all the criticism has been said, all the takes been had.
-
@jens @jonny At work "everyone has to use AI" according to C level people, despite all examples of things going wrong or at best "only" wasting lots of resources. Full hype mode.
LLMs are a tool. I just haven't figured out what for they are a reliable and reasonable choice.
For "How could problem X be addressed?", f.i. with a language I know very well, it might generate an answer with good points for me to look up and verify by myself in detail. Like a shortcut for a bunch of search queries.
-
So, look. One shot rewriting the whole test suite in another language is probably not great to do, but what happened here is so much worse than you are expecting.
https://github.com/RsyncProject/rsync/pull/903/
This does not "translate tests into pytest" or a unit testing framework, it writes its own testing framework where tests are whole python scripts that redefine basic test functions in every script. Surely there would be a single way to "run rsync and get the results" - nope, well, there is, but then every test file will randomly redefine its own
_run_and_capturefunction. So like now rsync needs a test suite for its test suite.If instead of telling an LLM to "rewrite the tests in python" you just searched "python testing" you would find the pytest docs. And then you would find examples. And then you could write fixtures to deduplicate all the prior shell script setup and teardown stuff, and so on. But since it was just "rewrite the tests in python" its now worse than before, and the odds of the rewrite actually being a 100% faithful translation are close to 0.
@jonny Ugh. I didn't realize at first which project this was. Then I looked at the repo. Yes, the road to hell IS paved with good intentions...
-
To a person, the whole purpose of the test is for it to fail when it should. That's an elemental part of writing good tests: they must fail before the patch, or else they provide no protection. We want protection from failure, that is good for us. We need tests to protect us because we can't possibly evaluate all the other parts of a complex system when we try to fix one part of it.
LLM slot machines change what tests mean - of course we still want the code to work good, but if we're not evaluating the code or the tests, then what the slot machine turns them into is just a high score and the jackpot condition. 130 new tests added, that means its good. They pass, that means I win.
The bugfix loop with LLMs defeats the purpose of automated tests and renders it no better than manual testing: you notice a bug, you yell at the LLM to fix it, you keep looking at the specific thing that's broken until its fixed, good robot, ship it. The changes don't have meaningful tests, and nothing else does either, so the slot machine loop repeats, bug->fix->win. Very velocity. Rocket fuel even.
@jonny It's a pattern I've been noticing all over.
Step 1: a process is created to measure something, like "does the software work right?" or "who do people want to be president?"
Step 2: There's an incentive for the people who perform and maintain the process to get a certain outcome, like good performance reviews or the guy you like being elected.
Step 3: Lacking the power to alter the thing being measured, the people in charge get creative with how they measure. -
@jonny Gambling doesn't attract me at all, never has, wonder if that's why I've been more resistant to it than others...
I'm kinda similar, the most I tried gambling was when I went to a gambling town for a wedding. I came out up but it just didn't do much for me.
Except poker, maybe because that involved people. That was fun.
-
@jonny It's a pattern I've been noticing all over.
Step 1: a process is created to measure something, like "does the software work right?" or "who do people want to be president?"
Step 2: There's an incentive for the people who perform and maintain the process to get a certain outcome, like good performance reviews or the guy you like being elected.
Step 3: Lacking the power to alter the thing being measured, the people in charge get creative with how they measure.@jonny It's those darn management cybernetics all over again. In order to make good software, we built a system of writing and passing tests—and (say it with me) the purpose of a system is what it does. The starting assumption that "did we make good software" would always be a critical test turned out to be false, because no system was constructed to keep it true.
-
@themipper @jonny
> It is a shame to see more and more foundational projects fall into the LLM trapThe one that breaks my heart is vim.
@JetSetIlly @jonny damn.
-
@jonny I have seen this pattern of which you speak when attempting to use LLMs to compare TCP Delayed-ACK implementations between BSD derived code bases. They generated output suggesting semantics that just weren't there, presumably based on how similarly named things were between each fork, but this was not obvious without reading the source for oneself in context. This went doubly for FreeBSD where there are multiple TCP functional blocks ("stacks").
@jonny Reminded of the "similarity quagmire" by changes Netflix guys are staging for FreeBSD TCP logs. My engineering notes from Feb, they did 80-90% of what I had in mind to do. The LLMs kept confabulating and conflating the meaning of "delayed_ack" with delacktime. "delayed_ack" is a variant ; its meaning is context dependent for macOS or between BSDs. it is neither a timer delta, nor is it an outstanding segment threshold counter; it can be all of those things. The LLMs did not infer this.
