The fact that AI dev tools haven't been wedded very well into language servers in IDEs after several years now is the strongest indicator out there that this tech still has a very long way to go, and we're overestimating its short-term impact while underestimating its long-term impact - ThreadSky

phillipcarter.dev • 12 days ago

The fact that AI dev tools haven't been wedded very well into language servers in IDEs after several years now is the strongest indicator out there that this tech still has a very long way to go, and we're overestimating its short-term impact while underestimating its long-term impact

Comments

mia-ddysgrl1.bsky.social•10 days ago

Even typing its not up on English vocabulary - always either “correcting” my word to something stupid or telling me it has no knowledge to offer.😶

kcieslak.io•11 days ago

People don’t listen to me 🤷

kcieslak.io•11 days ago

But FWIW I think Copilot is doing some of it - https://github.blog/changelog/2024-08-01-improving-github-copilot-completions-in-vs-code-for-c-developers/

And Agent mode is checking for errors after code was generated

brucejdragon.bsky.social•12 days ago

What would that look like in your view?

Asking as a compiler dev, trying to understand how my roadmap should change…

phillipcarter.dev•12 days ago

Some things I think about:

LLM generates code -> syntax and semantic checks -> feedback fed to LLM -> repeat until exit condition met

Write code -> LLM analyzes changes, suggests where to update -> user-LLM interaction kicks off

phillipcarter.dev•12 days ago

Refactor -> LLM kicks off probabilistic analysis to find "other worthy changes" -> kicks off some user-agent loop

LLM "tests" code by running it -> runs with debug mode, symbols on -> inspects real values for symbols, feeds back to LLM -> loop to emit more code (E&C), repeat

phillipcarter.dev•12 days ago

Basically I view a compiler service as a database of syntax and semantics about the codebase, and when augmented with project services and an embedding index, is such a wildly effective tool and making the LLM significantly more powerful within that codebase

brucejdragon.bsky.social•11 days ago

Seems like some of this is already happening via tool invocation. What’s the upside of pushing this into the language server? Faster interactions?

I see potential downsides to the source code no longer being the medium of shared “understanding” between human and LLM.

phillipcarter.dev•11 days ago

I think that with a language service as a database, you're now also playing in the space of "well what likely relevant data can I use for this problem?" and that's profoundly interesting to me.

But yeah, a lot of this could be done via tools as well. I'd argue that also has a long way to go!

Comments

Posting Rules

Reply