“Large language models do not reason. Their “training” concerns the statistical analysis of massive amounts of text to determine likely correct textual responses based on a given input. This is a simplification, but the basic gist is there: text is ingested, “tokenized” to identify important terms and positions, and based on many rounds of training, the model is eventually “fitted,” meaning it reliably produces output that makes sense given the input.”
The production of other kinds of Trained MOLE is similar. It is not compiled from source code, like traditional software.
Why do I need to repeat in each and every thread? Be polite, be sensitive and respectful to each other. There is a limit to my tolerating abuse in any manner to others. If there is any more from anyone, I will simply ban them. I cannot keep preaching the same thing to adults.
ban me in particular, ad hominem me all you want, but if you are gonna pretend that you have an argument while at it, imma gotta demolish that nonsense with zeal, no remorse, not gonna stop
“Vegetarians don’t force their eating habits onto everybody else. In the same spirit, regardless of the reviewer’s opinion about large language models (LLMs), patches written with the help of LLMs aren’t outright rejected as long as:
The patch author has a full understanding of the submitted code. The author has carefully reviewed the output of the tool used to generate the patch.
The patch (and assorted description) is indistinguishable from one fully written by a human. In particular, the submission doesn’t contain huge walls of unnecessary text or code.
The patch author agrees to the Developer Certificate of Origin. In particular, the author certifies that they have the right to submit the patch under the project’s open-source license. This can be achieved by adding a Signed-off-by trailer to the commit description.
The patch author discloses their use of LLMs in the description. This can be achieved by adding an Assisted-by trailer to the commit description.
The reviewer may still decide to reject the patch without further justification.”
This is a good example of the kind of policy I’d like see drafted for the F-Droid project, to address the various software freedom issues raised by generative models. That said, it only addresses one of those sets of issues; auto-generation of code.
Also, that policy doesn’t protect the project against incorporation of whole chunks of proprietary code coughed up by a MOLE, or code under incompatible copyleft licenses. Any software I maintained would follow the Gentoo and Chezmoi examples shared by @shuvashish76 , and require contributors to certify that no auto-generated code was used to prepare the patch, to protect against that risk. Claims that this is akin to forcing dietary preferences on others is about as valid as making the same claims about food safety standards, as if some people like food poisoning. In both case, it’s about protecting the community from potential harm.
Let me get this straight, so if you are not using an LLM then you are allowed to submit a patch that you don’t have a full understanding of, have not carefully reviewed, that does contain walls of useless code, and that might actually be an illegitimate copy of another work? Sounds like a massive brainfart, none of this should be acceptable regardless of LLMs, so this just amounts to the last point “oust yourself if you are one of those filthy loosers that uses LLMs”, while the rest is just pointless verbiage just to make it sound important.