Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Gitlab’s AI-assisted code suggestions

about.gitlab.com

Post reply on HN
Print view View original

151–160 of 160 posts

Re: Gitlab’s AI-assisted code suggestions

#151

Post by jgalt212 »

tallytarik wrote: post #131

GitLab was rather quick to board the AI hype train this year. Fine - they don't want to look like they're behind the curve, have to please investors, and whatever else. But IMO there are plenty of other places to add real value across the GitLab product with AI/ML features. Here, it just looks like they saw GitHub do something and felt a need to copy it. But two years late, and worse. As a longtime GitLab user (and o…

> have to please investor

I think it's more an internal itch. For a persistent money loser, the stock in holding there pretty well.

https://finance.yahoo.com/quote/GTLB?p=GTLB&.tsrc=fin-srch

Re: Gitlab’s AI-assisted code suggestions

#152

Post by koolba »

visarga wrote: post #136
jedbrown wrote: post #29

Earlier quoted context omitted.

Even MIT licensed code requires you to preserve the copyright and permission notice. If a human did what these language models are doing (output derivative works with the copyright and license stripped), it would be a license violation. When humans want to create a new implementation with clean IP, they have one team study the IP-encumbered code and write a spec, then a different team writes a new implementation acco…

Why can't AI do the same: copyrighted code -> spec -> generated code. ... and then execute copyrighted code -> trace resulting values -> tests for new code. AI could do clean room reimplementation of any code to beef up the training set. It can also make sure the new code is different from the old code at ngram-level, so even by chance it should not look the same. Would that hold up in court? Is it copyright launderi…

Isn’t the language model itself the spec?

Potentially for all of the inputs at once.

Re: Gitlab’s AI-assisted code suggestions

#153

Post by al_be_back »

Their service calls Google Vertex AI Codey APIs to generate suggestions [1].

It's fine to integrate Off-the-shelf solutions, but you don't have to implement every tech-hype going viral. I fear AI-assistant features are pushing more prescient features further down the backlog.

[1] https://docs.gitlab.com/ee/user/project/repository/code_sugg...

Re: Gitlab’s AI-assisted code suggestions

#154

Post by roflyear »

lolinder wrote: post #77
roflyear wrote: post #60

Earlier quoted context omitted.

People love to cargo-cult. It's wild. It was okay but I'm happier with ChatGPT for one-off things I don't know how to do "give me a sql query that does this obscure thing" etc..

It's okay that it didn't fit well into your workflow, but accusing those who use it of cargo culting isn't constructive. Other people don't have the same workflow as you do, and as simonw noted above, some of us actually adjusted our workflow to get the most out of Copilot.

Not everyone who used it cargo-culted but certainly some did.

Re: Gitlab’s AI-assisted code suggestions

#155

Post by jedbrown »

visarga wrote: post #136
jedbrown wrote: post #29

Earlier quoted context omitted.

Even MIT licensed code requires you to preserve the copyright and permission notice. If a human did what these language models are doing (output derivative works with the copyright and license stripped), it would be a license violation. When humans want to create a new implementation with clean IP, they have one team study the IP-encumbered code and write a spec, then a different team writes a new implementation acco…

Why can't AI do the same: copyrighted code -> spec -> generated code. ... and then execute copyrighted code -> trace resulting values -> tests for new code. AI could do clean room reimplementation of any code to beef up the training set. It can also make sure the new code is different from the old code at ngram-level, so even by chance it should not look the same. Would that hold up in court? Is it copyright launderi…

Language models don't understand anything, they just manipulate tokens. It is a much harder task to write a spec (that humans and courts can review if needed to determine is not infringement) and (with a separately trained tool) implement the spec. The tech just isn't ready and it's not clear that language models will ever get there.

What language models could do easily is to obfuscate better so the license violation is harder to prove. That's behavior laundering -- no amount of human obfuscation (e.g., synonym substitution, renaming variables, swapping out control structures) can turn a plagiarized work into one that isn't. If we (via regulators and courts) let the Altmans of the world pull their stunt, they're going to end up with a government-protected monopoly on plagiarism-laundering.

Re: Gitlab’s AI-assisted code suggestions

#156

Post by dustyharddrive »

cassianoleal wrote: post #8
dgrin91 wrote: post #2

I wish them luck. While I'm not a huge fan of copilot coding AIs, I understand how they are going to be a killer feature for many and I worry that it would create a big moat for Github. I hope that Gitlab is able to close that gap because I really like Gitlab.

