Ask HN: Best AI Code Assistant?

[surrTurr]

Hi HN.

I have been using JetBrains AI and GitHub Copilot over the last few months, as well as some lesser known ones like GitLab Duo and Tabnine.

In my testing I found:

* GitHub Copilot: most stable, highest quality code completions; however, the completions seem to have gotten worse over the months (using since many months now) * Intellij AI: Interesting IDE integrations, however the code completion often simply does not work * GitLab Duo: Only quickly checked out the code completion 2 months ago, quality was way worse than GitHub Copilot * Tabnine: Interesting features (& control), however quality of the generated code also is not on the level of GitHub Copilot

What are your experiences so far? Do you have any recommendations? Especially interesting would be solutions that work on large codebases, offer fine tuning or knowledge base integrations (e.g. confluence)

[yaro-works]

Try Supermaven. Suggestions are much faster than any copilot, and it has the largest context of your project. A 1 million token context window. Also, you get access to GPT-4o, Claude 3.5 Sonnet, and GPT-4 out of the box

[_andrei_ → yaro-works]

Not sure why this was downvoted, Supermaven is miles ahead of everyone else, and it was created by the same people that made Tabnine (which was the first of its kind, and that they sold later).

[mstaoru]

I tried Supermaven, and while it is certainly fast, it is more distracting to me since I have to read the suggestion, find it wrong, and type by myself (Python). Our FE guys who are doing React find it much more useful, probably because of the sheer amount of repetitive boilerplate React and JS require.

[alwinaugustin]

Most of these AI assistants are still works in progress. Sometimes, they provide correct suggestions; other times, they do not. This is understandable, as generative AI still requires significant improvement in understanding context and predicting the most appropriate design, architecture, and solutions. These tools can be more time-consuming than manually editing or refactoring the code. I use these tools when I find it difficult to understand certain logic, as they explain it well. However, we still have to wait to generate production-ready code.

[nreece]

Been hearing great things about Cursor, and its been impressive in my few days of use.

https://www.cursor.com

[yaro-works → nreece]

Cursor has 1,883 ms latency while generating hints, while GitHub Copilot has 783 ms, and the fastest one, Supermaven, has 250 ms. Cursor is nice, but it's really slow compared to its competitors.

[muzani → yaro-works]

That's like saying an airplane has lower acceleration lol. It saves hours in copy-paste, and even if we're just talking boilerplate, it writes code throughout the screen, instead of just at the cursor (ironically).

It takes more than 2 seconds to figure out where to paste code and often longer to debug it if you try to do it below 2 seconds. Cursor just does this on a CMD+Enter.

It also runs GPT-4o and Claude, so the suggestions are much higher quality than Copilot.

[yaro-works → muzani]

I'm a fan of Supermaven, and speed does make a difference. In our team, most of the devs switched to Supermaven after seeing how quickly and concisely it works. I used to work with Cursor, but after trying Supermaven, which has the largest context among any Copilots, I found it instantly reads the context of my project and suggests code in my project's style.

From what I've found on their blog (https://www.cursor.com/blog/problems-2023), Cursor's context is 500k tokens, which is much better than GitHub Copilot's 8k tokens, but still not a competitor for Supermaven's 1m tokens. So in the end, Supermaven learns from your project. The more you write, the more it actually matches your code style.

[muzani → yaro-works]

Thanks, I tried Supermaven. I'd still rank it at second after Cursor. Supermaven seems to be a good fit for those who are into the flow.

But I can't get any flow thanks to clean architecture, so Cursor works - it modifies the whole pipeline based on your feature.

[mraza007 → yaro-works]

If that’s the case I’m definitely going to try supermaven,

I have been a cursor for past 3 months and I’m loving it so far but you are right at times it does get slow

[bubblebeard]

I’ve looked at several alternatives over the past year and Copilot is the best I’ve seen so far overall when you factor in ease of use, price to performance etc.

There may be services which generates code better, but at the end of the day you still need to review and touch up anything that is generated.

Copilots chat integration in Jetbrain IDE’s has also been great at helping me discover language features/functions, especially when learning new languages.

[th3w3bmast3r]

I tried JetBrains AI and Github Copilot and ended up just removing them altogether.

The autocomplete when it works is wonderful but more often than not, I need to delete and type it again. Most of the time I know what I want to type and not using AI saved me the time it took me delete and try again.

I did like the chat window but it wasn't enough for me to keep it. Maybe you can't teach this old dog new tricks :(

[jaymee]

I'm using Codium VScode extension. It's free and it is doing its job.

[dualogy → jaymee]

Don't you mean "Codeium"? Because "Codium" is a vscode fork, as such itself not dealing with AI.

[purple-leafy]

I’ve completely turned copilot off, its suggestions are usually dogwater. The thing I miss most is when it would sometimes correctly write imports

[nailer]

I vote against Github copilot. The context window is too small for it to do anything meaningful.