Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

>They're probably right that building AGI will require a ton of computational power, and that it will be very expensive

Eh.

Humans have about 1.5 * 10 ^ 14 synapses (i.e connections between neurons). Assume all the synapses are firing (highly unlikely to be the case in reality), and the average firing speed is 0.05ms (there are chemical synapses that are much slower, but we take the fastest speed of the electrical synapses).

Assume that each synapse is essntially a signal that gets attenuated somehow in transmission. I.e value times a fractional weight, which really is a floating point operation. That gives us (1.5 * 10 ^14)/(0.0005)/(10 ^ 12)) = 300000 TFLOPS

Nvidia 4090 is capable of 1300 Tflop of fp8. So for comparable compute, we need 230 4090s, which is about $345k. So with everything else on board, you are looking at $500k, which is comparatively not that much money, and thats consumer pricing.

The biggest expense like you said is paying salaries of people who are gonna figure out the right software to put on those 4090s. I just hope that most of them aren't working on LLMs.



Inference compute costs and training compute costs aren’t the same. Training costs are an order of magnitude higher.


Training will be significantly cheaper and take less time once we have the correct software for the 4090s

Right now, the idea is that every time you build an LLM, you start the training from scratch, because thats all we know how to do.

Human like AI will most definitely not be trained like that. Humans can look at a piece of information once or twice and remember it.

Just like the attention paper, at some point someone will publish a paper that describes a methodology of feeding one piece of data back through the network only a few times to fully train it.


LLMs are just training on massive amounts of data in order to find the right software. No human can program these machines to do the complicated tasks that humans can do. Rather we search for them with Gradient based methods using data




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: