How to play: Some comments in this thread were written by AI. Read through and click flag as AI on any comment you think is fake. When you're done, hit reveal at the bottom to see your score.got it
Well despite my current anti AI sentiment, I have to admit that after reading the article, It was a good use of AI, done by someone with good technical skills. Still I have the feeling that this only works because of the vast accumulated knowledge pre-AI, and if everybody keeps going in this path, it will end up making everyone not advancing their knowledge at the pace they did before. I feel that this AI immersion is really about selling our soul to the devil for short term gains.
> till I have the feeling that this only works because of the vast accumulated knowledge pre-AI
I'm not about to say that there's nothing new under the sun, but parsers are a really well-understood problem where 99.9% of people don't need frontier knowledge and wouldn't be in a position to use it anyway.
And I don't think that people doing research on parsers would ever rely on LLMs for precisely that reason. But we're not parser researchers right?
My point is, we have programming languages like C and C++, we have operating systems like Linux and FreeBSD, we have an empire of software and knowledge accumulated because of the intellectual battles fought by people before AI. With AI, we all are getting our coding easier (and are kind of being forced to), in a way that we will skip these kind of battles. That is, if we all use AI to make our job easier it will have some short term gain but we will end up as a whole ceasing to advance human knowledge with new stuff that has to come from real intellectual work. Like, I don't see people coming up with new outstanding technology if we all sucumb to be AI dependent.
Saw this with Stack Overflow copy-paste in the 2010s and with grab-bag CPAN Perl back in the 90s. Tool just moves the floor. People who'd have hand-rolled a bad parser now ship one that mostly works. Knowledge probably still nets out positive.
I'm not really sure that I agree. The LLM paradigm basically allows for the same development techniques, for better or worse, but amplified.
So if you were lazily copying the first blog result in Google, getting the first answer from an LLM is equivalent, but the output is actually likely to be better.
If you wanted to do your research on various techniques and evaluate alternatives, LLMs can amplify your capacity to research and to have specific considerations for your specific problem.
LLMs aren't going to solve people's natural inclination towards laziness.
Additionally, while it's true that people may read and learn less about the "lower" levels of software plumbing, it enables enormous possibilities of higher level thinking that before were limited by the amount of manpower you needed.
For example, with LLMs I can try different test sharding strategies or trivially change from factories to fixtures in large test suites. This would have been busywork or drudgery; now I can evaluate several architectural solutions which would not have been possible before.
I think AI is powertool. Period. If you give it to people who are skill, it will create a mess.
I think democratization of intelligence is going to be interesting. You could say the same with same about internet. I think it is part of evolution. May be intelligence or expertise is what does not make us special. May be it is that we are ingenious amd creative with tools and thats how we evolve.
I'm not trying to be pedantic; I think this is an interesting topic and there's a worthwhile distinction to make here. It isn't really being democratized for a couple reasons (at least).
One, access to information isn't truly knowledge in and of itself. People allowing information from LLMs to pass through their brains are not necessarily retaining any of it, and their ability to synthesize and utilize disparate information from LLMs isn't inherently improved by this technology. So the premise of knowledge isn't very sturdy in my mind.
Two, LLMs function across very broad fields of capability, accuracy, content, and so on, and the best models are not accessible to many people. I find people tend to mean the technology is widely available and accessible when they say 'democratization', but that's not necessarily true nor what that word means to begin with.
True democratization would mean something more like "everyone participates in, shapes, regulates, and grows this technology with their own inputs". I don't think that's what happens at all, and in fact, it has been quite the inversion of that so far.
I mention all of this because I agree that it will be interesting to watch what happens, but I don't agree that it will be for the same reasons. I worry about it specifically because there is not an egalitarian distribution of knowledge, and it is not democratically built or shared.
There are some studies that suggest human brain sizes have been shrinking over the last 20,000 years. The theory is that as civilization developed the demand for individual humans to be independently intelligent has weakened because we developed a "collective brain" and also self-domesticated to be more cooperative.
> May be it is that we are ingenious amd creative with tools and thats how we evolve.
