[matt_heimer]

I'm building a Java HFT engine and the amount of things AI gets wrong is eye opening. If I didn't benchmark everything I'd end up with much less optimized solution.

Examples: AI really wants to use Project Panama (FFM) and while that can be significantly faster than traditional OO approaches it is almost never the best. And I'm not taking about using deprecated Unsafe calls, I'm talking about using primative arrays being better for Vector/SIMD operations on large sets of data. NIO being better than FFM + mmap for file reading.

You can use AI to build something that is sometimes better than what someone without domain specific knowledge would develop but the gap between that and the industry expected solution is much more than 100 hours.

[jacquesm]

AI is extremely good at the things that it has many examples for. If what you are doing is novel then it is much less of a help, and it is far more likely to start hallucinating because 'I don't know' is not in the vocabulary of any AI.

[Filligree]

> because 'I don't know' is not in the vocabulary of any AI.

That is clearly false. I’m only familiar with Opus, but it quite regularly tells me that, and/or decides it needs to do research before answering.

If I instruct it to answer regardless, it generally turns out that it indeed didn’t know.

[jacquesm]

I haven't had that at all, not even a single time. What I have had is endless round trips with me saying 'no, that can't work' and the bot then turning around and explaining to me why it is obvious that it can't work... that's quite annoying.

[dwaltrip]

Try something like:

> Please carefully review (whatever it is) and list out the parts that have the most risk and uncertainty. Also, for each major claim or assumption can you list a few questions that come to mind? Rank those questions and ambiguities as: minor, moderate, or critical.

> Afterwards, review the (plan / design / document / implementation) again thoroughly under this new light and present your analysis as well as your confidence about each aspect.

There's a million variations on patterns like this. It can work surprisingly well.

You can also inject 1-2 key insights to guide the process. E.g. "I don't think X is completely correct because of A and B. We need to look into that and also see how it affects the rest of (whatever you are working on)."

[jacquesm]

Ok! I will try that, thank you very much.

[dwaltrip]

Of course! I get pretty lazy so my follow-up is often usually something like:

"Ok let's look at these issues 1 at a time. Can you walk me through each one and help me think through how to address it"

And then it will usually give a few options for what to do for each one as well as a recommendation. The recommendation is often fairly decent, in which case I can just say "sounds good". Or maybe provide a small bit of color like: "sounds good but make sure to consider X".

Often we will have a side discussion about that particular issue until I'm satisfied. This happen more when I'm doing design / architectural / planning sessions with the AI. It can be as short or as long as it needs. And then we move on to the next one.

My main goal with these strategies is to help the AI get the relevant knowledge and expertise from my brain with as little effort as possible on my part. :D

A few other tactics:

- You can address multiple at once: "Item 3, 4, and 7 sound good, but lets work through the others together."

- Defer a discussion or issue until later: "Let's come back to item 2 or possibly save for that for a later session".

- Save the review notes / analysis / design sketch to a markdown doc to use in a future session. Or just as a reference to remember why something was done a certain way when I'm coming back to it. Can be useful to give to the AI for future related work as well.

- Send the content to a sub-agent for a detailed review and then discuss with the main agent.

[intended]

Eh… I am not sure if that translate to “I don’t know”.

IDK would require the LLM to be aware of the frequency of cases seen in its own training.

I can see this working as a risk ranking, which is certainly worth trying in its own right.

Does it actually say “I don’t know?”

[j45]

I don’t know can be added to the vocabulary depending on the technique being used.

There are so many overlapping and also unique approaches to software development beyond vibe coding and ai driven software development.

[mtrovo]

I think the main issue is treating LLM as a unrestrained black box, there's a reason nobody outside tech trust so blindly on LLMs.

The only way to make LLMs useful for now is to restrain their hallucinations as much as possible with evals, and these evals need to be very clear about what are the goal you're optimizing for.

See karpathy's work on the autoresearch agent and how it carry experiments, it might be useful for what you're doing.

[riffraff]

> there's a reason nobody outside tech trust so blindly on LLMs.

Man, I wish this was true. I know a bunch of non tech people who just trusts random shit that chatgpt made up.

I had an architect tell me "ask chatgpt" when I asked her the difference between two industrial standard measures :)

We had politicians share LLM crap, researchers doing papers with hallucinated citations..

It's not just tech people.

[withinboredom]

We were working on translations for Arabic and in the spec it said to use "Arabic numerals" for numbers. Our PM said that "according to ChatGPT that means we need to use Arabic script numbers, not Arabic numerals".

It took a lot of back-and-forths with her to convince her that the numbers she uses every day are "Arabic numerals". Even the author of the spec could barely convince her -- it took a meeting with the Arabic translators (several different ones) to finally do it. Think about that for a minute. People won't believe subject matter experts over an LLM.

We're cooked.

