📊 Today's Overview
Anthropic's IPO roadshow is the story of the day: pricing at $70–78 a share for a valuation up to $1.4 trillion, with Google and NVIDIA subscribing $26 billion at the offer price. Annualized revenue hit $810 billion — wait, $81 billion — and the filing lands against a 25% pullback in the Philadelphia Semiconductor Index. Meanwhile NVIDIA open-sourced Cosmos 3, an omnimodal world model, and Runway released Praxis-1, a 27B open-weight robot world model. CoreWeave's blowout Q3 sent AI infrastructure stocks soaring.
🔥 Trend Insights
- World models split into two camps: Runway's Praxis-1, NVIDIA's Cosmos 3, and MBZUAI's PAN 2 all landed this week — one camp bets on unified omnimodal backbones, the other on video-pretrained models with small action heads.
- AI eats its own CUDA moat: Anthropic's roadshow claims Claude ported to AMD in three days and wrote nearly 100% of the kernels, compressing a 6–12 month job into a weekend.
- Agent architecture gets a reality check: Amazon's controlled study finds no universal best design — multi-agent setups burn 1.4–4.7x tokens for gains that rarely clear significance.
🐦 X/Twitter Highlights
📈 Trends & Buzz
- Google launches Gemini Decisions API and Decision 1 model — No token-by-token generation. One call scores up to 1,000 questions: yes/no, multiple choice up to 255 options, and rubric grading. Inputs cover text, images, audio, video, and PDF. Priced at $0.10 per million input tokens, output free — Google says roughly 5x cheaper than Flash-Lite. Internal teams already use it for Stitch Rapid product flow design @sundarpichai (Google CEO)
- Axios exclusive: Trump authorized strikes on companies transshipping Nvidia chips to China, Commerce Secretary blocked it — A government official repeatedly stopped enforcement: Commerce Secretary Howard Lutnick @axios (news outlet)
- FT exclusive: Eric Trump-backed humanoid robot company in talks with IDF — Plans to supply humanoid robots to patrol Hamas tunnels under Gaza, with possible combat involvement @FT (Financial Times)
- Google DeepMind releases AlphaGenome 2 — Builds on AlphaGenome 1 to predict how different cell types read the same DNA, mapping genetic variants to cell-level disease causes @sundarpichai (Google CEO)
- LlamaIndex pivots to document OCR and applied research — CEO says the company is now an applied research lab for document processing. It built three eval suites (ParseBench, ExtractBench, ExtractBench Pro), trained its own OCRia-1 model, and invests in data synthesis, training pipelines, and agent tuning @jerryjliu0 (LlamaIndex CEO)
🔧 Tools & Products
- Runway open-sources world-action model Praxis-1 — A 27B MoE diffusion Transformer predicting video and actions jointly, Apache 2.0 weights. Runway says it tops both the world-action model and open-source leaderboards on NVIDIA RoboLab-120, completing the same robot task at 51x lower cost and 15.6x faster than frontier LLM agents. Live on LeRobot and vLLM-Omni; vLLM-Omni supports offline inference plus OpenPI/Positronic serving, runs on a single H100 80GB. NVIDIA released an NVFP4 checkpoint and Jetson deployment guide @runwayml｜@AndyChenML (Runway team member)｜@vllm_project｜@NVIDIARobotics
- Speridlabs open-sources Iris-Z image model — Two sizes, 9B and 32B, unifying image generation and depth estimation in one backbone, with depth-conditioned generation too. Speridlabs says its depth estimation beats dedicated models like MoGe-3. Weights, code, and demo all Apache 2.0 @speridlabs (image model team)
- Plurai launches agent simulation testing platform, free to use — Rebuilds a digital twin of the agent and tool environment from user code, generates simulated users and tasks, and checks what the agent actually does rather than what it says. Every failure comes with a reproduction run and a verifiable fix, and it can replay production failures to stress-test new versions. Integrates with Claude Code, OpenAI Codex CLI, and GitHub Copilot @ilan_kadar (Plurai founder)
- X open-sources For You recommendation algorithm — Also launches an Under The Hood report so users can check whether their account or posts are label-limited, and whether government demands or legal requests restrict reach in specific countries @Safety (X safety account)
⚙️ Engineering Practice
- SGLang unveils two inference system optimizations — A megakernel lifts Muse Spark Open from 517 to 1,109 tokens/s on 8×GB300, cutting kernel launches per decode step from 1,450 to 97. The other is a unified memory arena for hybrid models, letting full-attention KV, sliding-window KV, and recurrent state share one pool — no partition parameters, zero cross-type fragmentation @lmsysorg｜unified memory (SGLang maintainers)
- LeapQuant quantizes linear attention recurrent state to 4/8 bits — A training-free method matching FP32 accuracy on Qwen, Kimi, and GLM. Linear attention kernels run 4.25x faster, inference throughput up to 2x higher, memory down up to 56% @conlesspan (LeapQuant author)
