Jensen Huang Is Out. Now Congress Wants Answers!
- 01
Jensen Huang resigns
Nvidia said on November 12 that Jensen Huang had resigned as president, CEO and director, effective immediately, four days after he apologized for being unfaithful to his wife. CFO Colette Kress is interim CEO while the board, now with its first-ever chair, Harvey C. Jones, searches for a permanent one. The New York Times reported U.S. officials had concluded the woman was a Chinese intelligence asset, which Huang denies affected national security; the House Foreign Affairs chair says Huang agreed to appear, and the Senate Banking chair wants him in December.
- 02
Banned chips in weights
Will DePue says a model's freshly initialized weights depend on the exact GPU and seed, and that traces survive training. Searching over six trillion combinations, he matched six Chinese open-weights models, including DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3, to export-controlled Nvidia Blackwell systems, and reproduced DeepSeek-V4-Flash's starting weights bit for bit on a B300, while an H100 matched none. A match shows the hardware, not where it sat or whether any law was broken.
- 03
llama.cpp hits 1.0
llama.cpp and its ggml library reached v1.0.0, three and a half years after the project began. From now on, any break in the public C API means a major version bump. It supports more than 150 model architectures, with hardware for its CI runners from Nvidia, AMD, Qualcomm, Arm, IBM and Intel; Hugging Face's Clement Delangue framed the release as winning by building in the open.
- 04
Gemini 4 Hydrogen
Gemini 4 Hydrogen is Google's fast model for everyday coding and agent work: 2.5 times faster than Gemini 3.8 Flash at its default setting, with over five times fewer fabricated claims, and a 51 on the Artificial Analysis Intelligence Index (Argon: 50). It costs $1.50 per million input tokens and $7.50 per million output across a 1M context. Orchestrator in Antigravity switches between Carbon and Hydrogen, near Carbon's success rate with 24% less quota.
- 05
Claude joins your calls
Claude Meetings lets Claude join a Google Meet call as a visible participant, answer in writing and build docs and whiteboards while people talk, using each person's connected tools only with their permission. It saves a transcript, summary and screen descriptions but no audio, stays out of calls with outside guests by default, and is rolling out in beta on Team and Enterprise plans.
- 06
A bracelet that listens
Vexo is a Y Combinator-backed wristband that listens continuously (up to four days a charge) and turns conversations into actions, like texting family, booking calendar time, ordering groceries or prompting Claude. A button or "Vexo, stop listening" pauses it; the company says conversations are encrypted and never used for training, and its FAQ notes some places require everyone's consent to record. It ships January 30, 2027.
- 07
ChatGPT gets a wallet
ChatGPT Wallet stores a card so Dots, ChatGPT's agent, can pick and order things for you. Each purchase shows the item, total and card before you confirm with Face ID or a fingerprint, and merchants get a token instead of your card number. It starts with some U.S. Pro subscribers 18 and over; OpenAI plans to add spending limits under which Dots can buy on its own.
- 08
Claude decodes Gödel
Jan von Plato's Helsinki team had been decoding Gödel's 16 workbooks (1,009 pages, much of it in Gabelsberger shorthand) at about a page a week. Given the transcriptions, Claude produced a 650-page explanation in a couple of days and extracted 663 theorems; in the Coq proof assistant about 400 checked out as stated, about 200 after repairs and about 50 were false. It also found Gödel was one definition short of the axiom-of-choice independence proof he claimed in 1966.
- 09
Claude refuses Pokémon
Big Kid Nation, who has spent two years vibe-coding the free fan game Pokémon Arena 2D and says it cost up to $2,000 a month, posted a screenshot of Claude declining to write code for images it detected as official Pokémon designs. IGN covered the complaint; Anthropic hasn't said anything about a policy change.
- 10
AI GTA before GTA 6
A Reddit developer released lucidml/gta, an interactive world model trained from scratch on 800 hours of GTA 5 footage: give it a start image, hold keys, and it generates and streams the next frames to your browser. Its Summaryformer memory keeps the newest frames whole, squeezes older ones into summary tokens and never drops the start image. It runs on consumer GPUs, and the author says consistency and quality need more training.