These last couple of days, I've been trying to deepen my understanding of CRDTs. There aren't too many papers on CRDTs, so you can get caught up quickly by diligently reading the papers. Though I haven't read them all, I was surprised by my pace of understanding the papers with GPT-4. I created a GPT that helped me read papers called PaperMatey. (👈 Click to use it yourself)
I was able to ask it for interpretation of the logic equations, definitions on jargon, concrete examples to passages in the paper, all to help me understand what's being conveyed.
Furthermore, I'm able to ask about implications of passages in papers or certain concepts. I can also ask it to mashup two ideas from two different papers and ask it to hypothesize about its implications. It's not as good at this as I thought it would be, due to success in style transfer tasks. However, it's currently serviceable and it's a hint of what may be to come.
Disappointingly, it doesn't do a good job of using diagrams to explain things so far. I asked it to draw me a diagram explaining the construction of a merkle search tree. While the overall structure is there, it lacks coherency.
However, if I ask it to draw explanatory diagrams by writing python code using Matplotlib, it was able to do much better. This still has downsides, as it does sometimes hallucinate (perhaps 10% of the time?), and you do have to play spot-the-error in the illustrative diagram, just as you do in code. This can be challenging if you're still learning about a new concept. For now, I think of it as a pop quiz that teachers like to give out on occasion.
Here, for another paper, I asked it to show me a diagram of a join-semilattice with the join-irreducible nodes highlighted. As you can see, it's pretty much useless other than for some nice abstract wall-art. That said, I wish math and computer science diagrams had more of this aesthetic, without chart junk, of course.
Overall, this lets me read about a paper a day in this space, and the ability to ask questions to pontificate and clarify. I think it gives me a way to understand deeper than I would have otherwise, even with the limitations of GPT-4.