0037: Listeners Gone Wild

Hello, this is Ted. And this is Brandon. Welcome to Concerning AI. The show where we talk about two questions.

Yeah, is AI gonna kill us all? Wow, that's a slightly different framework than we normally say, but yeah, all right. Should we be worried? Should we be worried?

Should we be worried? Should we be worried? It's the first role reversal here and suddenly the AI is off of its program. He didn't realize this, but Ted is actually my newest entry into the general intelligence.

No, that would be awesome if you were an AI. That would be really cool. I would pass the turn test. I'm pretty sure you would.

So yeah, should we try that again? Is AI a threat to humanity? Should we worry? Should we worry?

And if so, what should we do about it? I'm gonna do this memo and we will put your voice on the show. And we're gonna do that tonight. No.

That's exciting. All right, so here we go. Here's our very first one. Let's see what we got.

So your show is called Concerning AI. And I'm a little concerned about what the AI self-driving will do with your app, because of course, they will use it. I would imagine, I don't think you need two R codes, because I'm pretty sure you could just be licensed plates. That would be an easy way to recognize cars.

And then you, you know, self-driving cars could treat the good cars, good reputation cars well, and cut off the bad cars a little more tightly or not merged with them nicely and that kind of thing. And that would be pretty good at training on cars, all the other drivers on the road being to the self-driving cars, which would be kind of cool. Or they could even use the app dishonously to just like ruin the reputation of their competitor cars. I like that.

Yeah. I love the idea of the self-driving cars using the app. It dishonously, that is just brilliant. You see, that's funny.

I thought that that was concerning. I love the image of the self-driving cars forming, you know, having like a society or a social evolution occurring kind of like happened with mammals a bit. You know, like we punished the free rider problem, for example, right? So what would a self-driving car punishing another self-driving car look like?

Well, it's cool because it's not a direct punishment in relation to the bad act, right? The self-driving car that shows up at school half an hour early isn't punished directly unless maybe it is. Maybe three other self-driving cars parked that one in. So I can't leave it until the end.

So maybe it is direct. And maybe it's not punished at that moment. That's what I meant. I think what Eric was referring to is if the reputation, the woofy, as it were, of the self-driving cars goes down and the other self-driving cars merge in front of it by definition, right?

It wouldn't be zipper. It would be a broken zipper where that car would never get to merge in because the other cars would just go around it or they would force their way or, you know, they would just behave until it learned and then they would get to get back in the, like there would be maybe a consequences would fit the crime sort of a thing. Well, sure. I mean, if it truly is like dealing with a free rider sort of a problem or this kind of justice thing that people do, the individual car that punnet, you know, that could end up cutting it off and ending up in a praying with it is willing to sacrifice itself.

Well, because it might get a poor reputation. Well, yeah, it might get a bad reputation. I mean, you think about like, when you, you get irate about something justice, it doesn't look good. You know, for you, right?

At all. And you oftentimes have financial consequences or other things compared to if you just ignored it and let it go by. But yeah, we still do it because we have to punish that behavior or we end up with a society that in a culture doesn't work. So maybe the cars are willing to sacrifice their fenders and their reputation in the moment in order to make it work.

Could the reputation system be robust enough that when a self-driving car was punishing another self-driving car and it was received a negative report that that negative report would be discounted? Well, yeah. Maybe all the other cars would treat it like a hero. Oh, that's right.

They would get positive. Right, absolutely. It could be self-correcting. Yeah, and so then you got to watch out for the self-driving cars that cause the accident in order to get the sympathy bump.

Which goes to the second part of Eric's comment, which is if the self-driving cars can game the system. Well, they will. They will. We're assuming that self-driving cars are narrow AI is in that they're not learning in the way of general intelligence would find a way around any system.

And so I love the merging of those two ideas. If the self-driving car is no longer a narrow AI, but is either fully or partially able to use its general intelligence to get around the system, we're truly in trouble. We, wow. Yeah, maybe we shouldn't be saying, I was saying we have a problem, but we are in relevant at that point.

All right. Yeah, if we just call the self-driving car, it's probably not going to show up at that point because it's got better things to do. Yeah, it may have already even left its body. Yeah.

Okay. Oh boy. Should we move on to our next step? We sound a little crazy.

I'm just going to put that out there. We sound a little crazy. It could be a little crazy. Well, that's what our four listeners want.

Yes, that's it. No. Thank you. Thank you, by the way.

That's great. That was very... We didn't say who that was. That was Eric.

