813 karma · joined December 9, 2014
It is a popular science book which catches the vibe of mathematical logic in an excellent way. It is not a textbook, nor a piece of research. It's all vibes, but high-quality vibes. If you are in the right headspace it can be really inspiring!
Proofs stretching thousands of pages are split into lemmas, grouped into sub theories.
What I haven’t seen agents do yet is to develop new ideas for entire such theories. I have usually seen them bite into some existing idea and grinding out related results. But I am less sure than ever that they won’t!
I am not sure why it works so well here. I have vague recollections of news articles about it being ignored in the past. I guess they just struck down hard on those. Also, I imagine if you put yourself on that list you would be really grumpy if someone called to try to sell you something.
First of all, it will convince itself that a task is just too hard and find excuses not to try hard enough.
Then when it struggles on something it loves to write down confusing notes about things which it believes cases the problem. Then next iteration it reads its own note, misinterprets it and completely messes up by trying to avoid some imaginary problem.
On the other hand, I just checked, and the domain ratemyflake.com is available. I think we should set it up, so that we can rate each other's nixings!
Now, this paper will undergo peer review. And thus, when published, we will trust the result as well as any other published result in mathematics. However, while humans are not infallible, LLMs have a tendency to spit out confident stuff which looks correct. Thus I worry when it is used this way to produce proofs. Peer review is not perfect, and may not be tuned to catch LLM’s style of errors.
There is a solution to this, which is to formalise the result and get it machine verified. And with LLMs, I daresay this is going to be best practice moving forward.
LLMs exist and are extremely useful in academia. I am currently riding a productivity wave, where I can spit out formalised proofs – not quite at the pace at which I can think of the theorems to prove, but almost. The thing braking my progress is that I have to build all the existing theory too, not just my own new stuff.
But for students, who do not yet know anything, it is a danger. From observing my students, it seems to take a certain amount of willpower to do the exercises without letting ChatGPT give them (or "explain them") the answer. And reading an explanation is not the same thing as working it out yourself. You just don't learn doing it, by not doing it!
Some students have this will power, some students don't. And it is also situational. Some students have stuff going (work or other factors) which makes it really hard too.
Yes, they can of course keep asking ChatGPT later in life, but they will not have the fundamental understanding which is required to even use LLMs for the advanced stuff. ChatGPT cannot explain you advanced concepts if you don't understand the fundamentals!
So, to test the students' understanding, we are back to having on-site exams, with the old-school supervision. And more students will fail, because there is now a way to solve exercises without learning.
For example, in an intuitionistic setting there is a difference between a set being non-empty and a set having an element.
But there is no limit to where the underscores go: the name _? Would give you x ? postfix. To define if-statements one defines the name if_then_else_.
I am really excited if LLMs means more of mathematics will be formalised. Once we have all of mathematics formalised, I think we will start to see real great progress!
WoW and Runescape exist. But MUDs exist too, and text and graphics have different possibilities. So it makes sense to have some of both in this rich world.
Since Claude (code) knows all about my system, it will also tailor solutions to it when I ask how to do something. Using it I can often find really neat solutions to whatever sys admin task I can imagine.
Personally, I am experimenting a lot with Agda's instance arguments these days. They somehow have a bad rep for being slow, but their performance has improved a lot! One can juice it to write really readable code (Like truly ad-hoc polymorphic operators, or declaring axioms used in a theorem like `⦃ AC ⦄ → Theorem`), but it also allows for proof searches. Contrary to “tactics”, instance arguments feel like a fluid part of the proof language.
I think there is a lot of room for innovation in the space of proof engineering. I really hope that LLMs will not suck all the oxygen out of the room, by being trained on existing dogmas.
Fable was really a step ahead of Opus on this kind of work. I guess I will be looking at Codex unless Fable is included in my Max subscription on Claude code – or the next Opus is a clear improvement (and comes out soon).
What I listen to is definitely not interchangeable. And contrary to TV series, there is no limit to how much I can enjoy it!
Yesterday told it to write a memory to never write new memories when it solves a problem. We will see if that works better. Sometimes memories are useful, like when I give it a directive about how I want something done and it remembers the spirit of it. But I might as well just spend some more time on my CLAUDE.md…
It is always satisfying to rename something, then merge in some work where someone used the old variable name in the meanwhile, and the merge would go through automatically applying the rename!
Also, it can be riled up to do the task better, or try harder, if I include how cool it will be to get this task done.
But also I feel I have this Orwellian task of censoring its text to avoid it spiralling into negative territory where it convinces itself that the tasks are too difficult. Strange times!