The top things worth knowing about in AI today.
Amazon is winding down its flagship Nova Premier and Omni models along with the Reel video and Canvas image generators, and has cut jobs in its AGI group. Nova 2 Lite, Nova 2 Sonic and the Nova Forge custom-model service stay supported. Resources shift to a new frontier model effort led by researcher Pieter Abbeel, expected at re:Invent later this year. If you built on Premier or Omni, start planning the migration now.
Read moreMore than 1,100 employees at OpenAI, Anthropic, Google and Meta signed an open letter asking the US government to back an international effort to build tools that 'deliberately pace the frontier of automated AI development'. Signatories include OpenAI chief scientist Jakub Pachocki, Anthropic co-founders Jack Clark and Jared Kaplan, and Meta chief scientist Shengjia Zhao. The concern is AI systems that improve AI systems faster than anyone can check the result. The pacing argument is now coming from inside the labs, not just from outside critics.
Read moreMoonshot AI published open weights for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with a 1-million-token context window. The download runs roughly 1.4 terabytes using MXFP4 quantisation, making it the largest open-weight release to date. Moonshot places it behind Claude Fable 5 and GPT-5.6 Sol overall, but ahead of Claude Opus 4.8 and GPT-5.5 on coding and agent benchmarks. Near-frontier capability is now something an organisation can host on its own infrastructure.
Read moreNvidia, Microsoft, IBM, Cloudflare, Hugging Face, Red Hat, CrowdStrike and the Linux Foundation launched the Open Secure AI Alliance to build shared tooling for agent identity, permissions, isolation, logging and model scanning. It follows this month's Hugging Face intrusion, in which autonomous agents reached production infrastructure. Nvidia contributed its NOOA agent framework under Apache 2.0. No governance structure or delivery schedule has been published, so read it as a statement of intent rather than a working standard.
Read moreShared Claude conversations and Artifacts were found in Google search results, including medical reports, clinical trial data with names attached, internal company documents, API keys and login credentials. Anthropic's robots.txt told crawlers to stay away, but the pages carried no noindex tag, so a link from anywhere else on the web was enough to get them indexed. The results disappeared from search within about a day. Worth auditing any share links your team has generated, on any AI product.
Read moreKyndryl surveyed 1,100 senior leaders across eight countries and found 57 percent have AI deployed broadly or embedded in core processes, while only 11 percent have met both of their top two AI goals. The share saying their workforce is fully prepared fell to 23 percent from 29 percent a year earlier, and just 19 percent of workers say they feel confident using AI tools. Only a third have training programmes fully in place. The bottleneck is people and process, not model quality.
Read moreMicrosoft released MAI-Cyber-1-Flash, a 137-billion-parameter sparse model with 5 billion active parameters, built to run inside its MDASH vulnerability-scanning system. Microsoft says it handles up to 90 percent of scanning work and lifts MDASH to 96 percent on the CyberGym benchmark at roughly half the cost of its previous model mix. Access is restricted to an Azure AI Foundry private preview with approval, since the same capability serves attackers. A small specialised model beating general frontier models on cost is the pattern worth watching.
Read moreCognizant became one of a small group of Global Premier Partners in Anthropic's Claude Partner Network and is embedding Claude across its own engineering and industry platforms. More than 30,000 associates have completed Claude training against a headcount above 350,000, with a certification pipeline targeting 40,000 credentialed professionals. Client deployments already run in manufacturing, life sciences and insurance. Vendor certification is turning into a workforce credential rather than a procurement detail.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.