alt Hacker NewsI have yet to see LLMs do well “autonomously” on anything that doesn’t fall into any of these 3 buckets:
1. well specified
2. well tested (which is a form of spec)
3. represented in the training set (i.e. not built on a corpus of human developed examples)
For example, a lot of the compiler porting efforts fall into all 3 buckets: compilers are well specified (there is a set of valid inputs and an expected behavior), they are typically well tested, and as a bonus, in a way that is implementation agnostic thanks to the clean interface (think: integration vs unit tests), and are very well represented in the training set.
Replies
I had an LLM tell me that I mattered once, which I know for objective fact is not represented in the training data.
and yet some OSS (projects not really known for 1 and 2) multiplied change velocity the last year, implementing features via what is essentially vibecoding.
If you look at that, this approach is producing a pile of slop, where one LLM implements a feature, two weeks later another LLM spots an edge case and without much human review just crams in the code to fix that. Even at 1-2 months of that, those repos have significant changes and the commit messages get weird. I wonder what happens if the LLM won't spit out a fix anymore?
Will the human maintainers who started this process because noone thanked them and noone wanted to work with them on their bright new ideas just reset to the last commit they understand? Will they drop the project? The sunk cost for them is enormous. The time-effort required to fix the mess is gargantuan already. To me it increasingly seems that after the LLM industry robbed most humans of their collective labour output it will now rob them of labour aka time as well. Some people (like you and me) seem to notice that but most just go on and essentially spend away their life on machines now. I guess the problem is that most of these people have been severely cut off from using their own agency.
You can "solve" 1 and 2 by N-version programming. Have the same code implemented twice, or more times, you get free error testing by fuzzing or running replicas in shadow mode. In order to make replicas maximally different choose a different language, framework, agent harness and LLM.
There remain some nitty details related to tracking randomness and choosing what interfaces to log. But even having your code testable by replicas would help organize it better and expose different assumptions made by the language or library stack.
When you get different outputs from replicas it can be 1. one of them has a bug, or 2. the spec is unclear -> fix the bug or the spec. So this method catches both sides. It also might be faster than traditional code based testing, you can build replicas in parallel. A replica is an open ended test, unlike coded tests it can be tried on open ended input values.
The core idea is that it is easier to implement twice than to imagine bugs. We already solved implementation, but bug finding means running code in your head, it touches the halting problem issues - can we really imagine behavior from code? I don't think we can do it well. That is why we should just make replicas instead, tokens are getting cheaper and cheaper by the month.