Ok, bye bye [LWN.net]

Ok, bye bye

Posted Jul 23, 2026 20:50 UTC (Thu) by dskoll (subscriber, #1630)
In reply to: Ok, bye bye by mb
Parent article: Codeberg: Protecting our FLOSS commons from LLMs

Soooo, if I had typed in exactly the same code in via keyboard, it would be fine?

Yep.

And the reason I say 100% AI-generated code should be treated as object code is that if it is indeed 100% AI-generated, then no human needs to look at it.

AI tools are non-deterministic.

So you think that Codeberg should waste its resources hosting code that's essentially spat out by a fancy random-seeded Markov-chain generator?

Just commit the prompt and the random seed and the non-determinism is eliminated, assuming the LLM you're using lets you get at the seed for reproducibility purposes.

to post comments

Ok, bye bye

Posted Jul 23, 2026 21:02 UTC (Thu) by mb (subscriber, #50428) [Link] (11 responses)

Well, quite honestly, today's AIs produce much higher quality code than Joe Average programmer.
I would rather trust AI generated code than code from an average programmer.

And I say that, because I looked at lots and lots of generated code.
And I used AI agents to find bugs in handwritten code.

One of my tasks in my paid day-job is to use modern AI agents to track down actual bugs in human written code that you probably trust daily, if you leave your home. A malfunction in this code can cause damage, accidents and serious injuries.
And you know what: AI agents do understand this code *MUCH* better than Joe Average 30-years-of-experience programmers.
AI agents *reliably* find actual problems in handwritten and highly tested code.

And the generated code, if you use AIs in that mode of operation, is equally good, if you use the tool in the right way.

Ok, bye bye

Posted Jul 23, 2026 21:04 UTC (Thu) by dskoll (subscriber, #1630) [Link] (10 responses)

But again, that doesn't address why you need revision-control for AI-generated code. The code is a generated output, not an input.

Ok, bye bye

Posted Jul 23, 2026 21:11 UTC (Thu) by mb (subscriber, #50428) [Link] (8 responses)

I do not get your point.
Handwritten source code is also an output. It's an output of my keyboard.
Source code needs version control.

To say it again: Review is an essential step of AI based code generation. For this alone version control of the code is needed. How else are you going to bisect a regression?
There is no difference to handwritten code.

Ok, bye bye

Posted Jul 23, 2026 22:14 UTC (Thu) by dskoll (subscriber, #1630) [Link] (7 responses)

The GNU GPL has a workable definition of source code:

The "source code" for a work means the preferred form of the work for making modifications to it.

In a 100% AI-generated project, the preferred form of the work for making modifications to it is tweaking the AI prompt.

Review is an essential step of AI based code generation

If you mean human review, then it's not a 100% AI-generated project. Then it's 90% or whatever.

Ok, bye bye

Posted Jul 24, 2026 0:34 UTC (Fri) by mjg59 (subscriber, #23239) [Link] (5 responses)

> In a 100% AI-generated project, the preferred form of the work for making modifications to it is tweaking the AI prompt.

No, I'm perfectly happy taking what's currently a 100% AI generated project and turning it into a 99.9% project. But also from a practical perspective, vendors don't host models forever and hosting a prompt for a model I don't have access to isn't helpful.

Ok, bye bye

Posted Jul 24, 2026 1:00 UTC (Fri) by dskoll (subscriber, #1630) [Link] (4 responses)

Isn't relying on vendors who host models a little dangerous, though? That's like using a build machine that anyone can change whenever they want. They can also change the pricing on a whim. So something that works pretty well for you one day might be much worse the next, or much more expensive.

Yeah, I was somewhat trolling by saying that all you need to version-control is the prompt, but that was in response to OP's claim of a "100% AI" project.

Anyway. This is a pretty divisive issue, and IMO if Codeberg wants to refuse to host mostly-AI projects, that's its right. I happen to agree with Codeberg, but those who don't agree have plenty of other free options.

Ok, bye bye

Posted Jul 24, 2026 4:20 UTC (Fri) by mjg59 (subscriber, #23239) [Link]

> Isn't relying on vendors who host models a little dangerous, though?

Well, yes, which is why it makes sense to provide the source code rather than just a prompt.

