On 25/6/26 23:21, Ludovic Courtès wrote:
Hi Greg,
Greg Hogan <[email protected]> skribis:
>
Somewhat continuing from above, I expect the project would only run
free models, and I can imagine a future where the user could request
insight from that model on failed builds similar to guix-cuirass-bot.
Or a QA-like system could evaluate logs for failed builds before
retrying multiple times, conserving resources. The user would be
taking action (such as following a link) to see the generative output.
I would rather not make a policy based on this hypothetical scenario
because AFAIK, “free models” (as in gratis) today perform poorly and
very slowly on commodity hardware.
Isn't this very close to what Thanos Apollo is already doing with
Apertus + Hermes (which both seem to be free as in libre)?
We should not have such large uncertainties about what we are even
talking about this late in the GCD process.
Haven't been able to confirm the performance of such free models myself
yet though, as there are about a gazillion dependencies needed that are
not yet packaged in Guix; working on that.
Also, I don't see how we can even discuss such a not-so-hypothetical
future under the GCD, because it would come awfully close to 'encouraging'.
