Comment by jorleif | Hacker News Reader

Comment by jorleif | Hacker News Reader

"only a tiny component of thinking is about prediction at all"

Here I really struggle to follow his argument. It seems, he is saying that thinking is not prediction, but how does he know that? As I see it, we don't know what thinking is, so is he then saying that thinking does not feel like prediction, and therefore it cannot be prediction? It is well known that we are only aware of our very highest level of thought. What goes on below the surface may very well be very much about prediction without us knowing about. When we think about "the world" we must have some internal representation of the configuration of objects that we think about. In my opinion this is implicitly a kind of prediction, because we create (at least mostly) configurations that are plausible, if not true (counterfactual, but possible).

It seems to me that creativity is mostly about finding good approximate solutions to hard (e.g. NP-complete) problems. How we do it is so far unknown, but it seems that it is not in principle so very special. The explanation for the calendar he speaks about is actually not very complicated inference at all, and is probably in reach for artificial systems. All he has to do is to generalize 19 as belonging to a number category, and then apply the predictive properties about how numbers work. Not all prediction is naively predicting that the world is as it was before. If change is predictable in the past, then change can also be predicted to the future. Prediction is difficult, but that does not make it less important.

I think the predictive framework is a fairly universal one. You could always as a thought experiment feed a predictor all sensory experience humans ever had, and provided this predictor works in a manner in some sense optimal it could answer whatever question at least as well as any person.

Likewise we can already simulate fairly large biochemically accurate neural networks, and it seems a scaling in power would lead to an AGI.

The problem with the previous models is their inefficiency. It does seem to indicate a lack of understanding of fundamentals -- i.e. you can brute force essentially every well defined problem in CS, yet we don't declare them solved modulo some time for computing power progress.

You criticize the author for making unfounded assumptions about the nature of cognitive processes and follow up with your own equally unfounded assumption about the nature of cognitive processes. I do agree with you, but it is still bad form.

What do you mean is unfounded? My suggestion for what creativity might be as be? I guess I was unclear, I don't mean to say I know what creativity is. What I do mean to ask, is whether creativity is something more than just solving optimization problems? To me that way of thinking about it explains many of my own experiences (a-ha moments, unknown time to completion of thinking process, thinking going on in the background etc.) using a formulation that I understand. It is free for anyone to prove it wrong.

Is there a difference between reasoning and predicting? It seems you don't think so and the author does.

I think there is a difference, but my view is that it is impossible to reason without sufficiently accurate prediction, and that the concepts we use for reasoning are largely formed based on how predictive they are. Consider playing a game where two players throw dice, and the one with the larger number wins. There is no predictability between turns, so there is no point in reasoning far into the future, unlike in chess, where the positions change slowly, one piece per turn and in predictable patterns. In chess you still don't know what the other player will do, but you can reason about it, because there is a level of predictability.

I would define reasoning as answering a question or optimization problem, e.g. What is my best move? What is a good move? This can be formalized as calculating some kind of score for a possible action, whereas prediction is just about assigning a probability value to an event.

I would define reasoning as answering a question or optimization problem, e.g. What is my best move? What is a good move? This can be formalized as calculating some kind of score for a possible action, whereas prediction is just about assigning a probability value to an event.

So would it be fair to summarize this as "reason is prediction with a goal"?

Bad example. I can predict the dice game's problem space. E.g. for a single normal die (1-6), I can predict that I'll never observe a 7 [P(7) = 0%]. I can also predict the probability distribution. E.g. for two dice thrown together, the probability of throwing a sum of 7 is 1/6, whereas the probability of throwing snake eyes is only 1/36. Yes, the game's output is random. But what you're talking about is control, not prediction. If random meant "impossible to make predictions about", then statistics wouldn't exist.

Regarding winning, there's no way to improve one's chances (besides cheating?). But merely knowing that the game is predicated solely on chance can be useful for other applications. E.g. realizing that the lottery is a scam.

Personally, I'm in that camp that says reasoning is a instrumental value towards prediction.

I think there's a strong argument to be made that the answer to this question depends on exactly what you mean by "reason" and "predict".

If you mean "predict" in the same sense as it's used in the phrase "predictive model", then the answer could be a resounding no - there's a hypothesis that the brain implements just one basic learning algorithm, and everything else is just emergent phenomena coming out of that.

Reply on news.ycombinator.com