The latest verified reporting and analysis across ai.
Beam activates 23 billion parameters at a time and targets coding and agentic work, but its weights, model card and technical report are still promised for later in October.
Read story
A task force led by U.S. intelligence chief Jay Clayton will examine AI risks, opportunities and the federal government’s role. Its short mandate raises an immediate test: whether it can turn broad coordination into verifiable policy.
Read story
The Eighth Circuit has paused enforcement of a first-of-its-kind state law aimed at AI tools that generate realistic intimate images of identifiable people. The injunction is temporary and does not settle the constitutional dispute.
Read story
Google's new frontier model is entering a controlled rollout to trusted cyber defenders before developers or consumers. Its capabilities are ambitious, but public evidence remains dominated by vendor benchmarks and internal deployments.
Read story
Dots extend ChatGPT from a session-based assistant into a persistent cloud agent that can work across projects, browsers and connected apps. The launch is consequential, but its safeguards and real-world reliability still need independent evidence.
Read story
Major technology companies have signed a nonbinding safety framework as a new executive order tells U.S. agencies to replace ‘artificial intelligence’ with ‘super intelligence.’ The two moves signal political urgency, but neither creates an enforceable technical standard.
Read story
The all-stock acquisition would bring Fei-Fei Li and World Labs’ 3D world-model research inside AMD. The strategic logic reaches beyond one product, but closing still requires regulatory approval and commercial payoff remains unproven.
Read story
Nvidia's Open Agent Safety Platform combines an Apache-licensed sandbox runtime with optional out-of-band hardware monitoring. It could make agent permissions more enforceable, but vendor claims about stopping past breaches remain unproven.
Read story
Stanford and Caltech researchers connected GPT-6 Astra to a reusable skill library and persistent spatial map, allowing a Unitree G1 to tidy an unfamiliar kitchen and retrieve an out-of-view object. The demonstration is promising, but it remains a research prototype with substantial setup, latency and hardware constraints.
Read story