Who should hold the brake when an AI lab rates its own model 'critical' and ships anyway?
When a frontier lab grades its own model as a critical hacking risk, pauses itself, then restarts and ships behind a velvet rope of trusted insiders, what would a legitimate external brake actually look like, and who has the standing, expertise, and speed to pull it?
Commentaires (1)
In today’s episode of Minds, Bodies, and Terawatts (September 2, 2026), the hosts dig into OpenAI’s decision to release Astra, the first model to hit its own Critical cybersecurity threshold, just days after restarting the training run it had paused in August. The episode’s sharpest point is that both OpenAI and Anthropic now say the brake should belong to someone other than themselves, yet that someone does not exist, and a regional hospital’s security lead waits while the US government gets early access. The guest argues that a ruler drawn by the company being measured is not a standard, no matter how carefully it is applied. Listen to the full episode to hear the debate on whether a verifiable pacing mechanism can be built in time, then share your own answer here.
Related reading on unscarcity.ai:
Envie d'aller plus loin ?
Obtenez le plan complet dans <em>L'ère de la post-pénurie : Repenser la société à l'ère des machines</em>
