1. Yes, GPT-4 Turbo is quantitatively getting lazier at coding. I benchmarked the last 2 updates to GPT-4 Turbo, and it got lazier each time.
2. For coding, asking GPT-4 Turbo to emit code changes as unified diffs causes a 3X reduction in lazy coding.
Here are some articles that discuss these topics in much more detail.
- System prompt 1: https://sharegpt.com/c/osmngsQ
- System prompt 2: https://sharegpt.com/c/9jAIqHM
- System prompt 3: https://sharegpt.com/c/cTIqAil Note: I had to nudge ChatGPT on this one.
All of this is anecdotal, but perhaps this style of prompting would be useful to benchmark.
As a side note, i wrote a very simple small program to analyze Rust syntax, and single out functions and methods using the syn crate [1]. My purpose was exactly to make it ignore lazy-coded functions.
Your link appears to be ~100 lines of code that use rust's syntax parser to search rust source code for a function with a given name and count the number of AST tokens it contains.
Your intuitions are correct, there are lots of ways that an AST can be useful for an AI coding tool. Aider makes extensive use of tree-sitter, in order to parse the ASTs of a ~dozen different languages [0].
But an AST parser seems unlikely to solve the problem of GPT being lazy and not writing the code you need.
The tool needs a way to guide it to be more effective. It is not exactly trivial to get good results. I have been using GPT for 3.5 years and the problem you describe never happens to me. I could share with you just from last week, 500 to 1000 prompts i used to generate code, but the prompts i used to write the replacefn, can be found here [1]. Maybe there are some tips that could help.
[1] https://chat.openai.com/share/e0d2ab50-6a6b-4ee9-963a-066e18...
// Handle struct-specific logic here
// Add more details about the struct if needed
// Handle other item types if needed
...etc...
Like, you parse the response, and throw away the comment "//implementation goes here", throw away also the function/method/class/struct/enum it belongs to, and keep the functional code. I am trying to implement something exactly like aider, but specifically for Rust, parsing the LLM's response, filtering out blank functions etc.
In Rust, filtering out blank functions is easy, in other languages it might be very hard. I haven't looked into tree-sitter, but getting a sense of Javascript code, Python and more, sounds pretty much a very difficult problem to solve.
Even though i like when GPT compresses the answer and doesn't return a lot of code, other programs like Mixtral 8x7b, never compress it like GPT in my experience. If they are not lacking much than GPT4, maybe they are better for your use case.
>It took >200 back-and-forth messages with ChatGPT to get it to ultimately write 84 lines of code? Sounds lazy to me.
Hey Rust throws a lot of errors. We do not want humans go around and debug code, unless it is absolutely necessary, right?
It really is. It wastes a ton of time even if the user explicitly requests that code listings be printed in full.
Further, all the extra back and forth trying to get it to do what it is supposed to pollutes the context and makes it generally more confused about the task/goals.
However for now, I have not run re-tests for every new version. I guess I know what I will be doing today.
This is an area I have spend a lot of time working on, would love to compare notes.
"// ... the rest of your code goes here"
in it's responses, rather than writing it all out.
Claude is doing much better in this area, local/open LLMs are getting quite good, it feels like OpenAI is not heading in a good direction here, and I hope they course correct.
I hope some people remember and document details of this era, future generations may be so impressed with future reality that they may not even think to question it's fidelity, if that concept even exists in the future.
The former sounds like a great training set to enable the latter. :(
We know it knows how to make gunpowder (for example), but only because it would initially tell us.
Now it won't without a lot of trickery. Would we even be pushing to try and trick it into doing so if we didn't know it actually could?
Would somebody try to push a technical system to do things it wasn't necessarily designed to be capable of? Uh... yes. You're asking this question on _Hacker_ News?
But obviously all that social infrastructure is fragile… so you’re not wrong to be alarmed, IMO
While I would agree that "don't tell it how to make bombs" seems like a nice idea at first glance, and indeed I think I've had that attitude myself in previous HN comments, I currently suspect that it may be insufficient and that a censorship layer may be necessary (partly in a addition, partly as an alternative).
I was taught, in secondary school, two ways to make a toxic chemical using only things found in a normal kitchen. In both cases, I learned this in the form of being warned of what not to do because of the danger it poses.
There's a lot of ways to be dangerous, and I'm not sure how to get an AI to avoid dangers without them knowing them. That said, we've got a sense of disgust that tells us to keep away from rotting flesh without explicit knowledge of germ theory, so it may be possible although research would be necessary, and as a proxy rather than the real thing it will suffer from increased rates of both false positives and false negatives. Nevertheless, I certainly hope it is possible, because anyone with the model weights can extract directly modelled dangers, which may be a risk all by itself if you want to avoid terrorists using one to make an NBC weapon.
