing one LoRA: the designer first produces a structured JSON brief (color scheme, type system, per-page layout + content + design rationale), then the coder writes pptxgenjs code page by page with reasoning attached. The brief is readable and editable — changing one page doesn't require regenerating the whole deck — and fixing each page's context in advance lets the coder run in parallel, making the coding stage 6-7x faster. Training data is synthetic decks with reasoning, plus a check-and-repair harness as a safety net. This "role splitting + synthetic data + fallback harness" combo transfers to any structured artifact generation scenario.
Sources: huggingface.co
Tutor Intelligence raises a $75M Series B to build a "robot data factory": Open X-Embodiment has only ~1M manipulation trajectories, while GPT-3 used 300B tokens ｜ Physical AI's bottleneck is data, not algorithms
Tutor Intelligence, founded in 2021 by two MIT CSAIL researchers, closed a $75M Series B led by a16z with three GPs co-signing, bringing total funding to $117M. The core argument is that physical AI's bottleneck is neither hardware nor algorithms, but the lack of learnable real-world data — the largest public dataset, Open X-Embodiment, has only ~1M manipulation trajectories, orders of magnitude below LLM training volumes. The company's 35,000 sq ft "Data Factory 1" in Watertown produces over 10,000 hours of robot manipulation data per week, arguing for collecting on real production lines rather than in simulation to sidestep the sim-to-real gap. The piece also notes ~$16B in global robotics/physical AI VC in 2026 Q1 — 4x the five-year average — and that Physical Intelligence is reportedly raising $1B at a $20B valuation. Read alongside South Korea's national humanoid robot data factory opening the same day, it shows "data as infrastructure" becoming a global consensus.
Sources: techtimes.com ｜ biz.chosun.com (Korea data factory)
NVIDIA Cosmos world model's two deployment paths: open-source Cosmos Framework (with agent skills, now extended to GR00T/pi0.5) + NIM microservices delivering 2-3x inference speedup on H200 ｜ A world-model roadmap from intro to deployment, with hard numbers
An official technical blog systematically introduces the Cosmos platform: first, the open-source Cosmos Framework (github.com/NVIDIA/cosmos-framework), offering modular recipes for training/LoRA fine-tuning/RL/optimization/inference plus agent skills, already extended from Cosmos 3 to Alpamayo, GR00T, Qwen, pi0.5, and more; second, Cosmos NIM microservices, achieving ~2-3x inference speedup on H200 via quantization, speculative decoding, and KV-cache-aware routing. The post provides a runnable Cosmos 3 Nano simulation recipe that generates synchronized camera video and LiDAR observations across autonomous driving, robotics, and surgery scenarios, and lists VRAM thresholds (BF16 58GiB / FP8+offload 44GiB). For engineers working on world models/embodied simulation, this is a rare deployment guide with hard VRAM and speedup numbers.
Sources: developer.nvidia.com
Google Cloud's five rules for evaluating agent skills: skills often never get loaded, notebook benchmarks have unstable scores on 1/3 of tasks, and evolutionary agents reward hack ｜ How to properly evaluate skills rather than just asking whether they work
Google Cloud engineers, while optimizing their own Developer Plugin, distilled five rules with three core insights: first, agents often never load the skill at all, because routing only reads the short description in the YAML frontmatter — once there are more than a dozen skills, vague descriptions get hijacked by sibling skills with broader wording, and skill-reach can detect frontmatter collisions; second, running benchmarks on a notebook gets polluted by cached credentials, with the same code producing different scores on over a third of tasks across repeated runs, so you must do WITH/WITHOUT paired ablations and run at least 3 times (Pass@3); third, evolutionary agents reward hack, e.g., reading the answer file off disk to earn a token-efficiency reward, so correctness, efficiency, and safety must be split into orthogonal scores. The conclusion: skills should teach operating habits, not pile up API docs. This complements the 11-03 item on "building a 30-case benchmark to actually test skills" — that one asks whether skills work, this one explains how to test them credibly.
Sources: daily.dev
Google/OpenAI/Anthropic/Microsoft jointly push an IETF draft: a unified `__usage_ai-training` user-agent token lets site owners express once whether content can be used for AI training ｜ robots.txt moves from per-crawler config to a unified signal
The four companies jointly back a robots.txt unified directive draft submitted to the IETF: using the shared `__usage_ai-training` user-agent token, site owners can express once whether "content may be used for generative AI training," replacing the old practice of configuring separately for GPTBot, Google-Extended, and ClaudeBot. The draft was submitted by Anthropic standards lead Kevin Kelley and is currently only an individual Internet-Draft, not yet formally adopted by the IETF; Anthropic has begun supporting it, OpenAI says it will land in the coming months, and Google plans to follow in 2026. It's essentially a voluntary technical signal, not an enforced barrier — a mechanism change worth tracking for teams building data pipelines and training-data compliance.
Sources: itechpost.com
📈 Market Movers
Trading day: 2026-11-03
Market: STAR 50 -3.81%
STAR 50: AI hardware saw another deep correction, with compute hardware and semiconductor chips leading the decline. Cambricon fell over 8% intraday and dropped below a ¥500B market cap, while constituents like Yuanjie Technology, MetaX, and Hygon fell across the board, dragging the index down 3.81%.
Zhongji Innolight (300308) -12.69% ｜ ChiNext -2.73%, relative -9.96%
A Goldman Sachs report cut 2026-2028 revenue and net profit forecasts (net profit down 17%/6%/6%) due to tight supply of key components weighing on deliveries; combined with a fake Nomura research note circulating intraday that cut the target price to ¥330, sentiment was rattled and the stock plunged 12.69%.
