I nearly adopted OCaml for my current math research, in no small part because of Jane Street contributions. I instead chose Lean 4, as LLMs became able to help code in Lean. I give up a factor of two in performance (a gap likely to close) for what reads to me as the clearest expression of each algorithm. Lean is a beautiful language that makes OCaml read like Old English.
Lean is designed to be optionally verified; proof is its primary (but not only) application. It seems an impedance mismatch for Jane Street to explore this direction without also migrating to Lean 4.
As usual, this paean to deductive reasoning (“formal methods”) leaves out its fundamental limit: how closely do the postulates and definitions fit the domain they purport to map? (“In theory, there is no difference between theory and practice. In practice ...”) My guess is that Jane Street maintains large code bodies where the mapping is 1:1, because the purpose of the code is to implement a deterministic algorithm. Many other coders work in such areas. But millions of us don’t: most UIs, most exploratory work, etc.
There is a movement parallel to formal methods to define acceptance criteria at high resolution but not logico-mathematically, which at least grapples with the mapping problem but can’t resolve it where the map isn’t the territory, which is most places. Has Google’s results page, with its extremely evolved internal optimization frameworks really hit an optimum? Could that prototype you whipped up to capture a hazy idea have better illustrated it? These questions are best answered by looking outside the system to what the system serves.
Formal methods are precisely for the domains where the semantics are well-defined. Logical circuits (a lot of CPU components get formal verification), kernels, protocols, parsers, compilers, cryptography, security frameworks, concurrency primitives, etc. all benefit a lot from verification.
In other words, because GEN AI a lot of code, the idea is to shift human value toward verification. Sometimes I think about what programming really is. In fact, learning programming itself is a huge challenge for a non English speaker like me. I have to rely on machine translation to understand English documents that have no translation. The materials in my language are about 5 to 6 years behind.
Now, since it's impossible to code review the tens of thousands of lines of code that AI produces, I see discussions about establishing an absolute rule like mathematical proofs. Reading this reminds me of Rust's borrow checker. In fact, after writing in Rust a few times, it often leads to bad practices where people use tricks to avoid the borrow checker.
Actually, when mathematical rigor goes too far, humans tend to find ways around it. An undereducated programmer like me is especially prone to that.
Looking back at this kind of attempt, it will probably result in writing code only for specific formalized answers. If it becomes that standardized, I'm not sure it will be able to respond to human needs.
I think these defensive programming attempts are fine, but I want to do offensive programming (I coined that term). You take risks, but you fix things quickly and ship. Believing that over time, it will become good enough. Of course, for established industries where accuracy matters and the scope of work is well defined, like Jane Street, the approach in this article is correct. In other words, because there is enough data to adequately model the market's demands
But for social losers like me trying to make money, constantly moving from place to place looking for gold mines, these kinds of methodologies seem like a luxury.Established businesses with mature modeling need highly educated and specialized personnel to continuously optimize. But I know that realistically, I can't keep up with that demand. So I look for places where modeling is unstructured, but I'm not sure if I can use this approach even then.
Whenever I read about formal specs it always seems to me like “write the same tests just in a different way”, or worse, “write the same implementation but in a different way”.
I guess doing things twice can help catch errors, but I fail to see what’s so special about formal specs if they can suffer from the exact same bugs as the tests/implementation.
I guess the root of the problem is if you want to formally prove that a program does something, you have to be very specific (heh) about what that something is, at which point you are basically just writing tests/implementation all over again.
I have been looking into this on and off for years now, and I keep feeling like I’m missing something, but I don’t know what it is.
> Whenever I read about formal specs it always seems to me like “write the same tests just in a different way”, or worse, “write the same implementation but in a different way”. [...] Can anyone enlighten me?
A big difference is that formal methods allow you to use the "for all" quantifier.
For example, you might write a unit test that says "foo('abc') returns a string with no trailing whitespace".
But with formal methods, you can prove that "for any input x, foo(x) returns a string with no trailing whitespace".
