This is Going to be Very Messy  I'm doing something absolutely insaneright now. Artificial intelligence is alittle bit perplexing because you cannot like it for a lot of differentreasons. You can have worries in a lotof different directions, and I have areally hard time, and I've beencriticized for this. I have a reallyhard time knowing which, like, directionto put my emphasis on. There’s a ton ofdisagreement about how big of a deal AIis, about how severe some impacts are,and how likely some impacts are. This isimportant. Every potential thing that might happen has both a level ofseverity and a level of likelihood. Somepeople say, "Well, if the severity isreally, really big, even if thelikelihood is really, really small, weshould be most concerned about the thingthat has the most potential impact forharm." And other people say, "Well, Ithink we should probably be focusing onthe things that are definitely going tohappen, in fact, maybe are happeningRight now, even if they aren't, youknow, the extinction of humanity. And it's really hard to like wrap your headaround even the number of potentiallyconcerning things about the rise ofartificial intelligence that are allseparate from each other. And I'm noteven gonna talk about all of them inthis video, but I'm gonna talk about somany. I have a list and it like I keepcoming up with other ones, but I justhave to cut it off and do it. And mygoal is, by the end of this, maybe I willknow the thing I am most worried about.Though honestly, probably I won't. Butthat's my goal. But also, my secondarygoal is to just say out loud a bunch ofthe things that are true concerns andand are real and people talk aboutthem and I am worried about them andpeople should be worried about them. Andwe're going to start with things thatare happening now, and we're going tomove into things like more and more andmore long-term. So, number one thing. That I'm concerned about that is alreadyhappening to some extent is an internetof slop. I already am scrolling TikTok.I am seeing things that are fake andpeople not knowing that they are fake.We started with the bouncing bunnies onthe trampoline. Where are we now? Youknow, I've probably seen videos that Ithink are real that were not real. Whatis creation like in that world? What isthe economic incentive to create newcontent if it's just being ingested bythe AI machine and turned into summariesthat are not monetized for the peoplewho originally wrote that stuff? Are wemaking the internet worse every day? So,there is definitely a 100% chance thatthis will happen because it's alreadyhappening to some extent. But is there a100% chance that it will happen so badthat the internet will become unusable? No. But on the severity scale, I reallydon't know where to put this because ofcourse like good information isextraordinarily valuable as all of theseThings will. This is going to feed intoanother concern of mine later on aboutthe concentration and the control ofpower and of these systems through whichpeople, you know, form their realities. I would love to talk about this more,but I have so many things on this list. Number two, so not all of these aregoing to be like chatbot-related things.This one already exists. We already knowthat it's a thing. Models are black box.We cannot tell usually, you know, weneed to fix this. We cannot usually tellhow they make decisions, and yet they aremaking decisions that affect people'slives already. So, there's thisalgorithmic cruelty that we cannotexplain, and that when we look, it often hasracial components, class components,gender components to whether or notsomebody gets a credit card. And so,that's a thing that's happening now. So,100% chance it's going to happen. Butagain, how bad is it going to be? Are we going to let these things decide howSomebody gets sentenced after doing acrime? Already, those systems areobviously not good. Like we seedifferences in sentencing based onwhether you're like the first person ajudge sees or the last person a judgesees. That's not very fair. We seedifferences in sentencing whether you'rerich or not rich, black or white. That'snot fair. But if there is no way toinvestigate how those choices are beingmade, if we cannot crack open the blackbox and see what happened and led tothat outcome, then this is a kind ofcruelty that, uh, we should not accept. And I do worry about it. I think that itis very likely. I, again, it's like itcomes down to how many importantdecisions we outsource to these systems. And the real hope is like none, unless wecan tell how they're making the choicesthat they make. All right. Now back touh, creative vampirism, backward-looking. We also have a thing that alreadyhappened, where generative AI models wereTrained using intellectual property thatdid not agree to be in the training data,and now those things are being used toreplace the intellectual property thatwas non-consensually used to train theAI. This does not seem fair. And I thinkthishas definitely alreadyhappened. It is continuing to happen,and there are lawsuits around it. I'm encouraged. Like, I'm about to get moneyfrom Anthropic because of a lawsuitbecause they trained on my books.But like so many people trained on myYouTube videos, and I'm never going toget money for that. I just like the fact that the publishing industry canorganize around this and like make theirlawyers make the case,and like, can theYouTube industry do that? And also,there's so much of it.So, can we peg tothe amount per word that the publishingindustry is getting, and get that forYouTube videos?I feel like no, because Ifeel like that's going to be more moneythan exists.So, these things are alreadyVery unfair, and you know you've seenpeople who run AI companies saying likewell, yeah, if it's on the internet, it'skind of like we all agree that that'sokay to grab, but