By now, it’s clear that the only way the tech industry can justify the cost of AI is if it replaces vast swaths of the human workforce with machines that run 24/7.
The bad news is that this situation has created a world-historic financial market that, by some metrics, is looking worse than the run-up to the Great Depression. The good news is that this future of an AI takeover is looking increasingly unlikely, at least at the industry’s current pace, a fact which is now dawning on some of the biggest rubes and dupes in the corporate world.
According to a new survey from “Big Four” accounting firm KPMG, a significant number of corporate executives are reeling from sticker shock over new usage-based AI pricing schemes. Though enterprises could once count on AI companies to subsidize the price of large language models via flat-rate contracts, that’s no longer a given, as the rising cost of computational power forces the entire tech sector into a defensive posture.
A common cause of this kind of problem is lots of people simply default to whatever the “best” model is and throw every problem at it. The best model can handle those problems, sure, because it’s the best model. But it’s also the most expensive model.
A better strategy is a multi-model agent that breaks tasks down into smaller sub-tasks and then uses the minimal model that can manage each of those tasks. The high-level program architecture can be figured out by Fable 5 or whatever, but then each of the functions and classes can be written by a cheap local model like one of the Qwens, for example. An AI that’s been told to find correlations in a large corpus of documents could make use of a smaller model to analyze each of the documents to filter out the ones least likely to be useful and only actually “read” the most promising ones. And so forth.
This is something that’s an area of very active development. The harness is going to be just as important as the model, if not moreso.