Re: LLM usage in Debian -- ballot options comparison

--0000000000006060c50657d62a88
Content-Type: text/plain; charset="UTF-8"

On Thu, 30 Jul 2026 at 14:54, Christian Kastner <[email protected]> wrote:

> On 2026-07-30 13:33, Aigars Mahinovs wrote:
> > But what I wanted to pick up was the analogy with the free software
> > movement. Free software did *not* succeed by absolutely rejecting all
> > commercial software and non-free hardware from the onset.
>
> Huh -- until you mentioned this, it never even occurred to me that
> someone, especially opponents to LLM use, might see this temporary
> embrace of hosted LLMs as anything other than a phase.


Hosted vs self-hosted vs local and proprietary vs open weights vs free are
two separate dimensions.

Unless you have a prohibitively expensive rig at home, the dataserver
provider will always be able to provide you with a more powerful platform
for self-hosting and a commercial service provider will be able to use even
more powerful hardware.It's just a question of capital cost vs utilization
percentage. I expect that local usage can get to "good enough" stage and
then lower the cost of reaching that stage, but self-hosted and fully
hosted solutions will alway be able to provide more capability.

On the freeness scale we might be able to shift towards the more free end
of the scale, especially once the thorny copyright questions are resolved.
One way or another. If copyrights survive training then non-free models
become basically impossible. If copyright does not survive training or if
training is fair use then free models do not need to constrain themselves
to non-copyleft materials only and can be much better trained. Either way
we get better free models.

-- 
Best regards,
    Aigars Mahinovs

--0000000000006060c50657d62a88
Content-Type: text/html; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr"><div class=3D"gmail_quote gmail_quote_container"><div dir=
=3D"ltr" class=3D"gmail_attr">On Thu, 30 Jul 2026 at 14:54, Christian Kastn=
er &lt;<a href=3D"mailto:[email protected]">[email protected]</a>&gt; wrote:<br><=
/div><blockquote class=3D"gmail_quote" style=3D"margin:0px 0px 0px 0.8ex;bo=
rder-left:1px solid rgb(204,204,204);padding-left:1ex">On 2026-07-30 13:33,=
 Aigars Mahinovs wrote:<br>
&gt; But what I wanted to pick up was the analogy with the free software<br=
>
&gt; movement. Free software did *not* succeed by absolutely rejecting all<=
br>
&gt; commercial=C2=A0software and non-free hardware from the onset. <br>
<br>
Huh -- until you mentioned this, it never even occurred to me that<br>
someone, especially opponents to LLM use, might see this temporary<br>
embrace of hosted LLMs as anything other than a phase.</blockquote><div><br=
></div><div>Hosted vs self-hosted vs local and proprietary vs open weights =
vs free are two separate dimensions.=C2=A0</div><div><br></div><div>Unless =
you have a prohibitively expensive rig at home, the dataserver provider wil=
l always be able to provide you with a more powerful platform for self-host=
ing and a commercial service provider will be able to use even more powerfu=
l hardware.It&#39;s just a question of capital cost vs utilization percenta=
ge. I expect that local usage can get to &quot;good enough&quot; stage and =
then lower the cost of reaching that stage, but self-hosted and fully hoste=
d solutions will alway be able to provide more capability.</div><div><br></=
div><div>On the freeness scale we might be able to shift towards the more f=
ree end of the scale, especially once the thorny copyright questions are re=
solved. One way or another. If copyrights survive training then non-free mo=
dels become basically impossible. If copyright does not survive training or=
 if training is fair use then free models do not need to constrain themselv=
es to non-copyleft materials only and can be much better trained. Either wa=
y we get better free models.</div><div><br></div><div><span class=3D"gmail_=
signature_prefix">-- </span></div></div><div dir=3D"ltr" class=3D"gmail_sig=
nature"><div dir=3D"ltr">Best regards,<br>=C2=A0 =C2=A0 Aigars Mahinovs</di=
v></div></div>

--0000000000006060c50657d62a88--