DSS 2026 / Day 1 / Pearl Hall
Formal Verification in the Age of AI
Benjamin Flores
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:00Hey.
0:01Hello everyone. Thanks for being here.
0:04So I'm Ben from Veridion, and today
0:07we're going to talk about formal
0:08verification.
0:10What we do first at Veridion is that we
0:12formally verify critical code.
0:14We work with great clients like Ethereum
0:15Foundation or Lido that I see here.
0:19And
0:20I heard a lot of things on Twitter
0:22recently about formal verification, good
0:24and bad stuff. Some people saying it's
0:26the it's going to fix all of our
0:28problems. Some people saying it's pure
0:29slop. So today I just wanted to
0:33re-explain where AI can help us on
0:35formal verification, where AI
0:38doesn't help us so much, where it
0:40changes things, and where we headed
0:42because we had access to
0:45Gemini Carbon recently that is like the
0:48next the last model from from Google.
0:51And they they gave us free free tokens.
0:54So we spent a lot and
0:57we we did a lot of benchmarks and some
0:58stuff to understand where formal
1:00verification is headed with the next
1:02frontier models. So I'm going to share a
1:04bit of that work after.
1:06Okay. So let's start.
1:08So just to explain what was formal
1:10verification before AI, I will
1:15summarize it in three steps. So the
1:17first step was
1:18like any software development, you write
1:21what you want to do
1:23in English specs. If you're Spanish, it
1:25can be English. If you're Spanish, it
1:27can be Spanish. Sorry. But the idea is
1:29that in a natural language you're going
1:31to write what you want to do.
1:33And
1:34this will be your specification, your
1:37architecture
1:38in English.
1:40Then the second step would be you're
1:42going to take this specification and
1:44you're going to try to formalize it
1:46under
1:47under a specification that is
1:49mathematic.
1:50So you're going to take for example a
1:52specification that is the beginning of
1:54all specification in
1:56industry which is for example my user
1:58doesn't lose money.
2:00And going to you're going to try to
2:01understand okay what does it mean uh in
2:04terms of variables in terms of
2:06mathematics my user doesn't use money
2:09uh so you're going to take for example
2:11oh it means that in my vault uh I do not
2:14create shares so you're going to write
2:16oh uh the variable share do not go up
2:19okay that's a that's a very easy example
2:21and this was pretty hard to do because
2:24you had to understand uh your whole
2:26system uh software engineer
2:29as a whole your solidity system or
2:31whatever system you you were building
2:33but also you had to had some I would say
2:36mathematical level to understand how can
2:38I translate this into a mathematical
2:40specification
2:42so this was pretty hard to do
2:43um then you had the proving side so now
2:47if I resume you have step one you have
2:49your specific specification in English
2:51then you have it in mathematics then you
2:53have to prove that your mathematical
2:55specification is correct on all states
2:58of uh of your code of your
3:00implementation and that was pretty hard
3:02to do um for several reasons so the
3:06first reason
3:07is
3:08uh that we had one main tool to prove
3:11code that was SMT solver uh and an SMT
3:14solver what it is is is a deterministic
3:16tool that will try to take your problem
3:19and um divide it into constraints and
3:23try to
3:24find a counter example to your
3:26specification so it's going to try to
3:28say your specification is false here is
3:31a counter example of your specification
3:33and
3:34the idea is that
3:36if it find a counter example then it's
3:38false you come back to the formalization
3:40you you you try you try to reprove again
3:43and if it passes you're good but the
3:45problem is
3:46I said hard or impossible because on
3:49non-linear code on difficult code
3:52it was pretty much impossible to uh
3:55uh decompose your problem into
3:57constraints you couldn't test all
3:59possibilities and find a counter example
4:01so you weren't able to fully formally
4:03verify your
4:05your code if you have some code that was
4:07that was difficult. So what you do is
4:09that in general what what was done is
4:12that you would adapt your code or at
4:14least the model of your code to the tool
4:17that you're using that is the prover and
4:19you're going to say okay this times out
4:21because it's too difficult let's make it
4:23a bit less difficult.
4:24So what you do actually what you we did
4:27in the past was it's too difficult let's
4:30make my code less difficult let's adapt
4:31the code to the test which is of course
4:34not the good way to do things. You want
4:36your test to test your code and not
4:39adapt your code to the test right?