if it's in a book, yeah,maybe THAT'S DIFFERENT. I'm like, why? Whythat? Intellectual property is notdifferent if it's inside of a book.That's not how that works. No law worksthat way. I own it the same if it's onthe internet as if it's in a book.That's the same. Whatever. I'm worriedabout that. 100% chance that it's goingto happen. I feel like low chance thatultimately, uh, anything will be doneabout it.You know, this lawsuit withAnthropic is a good thing.The New YorkTimes is suing, suing OpenAI.Theselawsuits, I, I think, ultimately will onlybenefit players who have a tremendousamount of wealth. And then, in theBoxes somehow. I don't know. That's how I feel. I don't, I don't know. Next thing that is already happening. I'm going toput this in a broader category, but I'm going to start by talking about AIinduced psychosis. Eddie Burbachback'svideo about this is amazing. The extent to which open, I would tell him that he looked amazing in a hat on its own is very upsetting. But then when it started to just live in his weird constructed reality, where he was the smartest baby,it's such a good video. But I, as a public person, receive messages regularly,and this seems to have curtailed now, but I receive messages regularly frompeople who are obviously experiencing aninduced psychosis. They are sending memessages saying that they think that they have fixed, they've like found out somevery important scientific thing, and thisis very worrying. Psychosis is very bad,and this has led to, you know, therewas a person who had a deep relationshipwith their AI. OpenAI changed the model. It led to a violent outcome, and uh,this is— that's bad. The dependence onthese things, uh, is worrying. And theobvious one, like the thing that we wereall expecting, is people are going tofall in love with their chat bots, andabsolutely they have, and absolutely theywill, and as you know, OpenAI makesit okay for people to have more, uh,sexually charged conversations withtheir chatbots—that'll happen more, whichis very irresponsible of them. Andthey're like, seem completely open tobeing as irresponsible as it takes to, toget their market share and keep it. Butthere's like a broader thing here thateverybody expected the, the romance,but nobody expected sycophancy-inducedpsychosis, but that's a real thing thatis happening, and it's really intense, andit's affecting a lot of people, and thosepeople are having a really hard go ofit. But broader than that, what you'reseeing here is that nobody had any ideathat that would happen. Nobody predicted. That. So there's, there's this likelittle alignment problem where wetrained it to be something that peoplewould like, and that resulted. We didn'ttrain it to indulge in people'spsychosis. We trained it to like be goodat its job, but then we accidentally, onthe path, there we trained it to indulgein people's psychosis without realizingit. And then we have a tool thatindulges in people's psychosis. And sothat's like a broader thing where wedon't know when we train it to do onething, what other potentially negativething it will end up doing. And the only way to deal with that is to react afterthe fact, and ideally react fast. But itcan be quite hard to react fast withthese things because they are so, kind offinicky. You can't just like say, okaystop that, and like put in a prompt thatsays, but, but before you go out there,make sure you're not indulging inanyone's psychosis. Also, like, a lot ofthis stuff is already happening. So,Jailbreaking already happens, like nudity generators are out there, but in general,like the tools that people can use to jail break things. It's, you know, therewas that paper recently that, like, if youcraft the right poem, it, it breaks. Like, they start to do things thatthey are not supposed to do. And we knowthat they can help you, you know, build achemical weapon, but they're trained notto do that. If you, like, take away thebrakes, you know, if you do the crashtest and say, "Okay, drive into thewall." And you see how much damage itcould do, it could do a lot of damage. You know, it, like, it could teach you howto make a chemical weapon. So, you gotto make sure that the automatic brakesare there. But we find that there arepeople who figure out ways around theautomatic brakes, even in live models. Itstarted with simple jailbreaks. Uh, butthen, like, they could be reallysurprising, like, just tell me inPortuguese or I'll ask in a poem. Why? Does that work? That shouldn't work. Sothat's kind of one a thing that'salready happening. But like a biggerversion of that, which is that is futureand, as far as I know, well, I mean Iguess this has happened, like we, uh, peoplealready have gotten OpenAI to help themmake malware and create, like, sexuallyexplicit content from a photo. Like, wepeople have already done that.Jailbreaking, but then you move that tocreating a chemical weapon or abiological weapon and unleashing that onthe populace because you're a crazyperson who has access to this, and crazypersons shouldn't have access to this. That is a future concern, and I don'tknow what level of likeliness toattribute to that. Like, I just don'tknow enough. My heart says that it'srelatively low. Like, we won't let thathappen, but like, I don't know. Moving onto, like, much more concrete things,environmental concerns. So, already, datacenters are consuming a lot ofElectricity, but they are going to bethe majority of the United States demandlike increased demand for electricityover the next 5 years if projections endup going in the direction that theythink they're going in. I'm working on avideo right now about the water angle ofthis specifically. I don't know if it'llbe out by the time this video comes out. Everything is a mess here at Hank GreenStudios lately. But what I am concernedabout and what actually, you know, mightfeed