Seems like the Chinese open source models are mostly keeping pace to be ~80%+ as good as the latest frontier models. If that continues I don’t see how the frontier labs could get these premium prices.
Aren't those models "skimming" off the frontier models? (forget the term, but basically training themselves by prompting Claude / OpenAI etc)
If that's the case then once the data center / GPU money train runs out, the whole thing craps out
I've speculated to myself alot that there will eventually be parity between compute price and human price. However, most things should also come down in price, leaving the best humans with much,uch more wealth and power.
We just need one open weights model that matches current frontier models and people will start building racks with Apple Studios at home and cancel all their subscriptions.
It doesn’t crap out. It causes more AI labs to work on efficiency, which the US labs currently have little interest in doing because they want everyone to depend on their giant trillion dollar data centers. We will get more efficient AI and our hardware back when this bubble bursts, hopefully sooner than later.
Well OK but "bubble bursting" means a massive change in how we should view AI. This means that companies that have fired employees to replace them with AI systems are going to get see cost/performance ratios of these systems skyrocket. It means "vibe coding" costs $200/mo to do anything useful.
The entire industry can't just pivot on a dime to "efficiency" while maintaining this level of performance of the models.