Lucas Nussbaum [19/Feb 10:14am +01] wrote:
> I believe we should focus on what is desirable specifically for the
> Debian project. We should agree to disagree on AI in general, and
> instead focus on how we can work together despite that disagreement.
>
> I don't want to rush things, but I also don't think that it's useful to
> spend a lot of time rehashing arguments.
I agree with you that now is probably an appropriate time to start
trying to find a way for us to agree to disagree wrt LLMs.
I don't understand your point that the specific technology does not
matter, because the LLM-sceptics are concerned with the specifics of how
LLMs, in particular, are produced, and then what happens when they are
used.
Frankly narrowing the talk to LLMs is IMHO already talking about yesterday's news. Like, there is an LLM at the core of Claude, sure. But actually it is a whole orchestra of different LLMs (and SLMs) trained on very different data in very different ways and then there are agentic layers and constitutional layers that are not LLMs at all. And multimodality adds Large Multimodal Models that are also not *actually* LLMs anymore.
I like the suggestion of focusing on the output, regardless of the technology - code generation, patch generation, automated refactoring, test generation, documentation generation, code reviews with suggestions.