[ThrowawayR2]

Kind of a tangent but that did make me curious about how numbers are written in Arabic: https://en.wikipedia.org/wiki/Eastern_Arabic_numerals

[tracker1]

I guess "Western Arabic" would have been more precise.

[tstenner]

The architect should have required Hindu numbers. Same result, but even more confusion.

[dvfjsdhgfv]

Man this is maddening.

[godelski]

Honestly I think we're just becoming more aware of this way of thinking. It's certainly exacerbated it now that everyone has "an expert" in their pocket.

It's no different than conspiracy theorists. We saw a lot more with the rise in access to the internet. Not because they didn't put in work to find answers to their questions, but because they don't know how to properly evaluate things and because they think that if they're wrong then it's a (very) bad thing.

But the same thing happens with tons of topics, and it's way more socially acceptable. Look how everyone has strong opinions on topics like climate, rockets, nuclear, immigration, and all that. The problem isn't having opinions or thoughts, but the strength of them compared to the level of expertise. How many people think they're experts after a few YouTube videos or just reading the intro to the wiki page?

Your PM is no different. The only difference is the things they believed in, not the way they formed beliefs. But they still had strong feelings about something they didn't know much about. It became "their expert" vs "your expert" rather than "oh, thanks for letting me know". And that's the underlying problem. It's terrifying to see how common it is. But I think it also leads to a (partial) solution. At least a first step. But then again, domain experts typically have strong self doubt. It's a feature, not a bug, but I'm not sure how many people are willing to be comfortable with being uncomfortable

[j45]

There’s a possibility the same people might believe anything they read on social media or via Google and it’s something worthy of attention.

[roncesvalles]

And the worst part is, these people don't even use the flagship thinking models, they use the default fast ones.

[closewith]

In my experience, people outside of tech have nearly limitless faith in AI, to the point that when it clashes with traditional sources of truth, people start to question them rather than the LLM.

[FpUser]

I am curious about what causes some to choose Java for HFT. From what I remember the amount of virgin sacrifices and dances with the wolves one must do to approach native speed in this particular area is just way too much of development time overhead.

[matt_heimer]

Probably the same thing that makes most developers choice a language for a project, it's the language they know best.

It wasn't a matter of choosing Java for HFT, it was a matter of selecting a project that was a good fit for Java and my personal knowledge. I was a Java instructor for Sun for over a decade, I authored a chunk of their Java curriculum. I wrote many of the concurrency questions in the certification exams. It's in my wheelhouse :)

My C and assembly is rusty at this point so I believe I can hit my performance goals with Java sooner than if I developed in more bare metal languages.

[nly]

"HFT" means different things to different people.

I've worked at places where ~5us was considered the fast path and tails were acceptable.

In my current role it's less than a microsecond packet in, packet out (excluding time to cross the bus to the NIC).

But arguably it's not true HFT today unless you're using FPGA or ASIC somewhere in your stack.

[atomicnumber3]

The one person who understands HFT yeah. "True" HFT is FPGA now and also those trades are basically dead because nobody has such stupid order execution anymore, either via getting better themselves or by using former HFTs (Virtu) new order execution services.

So yeah there's really no HFT anymore, it's just order execution, and some algo trades want more or less latency which merits varying levels of technical squeezing latency out of systems.

[matt_heimer]

Software HFT? I see people call Python code HFT sometimes so I understand what you mean. It's more in-line with low latency trading than today's true HFT.

I don't work for a firm so don't get to play with FPGAs. I'm also not co-located in an exchange and using microwave towers for networking. I might never even have access to kernel networking bypass hardware (still hopeful about this one). Hardware optimization in my case will likely top out at CPU isolation for the hot path thread and a hosting provider in close proximity to the exchanges.

The real goal is a combination of eliminating as much slippage as possible, making some lower timeframe strategies possible and also having best class back testing performance for parameter grid searching and strategy discovery. I expect to sit between industry leading firms and typical retail systematic traders.

[smokel]

> AI really wants to use Project Panama

It would help if you briefly specified the AI you are using here. There are wildly different results between using, say, an 8B open-weights LLM and Claude Opus 4.6.

[matt_heimer]

I've been using several. LM Studio and any of the open weight models that can fit my GPU's RAM (24GB) are not great in this area. The Claude models are slightly better but not worth they extra cost most of the time since I typically have to spend almost the same amount of time reworking and re-prompting, plus it's very easy to exhaust credits/tokens. I mostly bounce back and forth between the codex and Gemini models right now and this includes using pro models with high reasoning.

[tracker1]

Maybe a silly question, but why Java? As a C# guy, my experience with AI is it hasn't been great with it, and I'd suspect similar for Java. I'd probably go with Rust, which my own efforts with AI has done really well with, even if I'm far from a Rust expert.

