zlacker

This would be the first time ever that an LLM has discovered new knowledge, but the far reach is that the information does appear in the training data?

>>jakela+(OP)
They've been doing it for a while. Gemini has also discovered new math and new algorithms.

There is an entire research field of scientific discovery using LLMs together with sub-disciplines for the various specialization. LLMs routinely discover new things.

>>ants_e+N2
I hadn't heard of that, so I did some searching and the single source for the claim I can find is a Google white paper. That doesn't automatically mean it's false, of course, but it is curious that the only people ostensibly showing LLMs discover new things are the companies offering the LLMs.

>>ants_e+N2
Citation needed, and I call bullshit. Unless you mean that they hallucinate useless algorithms that do not work, which they do.

LLMs do not have an internal model for manipulating mathematical objects. They cannot, by design, come up with new algorithms unless they are very nearly the same as some other algorithm. I'm a computer science researcher and have not heard of a single algorithm created by LLM.

>>jakela+(OP)
https://medium.com/@deshmukhpratik931/the-matrix-multiplicat...

And it's not an accident that significant percentage (40%?) of all papers being published in top journals involve application of AIs.

>>rerdav+nr1
This article is about the same thing I mentioned in a sibling comment. I personally don't find an unreplicated Google white paper to be compelling evidence.

>>jakela+4P1
It's a fast matrix multiply! (A decades-old human problem). What exactly do you need to replicate??! Just count the multiplies, fer goodness sake.

>>rerdav+mq2
> What exactly do you need to replicate??!

The AI coming up with it? When Google claimed their Wizard of Oz show at the Las Vegas Sphere was AI-generated, a ton of VFX artists spoke up to say they'd spent months of human labor working on it. Forgive me for not giving the benefit of the doubt to a company that has a vested interest in making their AI seem more powerful, and a track record of lying to do so.