Zvi Mowshowitz's weekly roundup: cyberattacks, a leaked draft bill and interpretability

On 22 October, Zvi Mowshowitz published issue 191 of his weekly AI roundup. He argues that the day was probably the worst in cybersecurity history. He cites a critical PHP flaw, announced a day before the fix because of a leak, and other incidents.

He also covers the leaked draft bill that Anthropic had been shopping around Congress. In his reading it is similar to the FRONTIER Act but weaker. He thinks Anthropic's first explanations made things look worse, and that Dario Amodei's later response was better, though the situation remains troubling.

On interpretability, he relays a new claim: neural networks are cleaner than thought, and individual neurons mostly carry a single interpretable meaning. He calls it good news, but says early reactions overstated it.

These are Zvi's assessments, not facts we have checked.