"it mostly worked" is just a more nuanced way of saying "it didn't work".
Apparently the author did eventually get something working, but it is false to say that the LLMs produced a working project.
Well, yeah. It’s a more nuanced way of saying that because “it didn’t work” isn’t very useful nor descriptive.
What if it wrote all of the boilerplate and just let you focus on the important bit that deserves your scrutiny?
You could say I failed every single project I ever built because it took many iterations to get to the final deliverable stage with lots of errors along the way. So more nuance would be needed.
But when it comes to LLMs suddenly we get all gleeful about how negatively we can frame the experience. Even among HN tech scholars.
I tried to do this a few weeks ago, I tried to build a NIF around an existing C lib. I was using Claude Opus and burned over $300 (I didn't have Pro) on tokens with no usable results.
built my startup in elixir. love it but nifs are one of the few ways you can crash the VM. I don't trust myseld to write a nif in production. no way I'd do it with AI in c. Thank god theres projects like rustler which can catch panics before it crashes the main VM.
Why C instead of Rust or Zig? Rustler and Zigler exist.
I feel like a Vibecoded NIF in C is the absolute last thing I would want to expose the BEAM to.
I would never ever let an LLM anywhere near C code. If you need help from LLM to write a NIF that performs basic C calls to the OS, you probably can’t check if it’s safe. I mean, it needs at least to pass valgrind.
I've done this. The NIF worked as in that it ran and was a correct enough NIF. It did not work in terms of solving what I needed it to do. Iteration was a bit painful because it was tangled with a nasty library that needed to be cross-compiled. So when I made a change it seg faulted and I bailed.
I essentially ran out of patience and tried another approach. It involved an LLM running C code so I could check the library output compared to my implementation to make sure it was byte-for-byte.
The C will never ship. I don't have practice writing C so I am very inefficient at it. I read it okay. LLMs are pretty decent help for this type of scrap code.
It's interesting why the author used weaker models (like Grok 3 when 4 is available, and Gemini 2.5 Flash when Pro is), since the difference in coding quality between these models is significant, and results could be much better.
I once wrote a little generalized yaml templating processor in Python by using an LLM for assistance. It was working pretty well and passing a lot of the tests that I was throwing at it!
Then I noticed that some of the tests that failed were failing in really odd ways. Upon closer inspection, the generated processor had made lots of crazy assumptions about what it should be doing based upon specific values in yaml keys that were obviously unrelated to instructions.
Yeah, I agree with the author. This stuff can be incredibly useful, but it definitely isn't anything like an AGI in its current form.
So all this arose because you didn't read the docs and note that get_disk_info/1 immediately fetches the data when called? The every-30-minutes-by-default checks are for generating "disk usage is high" event conditions.
18 comments
[ 2.7 ms ] story [ 38.1 ms ] threadWhat if it wrote all of the boilerplate and just let you focus on the important bit that deserves your scrutiny?
You could say I failed every single project I ever built because it took many iterations to get to the final deliverable stage with lots of errors along the way. So more nuance would be needed.
But when it comes to LLMs suddenly we get all gleeful about how negatively we can frame the experience. Even among HN tech scholars.
I had a bunch of fun getting ChatGPT Code Interpreter to write (and compile and test) C extensions for SQLite last year: https://simonwillison.net/2024/Mar/23/building-c-extensions-...
I essentially ran out of patience and tried another approach. It involved an LLM running C code so I could check the library output compared to my implementation to make sure it was byte-for-byte.
The C will never ship. I don't have practice writing C so I am very inefficient at it. I read it okay. LLMs are pretty decent help for this type of scrap code.
Then I noticed that some of the tests that failed were failing in really odd ways. Upon closer inspection, the generated processor had made lots of crazy assumptions about what it should be doing based upon specific values in yaml keys that were obviously unrelated to instructions.
Yeah, I agree with the author. This stuff can be incredibly useful, but it definitely isn't anything like an AGI in its current form.