[colechristensen]

Then you list all of the things you want it not to do and construct a prompt to audit the codebase for the presence of those things. LLMs are much better at reviewing code than writing it so getting what you want requires focusing more on feedback than creation instructions.

[grim_io]

Wouldn't Java always lose in terms of latency against a similarly optimized native code in, let's say, C(++)?

[jacquesm]

Not necessarily. Java can be insanely performant, far more than I ever gave it credit for in the first decade of its existence. There has been a ton of optimization and you can now saturate your links even if you do fairly heavy processing. I'm still not a fan of the language but performance issues seem to be 'mostly solved'.

[nly]

"Saturating your links" is rarely the goal in HFT.

You want low deterministic latency with sharp tails.

If all you care about is throughput then deep pipelines + lots of threads will get you there at the cost of latency.

[matt_heimer]

You can achieve optimized C/C++ speeds, you just can't program the same way you always have. Step 1, switch your data layout from Array of Structures to Structure of Arrays. Step 2, after initial startup switch to (near) zero object creation. It's a very different way to program Java.

You have to optimize your memory usage patterns to fit in CPU cache as much as possible which is something typical Java develops don't consider. I have a background in assembly and C.

I'd say it's slightly harder since there is a little bit of abstraction but most of the time the JIT will produce code as good as C compilers. It's also an niche that often considers any application running on a general purpose CPU to be slow. If you want industry leading speed you start building custom FPGAs.

[jodleif]

As long as you tune the JVM right it can be faster. But its a big if with the tune, and you need to write performant code

[andriy_koval]

Java has significant overhead, that most/every object is allocated on heap, synchronized and has extra overhead of memory and performance to be GC controlled. Its very hard/not possible to tune this part.

[matt_heimer]

You program differently for this niche in any language. The hot path (number crunching) thread doesn't share objects with gateway (IO) threads. Passing data between them is off heap, you avoid object creation after warm up. There is no synchronization, even volatile is something you avoid.

[andriy_koval]

> Passing data between them is off heap

how exactly you are passing data? You can pass some primitives without allocating them on heap. You can use some tiny subset of Java+standard library to write high performance code, but why would you do this instead of using Rust or C++?

[matt_heimer]

In some places I'm using https://github.com/aeron-io/agrona

Strangely this is one of the areas where I want to use project panama so I might re-implement some of the ring buffers constructs.

You allocate off heap memory and dump data into it. With modern Java classes like Arena, MemoryLayout, and VarHandle it's honestly a lot like C structs.

I answered "why" in another post in this thread.

[tyingq]

Depends. Many reasons, but one is that Java has a much richer set of 3rd party libraries to do things versus rolling your own. And often (not always) third party libraries that have been extensively optimized, real world proven, etc.

Then things like the jit, by default, doing run time profiling and adaptation.

[andriy_koval]

Java has huge ecosystem in enterprise dev, but very unlikely it has ecosystem edge in high performance/real time compute.

[roncesvalles]

There are actually cases when Java (the HotSpot JVM) runs faster than the same logic written in C/C++ because the JVM is doing dynamic analysis and selective JIT compilation to machine code.

[not_kurt_godel]

I personally know of an HFT firm that used Java approximately a decade ago. My guess would be they're still using it today given Java performance has only improved since then.

[andriy_koval]

it doesn't mean Java is optimal or close to optimal choice. Amount of extra effort they do to achieve goals could be significant.

[jcgrillo]

Optimal in what sense? In the java shops I've worked at it's usually viewed as a pretty optimal situation to have everything in one language. This makes code reuse, packaging, deployment, etc much simpler.

In terms of speed, memory usage, runtime characteristics... sure there are better options. But if java is good enough, or can be made good enough by writing the code correctly, why add another toolchain?

[andriy_koval]

> But if java is good enough, or can be made good enough by writing the code correctly,

"writing code correctly" here means stripping 95% of lang capabilities, and writing in some other language which looks like C without structs (because they will be heap allocated with cross thread synchronization and GC overhead) and standard lib.

Its good enough for some tiny algo, but not good enough for anything serious.

[jcgrillo]

It's good enough for the folks who choose to do it that way. Many of them do things that are quite "serious"... Databases, kafka, the lmax disruptor, and reams of performance critical proprietary code have been and continue to be written in java. It's not low effort, you have to be careful, get intimate with the garbage collector, and spend a lot of time profiling. It's a totally reasonable choice to make if your team has that expertise, you're already a java shop, etc. I no longer make the choice to use java for new code. I prefer rust. But neither choice is correct or incorrect.

[mewpmewp2]

I would say that if AI has to make decisions about picking between framework or constructs irrelevant to the domain at hand, it feels to me like you are not using the AI correctly.

[LtWorf]

I've seen SQL injection and leaked API tokens to all visitors of a website :)