Post #1802285
2026-04-28 11:00 UTC
@frankreiff @ikuturso @chaucerburnt @sinbad
on explaining past actions however, I see humans as having some distinct advantages:
1. operating with an extremely broad range of conceptual units (unlike LLM wordpiece tokens) for highly efficient encoding
2. we can reprioritize notability of observations based on past events such that if we end up regretting missing some fact, we can figure out that we need to pay more attention to something in the future, which doesn't seem possible with LLMs
Replies (1)
-
@archo@mastodon.gamedev.place 2026-04-28 11:06
@frankreiff @ikuturso @chaucerburnt @sinbad 3. we have the ability to operate with order-sensitive concepts like time and operational preconditions such that we can fill in the gaps from our extremely flawed memories with reasoning about what is possible and what isn't, as well as combining different relative orders of events into a single timeline LLMs seem completely incapable of any of that, and even the ordering of input tokens appears to be a mere suggestion and thus unreliable