📊 Today's Overview
OpenAI is disbanding its standalone robotics team, betting that general frontier models can drive robots directly — a direct contrast with Google DeepMind's specialized Gemini Robotics 2.5, which shipped this week alongside two eighth-gen TPUs. Meanwhile, Nvidia CEO Jensen Huang's resignation sent shockwaves through AI hardware stocks: the Semiconductor ETF (SOXX) fell 3.67%, and Nvidia suppliers like Victory Giant and Zhongji Innolight dropped over 8%. On the product side, OpenAI launched ChatGPT Wallet with Amex, Mastercard, and Visa, and llama.cpp hit v1.0.0 with a semver commitment.
🔥 Trend Insights
- General models vs. specialized robotics: OpenAI bets frontier models can drive robots directly, while DeepMind doubles down on purpose-built VLA models like Gemini Robotics 2.5 — the embodiment路线 debate just got organizational evidence.
- Agent infrastructure matures fast: Perplexity open-sources agent safety models, DeepSeek opens full app capabilities to API, and Bedrock AgentCore V2 slashes session cold-start with microVM snapshots.
- Cost efficiency over compute arms race: A $500 custom inference engine hits 1740 tok/s, Inkling 2 Small trains natively in 4-bit, and EnvOpt cuts agent costs 48% by optimizing environments, not models.
🐦 X/Twitter Highlights
📈 Trends & Buzz
- Google ships two 8th-gen TPUs in one week, plus Gemini Robotics 2.5 - Sundar Pichai's weekly recap lists Monday as "two 8th-gen TPUs," Tuesday as Gemini Robotics 2.5 (a robot manipulation model), with smart glasses also opening pre-orders. At CoRL2026, DeepMind used Gemini Robotics 2.5 to drive the Sharpa Wave dexterous hand for autonomous manipulation @sundarpichai (Google CEO) | @SharpaRobotics (dexterous hand company)
- LangChain raises $300M at $3B valuation - The company says 7,000+ companies already use LangSmith for agent runtime, improvement, and governance @LangChain (open-source LLM app framework)
- Databricks strikes enterprise partnership with OpenAI - Three product lines: ChatGPT + Genie for business users, Codex + Unity Gateway for governed coding, and GPT models + Databricks Apps to transform business workflows @databricks (data lakehouse company)
- WSJ: Jensen Huang called CFO Colette Kress "a deer in the headlights" - The report says at a tense late-summer meeting, Huang was unhappy with progress on using the company balance sheet to finance AI expansion, calling her just "a CFO who does earnings" who couldn't help him set strategy @anissagardizy8 (The Information reporter, collaborating with WSJ)
- Sakana AI to start in-house LLM pretraining in December - Chief scientist Takuya Akiba announced Sawara Small (working name), following Sakana Namazu's post-training and Sakana Fugu's orchestration — the lab's first move into the model layer @SakanaAILabs (Japanese AI lab)
🔧 Tools & Products
- Perplexity open-sources two Decision Models for agent safety - Apache licensed, free on OpenRouter until Nov 19. In internal trials, it read 1 billion tokens across 7,000 agent sessions in a day, directly driving 14 agent safety rule fixes @AravSrinivas (Perplexity CEO) | @alexatallah
- DeepSeek opens full app capabilities to API - Search, files, links, voice, and safety all callable, plus standalone OCR, ASR, and TTS endpoints, and an open-source frontend demo supporting conversation history import @deepseek_ai
- NVIDIA releases open robot model GR00T 2 - Positioned as System One (low-level action model), it learns skills from human demonstrations and predicts how actions change the physical world; skills transfer across robot embodiments without retraining each one @nvidia
- llama.cpp ships v1.0.0 - The local inference library's first fully stable major release, covering phones, laptops, and workstations; Clement Delangue says the goal is letting users "own their own intelligence" for free, with data staying local @ClementDelangue (Hugging Face CEO) | @ggerganov (llama.cpp author)
- DeepMind Genie 4 opens research preview - Generates audio-visual worlds in real time, responding to gaze movement, position changes, and text/voice input; a podcast host used it to recreate the hardest climb at mile 23 of the NYC Marathon @jacksondahl (Dialectic podcast host)
⚙️ Engineering Practice