4:41But I'm not I'm not criticizing it was
4:44very difficult again it was something
4:46very very difficult
4:48and sometimes even impossible.
4:51So that was before AI. Now what does AI
4:54change in the whole process? This is
4:57tool agnostic by the way this is like
4:59I'm describing the whole process without
5:01any names of tools just for it to be
5:04to be clear.
5:06So as said
5:07like
5:08at the end the formalization and the
5:10creation of the spec is the most
5:12important thing because
5:13you can prove something that we don't
5:15care about
5:17and it doesn't say that you do not have
5:19bugs you can prove some things you we
5:21don't care about you can forget to prove
5:23things. So the most the most difficult
5:26part was still the specification process
5:28but everything was hard in the process
5:30overall.
5:32And then
5:33well I explained already but the proof
5:36was hard for several reason and
5:38sometimes it was impossible.
5:40So now with AI so I'm going to come back
5:43here.
5:44This I said easy because
5:47it's like it's still the same thing and
5:50it was easy before but
5:52depends on your system of course you
5:53going to continue to write your
5:55specification in English and now we call
5:57this a prompt basically
5:59you describe your system in English what
6:01you want to do what it the guarantees
6:03you must hold etc.
6:05Now,
6:06the formalization of your spec it's not
6:09easy, very very easy, but it's easier
6:11because you do not need a lot of skills
6:15in mathematics to actually
6:17ask an AI and go back and forth and take
6:20an English specification and make it
6:22mathematics.
6:23This doesn't mean it's like very easy to
6:26fully cover your protocol with all of
6:28the specs you need. It means that the
6:30process of taking English to mathematics
6:32is easier with AI
6:34than than it was before. That's for
6:36sure.
6:37Now, this is where everyone is talking
6:39about formal verification on on X and on
6:42Twitter. It's basically because it
6:44changed a big thing here on the proving
6:46side.
6:47It unlocks a new way
6:49to do proving. So before, if I go back,
6:52I said hard or impossible because we use
6:55SMT solvers that are deterministic tools
6:58that cannot formally verify
7:01I would say non-linear code or difficult
7:03code.
7:06Here,
7:07we can still use SMT solvers and they
7:09work very well in some easy cases, but
7:11we can use another tool that is writing
7:14manually a proof as a mathematician
7:16would do
7:17for a theorem.
7:19And you I'm sure you heard about OpenAI
7:23and they're doing a lot of this
7:24actually. When they prove all of these
7:26theorems, they're basically sending a
7:28bunch of LLMs and they're saying to
7:30these LLMs, please prove this theorem
7:32using this and this hint. Please try to
7:34prove that it's correct. And the LLM is
7:36going to use classical mathematical
7:38techniques and write, oh, if this, then
7:41we could infer that, etc. etc. etc.
7:43because the LLM has been trained on all
7:45of the mathematics of the world and is
7:47better than 10 PhDs at mathematics. So
7:50now basically what was impossible before
7:53because we didn't have infinite money to
7:55pay mathematical PhDs
7:57is possible now to prove using a
8:00classical mathematical proof way because
8:03we have infinite infinite money to pay
8:06PhDs in mathematics because they are
8:07very very cheap and they run in data
8:09center in Texas.
8:11So
8:12we can now use this new technique. So
8:14it's unlocks
8:16a new way to prove and this is why
8:18everyone is talking about formal
8:19verification on X. It's not because like
8:23formalization is now easier,
8:25it's because we can unlock a new way to
8:28prove things which means we can prove
8:30every code that is correct.
8:34And that's that's why everyone is
8:35talking about it. On the formalization
8:37part, on the spec part, it helps but as
8:39it helps on every software engineering
8:41task.
8:43And now what's left for us, the human?
8:47This is still, as said before, this is
8:49still the most complicated thing to do
8:51is to ensure that the specs
8:55that you've written in English and in
8:56mathematics
8:58have a full coverage of your code, that
9:00you completely understand your code and
9:02your system and you're able to describe
9:05it
9:06fully so you do not miss anything.
9:09And that's still the most complicated
9:11part. Now if if
9:14these models are capable of proving
9:16things that no mathematicians
9:18ever found,
9:19maybe they're also capable of doing
9:21that.
