Block introduced Dessa, a full‑duplex speech model trained entirely on human speech that the company says can run on a single NVIDIA GPU rack or consumer machines such as a MacBook M4 Max. Dessa is not publicly available; Block has invited safety researchers to test it.
OpenAI announced on October 21, 2026 that its frontier models and Codex will be integrated into Google Cloud’s Model Garden, letting customers run agentic workflows inside Google Cloud with its governance and infrastructure. General availability is planned in the coming weeks.
IBM’s Granite team published updates on October 21, 2026: a new granite-speech-5.0-2b and revisions to granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. The changes target higher ASR accuracy, faster on-device transcription, and improved speaker attribution.
NVIDIA presents Kumo Relational, a pre-trained relational foundation model that predicts directly on databases without task-specific training or feature engineering. The company reports it beats supervised baselines across 41 tasks and scales to databases with 500B+ rows.
Spotify reported 788 million monthly users and nearly €5 billion in third-quarter revenue. Co-CEO Gustav Söderström says the company is positioning Spotify to be the media service that AI assistants access for music and audio needs.
OpenAI announced the Decisions API is generally available (Oct 22, 2026). Built with GPT‑6 Luna, OpenAI says it can make bounded-answer decisions up to 10x faster than using GPT‑6 Luna through the Responses API, with GPT‑Live, caching, and US/EU residency.
Anthropic circulated a draft bill that would block AI model releases judged to pose a “foreseeable, imminent” risk of more than 50 deaths or a billion dollars in property damage. Independent evaluators registered with the U.S. Commerce Department would alert authorities and prompt a D.C. judge revie
FairStress evaluates demographic fairness beyond final choices by testing how models respond when decisions are challenged. The benchmark includes 300 scenarios across 29 domains and a 50K FairStress-Core subset, and reports asymmetric sensitivity to persuasive prompts across many LLMs.
NVIDIA presents Cosmos3-H-Dreams, a generative simulator trained on clinical cholecystectomy recordings and designed for live kinematic control. The system runs interactively on a single NVIDIA RTX PRO 6000 GPU and includes teacher-to-student distillation for low-latency rollouts.
Vercel published usage data showing TypeSafe AI’s Jev was adopted by 35% of paid teams within two weeks of its limited release on September 15, nearly doubling the reach of Sonnet 5.5. Vercel’s report also highlights rising use of open source models and growing spend on frontier labs.