-
@jonny Reminded of the "similarity quagmire" by changes Netflix guys are staging for FreeBSD TCP logs. My engineering notes from Feb, they did 80-90% of what I had in mind to do. The LLMs kept confabulating and conflating the meaning of "delayed_ack" with delacktime. "delayed_ack" is a variant ; its meaning is context dependent for macOS or between BSDs. it is neither a timer delta, nor is it an outstanding segment threshold counter; it can be all of those things. The LLMs did not infer this.
@jonny The really bizarre thing is, that even if LLMs were capable of actual reasoning, they'd be considered to be committing category error in how they parse source code; logical fallacies then follow from that. Deep learning proponents also appear to have been doing this. The terminology describing LLMs and deep learning invites category error and cognitive dissonance. Geoffrey Hinton has a lot to answer for with his "It might be conscious" emergence-based woo.
-
I think the modal situation here is that the people are reading none or very little of what is being generated by the LLM, so the tests have a special role: Tests function as the pull arm on the slot machine, you just generate until tests pass, and that's a jackpot. Obviously that's meaningless when the tests are meaningless, so tests take on a very different meaning and role in slot machine coding.
Previously we would write careful test conditions that were based off some real problem or an understanding of what the code under test did, and had a specific thing they were intended to protect against. Tests move slow and are designed to protect us against the things we know can go wrong. When we learn of a new wrong thing, we add a test.
LLM tests have the form of tests but don't do the same thing. They often test nothing, and are just expressions of truisms that the probabilistic text space explored while generating. They have strongly worded names but end up actually asserting that basic language features work as expected. Because it is not us writing tests for ourselves, where we only harm ourselves by making them weak, they function instead as a passively obfuscated justification for the code that the LLM generates. The user wants the tests to pass. The LLM provides.
The tests are theater: they are the play field for the slot machine. They are mild, surmountable, need to fail a few times to be plausible, but must eventually pass within the expected generation loop window to deliver the payout.
@jonny nailed it
-
@jonny (Un)charitable interpretation: it smoke tests whether the ensure_ffmpeg function is syntactically correct — which is a failure mode LLMs are actually concerned about.
@henryk
The smoke test for syntactic correctness is, thankfully, the interpreter. -
@jonny It's like everyone decided to take a bath in mercury and leaded gasoline.
-
Another thing it reminds me of, with tokens unlocking after x time so you better be ready to jump and start getting things done again, is my aunties Farmville addiction.
-
@elebertus
Ive read so much LLM code at this point, there are still patterns that are present but elude my understanding, but one thing that's clear is that there are foundational flaw categories that are not improved upon by model version and appear in wildly different projects using wildly different models and harnesses. Testing is a big nexus of those flaws. I am not close to what would be a satisfying explanation of the dynamics, but every project suffers fucked testing problems.I suspect testing has the same properties as translation. It’s moderately easy to build machine-translation systems that are kind-of okay. A mechanical dictionary is a reasonable approximation. If something goes through your post, looks up each word in an English-French dictionary (for example) and outputs the resulting text, it won’t be correct, but it will be vaguely comprehensible. If you build a dictionary of bigrams or trigrams (sequences of 2-3 words) this gets a bit better because now collocations are more likely to be translated correctly. It won’t be as good as a professional translator, but it will more or less look like the target language. Add more statistical modelling and you will get better up to a point. But there’s a cliff where you can’t improve without actually understanding the content. No amount of statistical modelling will let you accurately translate the things that are statistical outliers and the extrinsic knowledge necessary means that you can’t infer a correct translation from the text alone without understanding its context.
Tests have a similar property. Good tests convey the intention, but the intention is not part of the code and so can’t be inferred from it. Good tests cover the things that the test author knows are corner cases, but these can’t be inferred from the code either (a few can, if the language has explicit error-handling constructs) because they’re a property of the input data.
In both cases, LLMs try to compensate for the lack of understanding by having a lot of examples of similar things in their input. If the thing you’re translating is similar to a load of other things, you may not need to understand it to translate it correctly because the first dozen (or hundred, thousand, or whatever scale you need) people to translate something like that did the hard work and you can reuse it. If the thing you’re testing is similar to a load of other things that already exist, someone else may have done the hard work of identifying the common failure modes and expressing intent.
But commonly LLM-generated tests end up testing that the code does what the code does. And that’s not useful. If you want that, just use fuzzing in a harness that tests trace equivalence between two versions of the program (for the same sequence of inputs, do they generate the same output?). That is useful for no-functionality-change-intended patches (typically things that improve performance or simplify unnecessary complexity), but most changes to the codebase are there because you want the behaviour to change. Good tests will fail if you changed something that was part of an API contract but will not fail if you added new behaviour, but tests based on the code will change.
