AI news for builders and product teamsUpdated Nov 17, 2026, 15:00 UTC
AI news, straight from the source.
Releases, research, and ideas for developers and product teams. Updated every hour.
Across the feeds
Newest first
Apple and Harvard researchers show that the gating mechanism in state space models pushes training toward in-weights memorization first, delaying or preventing an in-context learning solution even when the architecture and memory capacity are not fundamentally limiting.

Hugging Face's Transformers library now includes DecisionPipeline, AutoModelForDecision, and DecisionConfig for a new class of "decision models" that return probability distributions over predefined answers instead of generating text. The models support three primitives: choice, score, and noul, and can be served via a local /v1/systemone endpoint.

Apple researcher Kunal Talwar formalizes "evaluability" for black-box models when test samples may have appeared in training, showing accurate evaluation is possible only for distributions close to small support, and that query access expands what evaluators can do.

Simon Willison's llm-anthropic 0.31 adds Anthropic's new Fable 5.2 and Mythos 5.2 models; Fable 5.2 costs $8/million input and $40/million output, a 20% discount on Fable 5.1.

Simon Willison used Claude Code to build Beep Locator, a browser-based system that triangulates a mystery beep using multiple devices as microphones and located the culprit to within about 40cm — an old carbon monoxide detector with failing batteries.

President Trump took questions at a White House briefing about what he knew regarding Nvidia CEO Jensen Huang's acknowledged extramarital relationship with Alice Huang, identified in reporting as a suspected Chinese spy. Nvidia has opened an investigation into its CEO.

Wired reported that Nvidia CEO Jensen Huang hired crisis communications firm Heller Inc. at least a year before publicly acknowledging an extramarital affair, and that Nvidia's internal communications team was told to copy the firm on press inquiries about him.

Nvidia CEO Jensen Huang resigned after The New York Times reported evidence of a second affair, with the board installing CFO Colette Kress as interim CEO. Reports had already linked his acknowledged affair with Megaspeed executive Alice Huang to US intelligence identifying her as a Chinese spy.

Culper Research, a New York short seller, published a May 13 report questioning Nvidia CEO Jensen Huang's close relationship with Megaspeed executive Alice Huang, months before she was publicly suspected of being a Chinese spy.

Nvidia shares fell more than 4% on Monday after its board said it would investigate CEO Jensen Huang's relationship with an executive U.S. intelligence suspected of spying for China, and securities law firms began announcing shareholder investigations.

Stripe built Kai, a company-wide AI agent for non-engineers, in one week using LangChain's open-source Deep Agents harness; it grew from 296 to over 5,000 users in roughly four weeks and now reaches 83% of Stripe weekly.

Anthropic released Claude Fable 5.2 and Mythos 5.2, a shared model foundation split by access tier, ahead of its IPO; Fable 5.2 tops the Artificial Analysis Intelligence Index v5 at 65 points, five ahead of Gemini 4 Carbon.

At COP31 in Antalya, Turkey, negotiators from more than 190 countries are weighing Turkey's Antalya Artificial Intelligence Pledge, a commitment to use AI for environmental benefit that The Verge critiques as hollow.

Google Research and Google DeepMind have launched Language Explorer, an interactive web platform built on the open-source LinguaMeta dataset that lets researchers, educators and language enthusiasts visualize linguistic diversity across thousands of languages. It is also installed as an interactive exhibit at Berlin's Museum of Communication.

AWS announced that Anthropic's Claude Fable 5.2 and Claude Mythos 5.2 are available on Amazon Bedrock and Claude Platform on AWS, with Fable 5.2 positioned as Anthropic's most capable generally available model for coding and scientific work and Mythos 5.2 restricted to Anthropic's verification programs for cyber security and life sciences.

NVIDIA's Nemotron 3.5 Super VL, a 120B-parameter vision-language model with 12B active parameters per token, is now deployable in Amazon SageMaker JumpStart in two variants: a default NVFP4 build on G7 instances and full-precision BF16 weights on P5en.

Google DeepMind released Gemini Robotics 2.5, a vision-language-action model that improves success rates by 30% to 100% over prior baselines across 200 tasks and helped an Apptronik Apollo 2 humanoid achieve a 96% success rate on a Mercedes-Benz production-line kitting pilot.

A MIT Technology Review reporter's visit to Data Center Alley in Ashburn, Virginia, argues the national backlash against data centers is driven more by electricity demand and its local consequences than by the culture war framing Big Tech prefers.

Microsoft's latest Windows 11 Insiders Preview adds inline actions to Windows Search, letting users open or toggle specific settings such as Dark mode and Bluetooth directly from a search result. The feature is in the Beta channel and requires Preview Build 26320.9876 or higher.

Anthropic released Mythos 5.2 and Fable 5.2, claiming gains in coding, reasoning, and vision, and cut Fable's input price to $8 per million tokens, now below GPT-6 Astra. The release follows a METR report suggesting Anthropic may be nearing recursive self-improvement.