Grit: rewriting Git in Rust with agents
16 points by typesanitizer
16 points by typesanitizer
from the “Licensing” section about the choice to release this under MIT:
This might be a little controversial, but ultimately I think it's defensible and more importantly, the best thing for the wider Git community.
there is a mailing list thread linked above this that is apparently meant to underscore how problematic GPL licensing has been for libgit2. clearly, the author has more depth and exposure to how the licensing has impacted contributions to libgit2 than i have. however, i am not really convinced that they have laid out, with the mailing list link or in their post, any evidence that permissive licensing is “the best thing for the wider Git community,” unless you already take for granted that permissive licenses are somehow inherently better than copyleft licenses. i guess some people consider that debate settled?
on a personal note: i can see how and why people find the ability to do things like this valuable, but the process that is described sounds like pulling teeth to me. i can not imagine finding any pleasure or satisfaction in completing a project like this. not to mention that i would have to save up for a year or more to have the spare cash for such an endeavor.
I find it dubious that models trained on the Git source code working with the Git integration test suite is not a “derivative work.” This sounds like spending $15k on tokens to dodge a license so corporate profiteering on the commons can continue.
What I also don't understand is how any of the powers that be in the domain of intellectual property are OK with this. I understand how a system might screw over individual volunteer developers, but surely the owners of multi-billion-moneys proprietary software or movies or music must understand that it's only a matter of time before the leopards eat their faces too?
If the legal system agrees that an LLM can wash away copyright from a FOSS codebase, then surely an LLM can also watch a bunch of Hollywood movies and create their own "inspired" movies free of any ownership? Maybe movies are a bit far off, but there's been plenty of code leaks of proprietary software – can we just feed that to an LLM and have it generate an MIT-licensed "rewrite"? If one can do it with git, then surely the answer is yes? And music is probably within reach soon. Not to mention the LLMs themselves! They all rely on the same legal IP framework that is being eroded so fast it would shock even my edgy teenager self who in the 90s gladly (and misguidedly) would say things like "information can't be owned, duuude!" Yes, information can be owned, and this is necessary to build an economy around intellectual work. What the hell are we doing?
How is any of this sustainable?
Sorry, my intention wasn't to say that GPL+linking was problematic for libgit2, but to point out that it was chosen so that code could easily be shared between the projects and it basically never happened. So we could have just gone with a more permissive license without additional issues. Also that the list is still having licensing compatibility conversations because it's GPL that would not be happening if it was more permissive and didn't have to tow this line so that people could link to it.
I am also confused. From what I'm reading out of that mail (and also as a summary for the casual comment reader), the problem seems mainly that libgit2 is not compatibly licensed for contribution with git itself, which means that it has to reengineer or ignore a lot of changes in git mainline.
unless you already take for granted that permissive licenses are somehow inherently better than copyleft licenses. i guess some people consider that debate settled?
That was never a debate to begin with for the question of adoption. Copyleft licenses intentionally chose to be restrictive at the benefit of keeping it a more even ground. The result has always been that when the need arose to embed it, then people built alternatives.
Copyleft licenses can only defend that position when the cost of maintaining an alternative is too high. For instance GCC could defend its license choice for years because it was too expensive to have an alternative. Nowadays we have LLVM and it won out all the use cases where the original license was an issue.
After reading the reasons why one would do this I'm still not convinced this was worth throwing $15k and months of opportunity costs at:
Am I missing something? It's a mostly-passing-tests, not-super-usable, legally questionable clone built over months at some expense.
you're not missing anything. this is parasite behavior and letting it go unchecked will severely damage copyleft's leverage against corporations.
This is incredible, what a great experiment, I don't think that there ever was a time when you could get a "working" git implementation for just $8k.
8k is not the total sum that was spent; that was a single week of Claude usage.
a few paragraphs earlier:
I have no idea exactly how much I spent between Cursor and Anthropic (the main providers I used) but it was probably somewhere around $10-15k.
Were they trying to waste as much money as they could? Claude max 20x costs $200 per month and practically gives you enough credits to run Opus on max effort 24/7.
they were operating at per-token (API?) pricing to start with, and iirc, seem to imply they were caught off guard by the cost explosion. if their $15K ceiling is correct, it seems somewhat less absurd that they spent about the same amount as their first week blunder throughout the rest of this “last few months” that was spent on the project. not to defend the expenditure, just to point out that it does seem like they recognized the problem and then did course correct, to a degree.
I wanted to use OpenClaw so I could do this remotely on a machine I could manually take over and test with easily. When I started, the OpenAI models were not doing good work in that framework and the Claude subscriptions were just banned. So I admittedly spent more than I needed to there.
Though, it's interesting - even $10-15k is what? A high end consultant for like 3-4 weeks of work? I don't see it as a totally unreasonable cost, all things considered.
Thank you for pointing that out, still impressive that we're able to get something this big for so little.
Couldn't you get one off Fiverr whenever you want?
Probably the best ad for Fiverr if someone actually finds a dev there that does this in Rust cheaper