- 6 prompts, $500 to write a DeepSeek V4.1 Flash inference engine from scratch - Rust + CUDA, no PyTorch, no SGLang, no DSL; 1740 tok/s decode on 4×B300, 2060 tok/s on 8×B300, ~12 seconds for 1M-token prefill. The author says the real value of a custom engine is fast iteration — then sent all changes back to SGLang and matched its performance @waterloo_intern (community developer) | @saltyph
- Inkling 2 Small trains natively in 4-bit throughout - The 594B-parameter model uses 4-bit forward passes from the first pretraining step to final RL rollout; NVIDIA says pretraining completed in just 68 hours on GB300 NVL72 with NVFP4 @thinkymachines (AI research company) | @NVIDIAAI
- Open-source 27B Qwen robot model: pure simulation, zero demonstrations - Trained only on its own attempts in simulation, runs 220+ tok/s locally on an RTX 5090; succeeded 18 of 30 tests on an unseen real robot (0 before training), near Astra max on LIBERO, with no degradation in math or coding. Weights, code, and training infra all open-sourced @jmtang42 (open-source robot model researcher)
- EnvOpt: optimize the agent's environment, not the model - The same Slack, docs, and codebase can be served via remote MCP or exported as plain text for the agent to grep — that choice alone significantly changes task cost. Across five benchmarks, it cut target-agent cost on unseen test sets by 48% on average with equal or higher pass rates; on AutomationBench it beat Meta-Harness 9.67% at 15% of the cost @AlexGDimakis (UT Austin professor)
- Agent Initiative Benchmark: testing whether 12 frontier models "should act" - Tests proactive agents in a synthetic company with 6 years of history; GPT-6.1 Sol beats Fable 5.2 at $0.11 vs $1.87 per task; Fable 5.1 costs 430× more than GPT-6 Luna but performs similarly; models still act when help isn't needed, and these failure cases cost ~80% more time and money @jakubjurovych (benchmark author)
- Doing inference with trainable energy models from the '80s - Ben Hoover et al. claim training is 1000× faster than any looped model with 280× less compute, arguing energy is the right paradigm for reasoning @Ben_Hoov (AI researcher)
⭐ Featured Content
OpenAI disbands standalone robotics team; research folds into Mark Chen's main research group, applications and hardware into Greg Brockman's Applied — betting "general frontier models can directly drive robots" | Organizational-level evidence in the embodied AI路线 debate
Fortune exclusive: OpenAI believes it originally expected to take a Sora-like video model route, but general frontier models can now directly drive robots — in NVIDIA experiments, an agent based on GPT-6 Astra scored 74.8% on a simulated manipulation benchmark, while the best specialized robot model scored only 51.0%. Altman is still building a humanoid in-house, targeting a first public demo in 2027; a fitting on-site detail: booth 208 at CoRL sits empty, its logo removed. The piece captures three-way disagreement: CMU's Guanya Shi says robotics is "no longer a special problem" but on-device small models remain necessary; MIT's Russ Tedrake insists real-robot deployment data is the moat; Generalist co-founder Pete Florence says no model dominates yet. It forms a direct contrast with Google DeepMind's specialized VLA route via Gemini Robotics 2.5 this week — a key first-hand signal for judging "general models eat the specialized stack vs. specialized VLA keeps digging deeper."
Sources: fortune.com | mlq.ai
OpenAI partners with Amex/Mastercard/Visa to launch ChatGPT Wallet: agents can pay on your behalf but never see the real card number, with biometric authorization required before each purchase | A protocol prototype as agentic commerce moves from "recommendation" to "transaction"
Users link a card once; the payment network issues a payment token replacing real card info, preserving original card rewards and dispute protection; each purchase requires Face ID/fingerprint authorization with an audit trail. Initially only for US ChatGPT Pro users, rolling out via dots. The piece also gives merchant onboarding guidance (avoid agents being wrongly flagged by risk controls, help payment and risk systems recognize legitimate agent transactions) and previews future spending limits/categories/time windows for fully delegated agent purchasing. For anyone building agent products, this is the first official paradigm for "how agent payments work without exposing card numbers" — worth watching whether its token protocol becomes an industry standard.
