Re: Nominations for Anthropic’s Claude for Open Source program

Adhemerval Zanella Netto <[email protected]> writes:

> gnulib guys seem to be experimenting with LLM for patch review recently [1],
> and it has uncovered real regressions [2]. And I recall that their 
> experience with Coverity has been subpar [3], and I think they even started 
> disregarding it.
>
> Personally, I see value in adding an LLM to the patch review phase, especially 
> since we lack manpower and initial triage can help newcomers and uncover some 
> issues.

I don't have access to Gnulib's coverity, but GNU coreutils shows a lot
of Gnulib stuff anyways. My feeling is that it is useful, but one has to
do some work to filter out perfectly legitimate code (e.g., the coverity
machine complains about type limit comparisons which may be different on
less common platforms).

I don't disagree that LLM review is useful. However, I do want to state
my *personal* opinions here as I did in that Gnulib thread. Anthropic
(and OpenAI, which has a similar program) tend to play this program up
to show how altruistic and great they are. However, the 6 month limit
makes it seem like a strategic plan to gain future subscribers.
Obviously, I understand their employees have to eat. They should
absolutely ban users abusing generosity. However, it feels a bit shitty
given that their models would be relatively useless if they could not
train on and then regurgitate, likely violating licenses, free software
written by maintainers that they claim to be doing favors for.

Collin