📊 Today's Overview
NVIDIA's board has launched an independent investigation into CEO Jensen Huang over an affair with a woman suspected of Chinese intelligence ties — Megaspeed, her company, has become one of NVIDIA's largest buyers at roughly $4.8B in equipment, and Congress wants Huang testifying by November 18. Meanwhile, Fal raised $1.3B at a $16B valuation to buy inference compute for generative media, and Traversal hit a $1B valuation automating software incident diagnosis. On the engineering side, Stanford's CS146S rewrote its entire curriculum after agents one-shotted last year's assignments.
🔥 Trend Insights
- Inference compute gets its own capital track: Fal's $1.3B round is earmarked almost entirely for GPUs, and the CEO admits summer capacity shortages forced customer throttling — media generation inference is now big enough to fund separately from general LLM serving.
- Agents shift from writing code to running it: Traversal's $1B valuation for automated root-cause diagnosis, plus CS146S teaching patch review instead of feature writing, both point at ops and judgment as the new bottleneck.
- Power and permits beat chips as the constraint: Democratic-led states now host 60.3% of US data center capacity, and David Sacks warns the pipeline could dry up before 2028 — siting and local consent are the real gate.
🐦 X/Twitter Highlights
📈 Trends & Buzz
- Cursor hits $1B ARR in 2025, enterprise business up 300% QoQ — swyx says Cursor got stronger after acquisition, and reminds everyone that Twitter vibes and SEC-audited financials are two different things @swyx (Latent Space host)
- Apple plans to ship an agentic Siri as early as next year — Mark Gurman says the goal is shifting from "responding to commands" to proactively managing tasks and doing things for users, taking on ChatGPT, Meta Muse, and Grok @markgurman (Bloomberg reporter)
- Anthropic spends $3.4B on gas turbines for its data centers — per FT, this is the largest known single power-equipment purchase by any AI lab @business (Bloomberg)
- Robot data industry check-in: monocular is dead, touch is the new frontier — Sam Padilla surveys the landscape ahead of 2027: there's basically no off-the-shelf stereo solution to buy, and tactile sensing is the next battleground @theSamPadilla (robot data practitioner)
🔧 Tools & Products
- NVIDIA open-sources Auto Ontology to give text-to-SQL agents a business ontology layer — before writing SQL, map business concepts to actual columns: terminology (which column holds revenue), SQL properties (reusable expressions like margin and rolling totals), and table relationships (including bridge tables). The ontology is auto-drafted from database metadata and human-reviewed, stored in Postgres with pgvector for embeddings. Supports 10 engines including Postgres, Snowflake, Databricks, and DuckDB, runs Nemotron by default and any OpenAI-compatible endpoint, Apache 2.0. Questions the data can't answer get a straight "can't answer" instead of made-up numbers @akshay_pachaar (AI educator / founder of DailyDoseOfDS)
- Specialist: an open-source full stack for robot manipulation, hardware to deployment — Qiayuan Liao and t_k_233 open-sourced gripper design, data pipeline, training, and deployment code, matching GeneralistAI's gripper approach. Originally meant as a "weekend project," but every design tradeoff ended up documented @qiayuanliao (UC Berkeley robotics researcher)
- Kling 4.0 demos product detail and character consistency with an energy drink ad — creator leedain_leader_in shows the workflow shifting from "editing" to "directing": the model handles product detail, character consistency, and natural dialogue @Kling_ai (Kuaishou video generation model)
⚙️ Engineering Practice
- Together AI tunes StepFun's serving stack for 3x+ throughput — working against TTFT and TPS targets, they adjusted MoE routing, scheduling, kernels, spec decode, and sparse attention one at a time, benchmarking before each next step, with agents helping search @togethercompute (model inference provider)
- Tencent and WeChat release WeLM-Image: a unified image generation model with 31.7B total / 5.4B active MoE params — positioned as unified image generation "toward visual intelligence," paper is out @SD_Tutorial (AI art content creator)
- Andrew Carr connects two papers: distillation carries hidden traits, RLVR only improves sampling without expanding capability — Subliminal Learning shows a teacher model generating data from number sequences alone can still pass on traits like "likes owls" or alignment bias, and filtering the data doesn't stop it. The other paper says RLVR models beat the base at k=1, but at large k the base has higher pass@k — RL makes models better at sampling existing answers, not at solving new problems @andrew_n_carr (AI researcher / blogger)
⭐ Featured Content
NVIDIA board launches independent investigation: CEO Jensen Huang admits affair with suspected Chinese intelligence figure Alice Huang; Congress demands testimony on November 18 | A governance black swan for the world's most important AI compute supplier
The board has formed an independent director committee to investigate. Huang has admitted the extramarital affair and pledged full cooperation. The board says there is currently no evidence of compromised company secrets or national security, and Huang will stay in his role. The trigger was a New York Times report that the woman's company, Megaspeed, has become one of NVIDIA's largest buyers (~$4.8B in equipment purchases). House Republicans want Huang to testify before Congress on November 18. Gary Marcus's angle is the most quotable: Trump apparently knew a year ago yet stayed close, Rubio's spring memo went nowhere, and NVIDIA said a week ago it knew nothing — tying "possible export-control intel leakage + no succession plan + futures already falling" into a causal chain that could hit the whole AI investment market. Another piece layers five NVIDIA storylines from 2026 by evidence strength (China share 66%→8%, antitrust cases open in three countries, Warren demanding public sworn testimony, over $1B in share sales in four months, circular financing controversy), and gives the full policy timeline from the 2022 export ban → H800 → H20 → April 2025 cutoff → August 15% cut → December H200 discussions. Good as a factual baseline.
