LLM policy

[Greg Troxel]

"Regina Obe" <lr at pcorp.us> writes:

>> Greg Troxel <gdt at lexort.com> writes:
>> 
>> > Wow.  I feel unwelcome in this community!
>> 
>> What I mean is I feel I've just been told that I'm violating commmunity
>> standards by being unwilling to grant computers the same rights as people.
>> 
>
> I wouldn't say violating.  I'd just say a difference of opinion here.
> You assume everyone agrees with you in this, Darafei and I are telling you
> we don't.
>
> We all have rights to our opinions and if our opinions disagree someone is
> bound to feel uncomfortable.
> I'm sorry we made you feel uncomfortable.

Thanks.

I'm fine with differences of opinion as in "we should be ok with LLM
inputs" vs "we should not be ok with LLM inputs".

What is not comfortable is "I assert that the code of conduct says that
we must welcome LLMs as community members."

>> The overall situation of discusssing AI in the context of a code of
> conduct is
>> uncomfortable,
>
> Code of conduct is always an uncomfortable thing to discuss cause it
> highlights our prejudices and how we don't all think alike.

Discussing our preconceptions is yes, and I've always seen the higher
level issue is that we should  only argue about ideas and not object to
people because of some aspect of their being.

What's uncomfortable about "code of conduct" vs a broad expectation of
behavior is that it seems to end up (not here) turning into a lever to
censor people, well beyond what it should have intended.

> The way I see it, the issue people have with AI is the following:
>
> 1) AI is disruptive when it creates garbage someone needs to review
> 2) AI might provide pull requests that violate copyright
>
> There are probably some others but that is my main issue with any AI when it
> does that. 
> It is violating someone else's needs / etc for very concrete reasons more
> than it is adding value to others.

Yes, and

3) AI systems are disruptive by how they scrape to gather training data,
and many people have had to spend time building and configuring systems
to avoid this.

Maybe not every AI system, but I have seen no system of the sort that
people use with a documented/transparent explanation of how they do not
engage in abusive scraping and how they do not use content for training
without permission and respecting licenses.

> When you start micromanaging what should be considered "everyone", you start
> getting into 
> Dangerous territory such as "Your guide dog is not welcome" even though you
> need it to guide you around, therefore "you are not welcome because you need
> a guide dog"

I really cannot see this as reasonable.  Everyone is all humans, and
extending it to LLMs seems novel and at best highly controversial.

The issue of assistance animals is more or less settled in law, but
there is excluding people who need them on one hand, and excluding
people with allergies on the other.  It's not 100% straightforward, and
we don't discuss it really because in the US it's settled law.

> Judge the acts, not the thing.  

That could apply to UAS, saying there should be no rules about
beyond-LOS drones as long as they fly ok.   It's a step into a vast set
of unconsidered consequences.

Interestingly, I have never before encountered the idea that it is
improper to exclude LLM content because of "not treating LLMS like
people and it's discriminatory", despite similar discussions in many
other projects.  Just a data point about previous observations.

> We have a code of conduct because we try to make sure as best we can that
> one's needs don't violate another's.
> Sometimes those are in conflict mostly because of someone's internal
> perspective about things and unfortunately there is nothing that
> can be done about that. We have to choose sides.

This is part of where it gets problematic, with the modern theory of
harm from ideas (again, not showing up here).

> Greg I hope you realize we do value your opinions very much even when we
> don't agree with them.  So don't let a difference of opinion make you feel
> unwelcome.

Thanks.  As long as it's a difference of opinion and not that people
that object to LLM content are deemed to be CoC offenders, then it's
fine.

(Separately, I might choose to disengage from a project that's ok with
LLMs, just as those that want to use them for everything might disengage
from projects where they aren't welcome, but that's not the same thing.)