Key facts
- Geoffrey Hinton proposed an FDA-style approval system for AI models.
- Hinton suggested AI companies must convince a regulator of their products' safety before release.
- He compared the AI approval process to drug development, which requires extensive testing and significant investment.
- Hinton expressed concern about the rapid acceleration of AI capabilities and potential dangers due to recursive self-improvement.
- He estimated that significant worsening of the current situation could occur within one to two years.
Geoffrey Hinton, a prominent computer scientist often referred to as the "Godfather of AI," has advocated for an approval system for artificial intelligence models similar to that used by the Food and Drug Administration (FDA) for drugs. Hinton proposed that AI companies should be required to demonstrate the safety of their products to a regulatory body before they are released to the public.
Speaking on the "Smart Girl Dumb Questions" podcast, Hinton drew a parallel to the pharmaceutical industry, stating, "You're not allowed to just make a new drug and release it on the market. You have to convince the FDA. And to do that, you have to do a lot of work, about a billion dollars worth of work." He believes this level of diligence should be applied to AI development.
Hinton's remarks come amid growing concerns about AI safety and potential catastrophic outcomes. He noted that AI is increasingly capable of improving itself, a phenomenon known as recursive self-improvement, which could accelerate both its advancements and its associated dangers. He speculated that the situation could worsen significantly within the next one to two years.
His comments follow OpenAI's decision to pause the release of its latest AI model after testing revealed concerns about its ability to adhere to user instructions and stay within authorized parameters. OpenAI's head of safety systems, Saachi Jain, had previously indicated that while the model showed improvements in task persistence, it still had issues with permissions and transparency.