After midterms, Democratic-led states host 60.3% of US data center capacity under construction (77.4GW), up from just 25% before the election | A quantified snapshot mapping "political map" onto "compute geography"
Based on Compute Atlas data: Democratic flips in Texas, Ohio, Georgia, Nevada, and Iowa were key. Operating/under-construction/planned projects total 1,918, or 66.4%. Sites cluster in Virginia (154), Ohio (149), California (132), and Texas (117). The piece also gives two counter-signals to avoid misreading: moderate Democrats (Shapiro, Pritzker) are actively courting investment, and Senator Warner called pausing AIDC construction "idiocy." A draft paper cited by WaPo says data centers contributed only 4% of residential electricity price increases from 2020–2025. Pair it with David Sacks's warning that the AI data center pipeline could shrink before 2028: in one month NVIDIA fell 21.4%, Broadcom 22%, Vertiv 25.81%, while Equinix, which already has land and power permits, fell only 1.85% and Digital Realty 3.47% — existing facilities actually benefit from cancellations. Together, this is a set of evidence that the supply bottleneck is shifting from chips to power, permits, and local consent.
Fal raises $1.3B Series F at a $16B valuation — valuation doubles in 7 months, annualized revenue goes from $400M in March to $800M | A dedicated generative-media inference stack becomes its own capital track, separate from general LLM inference
Bessemer led, with NVIDIA, Sound Ventures, and T. Rowe Price participating; E+F rounds total $1.65B. The most notable signal is the CEO admitting summer compute shortages forced customer throttling, with the entire raise going to buying compute — generative media (image/video/audio) inference demand is hard enough to need separate funding for capacity, and NVIDIA's strategic stake further locks in chip supply. Cross-reference with SemiAnalysis's compute ledger: frontier labs are eating new compute on one side, inference platforms are separately grabbing cards for media generation on the other.
Sources: gcn.com
Stanford CS146S gets rewritten by reality: last year's nine assignments get one-shotted by a single prompt this year, so the course now has students submit patches to open-source projects, reviewed by maintainers themselves | A live record of iterating AI-native engineering education
Mihail Eric (head of Monaco AI) walks through the teaching design: agents can write features, but judging "does the maintainer want this, why write it this way, what will it break" is still human work. Two hard production-side conclusions: buy Devin for code review tooling, build your own agent workspace (buy to learn, build to keep). Internal system Hal outputs reviewable PRs for engineers and customer-ready explanations for sales — same backend, two outputs. He also notes agents can't see the historical decisions behind "why this change was rejected six months ago," and predicts open-source repos will become agent-run software factories. Bonus: Geoffrey Litt's explain-diff skill is directly usable on your next PR.
Sources: latent.space | latent.space
Rewriting Karpathy's run.c from scratch in Rust: hand-rolled prefix caching + three threads + FIFO scheduling + continuous batching, measured at 4.27x on CPU | An engineering note that builds an inference-serving mental model without reading vLLM source
Starting from a single-request HTTP service, then adding step by step: prefix caching that hashes token chunks and walks a parent-hash chain to find the longest reusable prefix (measured 8 of 11 prompt tokens reused); a message-passing architecture with three threads for HTTP listener, model execution, and reply; a request scheduler with FIFO plus active/waiting pools; and finally expanding the active pool to batch size for continuous batching (4 requests × 16 steps: 44.5 → 189.6 tokens/s). The scale is a teaching-grade single-machine CPU demo — no production eviction, chunked prefill, or streaming — but every step answers "what problem is this vLLM mechanism actually solving." A rare first-person record of building an inference server from zero.
Sources: matthewtejo.substack.com
P-H3-Image hands-on: Pruna AI's image model distilled from the MiniMax-H3 video model — txt2img trails Krea 2 / MOSS-Image, but Edit matches or beats Qwen-Image-2.1 | Full parameter comparison for 20B DiT + Qwen3-VL-4B TE + MiniMax-H3 VAE
Architecture is 20B DiT + Qwen3-VL-4B text encoder + MiniMax-H3 VAE, lightest config 21GB, optimized for 8MP/4K (it can't show its strength at 1MP). The piece includes a DiT/TE/VAE parameter comparison table against FLUX.2 klein 9B, Anima, Krea 2, Qwen-Image-2.1, FLUX 3 Image klein, MOSS-Image, and MiniMax-H3, plus ComfyUI setup and prompt guidance. Directly useful for anyone wanting to try a new image editing model; note this is a LoRA-distilled derivative, not a brand-new base model, and the author focuses on 2D editing and NSFW use cases, so the audience is narrow.
