[sxp]

To add some math to the discussion:

- A human uses between 100W (naked human eating 2000kcal/day) to 10kW (first-world per capita energy consumption).

- Frontier models need something like 1-10 MW-years to train.

- Inference requires .1-1kW computers.

So it takes thousands of human-years to train a single model, but they run at around the same wall clock power consumption as a human. Depending on your personal opinion, they are also .1-1000x as a productive as the median human in how much useful work (or slop) they can produce per unit time.

[ncr100]

The math is simpler, 1 human is irreplaceable by AI.

Therefore its value is infinite. Therefore Altman's hypothesis is toilet paper thin.

[thenthenthen]

I remember when toilet paper was like ddr5

[cheeseblubber]

The human brain also is a product of billions of years of evolution. We branched off from our common ancestor 7-9 million years ago. We encode quite a lot of structure and information that is essential for intelligence. The starting point of just our life time of training is incomplete.

If you calculate 100W * 7 million years * 365 = 255,500MW to train.

[arcticbull]

If you really want to go down that path then AI's are the product of human ingenuity and labor so you have to amortize all of that into AI training. Then numbers become pretty meaningless very quickly. That sand didn't up and start thinking on its own you know.

[grogenaut]

That's the NRE of getting to where we are and having these llms