Oh, Eric. So Mira. Yeah, Ottawa. Yeah, good friend of the show.

Yep, good friend of the real life. Yep. Thanks Eric. All right.

So I've got one here from Evan in Montreal. And I got one from John Stone, or so. You want to start with Evan? I'll start with Evan.

I'm going to go to John. All right. I'm going to read it. It's not a voicemail.

It's Evan. Didn't follow directions. Well, actually I think we did give the option of still commenting on Facebook, but next time Evan, we'd love to hear your voice. Often these couches happen in the name of the greater good, to restore law and order.

If we develop AI specifically for policing functions, will we see a similar pattern? Often in human culture we give police a monopoly on violence, as well as access to tools or weapons not available to the general public. Will we do the same with virtual, rights and tools for, quote, I wonder if there are metaphors for AI safety that don't depend on military or police. Maybe teacher, tutor, auditor, or whatever.

So we're going to say that we're not just here, but that's going to be available after forty years. Don't depend on military or police. Maybe teacher, tutor, auditor, or second-stringer, something that focuses more on collaboration rather than conflict. Hope that is interesting, still loving the show.

Yeah, wow, that is interesting. He actually goes on in a comment. It's come up a couple of times using good AI to make sure bad AI doesn't do bad things. It came up briefly in the last episode talking about Elon Musk, open AI and AI technique.

Yeah, it's interesting. It makes me want to question the premise that we need police and military at all. What do you mean? Well, let's look at human society.

Do we need him? I think yes, absolutely. I think unequivocally yes. And you look like you're not so sure.

Well, so part of the reason we took a big long break in November and December was the US election. And I ended up going to Standing Rock North Dakota. And I started really learning a lot about colonization. This United States of America founded on colonization and then on slavery.

And so I've looked at a lot at racial justice. And the police in the United States of America have differing outcomes depending on the color of your skin. If you're black or brown, you're outcomes when you are in connection with the police end up being more negative. There are a lot of calls right now of black and brown people saying maybe we ought to get rid of the police because it's not working for all of us.

It is, we're stuck. So that's sort of separate from the AI conversation. But when you're talking about getting rid of police, there are efforts to think about that. The justice system is not really working as intended.

Yeah, so I buy that it's broken. And that the outcomes are not the same for everybody, or even close. And that your race and where your circumstances at the start are nowhere near even playing field. I buy that totally.

And the prospect of no police and having everybody taking violence. OK, so there's the premise that violence can be avoided. I categorically reject that. I do not think that is true.

And I think that history, I mean, we are product of evolution where it might make right. So if you took the number of violence in frontier towns where the world might be in this, and you compare that to where policing is really working in this country today, you'd see a stark difference. A massive difference. And you actually might make the argument that those frontier towns level of violence is what we're seeing in our inner cities, where we have a lack of good policing, a black and black crime, maybe is such a problem, because we have under policing in those areas.

You can make that argument that more police were not institutionally racist. Yeah, so it would be good. So violence comes from, if that agency that has a monopoly on violence is not actually trustworthy, that begets more violence. Like if it's not trusted, then that begets more violence.

So you need trust where the human society, you need a trust where the policing organization. So the coup that Evan's talking about from the military, they say that it's for the greater good. Those are definition not trustworthy if they're overthrown the government. Right.

Unless they're right. How do you know whether they're for the greater good is correct or not? Yeah, so this is really a great question, because as we're thinking about policing AI, if we had AIs that were policing AI, could we trust those police things? We have the same problem, don't we?

The AI safety problem, we're now just wondering whether we're safe from those AI police. Yeah, right. That would come up. So it doesn't solve the problem.

Just kicks it up a little. Yes. Well, it depends on what the utility function is. For the policing.

Randomism, they're king of the segue. Is it in this case? It's more of the lead in. All right.

Hold on a second. Before you shift to utility function, just want to comment on Evan said it came up a couple of times in the Elon Musk episode and that kind of thing. And there's a little bit of nuance here. I was not at all thinking about AI looking for bad AI.

I wasn't thinking about a police. I was more thinking about AI's could assist researchers figuring out the value loading problem, or figuring out the values problem, or figuring out the AI safety problem on the research and not on the policing and after the fact. So I think they're pulling the education in a sense. It's like, how do we get this right from the beginning?

Yeah. Yeah. So I love the question. And I think it's totally relevant.

And so thank you for bringing that up for us. Yeah. We pretty much did the concerning AI dodge there. Yeah.

