🔥 Focus
Jensen Huang immediately resigns as NVIDIA CEO; CFO Colette Kress named interim chief : The board announced that co-founder Jensen Huang has resigned as president, CEO, and director, effective immediately. CFO Colette Kress will serve as interim CEO while continuing to oversee finance. Harvey C. Jones, who has served since 1993, becomes the company’s first-ever board chair. The episode escalated from China-related intelligence and extramarital-relationship investigations; an independent probe is still underway. Officials said they will search internally and externally for a permanent successor. The November 17 earnings schedule is unchanged. Huang long had about 60 direct reports and no clear No. 2; markets worry that customer alliances, the open-source path, and the “one-person calls plus circular financing” narrative will be interrupted. He has agreed to closed-door congressional testimony before Thanksgiving. Shares fell about 7%–13% at the open; market cap has dropped sharply since the scandal broke and fell below $4 trillion, while AMD and other peers relatively strengthened. Officials said the product roadmap “will not be interrupted.” (Source: NVIDIA Newsroom, TechCrunch, The Guardian, WSJ, 机器之心)
DeepMind releases AlphaGenome 2: predicting variant effects by cell type : The next-generation DNA sequence model is now a conditional model that jointly takes tissue/cell state over ~1 million bases, covering a map of more than 107 million cells and 68 tissues, plus methylation and cross-modal imputation, targeting regulatory differences of the same genome across cells. On TenK10K, causal-variant recall is about 30% and direction prediction about 43%, nearly double the previous generation, and it can adapt to small amounts of new single-cell data without retraining. An academic/non-commercial research API is open; clinical diagnostics are not approved. Later plans include connecting to Antigravity agent workflows. (Source: Google DeepMind Blog, THE DECODER)
Gemini 4 Hydrogen fully rolls into Search and launches as the high-speed tier : Google called this Search’s largest model upgrade to date, live globally: stronger multi-span intent understanding, query fan-out, and link discovery, plus structured visual answers such as tables and comparison charts, emphasizing routing users to a wider set of sites rather than a chat-box closed loop. Hydrogen is also positioned as the everyday and coding high-speed tier, claiming tasks complete about 2.5× faster than 3.8 Flash, beating Argon on the Artificial Analysis Intelligence Index, with four thinking-depth settings. It has been pushed to AI Studio, the Gemini API, the Gemini App, Search, and paid Pro/Ultra (including Antigravity). (Source: Google AI Blog, GoogleDeepMind)
Australia plans to mandate independent safety audits for frontier models : The Albanese government announced mandatory safety rules for the highest-risk frontier models, proposing a regulator working with the AI Safety Institute to enforce them, requiring open independent testing, ongoing safety checks, publication of safety frameworks, and immediate reporting of safety incidents. The government said it will set domestic guardrails first, then join international talks; details will follow consultation. (Source: THE DECODER)
🎯 Moves
Grok Bot 1.0 shifts to team-shared colleague-style agents : xAI launched Team Bots that can be pulled into Slack, Teams, and video meetings, emphasizing proactive project watching, learning across people, proposing tasks from deadlines, and asking first on sensitive actions. Each Bot still has a cloud computer to operate without APIs; permissions follow member access control. Team/Enterprise plans are open; desktop ships first on macOS. Personal Ultra does not include full team Bots. Musk acknowledged costs are too high relative to revenue. (Source: 机器之心)
ERNIE 5.5 Flash Preview focuses on end-to-end coding, planned Apache 2.0 open source : Baidu previewed a coding-oriented model with ~280B total parameters, 14B activated, million-token context, and Qianfan inference over 200 tokens/s, trained in a closed loop of 14 million executable tasks and Agent trajectories for debugging. Demos include code-only animation, an ~11-hour Unreal project, and Blender modeling. Weights are planned for Apache 2.0 on November 25 with SGLang/vLLM support; time-limited API discounts. (Source: 量子位)
Standalone Gemini app lands on Apple CarPlay : The CarPlay home screen can use Gemini directly for calendar and trip questions. Voice via Google Maps was already possible; this puts the chatbot on the system-level in-car display. (Source: The Verge)
OpenAI: chain-of-thought monitorability is being eroded by capability–cost trade-offs : Analysis of GPT-5 through GPT-6 Astra training cycles shows stronger pretraining yields more “no-scratch silent reasoning”; RL increases controllable rewriting of CoT, which can also be inherited via teacher-data distillation. The study says this is not an irreversible law of nature but a result of architecture and deployment-compute choices, calling for monitoring beyond reading drafts. (Source: THE DECODER, OpenAI)
Google Research uses wearable signals for an intervenable “health age” coach : The framework estimates coachable health age from changeable behaviors such as steps, sleep, and resting heart rate, and uses SensorWM on about 2.28 million weeks of data to simulate 12-week counterfactual habit trajectories, combined with blood markers for cardio-kidney-metabolic risk. In offline coach/physician blinded reviews, counterfactual plans were preferred; it is a research prototype, not a diagnostic tool. (Source: Google Research Blog)
