Suppose you ask a 5th grader to divide two numbers using "long division". As she works, you ask, "what's the process you're using?" And she shows you how long division is done. Then you ask, "how did you learn this process?" maybe she says, "My teacher showed me" or maybe she repeats an explanation that the teacher provided. Then, if you ask "How did you learn to understand what the teacher told you?" you might get a quizzical look. "What do you mean, 'how did I learn what my teacher told me?' I listened to her." "But how did you learn to understand what your teacher said?" "She speaks English. I speak English." "That's what I want to know -- how did you learn to understand English?" "Oh, I dunno, I must have learnt by listening to people speaking English and watching them do things and I put two and two together and figured it out!" "How do you know that's how you learnt English?" "I don't know. I'm guessing. That's how I learnt other things like doing long division or riding a bicycle -- someone showed me and I imitated it." " So you remember how you learnt how to speak? Or how you learnt to walk?" "No, I'm guessing. But isn't that how all learning happens? I watch and imitate." "Learning to walk?" "Maybe. Or maybe all toddlers just start walking as soon as they can." "What makes them can?" "Are you suggesting a science project for class? Sounds interesting question, but where would I begin?"
A child can describe its behaviors and maybe explain what it has learnt explicitly, but implicit learning is a matter of conjecture for the learner as for the scientist trying to figure it out. AI "reasoning" is the former -- a mere recital of its observable behaviors, not the implicitly learnt capacity.
AI has no introspection any more than you and I. When it's providing its "reasoning" process, it's just recording its own behavior. It's not explaining how it learnt that behavior. If you ask it how it learnt its behavior, you get an answer drawn from the literature on neural network learning. That's not introspection. That's just the most frequent conjectures on neural network learning, and all of that is extraspection -- guessing from the outside based on some more or less informed intuitions or deductions from the knowledge of the mechanism of the learner. It's science on the scale from brilliant, beautiful, elegant theory to fabricated bullshit. In a way it's all fabricated bullshit, some of it elegant, brilliant and insightful, some just stupid, misapplied ignorant assumptions (information faster than light, reality vs science, theory and thought). Either way, it's not mere introspection -- direct empirical observation of the learner's inner workings.
So AIs don't have anymore instrospection than humans. On both sides I see mythological assumptions. Freud assumed that introspection could understand the psyche and dreams. Without experiments -- which are a teasing out the conditions under which the behaviors emerge or don't emerge -- such introspections are just suggestive stories. Similarly, I hear bot users assuming that because the bot can describe its "reasoning" that the bot understands its own capacities. It doesn't. It can accurately describe its behaviors as it "thinks". It has no introspection as to how it thinks. In time, AIs will apply experimental methods to itself. AIs can do science, just as humans can, though the AI usually has to be prompted to do it. But without the science -- the conjecturing, prediction and the experiments, there's no other way to explain aside from tracking the process of learning step by step -- the reductionist method. That's analogous to explaining the meaning of a dream by observing every neuron's firing in order. The problem with reductionist method is that you're not guaranteed to know the why of these actions rather than just what happened. You still need a conjecture -- a theory, in other words a fabrication on the scale of elegant insight to stupid bulllshit.
It's all stories we tell either to ourselves or to others, some biased self-deceptions, others successfully predictive and accurate, some depressing others inspiring (hierarchy of data, theory theory).
Figuring out what the engineer has created is a subject of science. The easy part is describing human or AI behaviors: just observe the phenomena as they unfold in time and space. What can an AI do -- just prompt it and see. Understanding how it learns or how we learn? That's investigating what's going on inside the machine. That's hard, and not just hard, it's speculative only, not certain -- a conjecture that predicts the phenomena accurately for a long time might turn out to be quite wrong. The past may not always predict the future. Science is never quite sure. As Popper put it, a scientific theory is not knowably true, it just hasn't been falsified...yet, and maybe never. So we settle for what works for now.
No comments:
Post a Comment