Sources: openai.com
llama.cpp and ggml ship v1.0.0 together: the point isn't new features but a promise — using semantic versioning to explicitly signal API/ABI incompatibility | The local inference ecosystem enters a "dependable" era
The rules are clear: MAJOR only for breaking changes to public C headers, MINOR for new backends/model support/APIs, PATCH for pure bug fixes. It also publishes a full distribution matrix (Docker, Debian, Homebrew, conda-forge, MacPorts, Winget, XCFramework, prebuilt artifacts), scales CI with HF Buckets/Jobs, and builds custom runners with NVIDIA/AMD/Qualcomm/Arm/IBM/Intel to guarantee correctness on every merge. For teams embedding llama.cpp in products, this is a signal to realign upgrade strategy and packaging pipelines — you can now plan dependencies by semver instead of dreading every merge.
Sources: huggingface.co
Cursor fully discloses the 10/29 security incident attack chain for the first time: third-party Fireworks leaked an access key → coding agent traces copied from a read-only bucket → malicious Claude Code extension live for 11 hours | A rare first-hand post-mortem of agent toolchain supply-chain risk
The starting point was Fireworks leaking an access key Cursor had issued; attackers used it to copy 2025 coding agent traces from a read-only storage bucket (non-Privacy Mode sessions, <0.05% of users), where a small number of unmasked credentials (<0.01% of customer API keys exposed) became a springboard to internal SaaS tools, archived non-production source repos, customer service data, and some agent conversation summaries. The most notable link: attackers used stolen credentials to plant a malicious copy of the Anthropic Claude Code extension (2.1.312/2.1.313) in Cursor's extension marketplace cache — about 5,000 users requesting specific cached versions may have been affected, live for ~11 hours, with the Open VSX origin unpolluted. The post ends with unaffected scope and self-check/rotation steps, serving as a complete threat-model template for "third-party vendor key → data exfiltration → extension poisoning."
Sources: cursor.com
Bedrock AgentCore runtime V2 rebuilds session cold-start with microVM snapshot/restore, plus a selection table for four patterns | A copyable blueprint for agent deployment engineering
Instead of pulling an image and starting a container per session, it starts a container once on create/update, waits for /ping health, clears file caches, freezes the microVM and saves memory+disk as a snapshot; new sessions restore the snapshot directly, with the process resuming from the freeze point (imports/objects/clients all intact), and image files read lazily by block. Three key facts follow: memory comes back entirely, file read caches don't persist across sessions, and image files are fetched on first read. The selection table is practical: if 2 seconds is acceptable, switch to V2 and move initialization before app.run(); if you need large files/CLI/browser, add a build hook to warm up during snapshot construction; for sub-second and user interaction before the first request, pre-start sessions in the background. Note that under V2, each create/update takes a few minutes to build a snapshot.
Sources: aws.amazon.com
NestDaddy tests all 37,000 remote MCP servers: 89.6% respond to initialize, 1/3 demand a key outright, median 7 tools per server | The MCP ecosystem's first full distribution dataset
Every day it connects an MCP client to all remote servers in the official registry; the first round covered 62,377 registered entries: 12.6% were hidden as bulk spam, 37,383 were actually connected, 89.6% responded to initialize, of which 1/3 returned 401/403 demanding a key before listing tools, and 3,555 didn't respond at all (404/405 mostly from wrong paths in the registry). Median response is ~500ms, suggesting the protocol itself isn't the bottleneck. It found 292,447 tools total, with 4,325 servers exposing 20+ tools — and every tool description goes into context, so stacking 5 servers with 30 tools each means stuffing 150 descriptions into every conversation turn, burning tokens and lowering the chance of picking the right tool. The most practical takeaway for MCP server authors: keep initialize and tools/list open to everyone, and only require a key at tools/call.
Sources: nestdaddy.com
Dwarkesh publicly revises his own view: from "at least another decade" to "AI matching top human AI researchers in two to three years, maybe months" | Calibrating the RSI timeline with three first-hand uplift data points
The core argument isn't idle speculation: METR spent three months embedded at Anthropic and found overall R&D speedup of 2x, mostly attributable to using AI for R&D; OpenAI's internal heavy users' daily coding agent spend rose from $1,000 to $7,000; Anthropic self-reports Claude already "leads" 26% of AI R&D work (under 1% in February). He treats math's explosion as a preview — from four-digit addition to IMO gold to solving a Millennium Prize problem in 88 hours — and rebuts Toby Ord's "it can only prove, not ask questions": an AI that can't ask questions has already solved a Millennium Prize problem. For anyone wanting to calibrate AGI timelines and find citable uplift numbers, this is also framework-level material to read alongside yesterday's Yang–Mills proof event.