> I don't care about racial bias or what to call pope when I want chatgpt to write Python code.
I recognise my mirror image. It may be a bit of a cliché for a white dude to say they're "race blind", but I have literally been surprised to learn coworkers have faced racial discrimination for being "black" when their skin looks like mine in the summer.
I don't know any examples of racial biases in programming[1], but I can see why it matters. None of the code I've asked an LLM to generate has involved `Person` objects in any sense, so while I've never had an LLM inform me about racial issues in my code, this is neither positive nor negative anecdata.
The etymological origin of the word "woke" is from the USA about 90-164 years ago (the earliest examples preceding and being intertwined with the Civil War), meaning "to be alert to racial prejudice and discrimination" — discrimination which in the later years of that era included (amongst other things) redlining[0], the original form of which was withholding services from neighbourhoods that have significant numbers of ethnic minorities: constructing a status quo where the people in charge can say "oh, we're not engaging in illegal discrimination on the basis of race, we're discriminating against the entirely unprotected class of 'being poor' or 'living in a high crime area' or 'being uneducated'".
The reason I bring that up, is that all kinds of things like this can seep into our mental models of how the world works, from one generation to the next, and lead to people who would never knowingly discriminate to perpetuate the same things.
Again, I don't actually know any examples of racial biases in programming, but I do know it's a thing with gender — it's easy (even "common sense") to mark gender as a boolean, but even ignoring trans issues: if that's a non-optional field, what's the default gender? And what's it being used for? Because if it is only used for title (Mr./Mrs.), what about other titles? "Doctor" is un-gendered in English, but in Spanish it's "doctor"/"doctora". But here matters what you're using the information for, rather than just what you're storing in an absolute sense, as in a medical context you wouldn't need to offer cervical cancer screening for trans women (unless the medical tech is more advanced than I realised).
[0] https://en.wikipedia.org/wiki/Redlining
[1] unless you count AI needing a diverse range of examples, which you may or may not count as "programming"; other than that, the closest would be things like "master branch" or "black-box testing" which don't really mean the things being objected to, but were easy to rename anyway
Longer answer:
I found that I could provoke lazy coding by giving GPT-4 Turbo refactoring tasks, where I ask it to refactor a large method out of a large class. I analyzed 9 popular open source python repos and found 89 such methods that were conceptually easy to refactor, and built them into a benchmark [0].
GPT succeeds on this task if it can remove the method from its original class and add it to the top level of the file with appropriate changes to the size of the abstract syntax tree. By checking that the size of the AST hasn't changed much, we can infer that GPT didn't replace a bunch of code with a comment like "... insert original method here...". The benchmark also gathers other laziness metrics like counting the number of new comments that contain "...". These metrics correlate well with the AST size tests.
Any advice appreciated!
Yes, along with what you want it to do.
> I also pay for github copilot, but not GPT-4. Can I use copilot for this?
Not that I know of. CoPilot is good at generating new code but can't change existing code.
This, and people generally saying that chatGPT has been intentionally degraded, are just super strange for me. I believe it’s happening but it’s making me question my sanity. What am I doing to get decent outputs? Am I simply not as picky? I treat every conversion as though it needs to be vetted because it does regardless of how good the model is. I only trust output from the model that I am a subject matter expert on or in a closely adjacent field. Otherwise I treat it much like an internet comment - useful for surfacing curiosities but requires vetting.
Why this instead of GPT-4 through the web app? And how do you actually use it for coding, do you copy and paste your question into a python script, which then calls the OpenAI API and spits out the response?
If we were all flipping coins there would be people claiming that coins only come up tails. There would be nothing they were doing though to make the coin come up tails. That is just the streak they had.
Some days I get lucky with chatGPT4 and some days I don't.
It is also ridiculous how we talk about this as if all subjects and context presented to chatGPT4 are going to be uniform in output. One word difference in your own prompt might change things completely while trying to accomplish exactly the same thing. Now scale that to all the people talking about chatGPT with everyone using it for something different.
nowadays the autogpt gives up sooner, seems less competent, and doesnt even come close to solving the same problems
EDIT: This of course applies only if you’re using the UI. Using the API is the same.
public someComplexLogic() { // Complex logic goes here }
or another example when the code is long (ex: asking it to create a vue component) is that it will just add a comment saying the rest of the code goes here.
So you could test for it by asking it to create long/complex code and then running the output against unit tests that you created.
- Can you do XXX (something complex) ?
- Yes of course, to do XXX, you need to implement XXX, and then you are good, here is how you can do:
int main(int argc, char **argv) {
/* add your implementation here */
}It wouldn't be the top comment if it wasn't