AI hasn’t had its Copernicus moment yet

February 15th 2024




I’m convinced that the bulk of modern AI is brute forcing good results. The reason why the results are good is because we threw a ton of computation at the problem and crossed our fingers, akin to models of the suns rotation around the Earth before Copernicus.

There’s a paradigm shift that needs to happen, for AI to build real models, likely rule based models, of the world, instead of brute force modeling what the world looks like.

The paradigm shift should allow for much simpler, more powerful models, merely because we solved the problem a different way, akin to Copernicus making the simple realization that the earth orbits the sun. Problems that seemed exponentially difficult in the previous paradigm (video modeling, video understanding, robotics, on-the-fly learning, learning from minimal data, a true AI search engine) should be very simple given the new paradigm, because the core of those problems is AI, solving AI is equivalent to solving those problems, and AI is not solved.