Has anyone tried to rewrite some popular open source project with IA? I imagine modern LLMs can be very effective at license-washing/plagiarizing dependencies, it could be an interesting new benchmark too
As the author, it's a stretch to say that JustHTML is a port of html5ever. While you're right that this was part of the initial prompt, the code is very different, which is typically not what counts as "port". Your mileage may wary.
Interesting, IIUC the transformer architecture / attention mechanism were initially designed for use in the language translation domain. Maybe after peeling back a few layers, that's still all they're really doing.
This has long been how I have explained LLMs to non-technical people: text transformation engines. To some extent, many common, tedious, activities basically constitute a transformation of text into one well known form from another (even some kinds of reasoning are this) and so LLMs are very useful. But they just transform text between well known forms.
And while it appears that lots of problems can be contorted into translation, "if all you have is a hammer, everything looks like a nail". Maybe we do hit a brick wall unless we can come up with a model that more closely aligns with actual human reasoning.
Note that it's not clear that any of the JustHTML ports were actually ports per se, as in the end they all ended up with very different implementations. Instead, it might just be that an LLM generated roughly the same library several different times.
Not me personally, but a GitHub user wrote a replacement for Go's regexp library that was "up to 3-3000x+ faster than stdlib": https://github.com/coregx/coregex ... at first I was impressed, so started testing it and reporting bugs, but as soon as I ran my own benchmarks, it all fell apart (https://github.com/coregx/coregex/issues/29). After some mostly-bot updates, that issue was closed. But someone else opened a very similar one recently (https://github.com/coregx/coregex/issues/79) -- same deal, "actually, it's slower than the stdlib in my tests". Basically AI slop with poor tests, poor benchmarks, and way oversold. How he's positioning these projects is the problematic bit, I reckon, not the use of AI.
Same user did a similar thing by creating an AWK interpreter written in Go using LLMs: https://github.com/kolkov/uawk -- as the creator of (I think?) the only AWK interpreter written in Go (https://github.com/benhoyt/goawk), I was curious. It turns out that if there's only one item in the training data (GoAWK), AI likes to copy and paste freely from the original. But again, it's poorly tested and poorly benchmarked.
I just don't see how one can get quality like this, without being realistic about code review, testing, and benchmarking.
Note that this is semantically exactly equivalent to "up to 3000x faster than stdlib" and doesn't actually claim any particular actual speedup since "up to" denotes an upper bound, not a lower bound or expected value. It’s standard misleading-but-not-technically-false marketing language to create a false impression because people tend to focus on the number and ignore the "up to".
Saying “up to” means that bound is the maximum value of the data set. It may be far from the median value, but it is included (or you’re lying). With any other interpretation the phrase has no meaning whatsoever.
I will concede, proactively, that "up to" could refer to some maximum possible bound, even if the current set doesn't include a value at that bound, though I would argue that's likely deceptive wording. For example, you could say that each carton of of eggs on a pallet contains up to 12 eggs, because that's the maximum capacity of the carton, even if none of the actual cartons on this pallet actually have 12 eggs in them.
I used one of the assistants to reverse and rewrite a browser-hosted JS game-like app to desktop Rust. It required a lot of steering but it was pretty useful.
