Hi Matthias, On 2026-08-12 18:32, Matthias Geiger wrote: > On Tue, 11 Aug 2026 21:27, Christian Kastner <[email protected]> wrote: >> Again, I do wish to respect other people's principles and decisions. But >> we need to have an informed vote, and for that we need to get the facts >> straight. > > Some models might be able to do this, yeah. This doesn't change the fact > that unless the model is constantly retrained (which takes a significant > amout of time and resources), it will have information that is out of date.
As I stated before, this was realized very early, and RAG, tool calling, etc. are all consequences of addressing this problem. My previous example (with clearly visible live-searches) demonstrated this. I acknowledge that training is expensive, which I see as least of the problems. Whatever my position appears like, I very, *very* much share the environmental and political concerns others have expressed here. I just believe that the status quo is part of a bubble will eventually burst. Look at China, and Chinese companies, for example: China has and is rolling out mind-boggling(!!) amounts of alternative energy. China doesn't want to be dependent on fossil fuels, which they have to mostly import. China has made solar dirt-cheap. And Chinese companies are shipping open-weight models that you can download right now, and which are just 6-9 months behind US frontier models. That's more than good enough for many, if not most. What nobody seems to talk about are the social *benefits*. For example, my parents are approaching 80, and are near computer-illiterate. But for about a year now, they have been solving tech problems guided by LLMs, and this has given them a sense of empowerment, accomplishment, and independence. And I can tell you stories like this en masse. It sucks that currently comes at such high environmental costs, but I believe that will be fixed -- as China demonstrates. Either way, too many people positively benefit from this technology for it to ever go away, and I don't want it to go away. I want to see its problems fixed. And I think Debian is in a position to shape its future, and we'd be wrong to stay out of this fight. > Furthermore, it is simply not possible to build an LLM that does not > hallucinate [0]. Yes, and? It's also impossible to design a 100% secure system, that doesn't mean security measures are useless. "All models are wrong, but some are useful" is the key phrase here. The trend of hallucinations has been a steeply falling one. Which is no surprise, because this was the #1 problem from day 1, and of course this is where tons of work were invested in. If one accepts the notion that for the big players, this is about maximizing profits, then one must also accept that the best way to get there is to have a product that is not just useful to people, but also more useful than the competition, so you even get an arms race between companies trying to fix this problem. And they have been succeeding. > This means that even if you get to 99% accuracy, at the > end of the day, I still can't trust the output. Which defeats the > purpose, if such validation is still needed. Then I can just go ahead > and write my own code; if I make any errors, that is on me, but at least > it is deterministic. I don't follow, can you please explain? Why would it be better to start at 0% than starting at 99%? Best, Christian
