Fantastic stuff. People speaking about what they've thought about in depth sure are articulate :) I'm glad for transcripts.

A fun/funny mnemonic suggestion:

If you want to learn about System 1 and System 2, or about Type 1 and Type 2 operations, really the same, think of it as System 1 and 2. it will develop a personality. There are certain things that it likes doing, that it's able to do, and there are certain things it just cannot do, and you will get that image. It's completely crazy. There is no such thing as these two characters but at the same time, I find it enormously useful. And it's quite funny, I'm losing friends over this.

Personify mechanisms!

Infants can recognize intention, and have a system that enables them to divine the intentions of objects that they see on a screen, like one object chasing another. This is something that an infant will recognize. An infant will expect one object that chases another to take the most direct route toward the other, not to follow the other's path, but actually to try to catch up.

Cool! (it's not clear if he means newborns, but he says <1yr old)

Base rate neglect:

There are two cab companies in the city. In one, 85 percent of the cabs are blue, and 15 percent of the cabs are green. There was a hit-and-run accident at night, which involved a cab. There was a witness, and the witness says the cab was green, which was the minority. The court tested the witness -we can embellish that a little bit - the court tested the witness and the finding is the witness is 80 percent reliable when the witness says "blue", and when the witness says "green", it's 80 percent reliable. You can make it more precise, there are complexities, but you get the idea. You ask people, what's your judgment? you've had both of these items of information, and people say 80 percent, by and large. That is, they ignore the base rate, and they use the causal information about the case. And it’s causal because there is a causal link between the accident and the witness.

Now have a look at a very small variation that changes everything. There are two companies in the city; they're equally large. Eighty-five percent of cab accidents involve blue cabs. Now this is not ignored. Not at all ignored. It's combined almost accurately with a base rate. You have the witness who says the opposite. What's the difference between those two cases? The difference is that when you read this one, you immediately reach the conclusion that the drivers of the blue cabs are insane, they're reckless drivers. That is true for every driver. It's a stereotype that you have formed instantly, but it's a stereotype about individuals, it is no longer a statement about the ensemble. It is a statement about individual blue drivers. We operate on that completely differently from the way that we operate on merely statistical information that that cab is drawn from that ensemble. So that's a difficulty that System 1 has, but it's also an ability that System 1 has.

Those really are the same (under the default assumption that in the first case, blue and green drivers are equally accident-prone).

System 1 ... generates an attitude. There are things, you like them more, you like them less, that's System 1, the immediate emotional response to something. Choice, which is an explicit comparison, always involves System 2, and is always effortful.

Choice always involves type-2 thinking (except for instinctive flinching-away avoidance?).

JARON LANIER: I just had another thought for a minute, from a computer science perspective. If you compare the task of averaging versus summation, the difference is that summation involves accumulators, there's this variable. And when we do machine learning, it tends to fall into two systems, its slightly analogous to this. I don't know how much. What is it in your system where there's running averages you don't create new variables, or what you would call models, and so you're just tracking some sort of traditional data set.

?????? Speaking before finishing a thought is dangerous :)

He goes on to make a reasonable point that slightly updating type-1 association strengths might cost less energy (be physically more routine) than building a new concept - changing weights in a graph as opposed to adding nodes or edges.

(more Lanier) [You can] start with a prototype of a facial map and you look and you scan an environment for facial elements that match a face. You track faces, but you don't add it to your model space.

If you represent the prototype, then it's a model of sorts. Unclear what he means.

If you, instead, define a task where I am going to show you arbitrary new pieces of information, and I want it to have emergent new categories, wow, that is an entirely different class of problem, and that's vastly more difficult, and essentially not understood.

"Emergent" :) Unsupervised clustering of data is pretty standard. Yes, it's a more interesting and potentially more nebulous problem. It doesn't take to add "undecided on total number of categories to learn" though; reward generalization to unseen data or minimum description length (model complexity is penalized; fitting observed data is rewarded). See e.g. The Infinite HMM.

My intuition is that this is something like this System 1 System 2 extension.

Okay. Maybe. He seems to base this mostly on cheap vs expensive. I don't see why new category formation couldn't sometimes be unconscious, though.