> The state of its Windows support is unclear to me.
yeah this one is a bit weird, you can run the linux binary using WSL and that should work. We have a window flavor build but its not as heavily tested yet (we are figuring out a better testing flow for windows)
hey I am the engineer who worked on the TUI and harness and happy to answer your questions.
We worked quite heavily on the TUI, its written in ratatui and we did try to go the extra mile to make sure that there won't be any regressions from moving to an alt-screen rendering. Lots and lots of small details to manage and get correct, we also tried to get vim keybindings correct while keeping it mouse friendly.
Hey I am the coredev on sidecar. The reason you see autogenerated PRs is cause I am using our agents to write the code for the agent lol
The big difference is the complete loop, each PR gets its own VM with the tool chains installed so the agent can run cargo check or cargo tests etc.
We do find the LLMs of today are not the best elite engineers but very very competent junior engineers. It's been a weird but eye opening workflow to use.
ngl the total expenditure was around $10k, in terms of test-time compute we ran upto 20X agents on the same problem to first understand if the bitter lesson paradigm of "scale is the answer" really holds true.
The final submission which we did ran 5X agents and the decider was based on mean average score of the rewards, per problem the cost was around $20
We are going to push this scaling paradigm a bit more, my honest gut feeling is that swe-bench as a benchmark is prime for saturation real soon
1. These problem statements are in the training data for the LLMs
2. Brute-forcing the answer the way we are doing works and we just proved it, so someone is going to take a better stab at it real soon
will keep that in mind going forward! The most recent ones which I have been hitting are around "higher-ranked lifetime error"
I know my way around this now, which is to literally binary search over the timeline of my edits (commenting out code and then reintroducing it) to see what causes the compiler to trip over (there might be better ways to debug this, and I am all ears)
Most of the times this error is several layers deep in my application so even tho I want to ticket it up, not being able to create a minimal repo for anyone to iterate against feels like a bit of wasted energy on all sides, do let me know if I should change this way of thinking and I can promise myself to start being more proactive.
Rust can be hard to get right because of the borrow checker.
I had a similar situation happen to me where I went about refactoring the code to make borrow checker happy ... until the last bit when things stopped compiling and I realized my approach was completely wrong (in the rust world, I had a self-reference in the structs)
Having said this, the benefits of borrow checker out weight the shortcomings. I can feel myself writing better code in other languages (I tend to think about the layout and the mutability and lifetimes upfront more now)
My rust code now is very functional, which seems to work best with lifetimes.
I would love to know more about the authors pain, I do hope rustc gets better at lifetime compilation errors cause some of them can be very very gnarly.
I am missing the link to the thread, but diffusion models also give a very consistent output when prompted with `IMG_{number}` part of the reason could be the training data distribution
Thats a fair point, a significant part of our 4 person team had to skill up on the VSCode codebase to be able to meaningfully make changes to it.
I would love to know your workflow, you mention CLI tool or VSCode plugin, which one of them work for you? Whats missing from them where Aide can fill in the gap
There is pinned context on the very top where you can pin the files which you frequently use.
We will start including the open file by default in the context very soon (one of the gotchas here, is that the open file could not be related to the question you have)
that's a good metric to aim for... creating a full local index for 600k lines is pretty expensive but there are a bunch of huristics which can take us pretty far
- looking at git commits
- making use of recently accesses files
- keyword search
If I set these constraints and allow for maybe around 2 LLM round trips we can get pretty far in terms of performance.
JetBrains is very interesting, what are the best performing extensions out there for it?
I do wonder what api level access do we get over there as well. For sidecar to run, we need LSP + a web/panel for the ux part (deeper editor layer like undo and redo stack access will also be cool but not totally necessary)
ohh inserting.. I tried it on couple of big repos and it was a bit of a miss to me. How large are the codebases on which you work? I want to get a sense check on where the behavior detoriates with embedding + gpt3.5 based reranker search (not sure if they are doing more now!)
yeah this one is a bit weird, you can run the linux binary using WSL and that should work. We have a window flavor build but its not as heavily tested yet (we are figuring out a better testing flow for windows)