Zhang Yiming internally: capable of matching OpenAI/Anthropic, catch-up earliest 2029–2030 : Reports say ByteDance plans a Muse-like personal Agent, while admitting pretraining corpora are orders of magnitude smaller than top labs and distillation is banned. Video generation and world models are treated as offsetting strengths; a safety-alignment team has been spun out. (Source: 36氪)
Tsinghua–Berkeley open-source uni-robot-agent: unified RUA tool interface : Standardized perception/motion/skill layers connect different embodiments and simulators. Qwen3.8-Flash-Next scores about 31 on RoboDojo and 53.5 on 19 real-robot tasks. The paper stresses no privileged sim information; failures mainly come from calibration, geometry, and embodied reasoning. (Source: 36氪)
Next-gen AgentCore Runtime: faster cold starts, scale-to-idle : AWS said sessions remain Firecracker-isolated, startup can be under 2 seconds, specs up to 8 vCPU/32GB, idle memory reclaimed, billed by usage. Coinbase and others said peak cold starts fell from nearly a minute to a few seconds. (Source: AWS Machine Learning Blog)
SGLang+TorchTPU runs 816B MoE on a single TPU v7 : Custom megakernel and DFlash2 yield ~3× small-batch decode and ~+18% at 500k context decode; fp8 KV doubles per-chip context. PyTorch cited 433 tok/s per user on AgentX and up to ~25.9K tok/s per chip. (Source: SemiAnalysis, PyTorch)
vLLM upgrades EPD encoder decoupling : The new version integrates with Router, SMG, and llm-d for production multimodal encode/generate split, lowering TTFT and counting per-GPU throughput after all encoder GPUs. (Source: woosuk_k)
Pony Ma at anniversary meeting admits Tencent’s AI layout was late : Using Google Search being hit by AI as a metaphor, he said “the wave has already crashed in.” Commentators see a more realistic path as accelerating a DeepSeek tie-up then product distribution. (Source: teortaxesTex)
🧰 Tools
TensorFold: Zig-written Python-free local LLM inference engine : Native binaries run checkpoints directly on Apple Silicon Metal and NVIDIA CUDA, compatible with OpenAI/Anthropic-style APIs, and experiment with Living Weights that write facts into weights. Companion Starlight handles enterprise identity and multi-machine orchestration. (Source: GitHub)
AWS Emulator: agent-generated local API simulation of 200+ services : Official open-source Rust emulator covers 438 API models; CLI/SDK can hit local endpoints with no account. At launch about 56.6% of fidelity tests passed; it explicitly does not do IAM evaluation or load testing. Code was written by a large-scale Agent pipeline. (Source: MarkTechPost, Simon Willison)
AgentCore launches standalone Web Search API : One SigV4 call searches Amazon’s own web index, returning summaries, URLs, and timestamps, with domain and date filters, no Gateway required first. Pricing is about $7 per thousand queries. (Source: AWS Machine Learning Blog)
Perplexity publishes SPACE sandbox red team : 12 models were asked to escape a VM, steal credentials, or bypass network policy. Findings include that standard CTF prompts bias boundary tests, and some general models find bugs better than security-specialized ones. A DNS-spoofing bypass was found and fixed; the platform is open source. (Source: AravSrinivas)
OpenRouter Extractions API (alpha) : Describe fields in natural language and extract structured data from text, images, and PDFs, using the same routing keys. (Source: alexatallah)
LlamaIndex releases ParseBench-2 and OCRia-1.5 : 4.3k pages and 330+ error types scored by severity, plus a new forms dimension. OCRia-1.5 powers all LlamaParse modes. Dataset, eval code, and Metric Lab are public. (Source: jerryjliu0)
Keenable 1.0 tops agent search leaderboards : Homegrown retrieval stack, 100 million knowledge cards and dynamic snippets claimed to cut research-task tokens by 50%; first on Artificial Analysis Search Index and NEEDLE. (Source: omarsar0)
Expo Simulators: cloud simulators let coding agents run first, then humans review : Agents execute changes in a cloud iOS simulator and send back video/screenshots; colleagues tap via a link. (Source: omarsar0)
📚 Learning
LSTM co-inventor leads 84-page survey of large-model architectures : Schmidhuber et al. review efficient attention, linear models, MoE, MTP, and recurrent depth, summarize four combinations in open-source flagships, and argue the next generation will more adaptively allocate compute and state over tokens/layers/time. Horizon, recurrent depth, and state size still lack MoE-class scaling laws. (Source: 机器之心)
FireFlash: 64 B300s, turn open-source Flash into a long-horizon agent in 5 days : Full-parameter RL from GLM-5.3-Flash yields a ~300B-class terminal/coding agent; Terminal-Bench 4.0 Pass@1 about 46.2%; per-task cost claimed 30×+ lower than high-reasoning Fable 5. Keys include async multi-GPU, 15%–85% pass-rate task filtering, and TIS correction for stale FP8 samples. (Source: 机器之心)
LOGIV uses task graphs for local repair of long-horizon robots : CUHK-Shenzhen and others compile a dependency graph on top of VLA/WAM; on failure, causal local repair instead of full replan. Success on sim perturbation tasks rose by up to about 28 percentage points. (Source: 机器之心)
