HN user
history on this: this is something that aman was sort of challenged to do in a weekend as a sort of bet. it didnt end up working in that weekend but ended up working out overall! the post is quite fun as it walks through why the problem is hard.
its also important because it is a key primitive for doing good reliable multi-step code systems :)
one of the devs here, this is really magical when you actually try it. if you have tried to get a model to edit code in a large file reliably, you will know how painful this is. actually getting the model to do it reliably makes it possible to do more complicated work in the background :) so really quite excited for this.
a founder from anysphere here. while we are funded by openai startup fund, we have no other formal relations by them. we love oss models and are exploring how we can use them for amazing code products. the post is our honest reflection into the tradeoffs we see :)
The commenting was for the model to learn what the types were so it could translate it to gather. Doing that one shot is surprising to me because it is somewhat hard to get pytorch tensors to work out to the correct shapes.
GPT-4 certainly is aware of go generics. Do you have an example you would like to test?
I think companies are getting very receptive to uploading code with the new 30-day policy. We have been able to negotiate this for financial firms. Reach out and we might be able to make something happen.
OP here. We're hiring! We're an early-stage startup, with funding from OpenAI, building the next-generation AI code editor. If you're the best hacker or designer ever, reach out to me at sualeh@anysphere.co. The problems we work on are really fun (extremely fast context retrieval from across big codebases, connecting language servers to language models, instruction-tuned embedddings) and we think the potential of what we're building is enormous (make all software engineers literally 10x more productive).
Feel free to email us, and we can make something work!
We are designing the IDE for enterprise users actually. I.e. we can do an on prem deployment.
We are trying to build this right in to the editor. The UI is in the tweet thread. Feel free to sign up at control.dev
It definitely should be able to. Any example you would like me to try. Put in an issue and we will do it.
Definitely! Happy to try it if you have a prompt.
email us at sualeh@anysphere.co :) We would love for you try it out.
I am on this side of the argument. It is letting me code better than ever before :) I have been able to pick up new languages/build systems much faster than in the past (Swift for example.)
Yeah, its really cool how good it is. I personally am addicted to it.
Definitely.
For e.g. you cannot say go to VSCode and add this feature. But, If we point it to the piece of code and ask it to append it and write tests, it does really well. (We want engineers to have control haha.)
And they delete completions within 30 days.
If you know of any other cool examples or would like us to try something out and add to the repo, please let us know.
If you have access, the repo should be fairly replicable.
GPT-4 is just really good at code. We documented some of the examples that we found were crazy amazing and worth hackernews. Including GPT-4's ability to handle bazel, kube, terraform and rust. A summary can be found on this thread: https://twitter.com/sualehasif996/status/1635755267739598848
I appreciate the caution. We are funded by the FTX Future Fund (https://ftxfuturefund.org/our-grants/?_search=Anysphere). Feel free to look through our github (which is why we are open source).