Ok, bye bye

Posted Jul 24, 2026 6:10 UTC (Fri) by Cyberax (✭ supporter ✭, #52523) [Link]

The world of AI development is in a weird state right now. There were attempts to just use prompts and keep the source code as an ephemeral artifact, but they are too impractical for large projects. It's like if every compiler run cost you $100-$10000.

And committing the entire dialog with the LLM is kinda like committing a recording of a GDB debugging session.

Ok, bye bye

Posted Jul 24, 2026 6:23 UTC (Fri) by mb (subscriber, #50428) [Link] (1 responses)

See, we can't even agree on what "100% AI" means, right?
For me: Thing comes out of AI agent -> It's 100% AI.
For you: Thing that comes out of an AI agent is reviewed without a change -> It's not 100% AI anymore.

The repository state is unknown (how much AI is in there?).
The platform thresholds for permissible amounts of AI is unknown or secret.

And under these circumstances I am supposed to agree to Codeberg's ToS?
Sorry, thanks but no.

>Isn't relying on vendors who host models a little dangerous, though? That's like using a build machine that anyone can change whenever they want. They can also change the pricing on a whim. So something that works pretty well for you one day might be much worse the next, or much more expensive.

Sure. But why is that your problem, the forge's problem or anybody else's problem? It's my problem.

Ok, bye bye

Posted Jul 24, 2026 12:33 UTC (Fri) by antacon (subscriber, #138885) [Link]

In a practical sense it's Codeberg's problem because while you can afford to generate code you'd otherwise not be producing you are flooding Codeberg with hosting costs. Multiply your impact by however many other users that do the same thing and those demands go higher, faster.

That in turn becomes every user's problem since they all enjoy free hosting currently but costs may eventually require paid subscriptions. The forge may shutdown if costs can't be kept manageable.

Here is the Codeberg 2026 budget - https://blog.codeberg.org/codebergs-budget-of-2026.html

I'd advise you to find LLM friendly forges in the meantime and set up remotes with them for your projects. At least this will be a good exercise for you in redundancies.

You can see the bigger brief of the decision here - https://blog.codeberg.org/protecting-our-floss-commons-fr...

You can also try joining the org to be more involved in these decisions - https://codeberg.org/Codeberg-e.V.

Ok, bye bye

Posted Jul 26, 2026 22:26 UTC (Sun) by MarcB (subscriber, #101804) [Link]

> In a 100% AI-generated project, the preferred form of the work for making modifications to it is tweaking the AI prompt.

Not really. The prompt is not what an AI uses to generate code, it is its context - of which the prompt(s) are just a part. Potentially a small one.
The context also contains documentation the agent read, specifications you provided, web search results, sample code you pointed it to, whatever tool calls your agent did, ... and most importantly: also the results of the AIs internal reasoning, which are in turn based on the previous context.

In other words: the pure prompt is essentially worthless for any non-trivial software.

To give you a recent example: I wrote a prompt that said "reimplent the ancient, unmaintainable Java micro-service at ../here_be_dragons in Perl Mojolicious, drop feature A and B, only keep C and D, preserve the exact API and get the config from a YAML file instead of properties and a database; see dump.sql". The prompt was maybe 1% of the context the AI used to generate the code, the rest was the old code, including its tests and the DB dump.

For more serious development, you'd typically use something like "spec-driven development" (not to be confused with the pre-AI thing of exactly the same name). This has a set of rules the AI follows and you use this to create a specification, in multiple steps and iterations. In the end, you get a list of requirements, rationale and tasks that contain code generation and testing; usually in the form of a large tree of markdown files.

This can be committed, and iterated on, on its own, but typically it is committed with any code that was already created. It is simply far to expensive and slow to regenerate the code from spec, and even then it is not deterministic; especially if the spec was not clean (i.e. there were additional, significant prompts that are not part of it) or if the model you use to recreate it is significantly weaker. This arguably is the preferred form, and it works exceptionally well even for large codebases (but there still is a limit where it gets tricky and/or too expensive; figuring out how to push this is ongoing work).

Ok, bye bye

Posted Jul 23, 2026 21:21 UTC (Thu) by mjg59 (subscriber, #23239) [Link]

Code being AI-written doesn't mean that it wasn't written iteratively, with changes made as a consequence of inspection of the code or reconsideration of features. It's still helpful for a human to be able to review that process without having to boil an ocean just to regenerate the same output.