- Unify AI releases Continual-ARC benchmark — Three agents (OpenClaw, Hermes Agent, Prime Agent) all failed to retain useful information from experience and couldn't generalize to new tasks. Directly training a 9B model's weights achieved both memory and generalization @DanielLenton1 (Unify AI founder)
- DeepSeek V4.1 Flash runs open-source on 4 DGX Sparks — 90 tok/s single-stream chat, 139 tok/s for code, 368 tok/s at 16 concurrent users. Supports 1M context and 10k tok/s prefill; image input is live @Picomoorhead (Moor Insights & Strategy analyst)
- Paper: one-sided expert-parallel All-to-All for MoE inference over NVLink — Authors include Li Bo, Anton Korzh, Noam Duvdevani, and others. Reworks expert-parallel communication with one-sided semantics @DPZ (CS paper feed account)
- vLLM upgrades EPD (encode-prefill-decode) disaggregation — vLLM says it beats DeepInfer's approach on TTFT and per-GPU throughput (counting all encoding GPUs). Integrated with vLLM Router, SMG, and llm-d @vllm_project
⭐ Featured Content
Anthropic IPO roadshow details land: $70–78 pricing, up to $1.4 trillion valuation, $81 billion annualized revenue, Google/NVIDIA subscribing $26 billion in non-voting shares at the offer price ｜ The pricing and subscription structure for the first "pure AI company" listing
825 million shares on Nasdaq under ANT, raising up to $64.4 billion at the top end — a valuation clearly below the previously rumored $2 trillion. The amended filing discloses an annualized revenue run rate of about $81 billion at the end of October, up from $71 billion at the end of September. The macro backdrop stacks three pressures: doubts about AI valuations, the Philadelphia Semiconductor Index down 25% from its October 6 high, and a post-midterm selloff in AI stocks over regulatory fears. Meanwhile NVIDIA-backed Firmus canceled a nearly $30 billion IPO and OpenAI pushed its listing to 2027. For practitioners, this is the hardest set of first-hand numbers for judging whether the AI capital cycle has peaked.
Sources: fortune.com ｜ mundoamerica.com
Anthropic roadshow video discloses compute strategy in detail for the first time: Claude runs on Trainium / TPU / NVIDIA / AMD simultaneously, Mythos ported to AMD in three days and wrote nearly 100% of the kernels ｜ On-the-ground evidence that AI itself is eroding the CUDA moat
Tom Brown introduces the "compute multiplier" concept — physical compute times stackable efficiency multipliers — and explicitly points to recursive self-improvement: future Claudes discovering multipliers on their own. The hardest evidence is the AMD migration: accelerator adaptation that normally takes 6–12 months and dozens of engineers was compressed to three days with almost no human intervention (measured on MI355X prompt processing). If the numbers hold, buyers can freely switch suppliers on price-performance in a supply-constrained compute market, directly hitting NVIDIA's 70%+ hardware margins — but note this is a one-sided investor presentation with no third-party reproduction. Simon Willison used MacWhisper to produce an unofficial transcript, making this rare first-hand material.
Sources: the-decoder.com ｜ simonwillison.net
NVIDIA Cosmos 3 omnimodal world model: a single mixture-of-transformers handles and generates language/image/video/audio/action, open-source code + checkpoints + datasets ｜ Key first-hand material on whether a "unified omnimodal backbone" holds up
It folds a VLM, video generator, world simulator, and world-action model into one framework. The abstract claims SOTA across understanding and generation: the post-trained model was rated by Artificial Analysis as the best open-source Text-to-Image and Image-to-Video model, and by RoboArena as the best policy model. Open-sourced under OpenMDW-1 at the Linux Foundation. For anyone weighing backbone choices for world models or embodied AI, this is core evidence on whether a single backbone can absorb multimodal tasks.
Sources: hackernoon.com
MBZUAI IFM open-sources PAN 2: a 29B dense MMDiT world model, 704×1248 / 24fps / 20 seconds, Apache 2.0 weights plus training and inference code ｜ An open baseline for physical AI's "simulate first, execute later"
It self-reports a 34.0% full-plan completion rate across 100 multi-step tasks, leading the six models tested, but trails MiniMax-H3 on single-step actions and human preference — Physics-IQ system default 37.48 vs 42.91 — and about two-thirds of plans still drop steps. Reading it against Cosmos 3 is more valuable: one is a unified omnimodal big backbone, the other a dense diffusion route focused on video→action. Both failure modes (step-drop rates) show that long-horizon plan consistency remains the main weakness of world models.
