@janl

There are some weird incentives at the moment. Vendors have spent around $3T on under $100 B of revenue so, on average, LLM compute is subsidised by a factor of at least 30x. If an LLM-based approach reaches 10% of the efficiency of another approach, it’s cheaper for the end user. More importantly, it’s cheaper for the LLM reseller than selling the more efficient solution would be.