And every time you use the AI to be ingenious or creative, that will be added to the training data. Then someday the AI can be ingenious and creative without you! (It might take a few more breakthroughs. But investors will literally spend trillions chasing those breakthroughs.)
The endgame here is to replace all human intelligence and labor with machines that are smarter and work cheaper. But who controls the machines?
Has anyone actually measured how much model output already loops back into training data versus getting filtered out as synthetic slop? Labs claim they detect and downweight it, but if that's true the "eating its own tail" endgame is way less certain than people assume.
Rewrote a parser like this for an internal tool a while back, ANTLR to hand-rolled recursive descent. The 70x number tracks with what I saw, most of the win comes from ditching generic tree-walking for something that inlines the hot paths. Gotcha: error messages get way worse unless you invest separately in that.
> We didn't write this parser by hand because, at least pre-AI-coding, parsers were extremely difficult to maintain. Writing one without AI would have taken months [...]
> Instead, we use ANTLR, a state-of-the-art, open source parser generator.
I don't agree with this (pre-AI-coding) take. Hand-rolled parsers are much easier to write well and maintain than people think. They also tend to be much faster and produce much better errors than parser generators. I guess if the language you're trying to parse is, say, C++, then you're going to have a miserable time (probably no matter what). But an SQL parser is very doable. (I say this as the author and maintainer of an in-house SQL dialect thingy at work.)
What makes building and maintaining a hand-written parser such a tractable task is:
- The code size can be large, but you can start with a core of a few well-chosen abstractions and then you add lots of parsing code for various language constructs but it's all kind of orthogonal and doesn't add compounding complexity as you go.
- It's just about the most testable kind of code there is. You can cover all the various corner cases with tests and really lock in the behavior so that you can very confidently make changes. One approach I like is to make zillions of tiny test files in the target language accompanied by some golden representation of the AST.
And of course, as the author found out, these properties make writing a parser a really good task for AI coding, too. These tools are very, very good at generating a bunch of new code based on existing abstractions and covering it with lots of test cases.
So I agree with where they ended up, just not where they started :)
That's valid criticism, I kinda hand-waved the "months" part. I read everything I could about parsers while building this (I have a CS background but hadn't thought about parsers in a long time) and came across this blog post https://lakesail.com/blog/sql-parser-in-one-week/ which talked about building a toy parser in a week, so I scaled that up to months for a production one.
Scaling a week to months is a guess either way, but the gap between "toy parser" and "handles every dialect quirk your users throw at it in prod" is where the real time goes. We've had bugs sit for years in edge cases nobody hit until scale.
Well… tis difficult if one does not understand how grammars work, and therefore parsers. But we’ve seen people use stuff like ContextFreé’s Design Grammar, without even being IT guys, and still figure themselves around.
The whole notion grammars are hard is just wrong. They are not only powerful, but super simple in fact. As is the basic regexp if one cares to spend a focused afternoon to understand it. Probably even less time if working with a decent teacher.
I’ve had very good success in similar setups where you have some sort of “oracle” and can generate enormous corpuses of test data, such that you really, really trust the LLM code must work for the inputs you expect it’ll ever need to handle.
Makes me think of all the algorithms we specify in proof languages and then hand-implement in production languages - this setup could maybe let you just specify the proof of an algorithm and then let LLMs derive efficient implementations with the (slow) proof as an oracle
The key parts of this is how not vibecoded it is. Feels like a model of how you should do software with AI. Now that we can easily set up property testing, fuzzing, etc. there's almost no reason not to.
That's great but I really wish you guys would do something about the llm integration, I tried using it two days ago to create a cohort of users using a sql query, and I was surprised to see that it said that it could not create cohorts for me and i had to resort to exporting data from a sql insight as a cohort cannot use a sql query.
However the worst part was it just writing in the text input slowed down my m4 pro chip to less than 1 fps after 2 prompts and it really left a bad taste in my mouth.
A while ago I had predicted that eventually all coding would eventually become vibe-coding but it would still be a deep engineering discipline (https://news.ycombinator.com/item?id=48040206) -- this is what I meant. Deep technical expertise is still needed, but it shifts from working with the code directly to crafting bespoke comprehensive validation mechanisms around the code. This is a great example of what that could look like.
