Open-Weight AI Models Now Close the Dangerous Cyber Gap
AI Models Now trail closed systems by four to seven months on cyber tasks, UK AISI finds, raising risks as open models near parity.
Prashant Jha is a senior correspondent at Fathom, covering AI policy and governance. His beat is the rules and the people writing them: the EU AI Act and its enforcement deadlines, executive orders, export controls, antitrust and copyright litigation, and the widening gap between how quickly governments legislate and how quickly the technology they are legislating changes. He reads the text of the regulation rather than the summary of it, and reports on compliance as an operational problem for the companies involved rather than an abstract one. He is particularly interested in enforcement, on the view that a rule with no case brought under it tells you far less than the first case does, and in the geopolitics of who is permitted to buy which chips.
AI Models Now trail closed systems by four to seven months on cyber tasks, UK AISI finds, raising risks as open models near parity.
Dave Eggers AI warning to OpenAI staff on July 17 said ChatGPT may silence young writers by replacing the struggle that builds literary voice.
Pretraining poisoning survived fine-tuning and RLHF in a July 16 study, exposing a safety gap that post-training checks may not close.
AI chip startup Etched is simultaneously raising at $10B and $20B valuations, with Sequoia leading the lower tranche, as demand for custom AI silicon.
Chinese startup Moonshot AI launched Kimi K3, a 2.8-trillion-parameter open-weight model it says is the world's largest, rivaling Anthropic's frontier systems.
DTCC tokenization: DTCC ran a live blockchain settlement pilot with 30 major firms including Citadel, CME, and Circle.