There are local LLMs, they’re just less powerful. Sometimes, they do useful things.
The human brain uses around 20W of power. Current models are obviously using orders of magnitude more than that to get substantially worse results. I don’t think power usage and results are going to converge enough before the money people decide AI isn’t going to be profitable.
The power consumption of the brain doesn’t really indicate anything about what we can expend on LLMs… Our brains are not just biological implementation of the stuff done with LLMs.
It gives us an idea of what’s possible in a mechanical universe. It’s possible an artificial human level consciousness and intelligence will use less power than that, or maybe somewhat more, but it’s a baseline that we know exists.
You’re making a lot of assumptions. One of them being that the brain is more efficient in terms of compute per watt compared to our current models. I’m not convinced that’s true. Especially for specialized applications. Even if we brought power usage below 20 watts, the reason we currently use more is because we can, not that each model is becoming more and more bloated.
There are local LLMs, they’re just less powerful. Sometimes, they do useful things.
The human brain uses around 20W of power. Current models are obviously using orders of magnitude more than that to get substantially worse results. I don’t think power usage and results are going to converge enough before the money people decide AI isn’t going to be profitable.
The power consumption of the brain doesn’t really indicate anything about what we can expend on LLMs… Our brains are not just biological implementation of the stuff done with LLMs.
It gives us an idea of what’s possible in a mechanical universe. It’s possible an artificial human level consciousness and intelligence will use less power than that, or maybe somewhat more, but it’s a baseline that we know exists.
Yeah, but a LLM has little to do with a biological brain.
I think Brain-Computer Interfaces (BCIs) will be the real deal.
You’re making a lot of assumptions. One of them being that the brain is more efficient in terms of compute per watt compared to our current models. I’m not convinced that’s true. Especially for specialized applications. Even if we brought power usage below 20 watts, the reason we currently use more is because we can, not that each model is becoming more and more bloated.