The U.S. Commerce Department said on Thursday it plans to provide $389 mln in R&D funding to four companies developing power semiconductor devices and critical materials to advance commercial fusion energy and related technologies. As a condition of

OpenAI said this summer it granted METR and Redwood Research three weeks of deep, embedded access to its internal monitoring systems in a pilot assessment. The teams’ findings helped strengthen safety controls during system scaling and will underpin future research cooperation with METR and other third‑party evaluators.

1. Token Consumption: Steinberger estimates he spent approximately $10 million worth of tokens last month running AI computing tasks. However, he emphasizes that not all token consumption generates real-world value. 2. Current Infrastructure Status: Steinberger believes a major challenge for many is that their infrastructure hasn't truly caught up, arguing why workloads are still running on local computers instead of being entirely in the cloud. 3. Low-Quality Contributions from Open Source and AI: Facing the phenomenon of some open-source project maintainers ceasing to accept pull requests due to the overwhelming amount of low-quality AI-generated code, Steinberger believes code review mechanisms need adjustment. Developers should explain their thought process and reasoning behind code modifications, rather than simply submitting the final result. 4. The Future Number of AI Agents: Steinberger doesn't believe everyone will be running hundreds of AI agents in the future, but he anticipates people will have intelligent systems capable of operating around the clock, thinking, and providing assistance, even predicting future user needs and providing answers before users even ask questions. 5. Trust in AI Agents: Steinberger states that determining which data to provide to an agent is a challenge. The level of user trust in the underlying model is crucial because many of the model's behaviors are not entirely constrained by the harness. As the model's capabilities improve, some of OpenClaw's early design decisions can be adjusted accordingly. 6. AI Agent Permissions and Security: Steinberger advises against granting all permissions to a single agent. He suggests introducing another agent to approve permissions, allowing agents to negotiate and assess the necessity of tasks and the existence of safer alternatives. This two-layer mechanism enhances user trust in AI's autonomous task execution.