Entropy in conversational AI: the second Perceived AGI paper
Paper two of the Perceived AGI series is on arXiv. When people talk, they hesitate, doubt, and let their mood show. The paper asks how an AI assistant could carry something like that too.
The second article in the Perceived AGI series is now on arXiv: Entropy in Conversational AI. The first one, Perceived AGI, has been up since July.
When people talk, information is only part of it. They hesitate, they doubt, they let you see how they feel. AI assistants are built for the opposite: a clear, logical answer to every question. That is useful, and it is also why talking to one can feel like nobody is home.
Turning up the randomness does not fix it. The model repeats itself less, but the variation comes from a dice roll, not from anything it carries from one reply to the next. What we want is variation you cannot predict from what was said so far, but which hangs together over time, the way a person’s does. The paper calls that entropy and defines it so it can be measured.
It then tests a small layer on top of an unmodified model. The layer remembers a little from one reply to the next and picks replies to fit that. On the small model I tested, most of what I predicted failed. The paper says so. What lasts is the definition and the test, which anyone can rerun on a bigger model.
Why bother? Because if we take the uncertain part of a conversation seriously, we get assistants that are more empathic and more human to talk to. And anything that looks like an inner life changes how people attach to it, so it has to be tested against fixed criteria, not declared. Whether people actually perceive a mind in it is the next study, with humans.
