Hacker News Reader

Comment by frabcus | Hacker News Reader

Yes, the very explicit plan of both OpenAI and Anthropic is to use the not particularly efficient LLMs to automate their own AI engineering. That seems to be going well - on coding front and model tuning front so far. They have more planned.

And then use those to find fundamentally better new architectures for AI - that perhaps are as efficient as the human brain.

It might not work, but I didn't think it'd solve maths problems... So it might work. And if it happens, they'd use the data centres to run millions of instances of it.

It's scary, TBH.

View on HN

I recall them saying they use models to write CUDA kernels and whatnot. Makes sense, and unsurprising that models are good at writing code.

But I think calling this “automating AI research” is misleading. I’m not sure there’s evidence yet that they do creative research work. Even in mathematics, but they are finding counter-examples by intelligent brute-forcing. Not to downplay the results, as they are incredible, but this is one very specific kind of proof and not the most creative type, which arguably requires generalisation.

Quite the gamble.

> but I didn't think it'd solve maths problems

Finding counterexamples is low-hanging fruit, the automation of which isn't shocking.

What about finding the 1st known complex structure over S^6, proving Ehrhart’s volume conjecture, proving a sharp "density" bound on primitive sets conjectured by Erdos >60 years ago?

> Finding counterexamples is low-hanging fruit, the automation of which isn't shocking.

It's not good to be confidently wrong the way you're being.

We've had mathematical problems solved by brute force in the past.

We've then improved that through systems similar to prolog intentionally searching a tree.

Then systems added heuristics for which paths in that tree are likely to be taken.

The LLMs are just using slightly more accurate heuristics for this task.

But the real measure of understanding are tasks that are not so strictly constrained.

If a foundation model company burns billions of tokens to brute force an LLM into finding a new training algorithm that e.g. allows recurrent networks without catastrophic forgetting...

I won't really care that it didn't have a "real measure of understanding". I'll care that it has made an even more dangerous technology, which needs work to make it aligned.

It’s a bird! It’s a plane! It’s…AI skeptics moving the goalposts at light speed!!

Breaking news: a machine can do a tedious repetitive task faster than a human.

Reply on news.ycombinator.com