Hacker News Reader

Comment by ChrisMarshallNY | Hacker News Reader

One of the nice things about LLMs/ML, is that they can pound away at something for a billion cycles, and do exactly the same things that you or I would do.

for _ in 0..<1000000000000 { do_something_complicated() }

View on HN

Isn’t that one of the nice things of computers in general not a feature of llm?

The difference is the complexity of the repeated task

But don't those complexities still boil down to machine level instructions??

Or can/do llms operate outside of a CPU? Thanks

I’m not getting into angels and pinheads, but modern ML has the ability to perform “fuzzy analysis,” and interpret results in a far more flexible manner, than ever before.

They may not be able to match an MIT Ph.D, at analyzing experimental feedback, but they can probably match a lot of research assistants.

It’s like having a billion RAs, all running experiments, and triaging the results. I understand that is how they have made such good progress on medicines, with AI.

> “I have not failed. I've just found 10,000 ways that won't work.”

-Attributed to Thomas Edison

With machine learning, not with llm.

This is true.

The CPU/GPU is to an LLM kinda like axons and dendrites are to the human brain: just a low-level implementation detail. The main crux of an LLM is what happens at a higher level.

The machine-level instructions being executed are just matrix multiplications. Billions of them. The complexity of LLM behavior is emergent from that.

Reply on news.ycombinator.com