LLMs and teaching, with a Zero-Knowledge-Proof example.

In the early days of search engines, learning to write just the right query, could make the difference between finding the result in a minute, and searching for hours.

One undeniable value of LLMs is education. Interactive knowledge at your fingertips.

Quite a range of answers can be obtained from an LLM session, by providing hints and insights to the LLM and helping it refine the answer. This is a super-power and a defining property of LLMs that was never possible in human history until now.

Here is an initial explanation of the Chaum-Pedersen ZKP protocol when one asks ChatGpt.

At this point, this is an answer one could easily find in someone's lecture notes or summaries. Along years of studying, I have developed the opinion that this is a (typical) answer one would give if they really did 'NOT' want to teach you the topic. It is actually a bad answer.

Searching for truly good answers has been a long quest of mine. I alluded to this opinion at the time where it struck me, slightly encoded, in my stackexchange profile. The sentence in the screenshot hints at the fact that most people will actually send you astray with their answers. Why? we all know why: in order to preserve the hard work that was put into learning (mining? store of value, proof of work ... hmmm)

After interacting with the LLM and giving it direction to modify it's answer by

1. Instructing it to focus on certain aspects and remove others
2. Thinking and finding the true key elements (based on the origial answer) that may have led to the authors to their protocol, then providing those as insights to the LLM and instructing it to use them
3. A bit of iteration by trial and error

We obtain an excellent answer, equivalent to spending time hunting for the best answers on math.stackexchange, and then collecting the result into a coherent note:

And even that last point that contains 'magic' we could improve into

Now that is a useful answer.