AI news for builders and product teamsUpdated Nov 17, 2026, 14:00 UTC
The Decoder
Reporting and perspectives from The Decoder. Headlines and excerpts link to the original articles.
Latest stories
Newest first
Mistral has released a preview of Mistral Large 4, a one-trillion-parameter model trained in its own European data centers, with weights expected at the end of October. It scores 38 points on the Artificial Analysis Intelligence Index, still well behind Claude Opus 5.5, and Mistral pitches cybersecurity work that closed US models refuse.

South Korea plans to invest 4.7 trillion won ($3.49 billion) through government-backed equity investments to develop a homegrown frontier AI model, as part of its proposed 2027 budget pending parliamentary approval. A second funding track would support adoption of Korean-built models, and an open competition will include startups.

Reflection has announced Beam, its first open-weight model, a mixture-of-experts system that activates 23 billion of its 501 billion parameters per token. It is set to ship later this month under the Apache 2.0 license.

Researchers at PhAI Labs, with collaborators from Stanford, Oxford and Princeton, extended Yann LeCun's JEPA architecture into JEPA-Anything, a world model spanning seven fields. The effort also produced a liver cancer treatment candidate tested in lab samples and mice, though the study does not establish whether it could become a therapy.

Deepseek is close to raising at least $12 billion in a new funding round, with CATL and Tencent contributing the largest shares, Bloomberg reports. The Chinese AI startup plans to restructure for an IPO in early 2027.

Cohere has announced North 2, an enterprise platform that orchestrates AI agents across multi-step workflows and works with any model. The company says the platform can run on-premises, in the cloud, or air-gapped, and is aimed at government agencies and regulated industries.

Meta and Microsoft are cutting their internal use of Anthropic's Claude. Microsoft reduced its cloud division's monthly per-employee Claude budget from $100,000 to about $10,000, while Meta halved Claude Code users from roughly 60,000 to 30,000.

Reka AI released a research preview of Rho-1, a 19-billion-parameter omni-model that processes and generates text, images, video, and robot control actions in one neural network. It trained on 320 H100 GPUs for about three months.

OpenAI will add invisible textGrain watermarks to ChatGPT and Codex outputs in the European Union over the coming weeks, while API watermarking stays opt-in worldwide. Internal tests showed detection around 95 percent on 400-token psychology passages, dropping to 17 percent after a quarter of words are replaced.

A Quinnipiac University poll found that 77 percent of Americans want AI development slowed or stopped until safety is verified. 86 percent support independent safety standards, and 74 percent have little or no trust in AI company leaders.

Anthropic employees donated $540 million in 2025, nearly five times the amounts given by the next-largest Fortune 500 corporate donors, according to documents shared with potential investors ahead of its planned IPO. A company program tops up employee share donations with extra shares, tripling contributions from early employees.

Aleph Alpha has released Kolibri, an open-weight German-English mixture-of-experts model with 78 billion parameters, about three billion active per token, trained on 768 B200 GPUs in Germany and Finland and published under the Apache 2.0 license.

OpenAI will test display ads on ChatGPT's image generation loading screen in the US later this month, showing product carousels below the generated image. The company is also expanding ad measurement, brand safety pilots, and negative keyword controls.

An MIT expert committee report warns that AI is eroding key parts of the college experience, including office hours, study groups, and undergraduate research programs. It also cites breaking trust between faculty and students, with some professors considering AI agents instead of student research assistants.

Donald Trump announced a "Super Intelligence Force," using "superintelligence" as his term for artificial intelligence. The unit, led by Director of National Intelligence Jay Clayton and other officials, will coordinate outreach to consumers, interest groups, religious organizations, critical infrastructure operators, and AI companies, reporting directly to the president and chief of staff Susie Wiles.

Google researchers propose RRSI, a method to stop self-improving AI agents from memorizing their test tasks. It improves scores on unseen benchmarks by up to 4.7 points while using about 30 percent fewer tokens than an unregularized version.

NASA and IBM Research have released the NASA-IBM Lunar Foundation Model, described as one of the first open-source foundation models for lunar science. Trained on nearly 2 million tile bundles, mostly from 17 years of Lunar Reconnaissance Orbiter data, it reduced polar ice deposit prediction error by up to 22 percent versus the best baseline.

An Aleph Alpha study found that Chinese AI models from Alibaba, DeepSeek, and Moonshot AI often repeat state doctrine or refuse to answer on politically sensitive topics, with only 17 to 41 percent of responses rated balanced. Western comparison models scored higher on the same benchmark.

Starting in October 2026, Google will limit Gemini access for personal accounts: free users get only Flash-Lite, and $4.99/month AI Plus subscribers lose Pro access. Flash and Pro will require AI Pro at $19.99/month or AI Ultra tiers.

OpenAI CEO Sam Altman warns that ascribing religious force or surrender of human judgment to AI models is a "real safety issue." His comments follow reports on Anthropic's outreach to religious leaders and a statement by Pope Leo XIV.