We’re seeing a general trend toward more efficient LLMs rather than smarter LLMs, so we may be near the limits of what our current models can create, but not how efficiently they can create it.
I’m very interested in these small and efficient models.
It’s less likely we’re at the limit and more likely that the companies producing ai/llm tools are aiming for profitability instead of research/next gen implementation. Consider a current popular known use case for ai: Taking a fast food order. Now current estimates put a single query to chat gpt 4 to cost within the realm of 20-40 cents. Not exactly profitable if you’re trying to sell a $1.50 hamburger. If you can pair down these same mostly capable models so electricity and hardware costs are negligible, suddenly you open yourself up to a world of profitability.