This is a trivial example, but you could imagine something more complicated, such as "for any program P, compile(P) has the same behavior as P".
Of course, you have to define what "has the same behavior as" means!
I don’t write software, but in asic and fpga design, formal method specifications enable the use of things like SAT solvers to determine if the underlying test article meets the specification.
You specify properties of the design and the tooling tests the design in a variety of ways to see if it can violate those properties.
Let’s say you have a system that controls turn signals. You can specify properties that say things like lanes that cross each other can’t both have a green light at the same time or even within some period of time of each other.
The toolling can exhaustively check the design for this and surface code traces that violate it.
It's more than just "write the same tests just in a different way", because each test is largely independent, or grows to unmanageably large, and you have to do the work of figuring out the completeness of your test suite by e.g. branch coverage or other relatively crude overlap methods.
Statically proven type systems let you do this in a way that each contributing piece is checked in advance against all the other pieces, guaranteeing not just "this test passes" but, in combination, "all these tests create a reasonable, coherent whole", and it disallows incoherent models from being implemented in the code. An example of this is like Rust's match, which requires complete coverage of all the possible branches, but writ large across an entire codebase.
You're right that it does nothing for you if you make a fundamental logic or theory error, it can't catch that kind of error, but you'd be surprised how much more robust it is than "merely" e.g. Haskell's type system + ad hoc functional testing + property testing - which I would consider a quite strong system overall, but formally proven e.g. cryptography is a much higher bar.
seL4 despite being formally verified contained gnarly bugs, like [1]. Thus, it's not a silver bullet as some people think. Yes, it improves the quality, but only one of the aspects of hardening software, like ASAN, SAST, fuzzing, strict languages, and so on.
> Since it is the unverified SMP config of the kernel
I don't disagree with your point (formal verification does not rid you of all bugs), but this is not the subject of the linked issue. This was a bug in an unverified path.
Hopefully we get more ergonomic ways to do this? Like of the tools listed in the post, dafny + iris are the closest to being industrial I think. And amzn S3 has a history of TLA use in-house I think. But we probably haven't seen the typescript in this space yet, a zero cost abstraction that drops into existing tools, and people genuinely prefer it to the old way.
(And custom linters are also still pretty bad to write. Like golangci-lint is a painful codebase, haven't tried semgrep but the rules engine seemed intimidating. I've yet to use an AST API that I liked)
I caught the religion on using types in conjunction with LLMs about eighteen months ago, but I only really got serious about `lean4` like six to eight months ago and now I wouldn't even consider using AI assist in software work with a `CIC` proof substrate that has practical C/C++ (and therefore, everything) FFI.
We've banned everything from JSON to Python, rewritten `nix` to have a compiler, and almost everything we write is not only property tested and multiply fuzzed to within an inch of it's life, but we have proofs in `lean4` that at a minimum drive property tests via `.olean` linkage, and when we have the bandwidth we prove exhaustiveness over the domain and property test that.
We skip the whole C++/Rust thing because all of the fast stuff is generated from `lean4` and so it doesn't really matter (C++ has advantages when bugs in it are actually bugs in `lean4` code, but you could go either way).
This is a big departure, and I certainly don't blame anyone who is skeptical: "ban JSON and Python wtf?!?!", but we've done millions of lines (checked) of this stuff and AI + formal systems is a dramatically bigger leap than no AI -> AI and Python. The latter in our experience is not monotone in progress, the former is almost always monotone in progress.
And you can do wild shit, this is a formal proof of the polyhedral tensor calculus that is modeled by things like ISL and CuTe, and using that we can get swizzling and tiling using `mdspan` in C++23 on device and prove it's right (up to some L'Hopital arugment about the coverage which this example doesn't demonstrate well: https://github.com/b7r6/mdspan-cute
That in turn, well, it goes real fast. On the first try.
I wonder how formal methods can help us move faster with GenAI.