This isn’t limited to LLMs. Some of the LLVM tests are just ‘run this command, the output should look like this’. People typically reject these in review now because long and painful experience showed us that it was hard to refactor when a change broke a test and the change author couldn’t tell if the difference in output came from something we actually cared about or just something that happened to be part of the old version’s output. But humans can, at least, tell the difference in the tests because they understand what it is that they intend with the change that introduces the test.
-
@davey_cakes
@ainmosni
Exactly the same, I recognize that feeling too -
I suspect testing has the same properties as translation. It’s moderately easy to build machine-translation systems that are kind-of okay. A mechanical dictionary is a reasonable approximation. If something goes through your post, looks up each word in an English-French dictionary (for example) and outputs the resulting text, it won’t be correct, but it will be vaguely comprehensible. If you build a dictionary of bigrams or trigrams (sequences of 2-3 words) this gets a bit better because now collocations are more likely to be translated correctly. It won’t be as good as a professional translator, but it will more or less look like the target language. Add more statistical modelling and you will get better up to a point. But there’s a cliff where you can’t improve without actually understanding the content. No amount of statistical modelling will let you accurately translate the things that are statistical outliers and the extrinsic knowledge necessary means that you can’t infer a correct translation from the text alone without understanding its context.
Tests have a similar property. Good tests convey the intention, but the intention is not part of the code and so can’t be inferred from it. Good tests cover the things that the test author knows are corner cases, but these can’t be inferred from the code either (a few can, if the language has explicit error-handling constructs) because they’re a property of the input data.
In both cases, LLMs try to compensate for the lack of understanding by having a lot of examples of similar things in their input. If the thing you’re translating is similar to a load of other things, you may not need to understand it to translate it correctly because the first dozen (or hundred, thousand, or whatever scale you need) people to translate something like that did the hard work and you can reuse it. If the thing you’re testing is similar to a load of other things that already exist, someone else may have done the hard work of identifying the common failure modes and expressing intent.
But commonly LLM-generated tests end up testing that the code does what the code does. And that’s not useful. If you want that, just use fuzzing in a harness that tests trace equivalence between two versions of the program (for the same sequence of inputs, do they generate the same output?). That is useful for no-functionality-change-intended patches (typically things that improve performance or simplify unnecessary complexity), but most changes to the codebase are there because you want the behaviour to change. Good tests will fail if you changed something that was part of an API contract but will not fail if you added new behaviour, but tests based on the code will change.
This isn’t limited to LLMs. Some of the LLVM tests are just ‘run this command, the output should look like this’. People typically reject these in review now because long and painful experience showed us that it was hard to refactor when a change broke a test and the change author couldn’t tell if the difference in output came from something we actually cared about or just something that happened to be part of the old version’s output. But humans can, at least, tell the difference in the tests because they understand what it is that they intend with the change that introduces the test.
@david_chisnall
@elebertus
Well said -
Here's an example from some code that was thrust at me this week. The rest of the tests try a bit harder to look like tests, but this one is perplexing.
What does it test? The function name suggests its a smoke test. LLMs love to call things smoke tests. That would suggest this would be an early-run test that fails loudly if some basic precondition - like having ffmpeg - fails. Or, I guess we are smoke testing the
ensure_ffmpegfunction? Anyway who knows. However we first check if ffmpeg or ffprobe are present, which is exactly whatensure_ffmpegdoes. If they aren't present, a warning tells us that ffmpeg/ffprobe are required for the video tests, which makes it seem like this should be a parameterizing test that controls which tests are run, which of course it does not do.So the test literally does nothing and cannot possibly fail, but says it does at least two things, because to an LLM something saying it does something is the same thing as it actually doing that thing.
@jonny “to an LLM something saying it does something is the same thing as it actually doing that thing” bears pondering
-
RE: https://hails.org/@hailey/116657391001259044
all the criticism has been said, all the takes been had. the only metaphor i have been finding consistently useful for understanding what is happening with people and "AI" is addiction, and specifically gambling addiction.
@jonny I agree that it's gambling addiction.
