AI news for builders and product teamsUpdated Nov 17, 2026, 21:01 UTC
AI news, straight from the source.
Releases, research, and ideas for developers and product teams. Updated every hour.
Across the feeds
Newest first
Meta's updated commercial terms, effective October 30, authorize automated scraping of business accounts' linked websites and content, with no opt-out for businesses. Legal experts warn the terms could let Meta train AI on publisher content without licensing fees, though Meta says crawling practices are unchanged.

Databricks researchers propose injecting domain knowledge into large language models to improve triage of malicious npm packages, targeting a blind spot in binary and obfuscated code analysis.

Vercel data shows Jev, a limited-output decision model from TypeSafe AI, reached 35% of paid teams within two weeks of its September 15 limited release, nearly double the reach of second-place Sonnet 5.5 — the fastest two-week adoption Vercel has recorded.

Reflection AI released Beam, an open-weight sparse mixture-of-experts model for agentic coding with 501 billion total parameters and 23 billion active parameters, now deployable on Amazon SageMaker JumpStart and HyperPod Inference.

Google DeepMind's Gemini 4 Argon scored 97.5 percent on ARC-AGI-3 using a vendor-specific test harness, making Google the second lab after OpenAI to cross 95 percent. Under the standard harness the score drops to 49.8 percent, and Argon trails rivals in math and software engineering.

Mark Wahlberg said at TechCrunch Disrupt that he supports AI in filmmaking when it enhances existing work rather than replacing storytelling or actors, and that he is open to an AI likeness carrying on his legacy.

An InfoQ webinar on running agentic systems in production featured practitioners describing silent failures such as semantic and embedding drift, and recommending automated review, continuous evaluation, and per-agent access controls. The panel framed deployment approval as a strategic business decision rather than a purely technical one.

Meta's updated commercial terms, effective October 30th, grant it permission to crawl and scrape websites "associated" with businesses using its platforms and use that content to train AI, according to Nieman Lab. Meta says it is only consolidating ad terms, but legal experts warn the language may override site-level opt-outs like robots.txt.

National Geographic's "Super Scam Me," premiering Saturday at 9 pm ET and streaming on Disney+ and Hulu, follows journalist Mariana van Zeller as she poses as a fake online dating profile to expose romance scammers. The FTC says such scams cost Americans nearly $1.5 billion a year.

Veritas has launched a second-generation authentication chip that works on metal surfaces and can be scanned through glass, expanding its anti-counterfeit technology to jewelry, trophies, and glass items like fragrance bottles. The startup also announced a partnership with ATHLOS to embed chips in all 18 championship crowns.

A new essay argues neural networks may be far more legible than assumed, with roughly 90 percent of neurons serving a clear function, a claim researchers are still debating. If it holds, dangerous behavior could be spotted directly in a model's circuits.

Anthropic has granted six METR staffers company hardware, office space and internal tool access to independently audit its AI models and safety practices, a deeper arrangement than it has offered other outside reviewers. METR will publish findings independently, with redactions possible only in predefined categories.

Cloudflare has introduced a continuous behavioral accountability system for Verified bots, scoring whether they honor robots.txt preferences and back off after HTTP 429 rate-limit responses. Scores will appear publicly on Radar in the future, and customers will eventually be able to set their own trust thresholds for handling Verified bot traffic.

At TechCrunch Disrupt in San Francisco, Replit CEO Amjad Masad described a "self-driving company" where AI agents do much of the routine work while humans supply judgment and taste. He pointed to Replit's own support workflow, where an agent triages tickets before a person decides how to proceed.

Grafana Labs proposes using evaluations, service-level objectives (SLOs) and error budgets to measure whether AI agents behave acceptably, turning hallucination, task fulfilment, toxicity, cost and prompt injection into monitorable targets. An example SLO requires 95% of conversations judged fulfilled over 30 days, with the remaining 5% as error budget.

The Center for AI Safety has published a Wage Loss Dashboard that estimates over 117,000 AI-attributed US job cuts this year and about $8.5 billion in annual wages lost, broken down by state and per-person cost.

AWS has open-sourced Strands Decider 2B, a two-billion-parameter model that picks between predefined options with confidence scores, alongside Strands Box, a developer-preview kernel sandbox that enforces execution policies outside the model.

Microsoft will scale its Maia accelerator family to multi-gigawatt deployments from 2027, making Maia the primary inference system for Microsoft products. Maia 300 puts scale-up and scale-out on one Ethernet fabric and moves KV state directly between prefill and decode workers, but practitioners questioned the tokens-per-watt efficiency claim.

Meta released Muse Spark Open, an 816B-parameter open-weight mixture-of-experts model with a 1M-token context window, scoring first among US open-weight models on Artificial Analysis' Intelligence Index but trailing GLM-5.3 and Kimi K3. Separately, Goodfire's Nick Cammarata argues that cases of polysemanticity in neural networks dissolve under further agentic analysis.

Goodfire's Nick Cammarata published an essay arguing that most neurons in models up to trillions of parameters are already legible, and that many famous polysemanticity examples were misread; his critics say it does not resolve their key uncertainties.