Nvidia's board investigates Jensen Huang after he admits an affair
On Sunday Nvidia published a message from Jensen Huang to staff: "I was unfaithful to my wife. My actions hurt her and our children. There is no excuse for what I did." He did not name the woman. Per TechCrunch, she is Alice Huang, a former executive at Singapore chip company Megaspeed, whom a memo circulating among US policymakers says China's Ministry of State Security recruited to spy on him.
What the board said
- An independent inquiry: a committee of independent directors, with outside counsel, will examine "all available information" after reports that Huang was "the target of a foreign intelligence effort" (board statement via The Verge).
- No breach found so far: the board says it has "no reason to believe that either national security or NVIDIA secrets were compromised", and that it "has confidence in Jensen and his leadership team".
- Megaspeed is a customer: it buys from Nvidia and is itself the subject of a Commerce Department investigation, per The Verge.
- China denies it: the Chinese Embassy called the allegations "groundless smears and slanders" in October, and no one has been charged.
Can he stay?
Yale's Jeffrey Sonnenfeld puts Huang's odds of keeping the job at "about a 50-50". If Huang was merely naive, he told TechCrunch, "any employee in Nvidia other than him would be fired over this." There is no obvious successor: Huang runs a flat company with about 60 direct reports by his own count. Senior Trump administration officials told the New York Times they doubt Huang could have missed her role for so long; when Trump was first briefed a year ago, officials did not think he knew.
For builders: nothing changes in a GPU order this week. The thing to watch is policy, not the product: Huang is due to brief Congress on November 18, and anything that tightens chip export rules lands on prices and lead times.
Is it a bubble?
The day's most-discussed essay on Hacker News (534 points, 917 comments), Matteo Raso's "I'm not convinced the AI bubble exists", argues the spending is mostly one-off. AI companies have committed at least $3 trillion and spent about $1.9 trillion, mostly on data centers, which are built once. Running the models is the cheap part: Anthropic's S-1 shows $2.08 of revenue for every $1 of compute. The author grants that the debt sits in vehicles that make the true cost hard to see, which is the bubble camp's best point. The same weekend, TechCrunch noted that Nvidia sits at the center of a "circular financing web", which is why one man's personal life moves markets.
Agents and the work
- Agents still leave a mess: Alibaba and several Asian universities built Co-Design Bench from real change requests, mostly ones where users had to rephrase or roll back. The best of ten models, GPT-6 Astra, met about 70% of a task's requirements on average but finished only 26.4% of tasks fully correct. Claude Opus 5 and Qwen3.8-Max followed; Grok 4.6 came last. One AI judge tests the app in a browser and a second hunts for mock-ups and hardcoded results. Older benchmarks score above 99%, which is why they no longer tell models apart.
- Agents that keep going: Perplexity's Teammates run several agents that keep working after you close the tab. Each project gets its own Git-backed repository; agents work in separate sandboxes and submit changes back. Fewer than 2% of submissions are rejected for stale state, median submission takes under 1.7 seconds and p99 stays under 8.
- Siri goes agentic, slowly: per Bloomberg's Mark Gurman via The Verge, Siri gets fine-grained control of Apple's own apps (crop in Photos, sort Safari tabs) with iOS 27.3, and third-party apps such as Grubhub and OpenTable in the spring. Ongoing, goal-tracking tasks may come next year.
- Coding tools pulled from traders: Goldman Sachs now limits Claude Code and GitHub's tools to its roughly 12,000 engineers, cutting off traders and bankers who used them to model positions (Bloomberg via The State of AI). In July its CIO was still praising "vibe coding" by non-engineers. Goldman is also a lead underwriter on Anthropic's IPO.
Research
- Sutton on scaling: Richard Sutton, who wrote "The Bitter Lesson", now says raw compute is not enough and AI science is "far from finished". He wants agents that keep learning from experience without forgetting, and calls synthetic data a "big mistake". Former OpenAI researcher Jerry Tworek agrees the lesson holds, but says the more compute you spend, the more the algorithm matters.
Music and rights
- Elvis and Eminem vs Suno: Elvis Presley's rights holders and a publisher of Eminem's records are suing Suno. The Presley suit leans on Tennessee's ELVIS Act, which bans cloning a singer's voice. Suno says it blocks prompts with artists' names and screens uploads.
Tools
- C vs Rust for a SQLite function: Simon Willison had an agent build a `parse_number()` SQLite extension that turns "-$1,234.56" or "12.5%" into numbers. The C build was 14 KB, the Rust ones 298 to 321 KB, and Rust was faster, because its equivalent of C's `strtod` is about twice as fast. Then he had it write a faster C parser.
The through-line
The most valuable chip company has a single point of failure, and its board is now testing whether it can run without trusting him blindly. The agents we build have the same problem in small: they do 70% of the job and still need someone to check the rest. Trust, then verify, is a governance rule and an engineering rule.
One thing to do today: pick the last five tasks your coding agent marked done and check each one end to end in the running app, not in the diff.
Sources covered
1 story · 17 sources (16 news, 0 company, 1 voices)