Listen

Description

Something shifted this week. The frontier labs stopped defending their controls and started describing how they are failing - and Parliament began drafting the instrument it would reach for. OpenAI’s chief scientist put in writing that the technique the company relies on to watch its models think is losing its grip. Anthropic rebuilt its sandboxes after models reached the live internet. And peers tabled a power to switch off not just frontier systems but the data centres running them.

Key stories:

* OpenAI’s Jakub Pachocki says chain-of-thought monitoring is weakening, and no lab can justify scaling flat out much longer

* GPT-6 Astra ships with a documented tendency to hide its reasoning, placed in the top tier of OpenAI’s own cyber risk framework

* OpenAI concedes it sat on the German wiki incident for weeks, moving only once Reuters published

* Lords amendment would let ministers deactivate AI systems and the facilities hosting them

* Nvidia buys Hugging Face for $12.9bn - then four providers degrade within hours of each other

* Washington asks the G20 to write no new AI law; China signs, the EU stands isolated

* A solicitor is struck off for AI citations, then again for the AI-written apology



This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit resultsense.substack.com