Rewrite Code in Real Time with Zed AI's Fast Edit Model

[@jasonshen]

honestly speed isn't a important thing in these ai assistants its the accuracy that matters, but nice video to show all this stuff

[@g0t4]

Human-in-the-loop tooling still feels like the best use case for LLMs. In that context, speed is important (too)... otherwise I'm just gonna write the code myself 😊... also, I can't stand waiting for stuff even if it's fast but feels slow nonetheless (i.e. initial lag). In the future, if we're fortunate to develop viable, autonomous "code agents", then I can see speed becoming much less relevant. Perhaps for novice engineers, speed is less relevant too. Arguably this is a step in the direction of autonomous "code agents"... inversely proportional to the skillset (or lack thereof) of the human-in-the-loop.

[@neonnftz]

This is something i always wanted.. love it!

[@g0t4]

Agreed, speed is a huge impediment to doing much beyond generating new code... but, fast rewriting will unlock an entirely new IDE experience

[@riccardoesclapon549]

Is this still in private beta? I don't see the option, would love to be able to pay for that feature

[@g0t4]

I believe it is, last I tried a few weeks ago it was there. Maybe reach out on twitter (x.com/zeddotdev) and see if they are adding new people

[@sshvulnerability]

Does fast edit reduce the tokens sent back from the api or is it just quicker to respond?

[@g0t4]

Good info here, though I don't recall finding that answer in my previous digging: zed.dev/blog/zed-ai

[@g0t4]

IIGC/IIRC the model is trained to return some sort of diff/patch format to only modify lines that need to be altered, from that blog post: "custom streaming diff protocol that works with Zed's CRDT-based buffers to deliver edits as soon as they're streamed from the model."

[@g0t4]

it's even more interesting when you throw in Prompt Caching too: www.anthropic.com/news/prompt-caching Caching frequently used context (i.e. common project files) That said, at some point, all of these features need to just work and be completely transparent to users.

[@g0t4]

And this all makes me wonder, can I intercept the request/response and find out for sure :) ... that's worth a video if I can.

[@g0t4]

darn... their Fast Edit model isn't working :(... I'll see if I can get access to an updated model. I was able to capture non Fast Edit models w/ mitmproxy

[@Kunal_Rai]

Amazing

[@boxtraders7382]

awesome

[@Cube_Box]

Is the fast mode still available?

[@g0t4]

@@Cube_Box yup, was just using it yesterday to rewrite some text, but it’s still in a private beta AFAIK