I'm bearish on LLMs not because they're not helpful; they are. I'm doing theoretical physics, and gpt-4 is useful multiple times daily. It didn't replace Google

I'm bearish on LLMs not because they're not helpful; they are. I'm doing theoretical physics, and gpt-4 is useful multiple times daily. It didn't replace Google or something, but it is a useful additional tool. It just does not feel like "90B valuation without profit" useful or "multiple unicorns doing more or less the same" useful. Of course, gpt-5/6/7 can become more valuable to end-users, but that's the second reason I'm bearish. LLMs are powered by exponential growth, and no exponential growth is infinite. We are already using a significant part of all existing data, and going up more than 1-2 orders of magnitude in either data or compute feels unlikely unless there is some breakthrough. Breakthroughs are hard to predict, but they are generally rare, and likely there won't be one soon.