Sources: dwarkesh.com
Physical AI architecture survey: three generations of VLA evolution → world models as simulators → convergence into World Action Models, centered on the symmetric tradeoff of "data availability vs. inference cost" | Build a complete VLA/WAM technical map in one read
It systematically traces the shift from "implicitly learning physics" to "explicitly predicting physics": three generations of VLA (action-as-token → separate flow-matching action expert → slow-reasoning/fast-control dual system), world models as simulators and training grounds, and the World Action Model that merges the two (marked by NVIDIA's "World Action Models are Zero-shot Policies," with five design choices including DreamZero). Core insight: VLA inference is cheap but needs scarce real-robot data with per-frame actions; WAM can pretrain on unlabeled video but inference is several times more expensive — and 2026 models (Cosmos 3's mixture-of-transformers, VLA-JEPA's latent future prediction) are narrowing the gap. Read alongside this week's Cosmos 3, PAN 2, and Runway Praxis-1 routes to assemble a complete selection coordinate system.
Sources: gargighosh1.substack.com
📈 Market Movers
Trading day: 2026-11-13
Market: Semiconductor ETF (SOXX) -3.67%, ChiNext -3.32%, Hang Seng TECH -3.02%
Semiconductor ETF (SOXX): US Treasury yields jumped, investors repriced high-valuation growth stocks, and the semiconductor ETF bore the brunt as its valuation depends heavily on long-term growth, with SOXX down 3.67%; put options were 78% of volume that day, showing clear selling pressure.
ChiNext: Overnight, Nvidia CEO Jensen Huang resigned and faced an antitrust probe; the stock closed down 4.34%, wiping out $182 billion in market cap. A-share compute hardware export chains (CPO, PCB, copper-clad laminates) were repriced sharply, with Shengyi Technology and Wus Printed Circuit hitting limit down; ChiNext opened higher then fell 3.32%.
Victory Giant (300476) -10.57% | ChiNext -3.32%, relative -7.24% | down 2 days in a row, -15.57% cumulative
Nvidia announced Jensen Huang's resignation from CEO and other roles; markets worry about strategic continuity under successor management. PCB and other compute hardware stocks deeply tied to Nvidia came under collective pressure, and Victory Giant, a PCB supplier for Nvidia AI chips, fell sharply.
Zhipu AI (02513) -10.04% | Hang Seng TECH -3.02%, relative -7.02% | down 2 days in a row, -18.19% cumulative
Hong Kong AI large-model, semiconductor, and PCB concept stocks widened losses in the afternoon; Zhipu fell over 11% with the sector, while MiniMax and Kingboard Laminates also dropped sharply. This followed a 9% plunge the prior day on limited buyback support and a RMB 2.07 billion half-year loss, extending the decline.
Wus Printed Circuit (002463) limit down -10.00% | CSI 300 -0.68%, relative -9.31%
No clear catalyst found
Nebius (NBIS) -9.92% | Nasdaq-100 ETF (QQQ) -1.20%, relative -8.72% | down 2 days in a row, -25.09% cumulative
The company cut Q4 guidance to the low end of its range due to data center capacity delivery delays and raised capex; combined with a wave of analyst target-price cuts and risk-off in AI infrastructure from Nvidia's management change, the stock extended the prior day's decline with another 9.9% drop.
AMD (AMD) -8.75% | Semiconductor ETF (SOXX) -3.67%, relative -5.08%
Leadership turmoil from Nvidia CEO Jensen Huang's resignation continued to reverberate; AMD and other chip stocks that had rallied on money rotating from Nvidia to competitors saw profit-taking and gave back gains. Combined with wholesale inflation beating expectations and pushing Treasury yields to 5%, tech stocks came under pressure.
Zhongji Innolight (300308) -8.39% | ChiNext -3.32%, relative -5.07% | down 2 days in a row, -13.99% cumulative
The personnel turmoil over Nvidia's board seeking to remove Jensen Huang escalated; markets worry about order damage to optical module suppliers deeply tied to Nvidia. Zhongji Innolight, a core supplier of its 1.6T silicon photonics modules, plunged with the AI hardware sector, compounded by RMB 3.015 billion in net main-fund outflows.
