AI Brief, November 12: Gemini Decisions API, Nvidia board moves on Huang, AlphaGenome 2

Google launches the Gemini Decisions API, Nvidia's board reportedly prepares to oust Jensen Huang, DeepMind ships AlphaGenome 2, and United blocks AI agents.

Google opened a new front in the decision model category with the Gemini Decisions API, the lead story for builders today. Alongside it: Nvidia's board reportedly moving to oust Jensen Huang, DeepMind's AlphaGenome 2, an AWS emulator written mostly by agents, and United shutting personal AI agents out of bookings.

Google launches Gemini Decisions API for fast, calibrated closed-set choices

Google introduced the Gemini Decisions API, backed by a new model, gemini-decision-1, available through the Gemini API in Google AI Studio. It handles yes/no checks, multiple choice with up to 255 options, and ordered rubric scores, taking one multimodal input across text, images, audio, video and PDFs up to 256K tokens plus as many as 1,000 questions per request. Google says it scores candidate options in a single forward pass rather than decoding token by token, returning probabilities and confidence scores in 150-250 ms end to end, with no output tokens billed and a price cut relative to Gemini 3.5 Flash-Lite with the Interactions API. Simon Willison notes it mirrors TypeSafe's Jev, reusing Jev's term "Noul" for yes/no questions, but is fully multimodal where Jev is text only; he found Gemini returns less confident probabilities than Jev and OpenAI, which answer 1.0 or 0.0.

Why it matters: Tool selection, policy checks, model routing and context grading can move off full generative calls to a cheaper, lower-latency endpoint that returns confidence scores you can threshold for human fallback.

Sources: Google AI / Simon Willison / Latent Space / Hacker News discussion

Nvidia's board reportedly prepares to oust Jensen Huang after second affair report

The Verge reports, citing The Wall Street Journal, that Nvidia's board is preparing to dismiss CEO Jensen Huang and could meet as soon as Wednesday to decide his future. The move follows a New York Times report, based on five people who worked at Nvidia, that Huang had an extramarital relationship with Kimberly Powell, Nvidia's vice president of health care, who reports directly to him; it comes days after he admitted a separate affair with Megaspeed executive Alice Huang, whom US officials believe is a Chinese spy. Nvidia's stock fell more than 7 percent in after-hours trading Wednesday. TechCrunch reports the company has no apparent succession plan, and that Huang has about 60 direct reports.

Why it matters: Nvidia supplies the GPUs behind most AI infrastructure, so a leadership vacuum and an ongoing federal investigation add uncertainty to hardware roadmaps and supply planning.

Sources: The Verge AI / TechCrunch AI / TechCrunch AI

DeepMind's AlphaGenome 2 predicts DNA variant effects in specific cell types

Google DeepMind released AlphaGenome 2, a conditional DNA sequence model that predicts variant effects for one tissue, cell type or cell state at a time rather than averaging across tissues. DeepMind says it was trained on more than 107 million individual cell profiles across 68 tissues, adds DNA methylation to the readouts it predicts, and identifies nearly twice as many variants that change gene expression as its predecessor; on the TenK10K immune dataset it reports identifying 30% of causal variants versus 16%, and 43% of directional gene effects versus 23%. The model can also condition on unseen single-cell experiments without retraining. It is available for non-commercial research through the AlphaGenome API and in private preview for commercial research, and DeepMind says AlphaGenome 2 Skills will come to its Google Antigravity agentic platform.

Why it matters: Teams building computational biology tools get a cell-type-resolved variant model with an API and precomputed predictions, though it is research use only and not validated for clinical or diagnostic purposes.

Sources: Google DeepMind / The Verge AI

AWS ships an official local emulator written by a swarm of agents

Simon Willison flags aws-emulator, an official AWS emulator that lets you run tests against a local clone of AWS services without provisioning a paid API key. He notes it is written in Rust by a large swarm of agents, with 29,684 commits in the project's first three weeks, and currently requires Docker to run on macOS or Windows. Swaminathan Sivasubramanian described how it came together in a guest post on Werner Vogels' blog, framing it as a test environment for coding agents.

Why it matters: Local emulation removes a standing friction point in testing AWS-backed services, and the commit volume is a concrete data point on what agent-written codebases look like at scale.

Sources: Simon Willison / allthingsdistributed.com / Hacker News discussion

Huawei open-sources MACaQ, a model-agnostic quantizer for local inference

Huawei published MACaQ, a quantization toolkit under the Apache License 2.0 that targets llama.cpp, vLLM, Diffusers and PyTorch, covering text, image, video and audio models. The project says it can quantize a 122B-parameter model in 64 minutes using 32.5 GB of device memory, and can match the best Llama-3-8B perplexity of MR-GPTQ and DuQuant++ with 11 times less quantization time. It supports 2-, 3- and 4-bit GGUF for text generation and W4A4 for diffusion and speech models, and ships MACaQ-R, which refines existing GGUF files while keeping them runnable with a standard llama.cpp install.

Why it matters: It is another option for shrinking models for on-device or self-hosted serving, including refining GGUF checkpoints you already ship.

Sources: github.com / Hacker News discussion

United blocks third-party AI agents from booking its flights

The Verge reports that United will not let personal AI assistants complete ticket purchases, citing privacy and regulatory concerns; agents can still show users United flights. United says sharing account credentials with third-party assistants can expose sensitive information and lead to unintended purchases, account changes or use of MileagePlus miles, and that agent-driven check-in could conflict with federal rules such as attesting that bags do not contain banned substances. CEO Scott Kirby said he does not want to create another middleman, according to The Wall Street Journal. CBS News reports the restriction extends to travel agencies and aggregators, while Southwest has taken the opposite approach and made its flight inventory viewable to ChatGPT users.

Why it matters: Agent products that transact on a user's behalf need to plan for sites that refuse them, and credential sharing is emerging as the specific objection to design around.

Sources: The Verge AI / cbsnews.com / Hacker News discussion

OpenAI quarterly revenue reportedly jumped almost 50%, Bloomberg says

TechCrunch reports, citing Bloomberg and anonymous sources, that OpenAI's quarterly revenue reached a new high on strong demand for its recent models, an almost 50% jump from the prior quarter and a sharp acceleration from roughly 18% growth between the first and second quarters. The report says OpenAI's adjusted operating margin improved from -150% in the second quarter to -109% in the third. TechCrunch attributes the gains to tighter pricing, discounts, improving models and more enterprise-focused products and agents, and notes Anthropic has reported higher third-quarter revenue while leading in the enterprise market. The two companies calculate annualized revenue differently: Anthropic counts cloud partner sales, while OpenAI reports only its share of some partner sales.

Why it matters: Demand for frontier model APIs is still accelerating, which matters when you are betting on a single vendor's capacity, pricing and enterprise roadmap.

Sources: TechCrunch 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.