AI news for builders and product teamsUpdated Nov 18, 2026, 02:00 UTC
AI news, straight from the source.
Releases, research, and ideas for developers and product teams. Updated every hour.
Across the feeds
Newest first
At the AI in Science Summit in Dublin, Nvidia backed the European Commission's new AI in Science Enablers Initiative, tied to RAISE, offering open models, training, compute access and startup support. Nvidia says 35 supercomputers are in development across Europe, to serve more than 3 million researchers.

Anthropic reported $16.8 billion in third-quarter revenue, up about 45 percent from Q2 but far below the previous quarter's 140 percent growth, as some investors expect a November IPO valuation below $1.5 trillion.

Apertus 1.5 adds native image and audio understanding, an optional thinking mode, a 262,144-token context window and improved tool use to the 8B and 70B models, with weights released July 24, 2026 under Apache 2.0.

NVIDIA detailed its expanding Nordic footprint, including a Swedish consortium's Sferical One supercomputer with 1,152 GPUs, an expected 1 GW of regional AI cloud capacity by the end of 2027, and deployments by Spotify, Volvo Cars, Einride and several Swedish startups.

Figma announced Figma Specs, a Markdown-based editor for defining a company's design language that teams and AI agents can reference, plus public beta release of its code layers feature, at its first Config event in India.

MIT economist Eric So's new book "The Collision: What AI Does to Us" argues heavy AI use causes cognitive atrophy by removing the productive struggle needed to build mastery, and offers four personal principles for navigating it. It cites an Anthropic study finding about half of college students' queries were low-effort requests for simple answers.

Simon Willison built a LaTeX rendering tool that stores the entire document in the URL fragment, making every rendered document a shareable, bookmarkable link, inspired by Tiny Little Office.

Databricks engineers built an internal search app called Slide Hub over 4 million passages and shared lessons on sizing Lakebase for it. The write-up details how they tuned the database to handle the app's search workload.

Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model with roughly 816 billion parameters and about 42 billion active per token, supporting text, images, audio and video with a context window up to one million tokens.

Gudea, a startup tracking how stories and rumors spread online, pitched its AI-driven narrative analysis platform at TechCrunch Disrupt 2026 and recently raised $7 million in seed funding, bringing total funding to about $15 million.

Meta's open-weight Muse Spark Open model is now available on Together AI as a Day 0 launch partner, offering a 1M-token context window on serverless and dedicated infrastructure for agentic coding workflows.

Apple released a new version of MLX that adds distributed inference across multiple Mac Studios using an RDMA over Thunderbolt driver, showing an 816B-parameter mixture-of-experts model running at over 36 tokens per second on four M5 Ultra machines. Single-machine runs of smaller dense models see up to 4.6x faster time to first token and 1.6x faster decode versus M3 Ultra.

Sullivan & Cromwell worked with OpenAI to build Agreement Analyzer, an AI tool that identifies issues in draft agreements, bringing the firm's legal expertise into the contract review process.

Apple researchers compared Retrieve-then-Plan and Plan-then-Retrieve tool discovery on ToolBench, finding a complexity-dependent crossover: RtP wins on single-tool queries while PtR wins on multi-tool queries, holding across retrieval budgets, registry sizes and two model families.

Apple-affiliated researchers introduce Selection-based Structured Reasoning (SSR), which turns agent reasoning into a choice among pre-written natural-language candidates instead of freeform generation. Evaluated on seven multimodal search benchmarks with 2B and 4B models, it cuts per-turn reasoning latency by over 90% and per-question inference latency by 28–54% without losing task performance.

Databricks has introduced Predictive AI Functions, a new capability that lets data teams generate forecasts and insights directly within their existing workflows. The company says the tools answer recurring business questions such as what sales will be next quarter.

FactSet reduced its Text-to-Formula model training and deployment cycle from days to hours using Databricks AI Runtime, according to a Databricks customer story.

McKinsey & Company plans to have some entry-level business analyst candidates hold a 10-minute conversation with an AI-simulated client during interviews, per The Wall Street Journal. The tool, being tested with undergraduates in some offices, covers roughly 80% of scenarios consultants might face, and humans, not the AI, will assess candidates.

At AIE New York, OpenAI's Joseph Kim described a "threshold of usefulness" where models become reliable enough to save analysts time, and predicted agents will handle investment idea generation and trading by 2027.

Flock CEO Garrett Langley said onstage at TechCrunch Disrupt that the company's gunshot-detection microphones detect "just gunshots," while Flock's own website says the audio systems also alert on car crashes, fireworks and sideshows and can expand awareness beyond gunfire.