Is it that they can help write formulas faster? That they can help ensure formulas match the system they're modeling faster? If the problem you think formal methods will help with is sloppy code, isn't the verification code going to be sloppy as well, unless some (not sloppy) intelligence carefully confirms that the specification matches the target system, which was the labor that previously made formal methods too expensive? I guess I don't understand how the argument works if code was previously less sloppy and verification was too expensive, and now code is more sloppy, and there's more of it, but somehow the sloppy intelligence will make verification move fast enough to make it worthwhile.
Unless we have some non-sloppy intelligence that's less of a bottleneck on verification than humans, how are we in a better place?
Maybe it's that investing that huge amount of labor of verification by human experts is now worth it because so much code will be produced that uses the verification systems that the investment will now pay off. But that requires creating pretty general verification systems, such as type system verification or something (which is what they seem to be aimed at), rather than individually verifying software systems like the micro-kernel example.
In other words, maybe the play is to invest in reusable verification systems that can be run tons of times on new code and systems. If so, it's surprising that this wasn't always the strategy.
The general idea is that formal methods are self-verifying, up to a point. Sloppy formal methods simply won't prove.
The point up to which formal methods stops is: do the formally encoded requirements actually encode what I want to be true?
One can make the argument that the requirements is a much smaller surface to verify than that of the entire program.
The counter argument is that figuring out what you want the program to do has always been the hardest part of programming, and that programming in itself is a journey to discover latent requirements.
> One can make the argument that the requirements is a much smaller surface to verify than that of the entire program.
This argument is unfortunately empirically false for any program of any meaningful complexity (>1000 lines, probably even as low as >100 lines ignoring well-defined algorithms and data structures) using current formal methods.
Complete formal specifications are usually multiple times larger than the corresponding source code and encode esoteric propertys necessary for the proof, but which are largely even more impenetrable than a undocumented codebase.
So, it is both harder to figure out if you encoded the desired requirements and it is more complex. Your only advantage is confidence that what you wrote down is proven.
> isn't the verification code going to be sloppy as well
The beauty of formal methods is it doesn't matter if your proof is sloppy. As long as it passes verification, it is correct. And unlike in pure math, the proof that a software system is correct is usually a huge mess of special cases, loop invariants, proofs by induction, and boilerplate that requires a large amount of human labour while providing no insight.
Proofs are also brittle: a tiny change in the code can force you to throw your proof away and start from scratch.
To me, the exciting thing about formal methods in the LLM era is it allows humans to offload the difficult and tedious work of writing proofs to a computer. Taken to an extreme, the human could live entirely in the world of a formal specification, and the LLM could generate 100% of the code. The code may be a mess, but if the system proves it satisfies the spec then it can't be wrong.
The problem is that generating either code or proofs with LLMs is very expensive, and generating good proofs (I don't mean elegant, I mean proving the most important properties) is probably not very fast, either. I.e. the time of verifying a 1MLOC program could go down from 100 years to merely 10 years and the cost could go down from $1bn to $100m, but that's still not practical.
Things can be improved when people help guide and focus the LLMs, but these people still need to be formal methods experts.
So, formal methods produce runnable systems, but communication remains the challenge.
If a formal spec is messy, then it's a proof of ... what, exactly?
A formal specification that bridges tech and product, that lets non-technical contributors read and discuss all the logical nuances, directly as operational code, at product's level of abstraction of interest, would transform a lot.
It's no longer a challenge to create code, it's a challenge to create business requirements and translate them into systems.
I have been testing formal verification methods with multiple products. It will be great to also understand more about what’s tried and how it was done. For example attempting to verify the spec is what I have been trying to implement.
I nearly adopted OCaml for my current math research, in no small part because of Jane Street contributions. I instead chose Lean 4, as LLMs became able to help code in Lean. I give up a factor of two in performance (a gap likely to close) for what reads to me as the clearest expression of each algorithm. Lean is a beautiful language that makes OCaml read like Old English.
Lean is designed to be optionally verified; proof is its primary (but not only) application. It seems an impedance mismatch for Jane Street to explore this direction without also migrating to Lean 4.
