>>17072777
Maybe first pick an area? Offline LLM is probably negative utility honestly. Realistically, LLMs are only useful because they're trained on a specific area. If it's just some random model then odds are you can't expect it to have the relevant info more than like 1% of the time. It's like having an encyclopedia with CTRL-F, except encyclopedias don't hallucinate.

>>17072777
Start by defining what a scientific "breakthrough" even is in this context.
There's several cases where people discovered entirely new species in their back yard just by being observant and cross-checking with known species. If your LLM is capable of image processing, it could help find best fit matches and if they don't line up then it's a promising candidate for further research.

What you won't do is discover anything paradigm shifting.

>>17072777
Maybe you can download genomes from GenBank and find a way to analyze them that's more thorough than just sequence comparison? Just looking at taxonomy browser, in my opinion there's a lot of errors on that website that could stand to be corrected. I think it's probably not that hard to find a handful of examples of things that need fixing and make some phone calls about it?

>>17072831
if OP discovers anything novel that would be impressive, let's shoot for that

>>17072777
Live in The Year of Our Lord 1543 and publish a revolutionary scientific book using only math and your own recorded observations.