On November 8, Nvidia's board opened an independent investigation into CEO Jensen Huang's relationship with a woman suspected of being a Chinese spy. Huang stepped down, and Congress has subpoenaed him to testify. The company sits at the chokepoint of both global AI compute supply and US export-control enforcement — the scandal drags both into uncertainty at once.
In the same period, a METR report said Anthropic's internal R&D had sped up 2x, and on that basis Dwarkesh moved his AGI timeline from ten years to two or three. Amazon's controlled experiments found that multi-agent coordination costs several times more tokens for gains that are rarely significant. GitHub turned repeated judgment calls into SKILL.md files; at Stanford, after a single prompt demolished the coursework, assignments became patches to open-source projects. Gemini 4 Hydrogen leads on speed and low price, while Claude Fable 5.2 still leads on benchmarks both sides report. Google opened a separate decision-model track that returns calibrated probabilities instead of generated text. Agents like OpenClaw cannot learn from experience; training the weights of a small model can. Thinking Machines trained a 594B model end-to-end in 4-bit with NVFP4. Together uses an agent compiler to continuously recompile inference, beating the best open-source engine on throughput. On the embodied side, NVIDIA GR00T 2 learns skills from human video, and Google Gemini Robotics 2.5 is already on a Mercedes-Benz production line. World Labs opened Atlas to everyone, and Genie 4 can generate an entire marathon course in real time. In the markets, Nvidia fell on the CEO scandal while AMD and Broadcom surged against the trend.
Industry & Ecosystem
Automated AI R&D and the RSI call
Three pieces of material, from three different vantage points, point at the same thing: AI is accelerating AI research, and for the first time that acceleration is quantifiable.
Anthropic IPO roadshow video — Simon Willison produced an unofficial transcript with MacWhisper. In it, co-founder Tom Brown laid out the compute strategy systematically for the first time: Claude trains and runs inference across four chip families at once — AWS Trainium, Google TPU, NVIDIA GPU, AMD GPU. When AMD GPUs were first brought in this year, the Mythos model was ported in three days with almost no human intervention, and nearly 100% of the kernels were written by Claude. He introduced the idea of a compute multiplier: physical compute times the efficiency multiples that can be discovered, where the multiples compound — and he was explicit that Claude will find those multiples itself in the future.
Dwarkesh publicly revised his judgment from six months earlier, from "at least another ten years" to "two or three years, possibly months, until an AI matches top human AI researchers." The argument is checkable uplift data: METR spent three months embedded at Anthropic and found overall R&D speedup of 2x, mostly attributable to using AI for R&D; heavy users inside OpenAI went from $1,000 to $7,000 per day in coding-agent spend; Anthropic self-reports that Claude already leads 26% of AI R&D work, up from under 1% in February.
Zvi takes apart the METR report: the confidence interval is 1.3x–3x, no acceleration is visible from public models alone, and the estimate only moves from about 1.7x to about 1.9x once Anthropic's unreleased internal models are included. According to Forecasting Research, METR attributes about 80% of the acceleration to AI.
What the three share is not a conclusion but a type of evidence: the basis for judgment has shifted from extrapolated curves to internal measurement. The missing piece is equally clear — Zvi notes that the most informative bottleneck analysis was redacted by Anthropic, while Benjamin Todd's extrapolation suggests Claude could lead internal research by mid-2027.
Nvidia CEO turmoil and the political risk in chip export controls
This is not a corporate-governance story. Nvidia sits at the chokepoint of both global AI compute supply and US export-control enforcement, and the CEO's downfall drags both into uncertainty at once.
Hard Fork brought on NYT reporters Ana Swanson and Ryan Mac and biographer Stephen Witt to trace Huang's relationships with subordinate executives, Megaspeed and the halted Commerce Department chip investigation, and Nvidia's government lobbying. According to CNN, Megaspeed bought more than $4 billion of chips from Nvidia; companies like it can become a hole in export controls by renting compute to Chinese firms.
Gary Marcus asks who knew what, and when: Trump appears to have known a year ago yet stayed close; Rubio's spring memo went nowhere; Nvidia still claimed ignorance a week ago; Congress is about to investigate. He strings together a causal chain — possible leakage of export-control insider information, no succession plan, futures already falling — that could hit the entire AI investment market.
Zvi fills in the timeline: Culper's May short report had already named Alice Huang's close ties to Megaspeed, and there were public questions in October; Nvidia intervened to stop the investigation into Megaspeed; the stock is 7% behind the Nasdaq. He argues the real issue is not private conduct — when the CEO has an undisclosed personal relationship with a customer, the credibility of his argument against export controls is undermined at its foundation.
All three converge on one judgment: limited technical impact, policy risk that cannot be priced.
Agents: Technology & Applications
Agent architecture: cost and effectiveness benchmarks
Three independent evaluations point at the same thing — the performance bottleneck for agents is not architecture choice but implementation quality and cost structure.
The Amazon and University of Washington paper is the cleanest version: fix the model, task and tools, vary only the architecture, and compare 8 arms across 13 enterprise SOPs. Three conclusions. There is no universally best architecture — on the same SOP, task success ranges from under 10% on the hardest to near 100% on the easiest, and at the aggregate level no architecture wins with statistical significance. Coordination architectures must prove their cost with gains first: every multi-agent or skill-augmented arm spends 1.4–4.7x the tokens and 2.6–5.8x the latency, with gains usually small and rarely significant, so single-agent remains the efficiency default. Implementation quality outweighs the single-vs-multi-agent choice — high-quality skills lift success from 67.4% to 73.1%, and one targeted redesign made multi-agent the only arm to significantly beat the baseline (75.6% vs. 67.4%).
EnvOpt changes a different variable: it touches neither the model nor the harness, only how the environment presents information to the agent — for example, whether Slack goes through an MCP remote service or is exported as plain text for the agent to grep. Across five benchmarks it cuts cost by an average of 48% on unseen test sets with pass rates equal or higher; on AutomationBench it beats Meta-Harness by 9.67 points at 15% of the cost. The optimized environment transfers to other harnesses and models, and stacking it on Meta-Harness adds another 16% pass rate.
本期为订阅内容
订阅 RecsysFrontier，在 Substack 阅读本期全文
- 每天的 AI / 推荐系统前沿日报，第一时间读到
- 每周周报，一周重点一次看完
- 订阅期内所有日报周报全文
本期将于 11 月 21 日免费开放