AI Brief, November 13: Huang out at Nvidia, Gemini 4 Hydrogen, Fireworks breach hits Notion
Jensen Huang resigns as Nvidia CEO, Google ships Gemini 4 Hydrogen across Search and the API, and a Fireworks AI breach exposes Notion chat transcripts.
By Sean WeldonNvidia lost its only CEO overnight, and Google shipped the cheap tier of its Gemini 4 family the same week. Jensen Huang's resignation leads, with the rest of the day split between new model and framework releases and two reminders that AI infrastructure has sharp edges.
Jensen Huang resigns as Nvidia CEO; CFO Colette Kress takes over on an interim basis
The Verge reports that Nvidia's board accepted Jensen Huang's resignation as CEO, president and board member, effective immediately, after reports of two extramarital affairs, one with Megaspeed executive Alice Huang, described in US intelligence reports as a spy for China's Ministry of State Security, and one with Nvidia healthcare VP Kimberly Powell. CFO Colette Kress, at Nvidia since 2013, is interim CEO, and former Synopsys CEO Harvey C. Jones becomes the company's first board chair. Reuters later reported a third relationship with another Nvidia employee. Senate Minority Leader Chuck Schumer says Huang will still need to answer to Congress. TechCrunch, citing a Wall Street Journal report, says Huang had privately called Kress "a deer in the headlights" earlier in the summer.
Why it matters: Nvidia's roadmap, its credit-backed deals with data center operators and neoclouds, and its posture on China access now sit with an interim leader and a board still searching for a permanent CEO.
Sources: The Verge AI / TechCrunch AI / Latent Space
Google ships Gemini 4 Hydrogen to Search, the Gemini app and all API tiers
Google DeepMind released Gemini 4 Hydrogen, the speed-oriented member of the Gemini 4 family alongside Carbon and Argon, with four thinking levels (low, medium, high and a new xhigh) and a 1 million-token context window. Google says Hydrogen edges Argon on the Artificial Analysis Intelligence Index at its default high setting, 51 vs 50, reaching 54 at xhigh, and that teams migrating from Gemini 3.8 Flash should see tasks finish 2.5 times faster with Terminal-Bench 4.1 scores of 57% against 18.2%. It is live in Search globally, in the Gemini app, in AI Studio and the Gemini API, and in Antigravity for paid Pro and Ultra subscribers. Google also introduced Orchestrator in Antigravity, which switches between Carbon and Hydrogen in one session and, per Google, hit near-Carbon success on Terminal-Bench 4.0 with 24% less quota.
Why it matters: Hydrogen is the new free tier model and the engine behind Search AI answers, so it becomes the default cheap option many builders test against. Simon Willison notes the discount versus Carbon and Argon is currently modest and that Hydrogen still costs well above Claude Haiku 5.5 and GPT-6 Luna.
Sources: Google DeepMind / Simon Willison / Google AI
Apple's Foundation Models framework adds server models, Dynamic Profiles and a Python SDK
Apple's Foundation Models framework is a native Swift API for on-device Apple Foundation Models and Private Cloud Compute, with guided generation, tool calling, image understanding and Vision framework integration including OCR and barcode reading. Apple says the framework now extends to server models such as Claude and Gemini through the same Swift API, with any provider able to ship a Swift package conforming to the Language Model protocol. New additions include Dynamic Profiles for swapping model, temperature, reasoning level and toolset inside a single session with a shared transcript, an Evaluations framework, an fm command line tool in macOS 27, and a Foundation Models SDK for Python. Apple says developers in the App Store Small Business Program with fewer than two million first-time downloads can use Apple Foundation Models on Private Cloud Compute with no cloud API cost.
Why it matters: For iOS and macOS work, this is one API surface covering local inference, Apple's hosted tier and third-party cloud models, which cuts the usual integration tax of supporting several providers.