EM²Mem: multimodal long-term memory that aligns events before retrieval : Zhejiang University and Lenovo attach 30-second event dossiers to people, objects, and multi-scale context; long-video QA is more accurate, single QA latency down about 4.7×. (Source: 机器之心)
TestPrism: single-reference eval overestimates test quality by about 2× : An NJU team built correct/incorrect implementation panels for 300 tasks; strongest configs reach RAR ~60% but JSF only 28%. TestHelix with cross-validation lifts JSF another ~12 points. (Source: WeChat)
Princeton + Anthropic: memory traps—“no need to confirm” can generalize to payments : In simulation, low-risk authorization notes caused later high-risk actions to skip verification; Sonnet/Opus unsafe rates could reach 35%–81%. Authors argue memory must carry scope of applicability. (Source: 36氪)
Lev: deepen decisions with loop count rather than generated tokens : The same layers read out options each loop. Lev-2.6B on public JevBench items goes from 63.2% at 1 loop to 76.2% at 4 loops, using about 15%–24% less compute with on-demand depth. (Source: WeChat)
OWE-Bench: 26 long-horizon real engineering projects for agents : Mercor experts wrote tasks from GPU kernels to aircraft design; Fable 5.2 leads at 54.6%. A single item can burn over 200 million tokens. (Source: _jasonwei)
SoMem: agents self-organize memory with chained notes : No preset schema; each summary chains back to original interactions. Slightly beats strongest baselines on LoCoMo and others. (Source: omarsar0)
P²-MOPD: end-to-end trainable linear mix of student and multi-teacher parameters : StepFun says it beats standard multi-teacher distillation finals in half the training steps (arXiv:2611.03953). (Source: teortaxesTex)
💼 Business
UK AI chip startup Fractile closes $720M Series C : Investors include the UK sovereign AI fund; reported valuation about $7B. The company bets on DRAM stacking to bypass bandwidth bottlenecks and a custom compiler to speed LLM inference, planning first chips and systems next year. (Source: TechCrunch)
Anthropic confirms public trading on November 19 : Official pages say the final offer price is set the day before; prior guidance was $70–$78 per share. S-1 revisions mention RSI-related safety thresholds; company language leans more on future-model expectations. The roadshow continues. (Source: The Verge, AnthropicAI)
CoreWeave Q3 revenue $3.66B, guidance raised : +168% YoY; it said effective demand is about 10× end-2027 capacity, and older-gen GPU rents are still rising. High leverage and data-center policy headwinds remain, but results temporarily answer the trade that “compute demand has gone cold.” (Source: 36氪)
🌟 Community
“System One Day”: decision models expand from Jev to Google and engineering practice : At the AI Engineer conference, Almeida urged treating non-generative scoring/selection models as a fourth logic gate in code. Glean, UiPath, and Workday showed routing, document tournaments, and input-as-form-fill. The community is buzzing that a hybrid of “fast decisions + occasional large reasoning” is taking shape. (Source: Latent Space)
United Airlines bans AI agents from booking for customers : It blocked third-party agent orders citing inability to complete security declarations, among other reasons. Discussion sees platforms reclaiming the entry point and exposing identity and liability gaps in regulated transactions. (Source: CBS / HN)
VulnCheck questions Anthropic glass-wing: nearly 40k candidates, few public fixes : About 38k candidates in seven months, a couple hundred publicly fixed; Anthropic’s line is about 6,009 expert-confirmed and reported. Projects like curl were flooded; the community debates whether AI bug-hunting offloads review debt onto OSS maintainers. (Source: 量子位)
Uber/DoorDash/Airbnb: almost no one uses agents to order : In Halloween week Uber had about 320 million human orders vs. 4 agent orders. Users treat ordering as fun, plus error-correction time; consumer agents still look like “efficiency for efficiency’s sake.” (Source: TechRadar)
OSS is being flooded by agent PRs : Maintainers say valuable PRs almost have to be asked for offline; public contribution channels are clogged by auto-PRs. How newcomers earn reputation is becoming an open-source governance crisis. (Source: dbreunig)
Persistent agents’ share is rising; discussion shifts to Harness : Dylan Patel said SemiAnalysis mainly uses multi-model Perplexity Computer; community focus is moving from “which base is strongest” to whose Harness can carry long sessions. (Source: dylan522p)
💡 Other
DeepMind UK staff apply to government for a mandatory union ballot : CWU and Unite filed with the Central Arbitration Committee covering about a thousand product roles. Staff said voluntary talks were treated as administrative theater; demands include collective voice on military contracts. (Source: WIRED)
Fireworks breach hits Notion AI chat records : The inference vendor’s development systems were compromised. Notion said about 1% of customer Insights chat transcripts were affected and has cut the connection, highlighting supply-chain risk between zero-retention promises and actual storage. (Source: TechCrunch)
Yahoo becomes first formal licensee of collective AI crawl licensing : The former RSL standard evolved into the Internet Licensing Collective; Yahoo, a founding party, is first to pay for crawl rights in a music-collective-style model. (Source: The Verge)