9:22And that's why we try to understand by
9:26partnering with with Gemini. Thanks to
9:29them for giving $400,000
9:31worth of tokens.
9:33And
9:36now where where we headed? So as said,
9:38we got the future access on the on on
9:41every Gemini models in the last few
9:42months and we tested with a lot of
9:45tokens if these models are good at
9:48proving we already know. OpenAI has
9:50found that, we has found that. It's
9:52really really easy to prove but are they
9:54good at completely
9:56understanding the whole
9:58the the the whole
10:00the whole architecture you have and
10:02speaking everything that you're doing.
10:05It's actually like close to very very
10:07good and close to like human experts to
10:10be honest, especially if you
10:13contextualize it in a way that makes
10:15sense. Um, so I I do not have data here
10:18unfortunately because
10:19well they gave us the tokens but they do
10:21not gave us the the right to spend to to
10:24share the information yet, but I I can
10:26share orally. Um,
10:28so it's very good at finding gaps in
10:31existing specification. So if you give a
10:34context and say to the model with a
10:35freight model you're going to give oh
10:37look at that, look at all of this
10:39specification for this exact freight, do
10:41you think it's it's like full and
10:43correct? It's going to find gaps.
10:46In general these gaps are are are very
10:49good and it's also very good writing
10:52specs from scratch by just reading the
10:54code. Ideally you want to to write all
10:56the specs before doing any code and
10:59ideally you you have like your whole
11:01spec
11:03ready before doing any code and you
11:04launch that to the AI
11:06but
11:07it's still good at like going on the
11:10other way around which is reading your
11:11code, understanding what you're trying
11:13to do, writing the specs out of that
11:16and this is like very good at
11:18understanding
11:19misalignment in the architecture.
11:22So the the conclusion of that for us is
11:25that
11:26to us at least the the future of vibe
11:29coding or the future of coding with AI
11:31includes formal verification in some way
11:34because
11:35these models are really really good at
11:38implementing but the bottleneck is the
11:41review, is the fact that when these
11:44models are creating a bunch of slope you
11:46do not know exactly if this slope is
11:47correct. You can do some end-to-end
11:49tests but you will never
11:51you will never get the full length of it
11:54but with formal verification you can
11:55just shrink enormously the review by
11:58just reviewing the spec
12:00which permits you
12:02to then have a proof that the
12:03implementation follow the spec so you
12:05can have trust in your code and trust on
12:07the output of the AI. So we think at the
12:10end we're going to
12:12rewrite our
12:14at least um use
12:17formal formal verification native
12:19languages or make our current languages
12:21native for formal verification in some
12:24way so we can vibe code like this which
12:25is are giving all the specs to the AI.
12:27The AI doesn't even like give me the
12:30code but just give me a proof that it
12:33matches the specification I give I gave
12:35the AI.
12:37Um and we are doing a tooling that is
12:41like following this vision that's called
12:42Verity as the name of the company and
12:45it's a small lean eDSL so basically lean
12:48is the programming language for
12:49mathematics.
12:50Um and the idea is that it's following
12:53all of the vision that I had described
12:54you before. Uh you can either import
12:56your Solidity deterministically and try
12:59to formally verify it with this new
13:01method of proving not with SMT solvers
13:03uh or you can directly vibe code using
13:06Verity and it's going to compile to
13:08bytecode and it's going to
13:10Verity which is lean at the end of the
13:11day. It's going to compile to bytecode
13:13and you can deploy your smart contracts
13:15directly in a fully uh
13:17lean native way or formal verification
13:19native way.
13:21Um and yeah that's uh that's it. Thank
13:24you for listening and uh have a good
13:25day. Be safe out there.
Source and accuracy details
- Names, organizations, technical terms, code and numbers may be misrecognized.
- Source wording and disfluencies are preserved; no editorial corrections were applied.
- Export contains whole-second cue starts only; cue ends and subsecond timing are unavailable.
- A full caption-track export does not establish that every spoken word is represented.
- Automatic YouTube captions; not human-verified against the audio.
Source export SHA256: be430943ac68cec879aac41c29c6e434d51b5363cd95ccf3a4882690b74a3cd7