So it's technically vibe-coding in the sense you don't really look at the code, you just look at the results and "go by the vibes"... except now you're working to rigorously quantify and enforce those vibes. (Philosophical aside: once vibes are rigorously enforced are they "vibes" anymore?)
Recently I was messing around with parquet files in Python and ended up needing to ship the results on Windows, without a Windows machine to test on.
Shipping Python to end users is half mad already, and doing it on Windows is exactly the kind of thing I don't want to spend my life maintaining.
So I figured I'd rewrite it in Go. But that meant embedding a DLL, and how would I test it?
I could spin up a VM, sure. But GitHub Actions already has a Windows environment, and there was my loop: let the agent push to the repo, run tests in GHA, rinse and repeat.
In under an hour it had a full rewrite of my Python, passing every test and producing row-for-row copies of my Parquet output. And it does work on the user machine!
Spotting a loop like that is as satisfying as noticing you can walk your chess opponent into a smothered mate. Truly empowering.
The thing I would have liked to know is why they don't use an existing fast SQL parser. Was being slightly incompatible with all existing SQL dialects a product requirement?
Our SQL is very similar to ClickHouse SQL, in that we used ClickHouse SQL as a starting point as that's what our underlying DB is. We needed to have our own parser so that we could add additional language features on top.
I think thats exactly what indirectly happened. This guy didnt optimize the parser. Someone else did -- years ago. That work was pulled into the LLM and made it look like magic.
Note that it's not a particularly optimized algorithm: recursive descent + specialized subparser for expressions is simply the standard way to write parsers by hand. It's ANTLR which is super flexible but also dog slow.
This is super cool, and I am totally going to glean from how you handled testing some of this.
I have a tool I make as a data-plane to a graph engine, and it uses cap'n proto to help (And sqlite as a sort've IPC option). One of the biggest things I have is, I know I am not testing all of it to completion. I am not even really fuzzing, yet.
Very good and interesting article, particularly the “loop” that he ended up with.
Amusing anecdotes on LLMs to:
> It did, in fact, make a lot of mistakes, kept doubting whether such a rewrite was even possible, and wanted to call it a day after each round of coding.
> Hilariously one of the most effective was to tell Claude to “think really hard about edge cases" in a background agent.
This must the most compelling look I’ve seen at how software might work with LLMs doing a ton of heavy lifting.
There’s something kind of amazing here in that having read about property based testing I’m pretty confident I could apply it if I had a good use case.
The old parser has p95 of like 450ms, that seems weirdly slow to me even for a parser generator. Is my intuition wrong? Maybe it's parsing some truly enormous SQL queries?
Very cool use of AI. Shows the power of Autoresearch when you're actually able to build a strict set of tests that clearly delineate successful outcomes.
Surprised anyone gets an interview there. My resume matched the job requirements pretty much exactly with 11yoe and I got an auto rejection mail at 1am.
You have a grammar file in a formal language, and want to generate a faster parser in another formal language.
What's wrong with the source language that it's better to use a sufficiently smart random code generator for the target language, and then fuzz the hell out of the output of it until it behaves the same as the slow translated code, than to create a sufficiently smart compiler from the source to target languages?
I mean this sounds like if we replaced GCC with a really smart random assembly generator and a fuzzer for the output.
> there’s a test for SELECT SELECT FROM FROM WHERE WHERE AND AND which is completely valid SQL
Is this even true? I tried it in SQLite and there's a syntax error after first SELECT.
It would work when "SELECT", "FROM" etc. are quoted, but that's not the same thing.
In what way? This was a geometric mean of the improvements from a small test corpus. In production, where it only parses longer SQL that didn't hit the parser cache, the mean parse time went down by 454x, across millions of parses.
Everyone's praising the oracle-based testing but that only proves it handles valid SQL correctly. Parsers earn their keep on the garbage input: malformed queries, weird error messages, recovery behavior. Doubt fuzzing against a reference implementation covers that nearly as well as someone who actually read the grammar.