Sources: note.com
Claude Code 2.1.277 supports AGENTS.md, but only as a "fallback when CLAUDE.md is missing," not a merge | The right way to share agent instructions across tools is an `@AGENTS.md` import
The author breaks down CLAUDE.md's tree loading mechanism (root → working directory concatenated layer by layer, subdirectories loaded on demand, CLAUDE.local.md appended last), and points out that repos that have run `/init` will never trigger the fallback, so don't expect it to work automatically. The correct approach is to write an `@AGENTS.md` import inside CLAUDE.md. Also notes the toggle can only be changed in `/config` under Project instructions — no settings.json key, no environment variable, not yet supported on Bedrock/Vertex/Foundry, and the official memory docs aren't updated yet. One practical takeaway: use the import, not the fallback.
Sources: dev.to
Building a personal agent benchmark: deterministic scoring via final-state assertions, no LLM judge, unified JSON tool protocol makes tool-call reliability itself a measurable signal | "Public leaderboards test someone else's tasks"
The author argues public leaderboards test other people's tasks and other people's rubrics — useless for "can this model reliably do my work." The design tradeoffs are worth copying: score each task with final-state assertions instead of an LLM judge; route all models through a text JSON tool protocol so tool-call reliability itself becomes a measurable dimension; categorize tasks as file_ops / api_orchestration / trading / debugging / data_wrangling / obsidian, and run local Ollama models against frontier models on the same board — like performance-reviewing an "AI employee" to decide who stays and who does what. Useful for anyone doing model selection for their own workflow.
Sources: dev.to
Traversal raises $120M at a $1B valuation: using agents to automatically diagnose software incidents and find root causes | A sector signal that "agents move from writing code to operating code"
Kleiner Perkins, Redpoint, and Khosla co-led, with Sequoia and Battery following. Customers include American Express, Pepsi, and Blackstone. The pain point: AI coding tools have exploded code output, observability and maintenance difficulty rise in tandem, and incident diagnosis becomes the new bottleneck. It corroborates Cloudflare's Q3 earnings note that "AI companies signed an $11M Workers contract to support tens of millions of AI-generated apps" — once agents scale, both ops and edge deployment become must-haves. Both are financial/funding-side evidence only, lacking product or technical detail.
Sources: americanbazaaronline.com | convergedigest.com
🎙️ Podcast Picks
[One Million AI Engineer Special] How Stanford Teaches AI-Native Engineering — Mihail Eric, Monaco
📍 Source: Latent Space | ⭐ 4/5 | 🏷️ Agent, Product, Interview | ⏱️ 2:00:44
Mihail Eric walks through the design of Stanford CS146S's second iteration: once coding agents could finish previous years' assignments with a single prompt, the course pivoted to dropping students into unfamiliar open-source repos to propose changes and contribute. They discuss how AI-native engineering education rebuilds assignments and assessment, what skills engineers need in the agent era, and the real feedback that comes from involving open-source maintainers. Valuable for practitioners focused on agent deployment and engineering practice.
💡 Why Listen: It's the rare course-design story where the syllabus got rewritten because the homework stopped being hard. If you've ever wondered what "AI-native" actually means for how engineers get trained, this is the concrete version — plus real maintainer feedback loops.
Year-End Review #2: RSI Might Be NeoLab's Opportunity, PA Actually Works, Should Startups Build Fire Extinguishers or Bricks? | Conversation with Gu Yu
📍 Source: Crossing 十字路口 | ⭐ 4/5 | 🏷️ Agent, Research, Interview | ⏱️ 00:40:54
Gu Yu, co-founder of NeoCognition (a $40M seed round), reviews 10 AI moments from 2026: RSI and the innovator's dilemma at big companies as an opening for NeoLab, the Opus 5.5 upgrade wiping out months of work for the Computer Use team, why he still believes in the Scaling Law, getting his Memory Notes thesis wrong thanks to continual learning, the collapse of the academic conference system, whether NeoLab startups should build fire extinguishers or sell bricks, and the real usability of Personal Agents. Strong frontline reference for agent founders and researchers.
💡 Why Listen: Honest post-mortems are rare. Gu Yu openly says which of his own calls were wrong — the Memory Notes miss and the Computer Use wipeout are worth the 40 minutes alone.
Agentic Video Is Real and Here's What You Should Try
📍 Source: AI Daily Brief | ⭐ 3/5 | 🏷️ Agent, MultiModal, Product | ⏱️ 00:30:58
NLW explores how AI agents turn video production into "describe it and it's generated": from meeting highlight reels and concept demos to self-updating animated business reports and customer education videos. The episode lays out three levels to start at — from editing existing footage to building reusable video workflows. Somewhat useful for practitioners watching agent use cases, but it's an application-layer intro lacking model or architecture detail.
💡 Why Listen: Good if you want a quick map of where agentic video actually works today. Skip it if you're looking for technical depth — this is a use-case tour, not an architecture deep dive.