If models are requiring larger and larger infrastructure buildouts, does anyone have a clear sense of what users will have to pay in order to make the businesses profitable?
The trouble is that everything is changing so fast that any kind of forecasting is extremely error prone, especially when one forecast builds on another. First you need to guess how much more capable the models are going to get (at things where people will pay more for better performance), in what time frame, then guess what level of demand (inference volume) will exist with that level of capability...
The LLM developers like OpenAI and Anthropic like to tout things like Math Olympiad and Competitive Programming results as signs of progress, but there is no guarantee that they will be equally successful in applying RL to more general areas of commercial value where RL rewards are harder to define.
These companies also like to talk about "scaling laws" is if there was some inevitability about investing more money & compute and getting better results, but this only works until it does not, and they replace one broken "law" with another. Right now it's all about scaling of RL-training and test time compute, but how long will that last, and what type of problems will benefit?
The profit model here seems a bit like the Drake equation for calculating the probability of other intelligent life in our galaxy... it may be possible to define the equation, but the outcome depends on having the right values for all the variables, which are largely unknown.
guesses in the past look better because we tend to pay more attention to the correct ones
I think the surprise was the degree to which ad revenue would eat the world. Maybe it will be the same this time.