Home | Notifications | New Note | Local | Federated | Search | Logout

Note Detail

Rob Ricci@ricci@discuss.systems (2026-09-04 13:08:58)
After thinking for a day about Thi and @viveksrikumar 's most recent Ethics in AI lecture, plus adding some thoughts of my own: 

* LLMs and other AI systems are inherently inductive machines. They are trained on data from the past, which we can observe, and produce "predictions" (next tokens) about the future

* Induction has the flaw, pointed out by Hume, that it builds in the assumption that things work the same in the future, which we have not yet observed, as they did in the past (which we observed and based our inductions on)

* This assumption works most (almost all?) of the time, so it's reasonable to build a lot around it (including much of science itself). But: it holds until it doesn't

* The bet on AI is fundamentally not a bet on a different future: it is a bet that the future is just more of the present

* We run the very real risk of making the above statement true if we rely too much on AI in shaping the future

* AI is a fundamentally a bet against a future that might be better