Codeium [0] seems like a worthy competitor to Copilot already. [0] https://codeium.com/

I’m spooked by this comment: https://news.ycombinator.com/item?id=34434995

Re: Gitlab’s AI-assisted code suggestions

#157

Post by moonchrome »

cassianoleal wrote: post #8
dgrin91 wrote: post #2

I wish them luck. While I'm not a huge fan of copilot coding AIs, I understand how they are going to be a killer feature for many and I worry that it would create a big moat for Github. I hope that Gitlab is able to close that gap because I really like Gitlab.

Codeium [0] seems like a worthy competitor to Copilot already. [0] https://codeium.com/

I've tried it for a few days a month ago and it was noticably worse than copilot, copilot gave much more one shot completions sooner while codeium was line by line on same code.

Way better than the AWS code whisperer or whatever it's called but still not worth switching. Especially since I trust GH way more than some random company (GH already has access to my code anyway)

Re: Gitlab’s AI-assisted code suggestions

#158

Post by ColonelPhantom »

moonchrome wrote: post #157
cassianoleal wrote: post #8

Earlier quoted context omitted.

Codeium [0] seems like a worthy competitor to Copilot already. [0] https://codeium.com/

I've tried it for a few days a month ago and it was noticably worse than copilot, copilot gave much more one shot completions sooner while codeium was line by line on same code. Way better than the AWS code whisperer or whatever it's called but still not worth switching. Especially since I trust GH way more than some random company (GH already has access to my code anyway)

I should try Codeium then, Copilot's occasional "heap of code" was one of the most annoying parts of using it to me. Why? Because it looks plausible, so I accept it without actually understanding it.

But using it as a better autocomplete is actually quite nice, for example when I have some more repetitive code to write. And a line-by-line AI completer will at least give me the time to read it piecemeal, which I find much easier. Just like I understood way better when my math teacher demonstrated how to solve a problem compared to reading it from the textbook.

Re: Gitlab’s AI-assisted code suggestions

#159

Post by ColonelPhantom »

DrakeDeaton wrote: post #143
lostmsu wrote: post #85

Earlier quoted context omitted.

Just today Copilot saved me close to 1 hour of typing and replaced it essentially with ~1 minute of pressing Tab. I was hooking IAudioClient COM class to capture and silence arbitrary app's audio, and as soon as I wrote signatures for its members Copilot was able to generate skeletal stub implementation with logging as well as hooking code totaling about 150 lines of Rust.

This is an example of why codebases with strict typing & clear conventions will benefit the most from copilot.

It would be even better if Copilot was able to make use of the strict typing to prevent generating code that doesn't typecheck in the first place, to be honest. Copilot as-is actually feels more useful for dynamic languages, because static languages get really far without any LLMs and simple static autocomplete gets you really far. So there's more ground to make up.

In addition, Copilot benefits from a large training set, which I guess is best for Python and JavaScript as two very popular languages. Strictly typed languages seem like they are generally less well represented, at least among public GitHub projects.

Re: Gitlab’s AI-assisted code suggestions

#160

Post by moonchrome »

ColonelPhantom wrote: post #158
moonchrome wrote: post #157

Earlier quoted context omitted.

I've tried it for a few days a month ago and it was noticably worse than copilot, copilot gave much more one shot completions sooner while codeium was line by line on same code. Way better than the AWS code whisperer or whatever it's called but still not worth switching. Especially since I trust GH way more than some random company (GH already has access to my code anyway)

I should try Codeium then, Copilot's occasional "heap of code" was one of the most annoying parts of using it to me. Why? Because it looks plausible, so I accept it without actually understanding it. But using it as a better autocomplete is actually quite nice, for example when I have some more repetitive code to write. And a line-by-line AI completer will at least give me the time to read it piecemeal, which I find…

I'm already in the flow with copilot so I'm pretty good at expecting when it's useful. Basically I use it as AI copy paste. When I have variations of methods or conditional logic copilot is great at one shot solutions and it's trivial to verify (if it's wrong somehow then I would have done the wrong thing as well - it just saves me a minute of copy/paste and refactoring) copilot was just better.
Post reply on HN