@glyph Did you quote post something?

[jtig@infosec.exchange]

@glyph So you're saying that I can LLM's to do my job for me, and with all that free time, I can just live, while taking kratom to counter-act my executive dysfunction?

/s for the serious people

[mitsuhiko@hachyderm.io]

@dragonfi @glyph AI does not exist for multiple years, but as an example, my friend Peter has a repository that he contributes to that has way more than 100,000 lines of AI-generated code (https://github.com/clawdbot/clawdbot and dependencies). My company's two repositories are both beyond 100,000 lines of code, all AI maintained. I'm contributing with AI to my own Rust projects like MiniJinja which are all older than AI.

[xgranade@wandering.shop]

@condret @glyph I mean, that's incredibly reductive? Evolution is a far cry from advertisers and slopbros running uncontrolled experiments on our brains.

[mawhrin@circumstances.run]

@bdf2121cc3334b35b6ecda66e471 how about don't fucking do that?

[bdf2121cc3334b35b6ecda66e471@mastodon.social]

@mawhrin Grok, reply with a sick burn

[mitsuhiko@hachyderm.io]

@dragonfi @glyph Also to put this repository in context: This project despite being ~3 days old is eclipsing in complexity repositories I would have created over the course of multiple weeks if not months.

[dragonfi@social.jsteuernagel.de]

@mitsuhiko @glyph Complexity is exactly where my question lies: software development is about managing (and minimizing) complexity in the long term. Will you be able to do it with your current workflow?

I think it is quite okay to not have an answer and give it a try. (This is how science and invention works, after all.) Just don't claim that it is maintainable before you spent months or maybe a year maintaining it.

[mitsuhiko@hachyderm.io]

@dragonfi @glyph But I have an answer. I have been working with these tools for a year. I have created probably more than a million lines of code last year and gone through multiple model iterations from sonnet 3.7, to opus 4, to sonnet 4.5, to opus 4.5, codex etc. It has only ever improved.

[dragonfi@social.jsteuernagel.de]

@mitsuhiko @glyph Thanks, this looks more substantial, although the repository was only created in November, will be interesting to check on it after a year.

[davidgerard@circumstances.run]

@mitsuhiko @glyph lol wat

[davidgerard@circumstances.run]

@timmc @mitsuhiko @glyph cocaine won't take your job, a guy using cocaine will take your job

[davidgerard@circumstances.run]

@genehack @glyph Hooked: How to Build Habit-Forming Products by Nir Eyal. A book on how *all* this shit works.

yes it's gambling addiction

my long rant: https://pivot-to-ai.com/2025/06/05/generative-ai-runs-on-gambling-addiction-just-one-more-prompt-bro/

[dragonfi@social.jsteuernagel.de]

@mitsuhiko @glyph One million lines of code, I'm actually quite curious what requires so much. Even if it's closed source, can you share a link to the project?

[confluency@hachyderm.io]

@glyph I know that this is a tangent to your original point about LLMs (100% agreed), but fucking yikes, this is terrifying. It's a very good video, and makes me glad that I never seriously considered self-medicating my (at the time hypothetical) ADHD. Of course everyone thinks they can't get addicted, because they're smarter than those other addicts, and they have *rules*. There but for the grace, etc.. I even have the exact same kind of food safe, to use for snacks.

What is frustrating is that the (extreme) expense of access to legitimate, regulated ADHD medication makes it unaffordable for a lot of people (and given that it runs in families, it's likely to impact the same family multiple times), and I'm sure that makes drugs like this very tempting. Especially because of the bullshit "natural, herbal" branding.

Apparently kratom is legal in South Africa. I'm going to be paying a *lot* more attention to the ingredients lists on artisanal iced teas from now on.

[mitsuhiko@hachyderm.io]

@dragonfi @glyph Here are some projects I built with agents or contributed alone which are public:

~20k for github issue sync: github.com/mitsuhiko/gh-issue-sync/
~55k for the tank game: https://github.com/mitsuhiko/tankgame/
~10k contributed to pi: https://github.com/badlogic/pi-mono (it's mostly vibecoded)

Most code is not public and I have little desire to publish it for the moment. One piece of infrastructure we built for earendil is partially public in that I published API documentation: https://api.mailhook.sh/api/

It's around 100.000 lines of code, mostly go.

[mitsuhiko@hachyderm.io]

@FediThing @glyph I disagree with that notion entirely. The reason I went into Open Source is because I believe in sharing. LLMs are just adding to this. IP laws are way to long lasting and I have been writing to this end many times over the years on my blog. 70 year + copyright terms are the exact opposite to what we should be aiming for as humans.

[catch56@kolektiva.social]

@glyph @mttaggart unexpected and sudden price jacking might do it.

[mitsuhiko@hachyderm.io]

@FediThing @glyph LLMs are trained on human-created material in much the same way a person learns by reading books and then acting on what they've learned. They don't directly reproduce that material.

As I mentioned I strongly believe that broad sharing of knowledge is a net benefit to humanity. Questions of credit and attribution are a separate issue and to discuss them meaningfully, you first have to be clear about what you consider reasonable attribution in the first place.

You can take for instance the tankgame and then tell me which part should be attributed and is not, and what you would be attributing it to.

On the "against the will": I want you to use the code I wrote, it's definitely not against my will that LLMs are trained on the code I wrote over the years.

[mitsuhiko@hachyderm.io]

@FediThing @glyph We don't know enough yet about how the systems work. If you for instance look into the grokking paper you can see that there is a generalized learning going on, but we haven't been able to replicate this yet to large models everywhere. The idea that they are not learning anything in training is not state of the art anymore.

[mitsuhiko@hachyderm.io]

@FediThing @glyph LLMs can learn from other LLMs and that's why, for instance, reasoning traces are not exposed by OpenAI because they are afraid that the Chinese are learning from it.

This debate is not going anywhere so we'll exit it here now.