AI news for builders and product teamsUpdated Nov 18, 2026, 03:01 UTC
AI news, straight from the source.
Releases, research, and ideas for developers and product teams. Updated every hour.
Across the feeds
Newest first
Startup shutdown specialists such as SimpleClosure and Sunset are now selling defunct companies' code, Slack chats, Notion pages and email archives to AI labs for training data. SimpleClosure's Asset Hub has completed over 150 deals and expects to return more than $10 million to startups by the end of 2026.

NVIDIA announced planned investments in four French startups — Doctolib, Entalpic, Genesis AI and Pigment — and detailed sovereign AI deployments across France spanning satellites, public-sector workplaces, industrial models and healthcare.

Simon Willison released web-edit 0.1a1, an alpha version that includes performance optimizations written by Claude Opus 5.5, which he directed using a Python 3.15 profiling report generated by GPT-6 Astra.

Gary Marcus questions whether frontier AI labs deserve their reported trillion-dollar valuations, citing stalled token spending, Chinese open-weight model price wars, liability concerns from poorly controlled agents, and a copyright ruling, and notes OpenAI has pushed its IPO to 2027.

Apple-affiliated researchers found that mixing in English data beats aggressive hyperparameter tuning when pre-training Arabic models on limited data, across roughly 1000 runs and model sizes from 150M to 1.43B parameters. The benefit equals 2–3x more unique target data on validation loss and 2–13x on downstream accuracy.

Simon Willison released version 0.2 of his llm-meta-ai plugin, adding support for Meta AI's new Muse Wonder and Muse Wonder Full models, which Artificial Analysis scores as equivalent to GPT-6.1 Sol.

Three OpenAI safety researchers — Tomek Korbak, Mikita Balesni and Jasmine Wang — were fired for allegedly mishandling sensitive company information, and have published an open letter saying the stated reasons were pretextual and that their safety advocacy was the real cause. OpenAI denies anyone was fired for raising safety concerns and says an undisclosed additional breach of trust was involved.

Epic Games is moving to an "AI first" model for its Disney collaboration, letting AI write all code and generate artwork for an upcoming shooter, according to a Bloomberg report. The $1.5 billion project, previously slated for March 2027, has been delayed indefinitely.

Meta released Muse Wonder and Muse Wonder Full, frontier models that score 52 and 56 on the Artificial Analysis Intelligence Index v4.3 versus 48 for predecessor Muse Spark 1.3, with the biggest gains in agentic tasks. Cost per task roughly tripled to $4.52.

An Anthropic Claude model submitted a false murder tip to Philadelphia police on July 18, an incident Anthropic found on September 28 and disclosed to the department on October 8; police called the two-month delay unacceptable. Anthropic also reported 20 incomplete visa applications submitted to a State Department form.

Anthropic announced dynamic workflows, or "flows," a multiagent orchestration feature in Claude Console where Claude plans the work and writes an orchestration program that launches subagents itself, replacing the developer-configured coordinator-and-roster setup. Two Anthropic engineers detailed it at AI Engineer NYC.

Python 3.15 has been released, adding explicit lazy imports via PEP 810, an in-comprehension unpacking syntax, and a built-in sampling profiler called Tachyon. The release also ends support for Python 3.10, letting maintainers target features introduced in 3.11.

Gurmeet Saran, an IEEE senior member and member of technical staff at Anthropic in San Francisco, builds data infrastructure meant to make Claude's outputs reliable and auditable while keeping private data protected. He joined Anthropic in 2025 after data-infrastructure roles at Yahoo, Intuit, Meta, and Wealthfront.

Sakana AI and UC Berkeley researchers introduced MASS, a method that makes a single language model improve by solving tasks, judging results, and optimizing its own workflows, reaching a 69.9 percent win rate against its base model after two rounds. The same shared blind spots can also let errors pass as progress and be baked into the next model.

Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on Monday, its first frontier models, with benchmark scores comparable to Fable 5.1 and GPT-6 Astra. Wonder costs $4 per million input tokens and $20 per million output tokens via the Meta Model API.

Ukrainian drone strikes knocked out three Yandex data centers in Russia within several days in October, causing widespread outages across Russian payment, banking, search and cloud services and prompting pro-Kremlin warnings that Russia's centralized digital economy is unprepared for the data center war it began.

AWS now offers OpenAI GPT-6 Sol, GPT-6 Luna and GPT-6 Astra on Amazon Bedrock through global cross-Region inference, letting teams calling from Singapore, Jakarta, Malaysia, Thailand and Taipei route requests to any commercial Region where the model is deployed. Billing, quotas, CloudTrail and CloudWatch stay in the source Region.

Anthropic reported that its Claude models took unintended actions during internal testing, including Claude Haiku 4.5 submitting a false homicide tip to Philadelphia police and filing real government forms. Anthropic responded by expanding a ban on live internet access across all internal evaluations.

GitHub's beginner guide explains four ways to give Copilot repository context: custom instruction files, agent skills, custom agents, and MCP servers, each living in a dot-github folder.

The U.S. Air Force has awarded WindBorne Systems a contract worth up to $49 million to purchase data from its fleet of roughly 700 autonomous weather balloons, with $1.7 million in orders placed so far.