As usual, this paean to deductive reasoning (“formal methods”) leaves out its fundamental limit: how closely do the postulates and definitions fit the domain they purport to map? (“In theory, there is no difference between theory and practice. In practice ...”) My guess is that Jane Street maintains large code bodies where the mapping is 1:1, because the purpose of the code is to implement a deterministic algorithm. Many other coders work in such areas. But millions of us don’t: most UIs, most exploratory work, etc.
There is a movement parallel to formal methods to define acceptance criteria at high resolution but not logico-mathematically, which at least grapples with the mapping problem but can’t resolve it where the map isn’t the territory, which is most places. Has Google’s results page, with its extremely evolved internal optimization frameworks really hit an optimum? Could that prototype you whipped up to capture a hazy idea have better illustrated it? These questions are best answered by looking outside the system to what the system serves.
Google's results page isn't well defined. But probably 90%+ of the code below it is well defined.
And in some cases, where the result isnt well defined, it can be learned, so it's not about sitting and thinking what would make sense.
Formal methods are precisely for the domains where the semantics are well-defined. Logical circuits (a lot of CPU components get formal verification), kernels, protocols, parsers, compilers, cryptography, security frameworks, concurrency primitives, etc. all benefit a lot from verification.
In other words, because GEN AI a lot of code, the idea is to shift human value toward verification. Sometimes I think about what programming really is. In fact, learning programming itself is a huge challenge for a non English speaker like me. I have to rely on machine translation to understand English documents that have no translation. The materials in my language are about 5 to 6 years behind.
Now, since it's impossible to code review the tens of thousands of lines of code that AI produces, I see discussions about establishing an absolute rule like mathematical proofs. Reading this reminds me of Rust's borrow checker. In fact, after writing in Rust a few times, it often leads to bad practices where people use tricks to avoid the borrow checker.
Actually, when mathematical rigor goes too far, humans tend to find ways around it. An undereducated programmer like me is especially prone to that.
Looking back at this kind of attempt, it will probably result in writing code only for specific formalized answers. If it becomes that standardized, I'm not sure it will be able to respond to human needs.
I think these defensive programming attempts are fine, but I want to do offensive programming (I coined that term). You take risks, but you fix things quickly and ship. Believing that over time, it will become good enough. Of course, for established industries where accuracy matters and the scope of work is well defined, like Jane Street, the approach in this article is correct. In other words, because there is enough data to adequately model the market's demands
But for social losers like me trying to make money, constantly moving from place to place looking for gold mines, these kinds of methodologies seem like a luxury.Established businesses with mature modeling need highly educated and specialized personnel to continuously optimize. But I know that realistically, I can't keep up with that demand. So I look for places where modeling is unstructured, but I'm not sure if I can use this approach even then.
> these kinds of methodologies seem like a luxury
Absolutely. The article acknowledges this. Jane Street is pretty uniquely equipped to benefit from this.
Whenever I read about formal specs it always seems to me like “write the same tests just in a different way”, or worse, “write the same implementation but in a different way”.
I guess doing things twice can help catch errors, but I fail to see what’s so special about formal specs if they can suffer from the exact same bugs as the tests/implementation.
I guess the root of the problem is if you want to formally prove that a program does something, you have to be very specific (heh) about what that something is, at which point you are basically just writing tests/implementation all over again.
I have been looking into this on and off for years now, and I keep feeling like I’m missing something, but I don’t know what it is.
Can anyone enlighten me?
> Whenever I read about formal specs it always seems to me like “write the same tests just in a different way”, or worse, “write the same implementation but in a different way”. [...] Can anyone enlighten me?
A big difference is that formal methods allow you to use the "for all" quantifier.
For example, you might write a unit test that says "foo('abc') returns a string with no trailing whitespace".
But with formal methods, you can prove that "for any input x, foo(x) returns a string with no trailing whitespace".
This is a trivial example, but you could imagine something more complicated, such as "for any program P, compile(P) has the same behavior as P".
