The most depressing thing about the current bubble is that so many of the flagship ‘look how amazing AI is!’ things are things where LLMs are a pretty poor fit and other techniques work much better, but the baseline for comparison uses well under 10% of the compute, often under 1%. If you threw 10% of the memory and processor capacity at the non-LLM approach, it would work much better than the LLM approach. Yet it’s easy for OpenAI or Anthropic to burn $10M on a tech demo and incredibly hard to get anyone to spend $1M on a good solution to the same problem.

As long as money is flowing to these companies, they’re starving interesting research of resources because there’s no incentive to build efficient systems when a company is willing to set fire to investors’ money to run the horribly inefficient version for you.