Here’s my bet: once the algorithm is cracked, the one that will “solve” AI, the hardware will need to change form to fit the algorithm. Building bigger and better matrix multiplication chips will only go so far. After a certain point it’s a waste of money, because once the algorithm is determined, bespoke hardware for it will end up being way cheaper and cost effective, and will yield more efficient processing.

The industry is high on the idea that they can make all of this happen by doing more of what they already do, only with more money processing compute hardware and people. It’s because they got this great success with ChatGPT, and the people in power, the ones with all the money, want to keep it going, and throwing money at problems is what they’re used to.