Deep neural networks (a narrow subset of “AI” algorithms) are here to stay, and there are massive gains in other hard sciences that have been made with the nature of how they make predictions based on a corpus of training data.
Conversely, in my opinion, there is most definitely an economic bubble regarding LLM (a narrow subset of DNNs) purveyors due to the revolving investments vendors make into each other’s businesses to prop up the illusion that “line go brrrrr forever”.
Both things can be, and I argue are, true simultaneously.
Edit: Downvote me all you want. I work with DNNs for a living; it doesn’t take an “insider” or a financial analyst analyzing EBIDTAs to understand the frailty of the house of cards that the “LLM industry” sits upon right now. I say this as someone who self-hosts and uses llama.cpp and Open-WebUI nearly every day.
Platform and systems Architect for a german managed Services provider here.
I second that. Currently, I am working on replacing external AI providers with 150W AMD hardware. Most of our teams used LLMs for stuff where classical ML is far more reliable and faster.
The few places where LLMs are needed, we still can do locally.
So, instead of spending millions in some tokens that take gigawatt of Power, we estimate that our 5k employees will need a total power of 5kw across Europe. We have 100x the solar distributed across our locations.
We conflate LLM or Diffusion with AI here, because most had never contact with other kinds of AI.
Deep neural networks (a narrow subset of “AI” algorithms) are here to stay, and there are massive gains in other hard sciences that have been made with the nature of how they make predictions based on a corpus of training data.
Conversely, in my opinion, there is most definitely an economic bubble regarding LLM (a narrow subset of DNNs) purveyors due to the revolving investments vendors make into each other’s businesses to prop up the illusion that “line go brrrrr forever”.
Both things can be, and I argue are, true simultaneously.
Edit: Downvote me all you want. I work with DNNs for a living; it doesn’t take an “insider” or a financial analyst analyzing EBIDTAs to understand the frailty of the house of cards that the “LLM industry” sits upon right now. I say this as someone who self-hosts and uses llama.cpp and Open-WebUI nearly every day.
Platform and systems Architect for a german managed Services provider here.
I second that. Currently, I am working on replacing external AI providers with 150W AMD hardware. Most of our teams used LLMs for stuff where classical ML is far more reliable and faster.
The few places where LLMs are needed, we still can do locally.
So, instead of spending millions in some tokens that take gigawatt of Power, we estimate that our 5k employees will need a total power of 5kw across Europe. We have 100x the solar distributed across our locations.
We conflate LLM or Diffusion with AI here, because most had never contact with other kinds of AI.