Ilya Sutskever Just Found Something New In AI and could lead to Superintelligence

[John Clark]

Recently about the only things humans can still claim to be better at than computers at are manual dexterity and, what Ilya Sutskever kept talking about, the ability to make good generalizations from very little data; for example it only takes a teenager a few hours to learn how to drive a car, but it takes a computer much longer and much more data to do the same thing. Sutskever is one of the most brilliant AI researchers in the world and he was the chief scientist at OpenAI when GPT-4o was developed. When he left OpenAI after a well-publicized break up, he started his own company called "Safe Superintelligence". He said his new company would concentrate entirely on creating a super intelligence and would not release an AI model to the public until they had achieved that goal. Sutskever complained that other AI companies were just scaling up old ideas, like the transformer recipe, and he thought it was time to get back to basic research and develop some new ideas. He said his  new company would concentrate entirely on fundamental research, finding a new idea, validating it on a small system, and only then use enormous computation to scale it up by a factor of 10 . 

Sutskever publicly admitted there was a good chance his new company would never make any money, nevertheless so great was his reputation that he managed to attract about $1 billion. That was over two years ago and since then very little has been heard from the company until just a few weeks ago when Sutskever told some people at Nvidia that they had found something new that needed scaling up and invited them to take a look at what they had discovered. A few days later Nvidia announced they were making a "substantial" investment in Safe Superintelligence which would allow the company to have access to 10 times as much computational capacity as they previously had. 

It's pretty clear to me that Sutskever believes that his company has found the next major step beyond transformers. We'll just have to wait and see if he's correct.

Ilya Sutskever Just Found Something New In AI

John K Clark    See what's on my list at  Extropolis

eew