This is a very bold move. they are making a bet that LLM’s will get good enough where you won’t write code anymore, but instead guide the AI. As someone who uses copilot regularly, they are not there yet. If they do get there in 5 years, dark lang could be in a good spot. If they don’t, they will instead have occurred a huge opportunity cost of not shipping features people can actually use.
Though I disagree with the idea that we would be better off moving away from “text editor” IDE’s to “AI driven” IDE’s. One thing that makes LLM’s so useful is that you can generate text! They can already integrate well with the universal format. I see “AI powered” IDE’s being similar to structured editors (like darklangs) where it sounds superior in theory but in practice you just want a full featured text editor. All these AI features can be integrated into something like VsCode.
We shall see. Dark lang may prove me wrong. It would be interesting to see a future where the question isn’t “how good is the lsp server”, but instead “how good is the AI codegen”.
So this is funny because I was trying to figure out what Darklang is and what its used for, so i was looking at its doc. The page for usecases/limitations [0] says
> Dark is really good for building web applications, APIs, and things that speak HTTP. Dark is not intended to be usable for embedded systems, to run existing applications in existing languages, or for code that needs to be extremely high performance.
> Similarly, Dark does not have support for many use cases such as bitcoin/crypto, AI/big data, image or video manipulation.
lol.
Anyways, some bold claims in here.
> Do we expect that in 3-5 years time developers will still be typing out artisanal code at a keyboard? I struggle to see it. With recent announcements like ChatGPT Plugins and Copilot X, even three years feel generous
lol, that's funny. Looks like we're going to have to make it good for some of those things! In particular, AI uses a bunch of memory and specialized machines/GPUs if we're going to be training models or running our own, and our previous interpreter model (running on small-ish VMs with a few gigs of RAM) are clearly not suitable for that. We might end up using services for this though.
Though I disagree with the idea that we would be better off moving away from “text editor” IDE’s to “AI driven” IDE’s. One thing that makes LLM’s so useful is that you can generate text! They can already integrate well with the universal format. I see “AI powered” IDE’s being similar to structured editors (like darklangs) where it sounds superior in theory but in practice you just want a full featured text editor. All these AI features can be integrated into something like VsCode.
We shall see. Dark lang may prove me wrong. It would be interesting to see a future where the question isn’t “how good is the lsp server”, but instead “how good is the AI codegen”.