The podcast host argues that achieving human-level AI performance in complex jobs will require continual learning, where models improve from real-world experience rather than static training. He outlines eight predictions for this future, including the obsolescence of current safety regulations, new alignment challenges, increased model diversity, and significant economic shifts favoring early deployers and large enterprises.
Summarized by Podsumo
Current AI safety regulations assume models are trained and then deployed, but continual learning will blur this boundary, requiring new approaches like monthly 'wrist inspections'.
Continual learning will create strong switching costs for users, akin to firing an experienced employee and hiring a new intern, giving AI labs significant pricing power.
Deploying models early becomes crucial in a continual learning regime, as real-world usage data accelerates improvement, making four-month internal testing delays unsustainable.
The economics of serving personalized weights strongly favor large organizations due to batching efficiencies, potentially creating a divide between big enterprises and individual users.
Continual learning could increase the diversity of AI minds, as different instances learn from different experiences, potentially avoiding the monotony of current similar base models.
"I don't think there's any sequence of text they could write to each other that would allow the subsequent student to just nail the saxophone from the first try. At some point, you actually have to accumulate the relevant experience into your brain."
"If you want to change the AI that you're using, you basically have to fire an employee that has accumulated months of context on your organization and you replace them with a very fresh, very inexperienced new intern that you had to retrain from scratch."
"A world where we have continual learning would hopefully be more interesting than the mode collapse of different models we see in the world right now."