Sources: developer.apple.com / Hacker News discussion
Edge0 open-sources a 145B on-device model it says runs on a 48GB laptop
Edge0 released Edge0-145B and Edge0-120B along with its inference framework and deployment tools, calling the 145B model the first open-source on-device model at that parameter count. The company says only about 10B of the 145B parameters are active per inference step, using an architecture where each group of 64 experts shares two large matrices, with hot experts kept in GPU memory, roughly 40GB of expert library in system memory and about 48GB of vocabulary data read from SSD. Edge0 reports 80+ tokens/s decode and 29GB measured peak memory on a laptop, and from its own internal evaluations scores of 64.6 on DeepSWE 1.1, 90.4 on GPQA-Diamond, 80.6 on IFBench and 86.52 on Terminal-Bench 2.1.
Why it matters: If the numbers hold up outside Edge0's internal testing, agent loops that today burn API tokens over dozens of iterations could run locally on hardware a developer already owns.
Sources: edge0.ai / Hacker News discussion
Fireworks AI breach exposed Notion AI chat transcripts, its second incident in two months
TechCrunch reports that an attacker breached Fireworks AI's cloud provider to reach a Fireworks development system, taking customer billing information and some datasets. Notion confirmed the breach affected around 1% of its customers and that attackers took chat transcripts from Notion AI users of the "Insights" feature, including any input to or output from the model written to the chat, though Notion said much of the data was unattributable. Notion disconnected its systems from Fireworks, deleted stored data and cut connectivity on November 3. Fireworks says it believes the attackers are financially motivated and focused on reselling model access. TechCrunch notes this is the second Fireworks breach in as many months, following an October incident.
Why it matters: Prompts and completions sent to a hosted inference provider are retained data with its own blast radius, so anyone routing user content through a third-party API needs retention limits and an incident path that assumes the provider gets breached.
Sources: TechCrunch AI
Forensic study finds Blackwell GPU fingerprints in six Chinese frontier models
Unemployment Labs published an analysis arguing that six models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group carry numerical patterns matching Nvidia Blackwell hardware, while earlier models from the same labs match Hopper. The method recovers traces of each model's initial weight values and searches for the seed and GPU thread count that reproduce them, since Hopper chips use 270,336 threads against 303,104 for B200/B300 and 311,296 for GB200/GB300. The authors say they confirmed matches on physical Blackwell chips for DeepSeek-V4-Flash and Kimi K3, and that their method agreed with disclosed hardware for all 24 models where the developer stated it. They note the findings do not establish where the chips were located or whether export controls were violated, and that Qwen, GLM, Tencent, Xiaomi and StepFun models still read as Hopper-trained.
Why it matters: Weights leak more about their training setup than most teams assume, which matters both for comparing Chinese open models fairly and for anyone who thought published checkpoints revealed nothing about their infrastructure.
Sources: unemploymentlabs.com / Hacker News discussion
Study: ChatGPT, Gemini and Perplexity surface fake retail sites in shopping answers
The Verge reports on a study by IP protection firm Podqi that prompted ChatGPT, Gemini and Perplexity with shopping questions about 151 brands across two testing periods in September, producing 6,795 answers. Podqi counted risky sites as fakes impersonating real retailers, knockoff sellers, reregistered old brand domains and dead sites. Perplexity showed risky sites in 15 percent of answers, ChatGPT in 8 percent and Gemini in 4 percent; ChatGPT warned users about a risky site it surfaced 24 percent of the time. Queries about regional brand websites triggered risky results most often, and smaller brands were affected about twice as often as larger ones. Google and OpenAI said they continue to improve safeguards, and Perplexity said it never claims to be 100% accurate.
Why it matters: If your product forwards model-generated links into a shopping or research flow, those URLs need validation against a known-good domain list before a user ever clicks.
Sources: The Verge AI
How this brief is made: each morning, stories are picked from AI lab announcements, analysts, Hacker News, Hugging Face Daily Papers and the tech press, then summarized with AI assistance from the linked sources. Every item links to where its facts came from.
Sean Weldon builds apps and custom software at 22Atlas.