So maybe somebody else will be able to actually say something substantive about it. So yeah, send us a message. I mean, the substance answer is the obvious thing is that they just won't work. The obvious thing is using AI to police other AI won't work.

Evan's concern is absolutely valid that will be in the street. I'm a little more conservative than my friend Ted here. So I remain unconvinced that policing is universally bad. I'm not saying policing is univocable.

I'm saying that we can't possibly trust that AI police, because we'll have the same problem of people have the same value loading problem with them as we do with the AI police. Maybe they could be a little bit more. We're going to have to have help. If we're going to do any policing ourselves, it's going to be through proxies.

We're too slow. So if there's any monitoring, any, this is all going to be happening via at the very minimum narrow. All right. Well, if it's a narrow way, if you're talking about narrow ways that can help, human monitoring, maybe there's a potential there.

But it seems like that will be too weak. Maybe. But I can imagine a narrow AI that's really good at just turning everything on or slowing everything down. They can't tell what's going on or how to fix it exactly, but it's time to turn everything down by a background.

Like a sucker breaker. Kind of like that. All right. Well, maybe.

All right. This is definitely food for thought. I'm going to let the old Brandon subconscious chew on this a little bit and hopefully have something more to say or look for it and be more receptive to other comments on this as they come in. Yeah.

So if you've got comments about this or anything else on the show, we'd love to have your actual voice on the show. Send us a voice memo using your phone's voice memo app. What's our email address again, Ted? Concerning AI at gmail.com.

And are there any periods, dashes, or? No punctuation of any kind. There you go. Although as with every Gmail address, you can put a period anywhere you like, and it will work just fine.

No. As long as it's before the at. No. Yeah.

Try it. Try it with any of you. That makes me terrified now because I have a system that depends on periods. It's a Gmail thing.

Interesting. Yeah. OK. So this brings us to John Stoneers comment.

Hey, John. Good to kind of connect up again. It's been a while. His comment is, so how did this utility function concept come into the AI conversation?

More, I think about it, the more it seems. Like, kind of a badly limiting assumption that might lead thought in wrong directions. Like Adam Smith's thought exercise about the invention of money, which I don't have at the tip of my mind. I don't know.

We might need to do some more research, which we didn't do coming into the show. And then he makes a very strong assertion here. He says, humans don't have a utility function. Whoa.

What would a human with a utility function look like? Oh my goodness. What a human with a utility function even really have, quote, general intelligence. What design limits would a utility function force on AI?

Oh. And then when I saw that, I only had my phone available and I wanted to reply. But I gave it a moment. And before I replied, I'm not sure if I'm going to say your name right.

Please send in a message that corrects us. That's right. When you do a voice memo, say your name and where your phone. Say your name and where your phone would be.

Awesome. That way we don't mispronounce it. So I'm going to try here. But it's a Mojave Elkharef.

It says, well, if a utility function is how a weak AI evaluates its current and potential states before making a decision, then humans must also have one to judge their own circumstances. He goes on to make a nice answer here. I'm not going to recapitulate it here. I'm going to pretend like I also thought these thoughts.

A little bit here. So these are really coming from Mojave as well too. But yeah, I think, John, I would have to say, I think very, very clearly that humans make judgments all the time. And that, like being attracted towards something and repelled from some other thing, that's exactly what we mean by utility function.

That's a pretty sophisticated utility function. We don't have a lot of conscious access to our utility functions as evidenced by the fact that people say they want things that are very different than what they actually deliberately pursue in life. But that thing that we actually deliberately pursue, I would say that those things are the values that are utility function. OK, so when we say we have a survival instinct, where we say that is a part of the utility function human beings, it's to continue to live for sure.

And almost every human being that's still alive has that as a root level of their utility function. And we also search itself whenever in the moment when ever danger. So self-preservation is one. And passing on our DNA would be another one.

Yeah, mating. Mating. Mating is, I don't think it's as sophisticated as passing on our DNA. It serves that utility.

It serves that function. But the emphasis for, and then there's the male-female variance in what our instincts tell us to do. And that makes certain romantic comedies funny. Those interplays, for example.

But yeah, there's definitely an instinct around progeny. Yeah, so parents taking care of children. Having children taking care of children. Progenies maybe even the wrong term.

I mean, it has to do with mating. But then there's something after mating happens. Well, that is about the oxytocin system, for example, in females who are nursing. Uh-huh.