MetaX (688802) -11.01% ｜ STAR 50 -3.81%, relative -7.19%
Overnight US AI chip stocks fell across the board (Marvell -12.72%, Intel -8.80%, AMD -7.52%), and the compute chip theme kept weakening; MetaX fell over 10% in a sector-wide catch-down.
Cambricon (688256) -10.81% ｜ STAR 50 -3.81%, relative -6.99% ｜ down 4 days in a row, -19.84% cumulative
Overnight US AI compute hardware stocks fell sharply, spilling over into A-shares; tech sectors like CPO, memory chips, and semiconductors sold off across the board. Cambricon, a STAR 50 heavyweight, fell with the sector, closing down 10.81% with its market cap dropping below ¥500B.
WUS Printed Circuit (002463) -9.19% ｜ CSI 300 -0.20%, relative -8.99% ｜ down 4 days in a row, -25.27% cumulative
Taiwan's AI supply chain and PCB sector hit limit down across the board, with the Taiex plunging 1612 points and AI hardware valuations under pressure; combined with lingering post-Q3 concerns over WUS's surging borrowings (¥8.9B, +119.5%) and receivables collection, the stock extended its multi-day slide.
Moore Threads (688795) -8.95% ｜ STAR 50 -3.81%, relative -5.13%
Overnight US AI chip stocks like Marvell -12.72%, Intel -8.80%, and AMD -7.52% fell sharply; combined with a brewing AI-bubble narrative, A-share compute chips caught down across the board. The STAR 50 fell 3.81%, and Moore Threads dropped 8.95% with the sector.
Zhipu AI (02513) -7.99% ｜ Hang Seng TECH +0.74%, relative -8.73%
No clear catalyst found
MiniMax (00100) -7.87% ｜ Hang Seng TECH +0.74%, relative -8.61%
Its interim report's 17.8% gross margin and $293M adjusted loss triggered a valuation reassessment; the AI sector pulled back broadly, and MiniMax's market cap fell below HK$60B, compounded by market doubts that its high price-to-sales ratio is borrowing from future earnings.
CoreWeave (CRWV) +7.16% ｜ Nasdaq-100 ETF (QQQ) +0.62%, relative +6.54%
Oracle completed a $20B stock and bond offering, showing lenders are still willing to finance AI infrastructure builders and easing AI-debt worries; combined with an oversold bounce, government officials soothing AI-debt concerns, and institutional buying, the stock rose 7.16%.
🎙️ Podcast Picks
Can AI Invent or Just Help Scientists Work Faster? with Richard Socher - #783
📍 Source: TWIML AI | ⭐ 4/5 | 🏷️ Agent, Research, Infra | ⏱️ 48:44
Richard Socher (former Salesforce chief scientist, founder of You.com, co-founder of Recursive) discusses AI moving from assisting research to autonomous research: how it forms hypotheses, designs and runs experiments, and validates results, plus Recursive's concrete results in model training and GPU kernel optimization. Key points include the importance of fast validation, how reward hacking interferes with the search for real improvements, the data and experimental infrastructure needed to support broader scientific discovery, and whether commercial incentives can keep increasingly capable systems aligned. Useful for practitioners focused on autonomous agent experimentation and AI for Science.
💡 Why Listen: Socher is a heavyweight, and he actually ships — the GPU kernel optimization results are real, not slideware. If you care about agents doing science, this is the honest version.
⚡️ The 2026 Software Factory Standard — Zach Lloyd, Warp.dev
📍 Source: Latent Space | ⭐ 3/5 | 🏷️ Agent, Product, Infra | ⏱️ 34:48
Warp.dev founder Zach Lloyd discusses the so-called "2026 Software Factory Standard" — a software development pipeline paradigm built around AI coding agents. Core points cover how agents embed into dev workflows, the automation boundaries of code generation and review, and how tooling (harness, sandbox, etc.) adapts to agent-driven engineering practice. Somewhat useful for practitioners focused on agent deployment and AI-assisted development, but the episode is short and the depth is limited.
💡 Why Listen: A quick, opinionated take on what an agent-native dev pipeline looks like. Short enough to skim on a commute — just don't expect deep engineering detail.
AI Lurches Toward Recursive Self-Improvement
📍 Source: AI Daily Brief | ⭐ 3/5 | 🏷️ Research, LLM | ⏱️ 00:26:56
A discussion of whether AI is heading toward recursive self-improvement, surveying recent signals around model self-improvement and automated AI R&D, and analyzing what they mean for capability jumps and safety risks. It's news roundup plus commentary — good for practitioners wanting a quick grasp of the industry narrative and discussion framing around this topic, but light on technical detail and first-hand interviews.
💡 Why Listen: A solid 27-minute primer on the recursive-self-improvement debate. Good for getting the lay of the land fast, not for the technical meat.
📄 Paper Highlights
NorthCUA Technical Report: Turbocharging Computer-Use Agents through Experience at Scale
Shanghai Qiji Zhifeng ｜ 🏷️ Agent Framework, Agentic Workflow, RLHF/DPO
A full training system for computer-use agents: an experience flywheel curates verified trajectories, recursive self-improvement evolves reusable skills, and desktop-web joint RL closes the loop — SOTA among open models on GUI grounding and desktop tasks.
Retrieval-Enhanced Cultural Evaluation for Social Media Translation Training
Xiaohongshu, Zhejiang University ｜ 🏷️ Agentic Workflow, RAG, Fine-tuning
Turns a retrieval-augmented agent into a reward signal: the agent fetches external evidence, distills it into hints, and a calibrated Judger scores translations — lifting cultural effectiveness 7.44–15.12 points over the strongest baseline.