Google DeepMind and the Gemini team announced on February 18, 2026 that Lyria 3 is rolling out in beta in the Gemini app. Users can generate 30-second tracks from text, photos, or videos, and all outputs include SynthID watermarking.
Greg Brockman posted that the first steel beams are now up at the Stargate site in Milam County, Texas. The update also names SoftBank and SB Energy as project partners.
Anthropic released Responsible Scaling Policy 3.0, adding a structured Frontier Safety and Security Framework and new roadmap and reporting mechanisms. The update emphasizes explicit commitments to pause or withhold deployment if risk thresholds are exceeded.
OpenAI and Pacific Northwest National Laboratory announced a public-sector AI pilot focused on permit review for clean energy projects. The program starts in Washington state and aims to reduce timelines from multiple years to under one year.
A high-engagement Hacker News thread highlighted Jane Street’s detailed write-up of an ML puzzle where solvers reverse-engineered a hand-constructed PyTorch network and traced it to MD5-style logic.
At CES on January 5, 2026, NVIDIA introduced new AI foundation models under Cosmos and launched DGX Spark and DGX Station. The announcement links synthetic world modeling for robotics with desktop-to-datacenter AI compute.
In a February 20, 2026 (UTC) X post, Cohere said conversations at the India AI Impact Summit focused on responsible frontier AI scaling and language accessibility. The company tied this to Tiny Aya and New Delhi commitments.
In a February 26 statement, Anthropic said it will keep supporting U.S. defense and intelligence deployments but refuses two uses: mass domestic surveillance and fully autonomous weapons.
In a February 27, 2026 joint statement, OpenAI and Microsoft said new funding and partner announcements do not alter their existing partnership framework. They reaffirmed unchanged IP access, revenue-share terms, and Azure exclusivity for stateless OpenAI APIs.
Anthropic posted a policy statement on February 26, 2026 outlining its Department of War engagement and two limits it says it will not remove. The company says it will continue defense support but rejects mass domestic surveillance and fully autonomous weapons at current reliability levels.
OpenAI and Pacific Northwest National Laboratory introduced DraftNEPABench to evaluate AI coding agents on NEPA drafting tasks. In expert assessments, agents showed potential to reduce drafting time by 1 to 5 hours per subsection, or up to roughly 15%.
A high-scoring Hacker News post spotlights FDM-1, a video-native computer action model trained on an 11-million-hour dataset. The release emphasizes automatic action labeling with IDM and large-scale forking-VM evaluation for long-horizon interaction tasks.