Of course, you have to define what "has the same behavior as" means!
I don’t write software, but in asic and fpga design, formal method specifications enable the use of things like SAT solvers to determine if the underlying test article meets the specification.
You specify properties of the design and the tooling tests the design in a variety of ways to see if it can violate those properties.
Let’s say you have a system that controls turn signals. You can specify properties that say things like lanes that cross each other can’t both have a green light at the same time or even within some period of time of each other.
The toolling can exhaustively check the design for this and surface code traces that violate it.
It's more than just "write the same tests just in a different way", because each test is largely independent, or grows to unmanageably large, and you have to do the work of figuring out the completeness of your test suite by e.g. branch coverage or other relatively crude overlap methods.
Statically proven type systems let you do this in a way that each contributing piece is checked in advance against all the other pieces, guaranteeing not just "this test passes" but, in combination, "all these tests create a reasonable, coherent whole", and it disallows incoherent models from being implemented in the code. An example of this is like Rust's match, which requires complete coverage of all the possible branches, but writ large across an entire codebase.
You're right that it does nothing for you if you make a fundamental logic or theory error, it can't catch that kind of error, but you'd be surprised how much more robust it is than "merely" e.g. Haskell's type system + ad hoc functional testing + property testing - which I would consider a quite strong system overall, but formally proven e.g. cryptography is a much higher bar.
seL4 despite being formally verified contained gnarly bugs, like [1]. Thus, it's not a silver bullet as some people think. Yes, it improves the quality, but only one of the aspects of hardening software, like ASAN, SAST, fuzzing, strict languages, and so on.
[1] https://github.com/seL4/seL4/issues/1199
> Since it is the unverified SMP config of the kernel
I don't disagree with your point (formal verification does not rid you of all bugs), but this is not the subject of the linked issue. This was a bug in an unverified path.
See previous kleppmann post https://martin.kleppmann.com/2025/12/08/ai-formal-verificati..., and yes, obviously anything that you can put in the typesystem or the linter, you should weigh doing so.
Hopefully we get more ergonomic ways to do this? Like of the tools listed in the post, dafny + iris are the closest to being industrial I think. And amzn S3 has a history of TLA use in-house I think. But we probably haven't seen the typescript in this space yet, a zero cost abstraction that drops into existing tools, and people genuinely prefer it to the old way.
(And custom linters are also still pretty bad to write. Like golangci-lint is a painful codebase, haven't tried semgrep but the rules engine seemed intimidating. I've yet to use an AST API that I liked)
Quis custodiet ipsos custodes?
Relevant keynote I wrote about
Formal methods won, now what? https://muratbuffalo.blogspot.com/2026/04/bugbash26-keynote....
I caught the religion on using types in conjunction with LLMs about eighteen months ago, but I only really got serious about `lean4` like six to eight months ago and now I wouldn't even consider using AI assist in software work with a `CIC` proof substrate that has practical C/C++ (and therefore, everything) FFI.
We've banned everything from JSON to Python, rewritten `nix` to have a compiler, and almost everything we write is not only property tested and multiply fuzzed to within an inch of it's life, but we have proofs in `lean4` that at a minimum drive property tests via `.olean` linkage, and when we have the bandwidth we prove exhaustiveness over the domain and property test that.
We skip the whole C++/Rust thing because all of the fast stuff is generated from `lean4` and so it doesn't really matter (C++ has advantages when bugs in it are actually bugs in `lean4` code, but you could go either way).
This is a big departure, and I certainly don't blame anyone who is skeptical: "ban JSON and Python wtf?!?!", but we've done millions of lines (checked) of this stuff and AI + formal systems is a dramatically bigger leap than no AI -> AI and Python. The latter in our experience is not monotone in progress, the former is almost always monotone in progress.
