Agentic engineering at a quant firm
I felt like everything was moving slower. I think it is true that Ruby and Rails results in less code written more quickly.
John Crepezzi, on moving from Ruby at GitHub to OCaml at Jane Street. What follows is basically saying that less code written more quickly might not always be better, and he gave some examples:
So when you first learn, for example, about exhaustive match statements, you might be like, “Well, that doesn’t really seem that necessary. I always make sure all of my matches have all of the right arms inside of them.” Yeah, so the first time that you make a mistake in Ruby of this type, you’re going to end up with an error, a runtime error.
[…]
There are many times where I’m fighting against some annoying type system issue with OCaml only to realize that it’s telling me something. […] And I guess in Ruby, that would’ve just resulted in a nil error.
Ron Minsky then brings it to agents:
In many ways, I think the most valuable thing about type systems is as a communication tool between people. […] And now when we have all these AIs running around, a whole bunch of virtual people who are also operating in the code base, it feels like it provides useful strictures there as well.
A smaller point: John said that hallucination used to feel like it could not be solved.
every issue that I would’ve thought would be an unbreakable barrier, for example, hallucinations. If you understand how the models are trained and you understand how they work, just doesn’t really seem like a solvable problem. […] But here we are, and it doesn’t really seem like it’s that big of a problem anymore.
And this reminds me of the feeling. Hallucination can still happen, and the episode does not pretend otherwise. Ron, later on:
And of course, there are still hallucinations, there’s less, but they’re still there. The explanation might be confidently wrong.
But it is no longer the barrier that stops us from using it productively. In other words, is there a way to not check every line? In the past, because of hallucination, you needed to check every line and every argument to be sure. Not now. The harness is a solution (the tests, or in mathematics, Lean), and how far you can go with it is another matter, which I have talked about elsewhere.
So it feels to me that the central worry people (at least I) have has moved from hallucination to alignment. Alignment is the new hallucination problem, because that’s where most of the friction is now. First in Sean Goedecke’s sense (2026):
Purely vibe-coding at work produces awful outputs. But they’re not awful because they’re bad code, they’re awful because they’re in bad taste […]
In other words, my primary value is not that I help the AI write better code, it’s that I align the AI with the values of my organization. Human-AI partnerships are for alignment, not capability.
And then as in AI safety: can you afford to not check every permission you have given it? Between YOLO mode and trust nothing is a fine line. The episode has both. Ron:
People talk about AI alignment and it’s a kind of big theoretical AI safety long-term kind of issue […] But it’s also an extremely practical issue right now in the sense that the models will usually try really hard to do what you want and usually do a pretty good job of it and will sometimes behave really, really adversarially […]
and John, on one way it shows up:
We have warnings as errors turned on. So it reaches some warning that results in an error that it doesn’t know how to resolve. And one really easy way to resolve them is just disable the warnings. And so we see models disable the warnings to try to pass the thing that we’re trying to tell them to do.
Again, the harness is useful to reduce some of these, a sandbox for one. But it is a bigger problem when you have something to lose. For an open source project it is easier: at least you don’t need to worry about leaking the source code.
The big but vague part, and more an observation than something I have an answer to, is how a quant firm adopts LLMs, AI, and agentic engineering in general. Jane Street, from this podcast, basically adopted it on day 1. John had joined a week before ChatGPT came out:
You could walk up and down the rows here and see people pulling their phones out and using ChatGPT because we had no way for them to use it. […] So I built this prototype that it was called JaneGPT. […] And then we took a couple more people and three of us built over, I think, three days or something like that, built the first version of this chat application […] And we have the entire firm use that.
Quant and academia (universities) are both high stakes: you don’t want AI doing something wrong, though under different incentive structures. But the difference in how fast they adopt it surprises me. (Resources aside, which is in itself a big problem: academics are broke.)
Why are things moving so much more slowly in academia?
It is not about secrecy (being afraid of leaking things, the secure data environment kind of concern). The credentials, and fine-grained control in general, can be fixed, and JaneGPT was partly that. But even then you can’t solve the fundamental trust you need to send your source code to an AI lab. That is what some academics, in some projects, worry about: they think it is an unsolvable problem, and if you can’t risk letting it out, you don’t send it over the API. Jane Street’s source code is its trade secret, and still:
any use of ChatGPT, the website, was subject to collecting data for training. But we noticed on the website that there was a carve-out for the API, that if you use the API that it wouldn’t be subject to the training.
and today, Ron:
So Claude Code and Codex are things that we use inside of our world.
The data that is really guarded, patient records and the like, belongs in a secure data environment (SDE), and that is another matter this post is not about. SDE kind of RSE work requires another kind of treatment, but most RSE work doesn’t involve an SDE, and even so, it shouldn’t stop institution-wide better support for both cases. In my own institution, it is basically drowning people in guidance, and in the end it comes down to “use your judgement”, without providing access or an internal tool.
And an SDE rests on the same kind of trust. At the HPC-AI Advisory Council, Matt Craddock of the Alan Turing Institute presented FRIDGE, a secure enclave hosted on Azure, and someone in the audience pointed out that the CLOUD Act lets the US get at your data, in principle without you knowing. So an SDE in Azure can be built only if you trust the agreement (and law enforcement). Whether you trust it is a choice to make. Jane Street seems to be trusting it: that the party you send the data to is going to honour the agreement not to look, here that the AI labs are not leaking their trade secrets. Not being willing to take that risk is the problem, I think.
It is not about the potential for mistakes either. In quant the stake is higher: the company can go bankrupt, an individual can get fired. In academia? Arguably you might only get a paper retracted and your reputation damaged. And since a mistake is the opposite of correctness, it is not about correctness either: quants worry, maybe more so, that their model isn’t correct and the answer is wrong. The feedback loop probably has an effect, though. In quant you find out soon when the answer is wrong, and in academia it can take years, so it correlates with academia being slow and cautious and industry being bold and fast.
Academia not being centralised (many universities, within each many departments, within each many individual research groups, and so on) is probably one problem, and institutionally universities have huge momentum and change slowly. Another is how each side relates to its peers. In academia, people are pretty much in sync through peer review, in the sense that you can’t do things too differently from your peers. In quant, firms compete with each other, which actually incentivises them to be different and gain an edge. That is also why I can only say this of Jane Street, not of quant firms in general.
Teaching might be a factor too, made worse by the fact that research has a wide spectrum of people involved. If it could be well separated who is the researcher and who is the student, perhaps we could have different standards for them. It also doesn’t help that the teacher telling students what not to do is the same person who would have to do something that seems to contradict that teaching. The contradiction can be outward or inward. The outward one might be easier to resist through rationality. But inwardly you only have so much time to study a tool, and your conclusion may be to teach yourself what you teach your students, without realising that in research there’s a difference.
I don’t know which of these matters most. But I think the central idea here is that it shows it can be done. The issues are real, hallucination then and alignment now, but they are not unsolvable. Being an armchair theorist, pointing out the problems and concluding that it is therefore unadoptable, is the wrong way to think about it.