Hi Wolfgang, Wolfgang Pfeiffer wrote on Sat, Nov 01, 2025 at 03:28:46PM +0100:
> what people call AI (I call it software), I strongly disagree with the your terminology. I do not think that what people usually call AI (in particular, large language models and even applications of artificial neural networks in general) can reasonably be called "software". One requirement i would impose before calling something "software" is that at least some humans can understand what it does, and how. If nobody can understamd what it does, it is not a tool, because the definition of the term "tool" is "a thing that helps with a set of well-defined tasks", and hence "software" is "a computer program that helps with a set of well-defined tasks". Nobody knows yet what AI is, but calling it a "tool" or "software" is incredibly naive and irresponsible. That's certainly compounded further by the fact that even psychologists start scratching their heads when you ask them what "intelligence" is, let alone how to measure it. A tool can be dangerous when mishandled: with a hammer, i might hit my finger. A chat bot can be dangerous even when used as intended. For example, recognizing that the bot is hallucinating is a task that can be as hard as whatever my original task was. Occasionally, it might be even objectively harder, and i'd go as far as guessing it will be subjectively harder most of the time when factoring in the various psychological biases that humans have. Even abusing an AI as a tool for a tasks where on first sight it seems relatively benign - for example, generating a list of candidate URIs in a web search when you then inspect those websites critically to find the information you need - is problematic to such a degree that i reject the designation "tool" even in such relatively carefully guarded cases. Every AI is known to contain biases, but what those biases are is unknown. Measuring those biases requires, among other tools, a non-AI search engine - but non-AI search engines essentially no longer exist, at least as far as i'm aware. As a particle physicist, i call that a systematic error of unknown and unbounded size, which is the physicist's way of saying that the whole measurement is a complete and utter failure that does not provide any information whatsoever. Is this relevant to OpenBSD? Yes, in one sense it is. Making software simple, reliable, and secure is among the chief project goals of OpenBSD, and encouraging code review is amomg the chief methods employed to further those goals. Go ahead and do some code review on your favourite AI. Good luck. Ingo P.S. OpenBSD is also known for its stance that documentation matters. Go ahead and write some documentation for your favourite AI explaining to users how to use it and which results it will produce. Again, good luck.