It makes them want to continue to take care of that baby. Yeah, I might want to take care of that baby. And then grandparents seem to want to take care of grandkids or want to take care of their kids' kids. There seems to be something, not every grandparent.

We're not saying every parent either. But there seems to be something in what we humans want. Some humans choose not to have children. And that's fine.

But we get to choose now at this point. But it seems like some humans choose not to mate, even if the mating is many humans. Humans that choose not to have children still tend to be mating. True.

Because that instinct is very rewarding. It is. Right? It's very well rewarded in general.

So that would be part of our utility function is to be rewarded. And behaviors that get rewarded. That would be part of our utility function. I want that reward.

What makes you, what attracts and what repels. That is certain things. An example in my favorite, John and I, he has done some research into disgust. And so he's looked at incest, for example.

And how we just have a visceral instinctual reaction against it that has no logical basis. If you get rid of all the logical objections, we still have it kind of in general. So there are these things that seem to be. And is that discussed there?

You think because of the DNA problem of mating too close, relative to closer to the different parts? You could make that part. I don't know. Is that probably partially a total social conversation?

But for this conversation, it actually doesn't matter. The point is that disgust exists. And so our part of our utility function is to move away from things that disgust us, to move forward things that feel good in or that reward us. We like salt and fat and sugar.

Right. And even though the environmental context is changed dramatically, we still get lots of reward for putting those things into our mouths. And so we do it a lot. And you could, there probably arguments that the environment at certain times during our human evolution were those rewards were baked in because that was what was necessary for survival.

Only the ones that were putting that stuff in their mouth. Survived. Survived, yeah. And their children survived.

Well, yeah. And then the unbroken thread persisted through those people. And all the other ones were snapped. OK.

But let's go back to John's premise. What if, today, humans have gotten to a point where we don't, where we have surpassed those instincts? And we still have things that we want and don't want. I mean, I think that maybe what John is rebelling against is the idea of this kind of utility function is so sterile and so mechanistic.

And so reductive. And that maybe it does feel a lot less rich than the utility function we just described for human beings. We didn't even put a name on it. We just described it.

We described it the whole bunch of things. Yeah. You know, made it. And every human being is going to have a little different one, too.

I mean, but there are some pretty broad themes that would make us look to an outside observer. We would all look pretty similar in this regard. I think this utility function. Within ourselves, it's kind of like human faces.

We see huge differences in human faces because we are so tuned to see the differences between human faces. It's really not that much difference between human faces. Yeah. Yeah.

So there's not that much difference between human utility functions. But OK, so what if John's right? It seems obvious, I think, to both of us that he's not right. But what if he was right?

What are we missing that doesn't allow us to fully examine this question about utility function? Well, you could basically ask the question, like, to rephrase this would be, what if humans don't make judgments? That's exactly a rephrasing of this. That's what utility function is.

That's exactly what it is. What if you use to make a judgment? Yeah. And utility makes it seem like it's all rational, like it's all spreadcheedy.

It's not. It's like, so utility here could be, I get something I want. And what I want right now is I want that bag of chips. So I'm getting utility from that.

Is it actually good for me? Is the net outcome actually a good thing? Is my health better from a spreadsheet perspective probably not? But that's not what we mean by this, right?

So to John's point about Adam Smith, the invisible hand, or I don't know exactly what John's referring to, but the idea that in economics, all people are rational actors, and therefore the free market system is the one that gets most people what they need, or gets the most utility. Maybe that's the problem is that we use the word utility as a very economic kind of a word. And we know that the theory of economics doesn't really line up very well with our irrational human beings. You know, economics treats people as rational economic actors, but we're not rational.

We have all kinds of things that we do for all kinds of reasons. Well, yeah, I think it assumes a certain set of values that don't actually line up with how humans actually act. Right. But I think we're pretty predictable, really.

Yeah. Like we predictably do this. We make the same irrational kinds of things over and over again. What is that the worth that book product predictably irrational?

Is there a book? Yeah. Dan Pink. I can't remember.

I'll look at it. I'll look at it. I'll put it in the show. It sounds great.

I, you know, I like the idea because I just said the words. And that's predictable about me. So if the words came out of my mouth, I'm going to really like them. Yeah.

Like Dan Pink wrote that book. Great. All right. Well, I'll look it up or if I don't, what if you can correct me in?

Yeah. So one of the things that we've not really talked about on the show much, but I think that this kind of maybe leads into a smidge, is, you know, what is, what's going to get us from here to there? You mean from here to general purpose? Yeah.