Sources: marktechpost.com
Runway Praxis-1: turns a 27B video foundation model into an open-weight robot world model, predicting future frames while outputting actions, open-source #1 on RoboLab-120 ｜ A scalable pretraining path that swaps scarce real-robot data for everyday video
Architecturally, a small action expert reads the video model's internal state to generate actions. After fine-tuning on DROID, real Franka success rates match the previous SOTA world action model. Three weight sets (base/droid/yam) plus an NVFP4 quantized version are released. Read alongside GR00T 2, PAN 2, and Cosmos 3, the world model race this week has formed two clear technical routes — "human video pretraining + small action head" and "unified omnimodal backbone" — worth weighing by data availability and real-robot migration cost when choosing.
Sources: paperswithcode.co
Meta Intelligent Delivery: reframes "spend each unit of compute on the highest-marginal-return request" as two ML problems — offline high-value request identification +20%, roughly 10% more ranking performance end to end ｜ A systems-level optimization blueprint for ad ranking models approaching LLM scale
The core is a Resource Core Model — 6B sparse + 20M dense — that runs once before retrieval and is shared across the funnel, then fine-tuned into specialized models predicting outcome, compute cost, and latency impact. This replaces rule-based bucketing by surface/context. The post also covers multi-layer resource scheduling across multi-stage ranking, hundreds of models, and multiple hardware generations, plus agentic AI for continuously detecting and responding to system drift. For anyone running large-scale inference serving and recommendation systems, this is a rare full architecture disclosure of "compute marginal utility modeling" — the ideas transfer to any heterogeneous hardware inference cluster.
Sources: engineering.fb.com
AWS publishes a silent data corruption detection and recovery loop on SageMaker HyperPod: three checks + node cordon replacement + recovery from a checkpoint before the suspect window, with a full code repo ｜ When GPUs compute "plausible but wrong" numbers, ECC/Xid/DCGM all fail to catch it
The background numbers come from Meta Llama 3 405B's 54-day pretraining: of 419 unexpected interruptions, 6 were attributed to SDC. The scheme adds three checks inside the PyTorch DDP loop — non-finite gradient checks, ring replay audit (each GPU recomputes another node's micro-batch and compares loss and gradient fingerprints bit by bit), and weight fingerprint comparison. On a hit, it maps rank→K8s node name and cordons the node. A table lays out exactly what ECC / Xid / DCGM deep checks / NVLink checksum / checkpoint checksum each guard, helping build a mental model of "layered integrity checks." Validated with fault injection on 32 H100s.
Sources: aws.amazon.com
AWS uses anti-unification for semantic diffing of formal policies: filters syntactic noise like reordered conjuncts, surfacing only real disagreements where and becomes or, with SMT counterexamples ｜ A comparison tool for when LLM autoformalization samples multiple times with no ground truth
It declares algebraic laws and domain rules on an e-graph, computes the most specific common pattern of two policies, shows the shared structure once, and preserves each point of disagreement. An SMT solver then finds a concrete assignment that satisfies one and violates the other as a counterexample. The core scenario: when the same requirement is sampled multiple times into multiple SMT-LIB / Cedar / Dogwood / OpenShell policy / CEL versions, it points out which parts are stable and which drift with sampling, so reviewers prioritize the disagreements. The authors frame it as a formal version of semantic entropy — not just a score, but preserved structural information. A reusable methodology for teams working on agent safety policies and autoformalization.
Sources: aws.amazon.com
📈 Market Movers
Trading day: 2026-11-11
Broad market: Semiconductor ETF (SOXX) +6.01%, ChiNext +3.11%
Semiconductor ETF: AI cloud infrastructure company CoreWeave's Q3 earnings beat almost across the board, with revenue up 168%, reigniting the AI trade. Constituents Marvell, Micron, and Intel surged; SOXX rose 5%+ premarket and closed up 6%.
ChiNext: Microsoft and other international tech firms urged the US to stop advancing import restrictions on Chinese optical transceivers. The market read this as an easing of external pressure on optical modules, and compute hardware names in CPO, PCB, and optical communications rallied broadly, driving ChiNext up 3.11% on higher volume.