-
I suspect testing has the same properties as translation. It’s moderately easy to build machine-translation systems that are kind-of okay. A mechanical dictionary is a reasonable approximation. If something goes through your post, looks up each word in an English-French dictionary (for example) and outputs the resulting text, it won’t be correct, but it will be vaguely comprehensible. If you build a dictionary of bigrams or trigrams (sequences of 2-3 words) this gets a bit better because now collocations are more likely to be translated correctly. It won’t be as good as a professional translator, but it will more or less look like the target language. Add more statistical modelling and you will get better up to a point. But there’s a cliff where you can’t improve without actually understanding the content. No amount of statistical modelling will let you accurately translate the things that are statistical outliers and the extrinsic knowledge necessary means that you can’t infer a correct translation from the text alone without understanding its context.
Tests have a similar property. Good tests convey the intention, but the intention is not part of the code and so can’t be inferred from it. Good tests cover the things that the test author knows are corner cases, but these can’t be inferred from the code either (a few can, if the language has explicit error-handling constructs) because they’re a property of the input data.
In both cases, LLMs try to compensate for the lack of understanding by having a lot of examples of similar things in their input. If the thing you’re translating is similar to a load of other things, you may not need to understand it to translate it correctly because the first dozen (or hundred, thousand, or whatever scale you need) people to translate something like that did the hard work and you can reuse it. If the thing you’re testing is similar to a load of other things that already exist, someone else may have done the hard work of identifying the common failure modes and expressing intent.
But commonly LLM-generated tests end up testing that the code does what the code does. And that’s not useful. If you want that, just use fuzzing in a harness that tests trace equivalence between two versions of the program (for the same sequence of inputs, do they generate the same output?). That is useful for no-functionality-change-intended patches (typically things that improve performance or simplify unnecessary complexity), but most changes to the codebase are there because you want the behaviour to change. Good tests will fail if you changed something that was part of an API contract but will not fail if you added new behaviour, but tests based on the code will change.
This isn’t limited to LLMs. Some of the LLVM tests are just ‘run this command, the output should look like this’. People typically reject these in review now because long and painful experience showed us that it was hard to refactor when a change broke a test and the change author couldn’t tell if the difference in output came from something we actually cared about or just something that happened to be part of the old version’s output. But humans can, at least, tell the difference in the tests because they understand what it is that they intend with the change that introduces the test.
great analysis! @david_chisnall @jonny @elebertus
-
@jonny I just lost my beer league hockey championship as the last shooter on a 14 round shoot out. I'm sitting in my driveway reading your thread. I'll need to read it again in the morning.
I don't remember why I followed you originally. But I love this thread.
This whole rsync thing is the most interesting thing that has come out of the ai bubble.
I had a negative feel for rsync after years ago reading a blog criticizing its sloppy design.
Yet I rely on it daily. I have so many questions.
@poleguy @jonny rsync is the fallback method for zxfer, a FreeBSD script-based tool for replicating ZFS datasets over the network; normally it calls out to zfs send/recv to do the hard work of shuffling deltas. I don't use it myself; I'm currently using a 3rd party backup agent with SFTP backing for Win11 clients but not rsync itself.
-
But its not just as simple as "OK if I read the tests I should be fine" because LLM code is often untestable. It writes code with function and class names that make it seem like a something does something, but they might just be flat wrong. Or there is some invisible fallback condition the LLM encountered while generating code and added to just make tests pass, but has entirely different behavior.
If you've watched an LLM generate a project over time, you see it generating its own private language, and ive even seen it reinvent language features like function definitions themselves. Its names form part of an increasingly inaccessible web of meaning that no human can penetrate.
Writing tests requires a kind of "information gap" where you can have enough intuition about what something does, but not how it does it, so you can a) know what it should do, b) make a strong assertion about that expectation, c) without mirroring the internal implementation's limits. That's hard! And really only possible when the foundation, (a) is true. Code must have an articulable purpose in order to be testable, that's tautological, that defines what failure is. But since LLM code increasingly detaches from any kind of stable description or expectation, even if the tests look very rigorous, you can't know if they are just tailored to the specific internal details of its function to eke out a pass, because it's hard to know what it should do anyway.
So really you have to read the test code, the code under test, and also all the other code that might call the code under test. Aka you have to read everything. And rather than reading something that was written to be read, you're wading through a slop swamp. So you can't. It takes more time than just writing it. The erosion of testing is just an intrinsic part of the loop that you can't escape without breaking the spell of the slot machine, and it is what drives the loop.
@jonny I suspect the "information gap" you mention here corresponds to known LLM kryptonite: they can't emulate human reasoning in the abductive mode (to best outcome or expectation, usually inherently non-boolean, and the domain of GOFAI expert systems). Human decisions about refactoring legacy code bases (Chesterton's Fence) might constitute an example of problem solving requiring abductive reasoning. I published https://burdentennis.com yesterday but didn't put my name to it directly.