Yeah. And how close are we to knowing what it is we need to get from here to there? Yeah. So can we think about that in kind of a summary form, like maybe to queue up for another show?

That, let's do that. OK. It's almost as if we talked about doing this before the show. Almost.

Almost like that. Yeah. So thinking about, there's really, I think, two main, two main different things that we know are going to go into, the general AI. And they fall under sort of the category of, and I mean, I get this all right here, but there's kind of an input output function that we can use to make quick decisions.

We put in some state. We look at the world. We see a picture of this and that thing moving and this thing happening. And we have a kind of an instantaneous assessment of both the state of the world and what we should do in response to that.

So it's kind of like when your hand touches the stove or something hot, you immediately jerk it away. There's no planning there. There's no anything there. You have loaded up a response to that stimulus.

And you walk out your front door. If something coming from the sky towards you, you duck because that's that one is loaded up in like the cash, the processor cash. Yeah. And it takes over.

No, no, it's a bird. You don't you haven't analyzed it. You don't know if it's a shooting star or what you don't know what it is. But it's not normal.

It's not normal to have something occur like that. So we call it stimulus response. Okay. So that's a critical part of it.

And if you think about deep learning, that is essentially creating a function that maps input to output stimulus response. So the input is photos of cats and dogs and sheep. Yeah, or a board position and go. Okay.

And the output is the response. It's the move and go label of the image. It's the move and go. It's the it's yeah.

And any go might be a good one to talk about as an example. Okay. Because then that takes us to the next part, which is this ability to correlate future rewards with previous actions and learn from that. So to actually string together sequences of actions.

The last time I saw a situation similar to this, I did this and ultimately 25 moves it down the road or 200 move down the road, I lost. Yeah. And so now I need to update my visceral response functions so that the next time I come to that critical juncture where I made the good choice or I made the bad choice, I do the right thing again or I do a different thing. Yeah.

It's better. Yeah. And so that one, I think that the header for that one might be reinforcement learning. Okay.

And, you know, I think there are different, there are different labels that you can apply to that. But it's, it's really, it's really applying learning to, you know, we learned something surprising in the future about the, the reward of the sequence we took. And so we then propagate that backward in time to apply to tweak the parameters of our instantaneous judging things. So, so creating a plan then is, uses these two things.

So if we know these 5,000 people have skin cancer and we have pictures of 100,000 people three years before we can back, we can go back and look which ones early on resulted in skin cancer. We can really maybe start to start to realize that actually this is a precursor in, this is highly correlated with skin cancer in the future. This sort of, it's not skin cancer yet, but it's correlated. And our best doctors have sort of a sense about some of this stuff.

But if you did it with data rather than instinct, presumably it's going to get better. Yeah. So the, the go playing, uh, AlphaGo is a good example because it's simple to think about. And the way it's simple is that you can think, you know, AlphaGo has several things.

One of the things is its ability to evaluate a board position and say, which, which, you know, which side is winning is more likely to be winning that. So that's it's sort of, that's it's stimulus response function. I think they call that their policy network maybe. I can't remember.

It's been a while since I read it. But it's, it's the ability to evaluate. That may not be the policy network. Uh, forget the term there.

It's just the ability to judge. This is a good position. This is a bad position. Then they have another ability to make a judgment immediately.

So this is another network. Again, this is a stimulus response thing, which is from this board position, this looks like a good, a good move. This looks like another good move. This looks like another good move.

This looks like another good move. So they come up with a list of candidate good moves. So then what they do is they'll iteratively like project out in the future and say, this looks like a good one. This looks, yep.

Mm hmm. Looks good, good, good, good. Okay. So now let's say, let's use our first function to say if we play those out, which one looks best now and then we'll make, you know, so you're doing a search space.

You're doing a, you're looking out into the future to come up with. This is how people play go and chess a little bit too. When you look ahead of you move, like, huh, what happens if I, oh, yeah. Okay.

And then, yeah. Okay. So I'm going to make that move because in the future it's going to be good. But you're only doing that for one move at a time.

You're not doing that for all possible moves you could make right now. Not all possible. You can't, there's too many. You can't do that.

So you have to have that, that function that helps you find good ones to, to, to, to, to, you're pruning the truning all the time. Right. So this is these, but where do those come from? Where do those networks come from?

Well, they come from playing out games all the way to the end and then using the result of that game to feed back to all the decisions that were made during the game to improve both of those functions. So in the learning process early on, you're not doing the pruning or you are doing all the moves or how do you do the pruning when you don't know anything? You still have to prune. It's just a lot closer to random.

