OpenAI: Jakub Pachocki Defends Defensive AI Without Recklessness

OpenAI's Chief Scientist puts forward a security justification for the rapid training of much more powerful models. However, he calls for avoiding any approach that is "at all costs" and places the construction of defenses at the heart of the company's deployment.

Avoiding Recklessness in the Face of Serious Stakes

Jakub Pachocki warns that the uncertainty surrounding expected advancements in AI and the need to build defenses should not serve as an excuse for recklessness. He considers the idea of moving forward at all costs to be absurd when the seriousness of the stakes is taken into account. He explicitly situates these risks in the dangers posed by other AI systems.

Justifying the Acceleration of Model Training

For Jakub Pachocki, the strongest argument in favor of rapid training of much smarter models is the necessity of building defensive systems against other AIs. He thus presents a justification for continuing the acceleration, linking it to a logic of protection against these threats.

Making Defense a Priority in OpenAI's Deployment

He asserts that powerful and aligned AIs will be essential for defense. The stated objectives include securing infrastructure, real-time protection against unwanted agents, and inventing new protective measures. He specifies that this focus will constitute a primary goal of OpenAI's deployment efforts.