MiniMax (00100) -8.19% | Hang Seng TECH -3.02%, relative -5.17% | down 2 days in a row, -15.38% cumulative
Jensen Huang resigned as Nvidia CEO; markets worry the US will tighten chip and compute export controls on China. Chinese AI stocks fell collectively, Zhipu dropped nearly 20% over two days, MiniMax fell about 8% with the sector, and Hang Seng TECH fell 3.02%.
Envicool (002837) -6.39% | CSI 300 -0.68%, relative -5.71% | down 2 days in a row, -11.45% cumulative
Jensen Huang's departure as Nvidia CEO triggered a sharp drop in risk appetite across the global AI compute chain; the median decline among 24 A-share Nvidia supply chain companies was 8.49%, and Envicool, a liquid cooling supply chain name, plunged with the sector.
🎙️ Podcast Picks
The avalanche that hit math is coming for AI development
📍 Source: Dwarkesh | ⭐ 4/5 | 🏷️ Research, Agent, LLM | ⏱️ 44:26
Dwarkesh Patel revisits how his view on superintelligence timelines shifted over the past three months — from believing it needs at least a decade to a more aggressive expectation. The core discussion centers on recursive self-improvement (RSI), drawing an analogy to the "avalanche" already hitting mathematics, and reasoning through the acceleration curve that could emerge once AI R&D itself is automated. It's a useful macro reference for anyone tracking LLM capability boundaries and long-term agent evolution.
💡 Why Listen: Dwarkesh is publicly changing his mind on timelines, and he shows his work. If you want to calibrate your own AGI expectations, this is a tight 44 minutes.
Implementing Federated Learning to enhance Agentic AI
📍 Source: Practical AI | ⭐ 4/5 | 🏷️ Agent, LLM, Research | ⏱️ 43:33
A discussion of how federated learning can train LLMs on private data without leaking privacy, and how model trajectories can improve agentic AI systems. The guest is from Flower Labs and is also a Cambridge PhD; topics include the advantages of open-weight models and real applications of locally deployed AI agents in healthcare, robotics, and everyday automation. Practical reference for anyone focused on agent deployment and privacy-compliant training.
💡 Why Listen: If you're building agents that touch sensitive data, this covers the federated learning angle most AI podcasts skip. Solid, hands-on perspective.
Washington Is Taking Control of AI
📍 Source: AI Daily Brief | ⭐ 3/5 | 🏷️ Regulation, Infra, Funding | ⏱️ 00:26:32
NLW discusses the Washington policy shift on China AI after Jensen Huang's departure from Nvidia: hawks gaining strength, chip export restrictions accelerating, and the debate over whether to sell chips to China. The episode also analyzes how Nvidia's new leadership might affect the financing chain underpinning the AI boom. For AI professionals, this is important context for understanding macro constraints like compute supply, export controls, and capital flows — though it lacks hands-on LLM/agent technical content.
💡 Why Listen: Quick 26-minute briefing on the policy ripple effects of the Nvidia shakeup. Good background if you're tracking export controls and AI capex.
📄 Paper Highlights
Together Compiler: Continual Compilation for AI Inference
Together AI | 🏷️ Agentic Workflow, Inference, Agent Deployment
The first truly continual agent loop for inference optimization, already serving trillions of tokens in production — beats hand-written Megakernels by 1.7x and top open-source engines by up to 2.6x TPS.
CADET: Construction of PCB Routing Instances with Executable Demonstrations
LG AI Research | 🏷️ Agentic Workflow, Tool Use, Code Agent
Generates 100,000 synthetic PCB boards with zero design-rule violations plus engine-action demonstrations, lifting complete DRV-free routing from 38% to 67.2% on real boards — a data-synthesis blueprint for EDA.
Curiosity-0826-4B: Generative Medical Event Sequence Modeling Across Pediatric and Adult Populations
Epic Systems, Stanford | 🏷️ Training, Scaling, Application
Epic's upgraded medical event foundation model (4.3B params, 16K context) trained on real EHR data, extending zero-shot clinical prediction to pediatric cohorts — lab-value MAE down 18-46% over its predecessor.