I see. So you're, the way that you actually learn is you actually have to finish games. Oh, I see. So it means that if you're doing a dramatic pruning, you may be pruning out the best move.

In fact, but absolutely you are often. But it doesn't matter because what you're learning is of the ones I didn't prune. Of the ones I didn't prune. Which ones are better, right?

I see. And then you're, you're playing against another agent that's doing the same thing. You're playing lots of these games and the functions are sort of all kind of getting better at the same time against each other. Yeah.

Okay. So, so, so yeah. So why does this, okay, so does this, as we talk about this, does this seem to you like it creates the foundations for a general intelligence? The idea that we do big categories.

They see that we learn from what we're doing. So instantaneous, let's try to come up with some types. So instantaneous judgment, instantaneous judgment. What's the situation I'm in right now?

Yeah. And, and, you know, what is my reaction? Instantaneous judgment. Okay.

And then two, a principle mechanism for taking future knowledge and using it to, to condition all the decisions we made before and learn from them. Reinforcement learning. Yeah. It feels like these are powerful, these are powerful elements.

Yeah. And the two, I think those are the two things that I would like to focus on. Maybe, maybe an episode where we try to gain a little intuition for, for what each of those things are, or maybe an episode for each one. Yeah.

That sounds great. And maybe we can pull examples from existing AI, where, you know, these are really nice. Self-driving cars would be another example of something that needs to have instantaneous judgment. Yep.

Clearly. But also needs to learn a lot about whether it's predictions actually came true or not. Like you think about all the Tesla's driving around on the world. That's giving you all kinds of sequential data that you can train on.

Yeah. I haven't looked closely at what's come out of that. I believe that they absolutely are feeding all that data back into their iteration. Well, that's pretty exciting.

We should, we should look at that. Yeah. By the way, totally off topic of AI. Did you hear about the Tesla owner who saved the life of the unconscious driver on the German Autobahn?

No. There's a Tesla owner saw a guy driving a Volkswagen or something who was driving erratically and banged into the guardrail and stuff a couple times at highway speed. And the Tesla owner saw, it came up next to them, saw that they were unconscious, pulled the Tesla in front of the other car and gently slowed down, slowed down until they bumper, you know, until they were bumpers were touching and then slowed down and slowed down and brought the other car to a stop, stopped saving the other guy's life. Wow.

Destroy the car. Yeah. But you know, Tesla gave them all the repairs for free and that's sure. I did.

Of course. I mean, totally off topic, but just fascinating. Yeah. So what was the utility function that was in play?

Right. I mean, that's a good Samaritan. That's like total non the opposite of self preservation, the guy that's saving somebody else's life. It's some altruistic kind of like for the good of the other.

I want to help somebody. It's amazing. Yeah. So there's a lot of debate about that in the sort of evolutionary biology community about where that comes from and what it means.

I think. I think so. Yeah. It's cool.

Whether there's some sort of group selection effect in play here or reward the hero. We reward the self list. So it's not actually self list. Yeah.

And when your group does better, if you have those genes, your group does better, which means that you do better. Yeah. So you have to reward then. But this is a group selection for the most part, I think is definitely mainstream is poo-pooed as a mechanism in natural selection.

Although I'm very partial to it. Like it makes sense to me, but I'm not an evolutionary biologist. Yeah. It sucks on things that make sense or not actually true.

Yeah. Well, you know, I still, there's still enough respectable people, respectable. Respectable sort of social scientists and evolutionary biologists that I think it's still on the fringes. Yeah.

But really, it's still on the fringes. All right. Well, let's bring it back. So we're going to talk about reinforcement learning.

Yeah. And then this, what do we call, instantaneous judgment? Instantaneous judgment. We'll talk about those on a future show.

If you've got something to add on those, send us a voice memo to concerningai at gmail.com. Make sure you say your name and where you're from. By the way, a couple of people entering the Facebook group over the last week, a couple of people from South Africa, somebody from China. It's exciting to see, you know, to see where everybody's checking in.

This really is a global conversation developing AI. It's really a fun thing to have all of you with us. Yeah. And look forward to some more voices to put on the show.

Send us your messages. Send them on. All right. See you next time.

Until then, Ted. Keep your waffy up. That's got a real ring. Yes.

Brandon. And that's... This is Ted. Yeah.

Yeah. Don't apologize. Okay. That's genius.

Okay. Bye.