Now Playing
Interpretability Breakthrough or Overhype? Microsoft Exits UK & Congress Acts
Audio is available on Spreaker — see link below.
AI safety's biggest interpretability claim in years is under fire, Microsoft pulls out of a UK data centre deal, and Congress advances three distinct AI regulation bills in a single week. This is everything you need to know from the last 24 hours in artificial intelligence.
Audio is available on Spreaker — see link below.
Nick Cammarata just made one of the most consequential claims in AI safety in years, and the field isn't sure whether to celebrate or push back. Cammarata, a researcher at Goodfire, published findings suggesting that large language models are far more interpretable than the AI safety community had assumed.
The important distinction is that this is a blog post, not a peer-reviewed study. Researchers at Stanford and DeepMind have already raised questions about the rigor of the findings.
Away from interpretability, Microsoft has pulled out of a data centre partnership with Nscale at a facility in Loughton, citing power grid constraints that would delay connection until the twenty-thirties. This follows OpenAI's earlier exit from Stargate UK plans.
OpenAI is signaling a serious move into physical AI. Sam Altman has teased a humanoid robot demo, and the company has posted twenty-seven robotics job openings focused on motion control and manufacturing.
On the regulatory side, the Senate Commerce Committee advanced several children's online safety bills, including Ted Cruz's CHATBOT Act, which would require parental consent and family accounts for AI chatbot services. Separately, Representative Greg Casar introduced the Stop AI Price Gouging Act, targeting AI systems that set individualized prices using personal data.
Anthropic committed five percent of its shares to a public benefit fund for societal impact projects, with an option to add another two percent for every fifty percent increase in valuation. The governance structure hasn't been defined yet.
The near-term watchpoints are clear. Cammarata's interpretability findings need independent replication before the safety community can responsibly update its assumptions.
Chapter summary auto-generated from the verified script. Listen to the full episode for the complete content.