Oddbean new post about login | logout Asking for your insights. What are the possible (temporary) causes for stalling or reversing the price-decline of llm tokens over the next years. EU and otherwise: - Mono-Oligolopies through regulatory capture - Energy prices increasing - Energy transformer shortages Any other attack vectors or cause Im missing? Any data points to clarify the likelihood? #asknostr #ai #llm
I don’t have any insights on the direct questions but I do think there are too many open source models now. So any price reversal or stalling from big players wouldn’t matter.
Check, yea seems like the right train of thought. However the efficiency of the models seem to have a natural limit with the current architecture though. So for costs I would think chips, transformers and energy are bigger drivers. Algos could be a potential driver if the eager beaver EU states block that out. Then the closed players could get some market power to influence prices. Just looking to get the total picture and see how I’m being stupid.
Open source is winning, seems unlikely
Yea agreed and as long as they flourish the algo’s wont be the determining cost driver but chips, transformers and energy. So market abuse won’t be an issue on the algo front for now. Still is in play for the rest of the drivers though. Any other drivers you see?
more powerful models that use multiple times more expensive hardware and much more energy becoming the common standard seems to be the only realistic possibility.
if some company released a model with 10x higher costs per token (due to the sheer hardware requirements for inference) but if it made gpt4 look like gpt2, you bet many people would gladly pay the extra price.