RFC: The Binutils and LLM generated patches

[Siddhesh Poyarekar]

On 2025-10-07 10:21, Nick Clifton wrote:
> Hi Sid,
> 
>>>    Generated-By: GNU-LLM version 1.0
>>>    LLM-Query: Create a program to solve all GNU problems
>>
>> Minor nit: due to the nature of LLMs, the query will likely 
>  > not result in the same output at all times.  I can see why
>  > you'd want the query though, but maybe also add in a note
>  > like so?>
>> ```
>> The LLM-Query needs to be recorded for auditing purposes, but is not 
>> expected to produce the exact same output due to the non-deterministic 
>> nature of LLMs.
> 
> OK, I can do that.
> 
>> Also maybe, replace LLMs with AI in general?
> 
> Ah but I am trying to avoid the use of the term A.I. in this
> document.  (As explained in the footnotes).
> 
>>   Or is binutils OK with patch submissions from other forms of AI,
> 
> No.  If the copyright ownership of a - legally significant - submission
> is unclear, then it is unacceptable to the Binutils.

OK, we agree then and have a permanent record of that agreement on the 
list :)

>> e.g. a reinforcement learning bot that has been trained to spot a 
>> buffer overflow?  Such a bot would also need code as training data, 
>  > but it may or may not have the same challenges as an LLM model 
> trained on,
> 
> Maybe, maybe not.  But until the lawyers sort this out and there is a 
> clear,
> legal definition of what can and cannot be copyrighted and who owns that
> copyright, the binutils (and open source projects in general imho) cannot
> afford to accept such submissions.

I agree, I was wondering if you want to make that explicit in this 
policy, but I reckon your position is to do this for LLMs for now, and 
not broadly for tools/transformers that ingest data that has 
unknown/unknowable copyright value to generate the content they intend 
to submit as their contribution.  The latter is assumed to be covered by 
the FSF assignment or DCO, where the submitter claims, minimally, the 
ability to submit changes for the copyright owner under the project 
license.  That's fine I think.

>> You could make a distinction (e.g. an AI model that has been trained 
>  > exclusively on data that has a clear copyright and the submitter is
>  > legally allowed to contribute it to binutils under said license) but
>  > that could be applied in general to an AI model without singling out
>  > LLMs.
> 
> I think that that would be something to look at once it becomes a reality.
> I do not know of any such models now, and I think that it is important
> that we get a policy document in place soon.  I am not saying that the
> policy is set in stone and cannot be revised, but I am saying that we
> need something clear and simple and in place today.

Ack, I agree.

Thanks,
Sid