And you can do wild shit, this is a formal proof of the polyhedral tensor calculus that is modeled by things like ISL and CuTe, and using that we can get swizzling and tiling using `mdspan` in C++23 on device and prove it's right (up to some L'Hopital arugment about the coverage which this example doesn't demonstrate well: https://github.com/b7r6/mdspan-cute
That in turn, well, it goes real fast. On the first try.
https://straylight.software/assets/lambda-hierarchy.svg
I wonder how formal methods can help us move faster with GenAI.
Is it that they can help write formulas faster? That they can help ensure formulas match the system they're modeling faster? If the problem you think formal methods will help with is sloppy code, isn't the verification code going to be sloppy as well, unless some (not sloppy) intelligence carefully confirms that the specification matches the target system, which was the labor that previously made formal methods too expensive? I guess I don't understand how the argument works if code was previously less sloppy and verification was too expensive, and now code is more sloppy, and there's more of it, but somehow the sloppy intelligence will make verification move fast enough to make it worthwhile.
Unless we have some non-sloppy intelligence that's less of a bottleneck on verification than humans, how are we in a better place?
Maybe it's that investing that huge amount of labor of verification by human experts is now worth it because so much code will be produced that uses the verification systems that the investment will now pay off. But that requires creating pretty general verification systems, such as type system verification or something (which is what they seem to be aimed at), rather than individually verifying software systems like the micro-kernel example.
In other words, maybe the play is to invest in reusable verification systems that can be run tons of times on new code and systems. If so, it's surprising that this wasn't always the strategy.
The general idea is that formal methods are self-verifying, up to a point. Sloppy formal methods simply won't prove.
The point up to which formal methods stops is: do the formally encoded requirements actually encode what I want to be true?
One can make the argument that the requirements is a much smaller surface to verify than that of the entire program.
The counter argument is that figuring out what you want the program to do has always been the hardest part of programming, and that programming in itself is a journey to discover latent requirements.
> One can make the argument that the requirements is a much smaller surface to verify than that of the entire program.
This argument is unfortunately empirically false for any program of any meaningful complexity (>1000 lines, probably even as low as >100 lines ignoring well-defined algorithms and data structures) using current formal methods.
Complete formal specifications are usually multiple times larger than the corresponding source code and encode esoteric propertys necessary for the proof, but which are largely even more impenetrable than a undocumented codebase.
So, it is both harder to figure out if you encoded the desired requirements and it is more complex. Your only advantage is confidence that what you wrote down is proven.
> isn't the verification code going to be sloppy as well
The beauty of formal methods is it doesn't matter if your proof is sloppy. As long as it passes verification, it is correct. And unlike in pure math, the proof that a software system is correct is usually a huge mess of special cases, loop invariants, proofs by induction, and boilerplate that requires a large amount of human labour while providing no insight.
Proofs are also brittle: a tiny change in the code can force you to throw your proof away and start from scratch.
To me, the exciting thing about formal methods in the LLM era is it allows humans to offload the difficult and tedious work of writing proofs to a computer. Taken to an extreme, the human could live entirely in the world of a formal specification, and the LLM could generate 100% of the code. The code may be a mess, but if the system proves it satisfies the spec then it can't be wrong.
The problem is that generating either code or proofs with LLMs is very expensive, and generating good proofs (I don't mean elegant, I mean proving the most important properties) is probably not very fast, either. I.e. the time of verifying a 1MLOC program could go down from 100 years to merely 10 years and the cost could go down from $1bn to $100m, but that's still not practical.
Things can be improved when people help guide and focus the LLMs, but these people still need to be formal methods experts.
So, formal methods produce runnable systems, but communication remains the challenge.
If a formal spec is messy, then it's a proof of ... what, exactly?
A formal specification that bridges tech and product, that lets non-technical contributors read and discuss all the logical nuances, directly as operational code, at product's level of abstraction of interest, would transform a lot.
It's no longer a challenge to create code, it's a challenge to create business requirements and translate them into systems.
I have been testing formal verification methods with multiple products. It will be great to also understand more about what’s tried and how it was done. For example attempting to verify the spec is what I have been trying to implement.
I am already seeing the uptake on Spec-Driven Development as the Rational Unified Process revenge.