[Jweb_Guru]

For example, if you are willing to accept undecidable typechecking, you can define inductive types in a much simpler theory (inductive types and dependent pattern matching are by far the most complex parts of CiC): https://homepage.cs.uiowa.edu/~astump/papers/from-realizabil... .

There are a few other examples but this is the one that immediately sprung to mind. The complexity of CiC is very much about decidable typechecking, it's not fundamental to writing a sound dependently typed proof assistant. And anyone who has actually worked on a kernel for CiC knows that while it is "minimal" compared to what people want to do with it, no efficient kernel is nearly as minimal as something like λP2 with dependent intersections.

[foooobar]

Thanks! But does that not already answer your question as to why Lean would use CIC and not a simpler metatheory?

That particular paper is from 2016 and the type theory presented there is still lacking a number of features from CIC, with those particular extensions being listed as further research. Work on Lean started in 2013 and the kernel in particular has not been touched in a long time for good reason. Around that time, CIC was already battle-proven with there being popular and well-tested reference implementations to learn from. Since then, all the fancy work has been done on the system surrounding the kernel.

I believe one of the goals at the start was that the kernel should be sufficiently simple and small so that it is easy to ensure that the kernel is correct. Despite the apparent redundant complexity that you suggest, the implementation still seems to be sufficiently simple to reach this goal, as no proof of false that was a result of a kernel bug has been reported yet. CIC being well-tested surely also contributed to that.

Another reason is that, as I understand it, it was not always obvious that Lean would break decidability of defeq. Afaik proof irrelevance was only baked in with Lean 3, while it was an option before that. By that point I don't see much of a reason to rewrite an entire kernel that has worked fine so far.

I also vaguely remember that there were arguments related to performance in favor of the current system (e.g. versus W-types), but I don't really understand these things well enough to make that point with confidence :)

[kmill]

How are the ergonomics of a proof assistant based on λP2? Lean does a lot of elaboration of metavariables and typeclass instances and such before handing things off for checking by the kernel -- I'm curious about whether λP2's undecidability is decidable enough, so to speak.

[Jweb_Guru]

That's currently an open research question, but the answer most likely is that the ergonomics could be made just as good, considering that elaboration of metavariables, typeclass instances, etc. already uses possibly-undecidable algorithms distinct from those used in the kernel in just about every proof assistant I'm familiar with (Lean included).

In terms of stuff that do go more or less directly through the kernel, some followup papers to the one I linked discuss how to translate inductive definitions etc. automatically into a form accepted by the simpler kernel, so I am even more confident that this would not present a problem.

[kevinbuzzard]

The fact remains though that Lean 3 worked well enough to do Scholze level mathematics. I never quite know what to make of people saying "if the designers had done it in some other way, it would have been better". Of course if someone comes along doing it another way, and then they go on to also achieve Scholze level mathematics, then I will be all ears.

[Jweb_Guru]

That's a very strange line of argument. Many great programs have been written in C. That doesn't mean that there is no way to improve on C. Perhaps to even the playing field, this hypothetical competitor should be given the resources of Microsoft Research as well :)

[foooobar]

I think Kevin is asserting that to make strong comparative claims about the quality of either system, it's good to have a large and functionally similar body of work to compare the systems with, not that there is no way to improve on Lean's core :)

MSR funds a single developer. The Lean community consists of at least 30 or so active contributors right now, with many more having contributed in the past. Most folks work in academia. Much of Lean's success is absolutely due to the work of that single developer, but I don't think that the playing field is uneven due to MSR's funding.