CoreWeave (CRWV) +24.71% ｜ Nasdaq-100 ETF (QQQ) +2.31%, relative +22.39%
Q3 revenue up 168% year over year with a smaller-than-expected loss; the company raised full-year revenue guidance, easing concerns about AI infrastructure debt and slowing growth. Combined with Nvidia's rebound lifting the AI sector, the stock rose 24.7%.
Eoptolink (300502) +16.98% ｜ ChiNext +3.11%, relative +13.86% ｜ up 3 days in a row, +24.93% cumulative
At midday the company said on the investor interaction platform that order delivery is normal, its Chengdu and Thailand plants are expanding at scale, and gross margins should stay stable going forward. Combined with Microsoft lobbying against FCC ban risk and money piling into CPO names, net institutional inflow exceeded 1.2 billion yuan.
Nebius (NBIS) +15.44% ｜ Nasdaq-100 ETF (QQQ) +2.31%, relative +13.13%
Peer CoreWeave reported better-than-expected Q3 results after Tuesday's close and raised its 2026 revenue guidance, easing concerns about AI infrastructure demand and lifting the neocloud sector. Nebius reports its own Q3 before the next open, and investors positioned ahead of it.
Zhongji Innolight (300308) +15.07% ｜ ChiNext +3.11%, relative +11.96% ｜ up 2 days in a row, +19.90% cumulative
Microsoft challenged proposed US import restrictions on optical modules, easing sanction fears and sparking a broad optical module rally. Combined with Anthropic's prospectus showing surging revenue and over $600 billion in compute commitments catalyzing the AI sector, Zhongji Innolight rose 15.07%.
Arm (ARM) +12.05% ｜ Semiconductor ETF (SOXX) +6.01%, relative +6.04%
Reason pending
Micron (MU) +11.89% ｜ Semiconductor ETF (SOXX) +6.01%, relative +5.88%
Reason pending
Envicool (002837) limit up +10.01% ｜ CSI 300 +0.41%, relative +9.60% ｜ up 2 days in a row, +15.77% cumulative
A Guosheng Securities report said high power density in AIDC scenarios makes liquid cooling systems a must-have, triggering a wave of limit-ups in liquid-cooled server names. Tongfei shares hit its 20cm limit up and Zhongji Innolight rose over 10%, with Envicool sealing its limit up alongside the sector, helped by a broad rebound in tech growth stocks.
Wus Printed Circuit (002463) limit up +10.00% ｜ CSI 300 +0.41%, relative +9.59% ｜ up 3 days in a row, +22.16% cumulative
PCB names rallied collectively in the afternoon, with leader Dongshan Precision hitting its limit up. Several PCB companies said AI orders are full and delivery is tight, and some announced capacity expansions for AI compute PCBs. Wus Printed Circuit sealed its limit up alongside the sector.
🎙️ Podcast Picks
Wall Street Has No Idea What Anthropic Is Worth
📍 Source: AI Daily Brief | ⭐ 3/5 | 🏷️ Funding, LLM, Product | ⏱️ 00:24:45
NLW breaks down the bull and bear cases ahead of Anthropic's IPO: surging revenue, massive compute commitments, and a new productivity suite aimed at Microsoft. He argues the valuation headlines are misleading and discusses what the IPO means for the whole AI trade. Also covers the Gemini 4 Carbon launch, Nvidia's China troubles, and OpenAI's revenue surge.
💡 Why Listen: Good for understanding how a top LLM lab actually makes money and what its compute bill looks like. Just don't expect deep technical discussion — this is news commentary, not engineering.
📄 Paper Highlights
Deep Personalization: Continuous Context Observation for Local Proactive Agents
NVIDIA, Stanford ｜ 🏷️ Agent Framework, Agent Memory, Agent Deployment
Introduces a closed loop where a desktop agent watches your activity, learns when to help from implicit signals, and runs entirely on-device — a practical blueprint for proactive agents that don't leak your data.
Agent Architecture Cost-Success Trade-offs: A Controlled Evaluation across Enterprise-Style SOPs
Amazon, University of Washington ｜ 🏷️ Agent Framework, Multi-Agent, Agentic Workflow
Holds model, tasks, and tools fixed and varies only the architecture across 13 enterprise SOPs. The punchline: multi-agent setups burn 1.4–4.7x the tokens for gains that rarely clear significance — implementation quality beats topology.
Comparative Analysis of Agent Evaluation Frameworks: Stability, Detection, and Discovery in Enterprise Analytics Agents
Amazon ｜ 🏷️ Agent Deployment, Agentic Workflow, Safety
Compares three evaluation paradigms on frozen execution traces. Trace-based judging hits F1=0.90 while text-only judging fires 5x more false alarms — and functional testing and security red-teaming found zero overlapping failures.