Today, we are going to discuss a highly controversial topic: can we reimagine computers from the ground up for AI' What got me thinking about this idea was the announcement of Unconventional AI which raised a considerable amount of money of work precisesly on this problem. Artificial intelligence today thrives on brute-force computation. Breakthroughs in deep learning, from chatty large language models to expert game-playing agents, have largely come from one approach: throw more matrix-math and more GPUs at the problem. This strategy of scaling models and data on ever-larger GPU clusters has driven impressive gains. However, it's also driving us straight into a wall of diminishing returns ' primarily in the form of exorbitant energy use and hardware costs. Recent events underscore this concern: a new startup called Unconventional AI made headlines by raising an unprecedented $475 million seed round to develop radically new computing hardware for AI. The motivation' Today's AI...
learn more
