[boccaff]

Don't we still need to verify that what the LLM proved is actually the stated system?If agents cutting corners (like deleting tests) is a concern, how can we be sure that the verification corresponds to the software and that every _load bearing_ assumption is true? I don't think that simply trusting the Lean core is enough.

[kolinko]

Of course, and correct verification assumptions were always a challenge, and they are a possible failure point - but 95%(?) of labour was then proving the code, and now that part can be largely automated.

Speaking of how difficult can it be to write a proper spec, I think a few years ago someone found a bug in spec in a helloworld-like example in a book about formal methods, and a bug in the code.

[nilkn]

Yes, this is a challenge, but I don't think it's as damning as it sounds.

First, the same kind of issue exists with any kind of testing. We all know that having unit tests and integration tests doesn't mean the software is correct and bug-free. Yet we still often believe it's worth investing in writing and maintaining a test suite. Formal methods is an additional level of assurance that I suspect we'd generally find worth it for critical software if it were easy enough to implement, even if it's not perfect.

Second, because formal verification has been so hard, nobody has really bothered making a serious attempt at solving the spec problem before (at least not to my knowledge). We've spent decades building frameworks that make writing tests really easy. We really have not come close to investing the same kind of effort into building frameworks and tools to make writing formal specs easy. I can imagine a world where we have sophisticated tools that help us create specs, plus real-time verification infrastructure that automatically reverifies code against the spec on each PR.

Third, and perhaps most importantly, there are many industries where this is actually low-hanging fruit: aerospace, medical devices, etc., actually already have formal specs that are verified manually by human engineers to satisfy certification requirements. I believe that's a more natural starting point vs. trying to scale formal methods for all software in general.

[jhanschoo]

While that is true, just having LLMs "think about it" may discover unintended behavior and exploits, and you can choose to what extent you want to verify that the claims are correct.

[jerf]

Just as an addendum to some of the other fine replies you've got, one of the things I'm interested in for formal verification is security assertions, and those don't necessarily need lots of comprehension to apply. If you can assert "this code never hits the network again" or "this code will never execute arbitrary code", the proof of that may be nontrivial but the comprehension of it is easy.

Proving that some complicated algorithm is actually what you think it is, even after you've proved that it does whatever it is it does, is hard. But to me, and me personally so not necessarily what everyone else thinks, that's not really what I'm interested in from proofs. I'm really more interested in these overarching declarations of what the code definitely doesn't do, or putting bounds on what it does, rather than proving that my billing code bills the customers precisely in the way I expect, when, frankly, the best specification of it I have is already what is in the unproved code base anyhow. I believe those have a very sensible story as to how both humans and AIs can use those assertions.

[andor]

The LLM does not prove anything (it cannot reason). It generates Lean code, and conveniently, in Lean, the code is also the proof. It’s not merely a model of the stated system, it is the system.

You still need to verify that the generated code is what you asked for, though.

[eru]

> The LLM does not prove anything (it cannot reason).

What is that supposed to mean?

> It generates Lean code, and conveniently, in Lean, the code is also the proof. It’s not merely a model of the stated system, it is the system.

Yes, and that's great. (Though, of course, the computer doesn't execute lean directly, it gets translated first.)

> You still need to verify that the generated code is what you asked for, though.

Yes, but you only need to read the theorems, not the proofs nor code.