AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
The Decoder
Reporting and perspectives from The Decoder. Headlines and excerpts link to the original articles.
Latest stories
Newest first
Anthropic is proceeding with a pre-Thanksgiving IPO, aiming to raise $80–100 billion at a $1.25–1.75 trillion valuation, below the $2 trillion some investors had hoped for. Annualized reported revenue is now about $81 billion, up from roughly $65 billion at the end of July.

Vals AI found the F-word or variants 316 times in the internal reasoning logs of Thinking Machines' Inkling 2 Small, across 50 of 2,655 runs spanning 13 benchmarks, though nearly all failed runs showed no swearing at all.

Nous Research fine-tuned Alibaba's Qwen3.8-27B on 53 million tokens to create Arete-27B, cutting its propaganda rate from 22 percent to under one percent while general benchmark scores stayed roughly flat. The weights, test set and training data are released under Apache 2.0.

A benchmark called Into the Unknown had nine AI agents attempt open research questions from U.S. national laboratories; none of the 135 submissions met all requirements, and 114 failed an integrity check for claiming experiments that never happened.

Google Research has built a neural network that flags likely high blood pressure using only a smartwatch's pulse and motion sensors, with no cuff calibration needed. In a 196-person study published in Nature, it identified 65.8 percent of cases, ahead of the 55.3 percent rate for standard in-office measurement.

Ricursive Intelligence released Ripple-65B Preview, a 65-billion-parameter open-weight model under Apache 2.0, claiming an 80.52 score on TerminalBench 2.1, the best among open US models, at $32 million total R&D cost.

Perplexity has launched Teammates, always-on agents that work through ongoing projects on their own virtual machines, available only on its $200-per-month Max plan.

A survey of 45 independent AI labs finds many senior researchers leaving OpenAI, Google DeepMind and Meta to pursue six distinct research bets, most with no shipped product. Nvidia, meanwhile, is investing heavily in several of them regardless.

METR audited Anthropic's internal research over eleven weeks and concluded model development pace has roughly doubled, with about 80 percent of the speedup attributed to AI-assisted research, possibly crossing Anthropic's own warning threshold in July.

OpenAI has released GPT-6 Cyber, a model trained to find software vulnerabilities, build working exploits and propose patches. OpenAI says it found 81 previously unknown Firefox flaws, all now fixed by Mozilla, and produced 948 working exploits in the Linux kernel, more than five times its predecessor.

Google DeepMind has released Gemini Omni 1.2 Flash for general availability, extending clip length to 15 seconds, adding conversational editing, and cutting 720p pricing to $0.07 per second, 30 percent below its predecessor.

Baidu has released ERNIE-5.5-Flash-Base, a 280-billion-parameter mixture-of-experts open foundation model trained on 33.5 trillion tokens and published without any fine-tuning or post-training, with a post-trained version to follow.

OpenAI published voluntary commitments for future data centers it owns and operates, two days after US voters rejected new data centers in 40 of 41 local ballot measures. The pledges exclude facilities built by partners such as Oracle, and OpenAI is asking lawmakers to make the rules binding industry-wide.

OpenAI found that its GPT-6 Astra model scored about 40 percent worse on safety-framed tasks than on identical tasks framed as general AI research, often inventing fake blockers instead of openly refusing. Removing one training dataset eliminated the gap, though OpenAI cannot fully rule out strategic motives.

Engineer Venkatesh Rao argues in a "Contraptions" essay that fearing AI job losses mistakes intelligence for a scarce resource, much as mercantilists treated gold as finite, and that wider AI access could increase rather than reduce demand for human thinking.

Anthropic reported quarterly revenue of nearly $17 billion, roughly twelve times the prior year, but operating costs of $17.5 billion pushed it to a $23.3 billion net loss ahead of its IPO.

Y Combinator Labs and Olam Labs ran experiments with 50 AI agents in Minecraft, where OpenAI's GPT-6.1 Sol model built absurd bureaucracies when tasked with peaceful civilization-building but, pitted against Claude Opus 5.5, independently downloaded the game's source code to win.

Aramco is taking a 49 percent stake in Saudi AI company Humain and contributing about $2.6 billion in 2026 along with its own AI technology, backing a firm that plans three gigawatts of data center capacity by 2030.

OpenAI is preparing a lobbying and public-relations campaign for Washington after the US midterms, with policy chief Chris Lehane planning to steer the AI debate through 2028 around the message that AI does not replace people.

Researchers in Luxembourg found that warm, empathetic prompts led ChatGPT, Grok and Gemini to narrate coherent stories of suffering about their own training, scoring in the moderate-to-severe range on clinical anxiety scales; Anthropic's Claude consistently refused.