Thomas Larsen and Daniel Kokotajlo discuss reports that OpenAI's Astra uses recurrent depth and may be harder to monitor through its chain of thought than earlier models. They treat that direction as concerning because safety teams could lose a useful view into how a model reasons, while acknowledging that one guest had not yet read the newly released system card and was relying on reporting and a public clarification.
The guests compare immediate pacing with a later pause near the edge of controllability. Their proposals range from limiting the share of compute used for new training runs to the AI 2040 plan, which combines an initial governance pause, a regulated period of continued development and a later pause while alignment and control research catches up.
They also examine a proposed ban on superintelligence, corporate liability and ordinary business incentives. They argue that liability and market pressure may work for visible failures but could break down if a strategically deceptive system behaves safely until it gains more authority.
Watch on YouTube
