Genie 4: DeepMind's world model adds native audio and real-time steering
Google DeepMind introduced Genie 4, a real-time interactive world model that now generates audio natively — ambient sound and character dialogue — alongside the visuals. Users can change a running simulation with natural-language prompts (weather, camera moves, new objects or people), and the model simulates aerodynamics and collisions. It is a research preview for trusted testers; access is coming to Google AI Ultra subscribers via Project Genie and to early-access partners in the Gemini API and AI Studio. DeepMind says detail and object permanence still drift over long sessions, and it gave no resolution, frame-rate or session-length figures.
Sources & depth
Meta locks in AMD and NVIDIA capacity, with both chipmakers investing in its equity raise
Meta announced expanded infrastructure deals with both major GPU vendors on the same day. AMD will invest $1 billion in Meta's equity offering; Meta is testing initial AMD Helios rack-scale systems, will deploy a custom Instinct GPU, and is co-developing the next EPYC generation (Verano) and informing the MI500 series. NVIDIA will invest $3 billion in the same offering, with Meta expanding Vera Rubin and Vera CPU deployments plus Spectrum-X and BlueField-4 networking, and the two will jointly optimize Meta's Muse models. Neither release gives deployment sizes or dates beyond "gigawatt scale" (AMD) and "near future" for broader Helios rollout.
Sources & depth
Vera Rubin NVL72 is generally available on Oracle Cloud
NVIDIA and Oracle said Vera Rubin NVL72 racks (72 GPUs per NVLink domain) and Vera CPUs are now generally available on OCI, with Oracle planning multi-gigawatt Superclusters built on them. The HGX Rubin NVL8 (eight GPUs, two 64-core Xeon 6 CPUs, 4 TB DDR5) is planned but not yet available. NVIDIA says OpenAI trained and served GPT-6 Astra on an OCI Supercluster of more than 100,000 Blackwell GPUs in Abilene, and is testing Vera CPUs for agent sandbox workloads. Rubin moving from announcement to rentable hardware is the news for anyone pricing inference capacity.
Sources & depth
Anthropic offers US governments 40% off Claude
Anthropic expanded its partnership with the General Services Administration: federal, state, local, tribal and territorial governments get 40% off Claude Enterprise, Claude for Government and the API, with no minimum-spend commitment, through GSA's Multiple Award Schedule now and via cloud providers in coming months. Anthropic estimates about $1 billion in savings and says roughly 500 government organizations already use Claude. Claude for Government is FedRAMP High, as is Claude on Bedrock and Vertex AI.
Sources & depth
Stanford professors say Anthropic trained Claude on pirated copies of their books
Six Stanford professors say Anthropic used pirated copies of their books to train Claude. Two of them, Tobias Wolff and Chang-Rae Lee, opted out of the $1.5 billion Bartz settlement and filed Cruz v. Anthropic in May 2026, alleging the books came from shadow libraries such as LibGen and PiLiMi and seeking up to $150,000 per work. The earlier Bartz ruling held that training on lawfully acquired books was fair use but that keeping pirated copies in a central library was not. The Stanford Daily report includes no response from Anthropic.
Sources & depth
Liquid AI releases LFM3, a family of ~2B open-weight models
Liquid AI released six open-weight LFM3 models around 2B parameters covering text, vision, audio and omni, plus two decision models. The 1.7B text model fits in about 1 GB at 4-bit with a 256K context. Liquid's own benchmarks claim it beats its 2.6B predecessor with 35% fewer parameters and beats Qwen3.5-9B on WebArena-Verified and AutomationBench, with a 128K KV cache of about 1.6 GB versus 4.3 GB for Qwen3.5-4B; larger models still win on agentic coding and knowledge. These figures are company-reported and unverified. Weights are on Hugging Face; the license is not stated in the announcement.
Sources & depth