Superintelligence Won't Happen Absent A New Approach To Gen AI
The podcast discussion with AI researcher and SSI co-founder Ilya Sutskever regarding the field’s shift from the recent “age of scaling” back to a necessary “age of research.”
Sutskever observes a persistent disconnect between current model capabilities, which score highly on evaluations, and their actual economic impact, a problem he attributes to models’ insufficient generalization when compared to human learners.
Sutskever suggests that current reliance on reinforcement learning (RL) creates systems similar to over-trained students who lack the foundational understanding needed for practical, robust application, necessitating major breakthroughs in areas like the use of a value function and continual learning.
Looking forward, Sutskever predicts that superintelligence, defined as a mind capable of rapid, ongoing learning, could emerge within 5 to 20 years, making the current moment crucial for developing superior training approaches.
Recall that Sutskever’s firm has raised $3 billion at a $32 billion valuation to straight shot superintelligence. No products will be rolled out between now and then. If that’s not a sign of an AI bubble, nothing is.