4 ms·
Even just the ability to reproduce a non-trivial bug, and then work out what general area of the codebase might be in seems out of reach for an LLM
by lokar 2y ago
Even just the ability to reproduce a non-trivial bug, and then work out what general area of the codebase might be in seems out of reach for an LLM
- skeptrune 2y agoYes! I think I'm going to make LLM bug reproduction my primary response to AI developer maximalists.
- beeflet 2y agoMaybe the solution isn't to train LLMs on human-written code, but to train some sort of text-to-AST generator on descriptions of programs and then the output of programs. That way it doesn't learn to include human-inspired bugs at least.
- mnky9800n 2y agoI always wondered why bother having an AI generate human readable code. Isn't it a machine? Why not have it generate machine code? of course, it wouldn't be readable, but would it be performant? Maybe. I guess my point is that if an AI is doing computing why would it doing computing resemble a person doing it? It's much closer to the machine then me. Why not take advantage of that?
- delapid 2y agoIsn’t that kind of what neural networks are in a way? Large blobs of computation we cant really parse fully
- mnky9800n 2y agoYea but there are a bunch of layers, lol, between that specification and the machine and I'm wondering if there's a way to reduce that assuming it's a bottleneck. But I'm not sure it is I'm just guessing.
- addandsubtract 2y agoBecause then you'd have no control over what the AI/code does. If you blindly want to trust the AI, you might as well just make the AI the black box between your input and output.
- mnky9800n 2y agoHaha what do you think people do with AI/ML? do you think that's currently in control?
- beeflet 2y agoThis actually gives me a funny idea, IDK if it would actually work. Something interesting you can do with LLMs is you can constrain their output. What if you just write a test case to encapsulate the bug then try to filter out the most "sane" LLM output that results in the code compiling and the test case passing? It doesn't seem like much, but it's like using more computing power to filter out all the "obviously wrong" solutions. Doing this in practice might require you to make too many test cases though.
- yoav 2y agoYou just invented a mid-level engineer. If you manage to have more breakthroughs to senior then you can maybe leave it unattended for a few months but not if the product team is inexperienced or strong willed. The code will still end up as unmaintainable spaghetti. So you need more breakthroughs to achieve staff or higher which requires a lot more agency and initiative on the part of the bot to push past what your team mates think they want and do what actually needs to be done modifying not only the code and the requirements but the process that leads to miscommunication and inefficiencies within the org. If you manage to invent a bot that knows what’s better for the people prompting it and acts on that then you’ve potentially also violated the first law of robotics and created a death bot. And boy oh boy would it be astronomically expensive to run.
- sureIy 2y agoThe instructions seem to have to be so specific and repeated that you need a full-time person to review and retry. My last struggle was creating an SVG arrow pointing to a specific corner. Both Claude and CGPT failed with many attempts.
- fragmede 2y agothe skill is in having it expose enough of the details that you only need to do a bit of messing with the output by hand and then your can be done with the ticket. there's no points for getting the LLM to spit out perfect code for you, the point is to get working code checked into git.
- sureIy 2y agoThis thread is about LLM replacing people, so yes you're confirming that you do need a skilled person to get working code out of it.