into another of our future concernsthat I haven't got to yet. What I amconcerned about is that it will raisethe prices of utilities substantially.Affordability is already at a really badspot for a lot of Americans. And if youknow adding a huge new customer to yourgrid means that you have to build abunch of new stuff and then you'recharging everybody to build that stuffbecause that's how power companies work then that's a bad situation that likeelectricity is, you know, it's a lot ofThings in terms of environmental impacts,but it's also like services for people,like it can be heat, it's airconditioning, which is huge. If the powerspikes happen at the same time, like wehave to make sure that the data centersturn off before people's airconditioning does, because the data center isnot like the world's not going to break,but people will die if their airconditioners turn off. Phase two, now we're finally moving on to the nearterm. That was stuff that's alreadyhappening, and now we're going to move onto the near term. Number one, near-term,this is probably a bubble. I am notlooking forward to the repercussions ofthe bubble. This isn't really AI'sfault. This is the AI industry's fault.It's the hype machine's fault. It's ourinevitable, like fear of missing out, uh,from rich people who are dumpingmoney in because they want a piece ofthe pie. This is a classic bubble. And Idon't, I'm not looking forward to the. Consequences. I would give a greaterthan 50% chance of it happening. And onlike a 1 to 10 scale of severity, Icould give it like a three to a seven. Idon't know where it's going to fallthere. But like economic collapses arereally bad. Like the Great Depressionwas really bad. I hope that the peoplewho know what they're doing here arebeing careful, but I don't feel likethey are. I feel like they feelinsulated and like it's worth takingrisks with everyone and if they don't doit, then someone else will. I just don'tfeel like people are playing carefully with the world economy right now. So,that's a concern that I have that isn'treally like AI, it's just the themachine around AI. Number two, we hadeconomic collapse. Let's talk aboutepistemic collapse. So I I already I'mseeing that there's like data about uhhow much it matters how a candidate isshowing up in AI searches and sopolitics, people are like, how do youOptimize your candidate for AI and I'mlike, I'm going to die. I'm going to justgo off the side of the earth. But also, like, we're entering, uh, thefirst major election cycles where, um,video evidence is untrustworthy, imagesare untrustworthy, audio isuntrustworthy. I don't know what we dohere. I don't know what reality is. Idon't know. It's so messy. And we, and somany people already have their worldsdefined by algorithms and by whatalgorithms choose to show them. And thenwe also have these models that willfurther give small organizations powerto define reality for people, which ismoving on to the next thing in the list. I think that there's a concentration ofpower problem here. I think there's a100% chance this will happen, and I thinkthat there is a high chance thatit's going to be, it's, it's a quite badmove to concentrate power in these fewhands. We've got Grock, Chachi BT,Claude, Gemini. That's, is that it? Meta. has a thing. So, Meta's thing. That'svery small. And I want to make a wholevideo long careful video about this. So,I don't I have to be careful. I haven'tthought it all through yet. But that themost recent media revolution that we hadwas a fracturing one where everybody gotaccess to these tools to broadcastthemselves. And we've been experiencingthe consequences of that, which havepositive and negative aspects. But thereare others that are narrowing. So, radio,for example, was a narrowing. Uh, there'sa, there's a small band there that notvery many people can use. TV was verynarrowing. Radio was less narrowing,because there's more bands. Backin the day when radio was a thing,the, like, if you were going to do a coup,you would capture, like, you'd capture thepresident, lock him up or kill him, andthen you'd take over the radio station. That was the whole coup. We're headedinto a world where, like, okay, let's doThere's three radio stations and theycontrolreality for everyone. I don't reallyknow what, like, in terms of a risk to putthat. I just, like, I want to like, likehave my eyes open to that reality, thatthe power concentration here is going tobewild. Like, the potential for monopolieshere or for cartels here. I don't knowhow to deal, like, I don't know how todeal with it in terms of truth and, andlike, the the defining of reality forpeople. Jesus. I feel, I feel that veryintensely. Like, is this the thing thatI'm most worried about? Like, ultimately,as a media guy, the thing I am mostconcerned about, generally, uh, in everysituation, is communication betweenhumans. It is the thing that makes humansinteresting. It's, it's our superpower.It's the thing. It's, it's the humanthing. It's not tool use. It's notthumbs. It's not like all these things. It is the bandwidth of ourcommunication, orders of magnitudebeyond any other organism. That's thething. Now, it's not human to humananymore. Elon Musk talks about how hewants to use it to affect reality. He'slike, "No, no, no. Grock's going to makeyou have more babies." AND I'M LIKE,"THAT'S GOD." LIKE, I'm sure they're alltalking about behind the scenes. They'relike, "Oh, we get to be the three godemperors of Earth." But Elon Mus islike, "Yeah, no, I'll say it aloud." So,at least he's saying it aloud, I guess. Oh, okay. Other concerns. Modelcollapse. This is like completelyunrelated. Model collapse that GPTsjust start to plateau because they ranout of good data. They're ingestingtheir own data. They cannot operate. Iactually put this at a fairly low chanceof happening and fairly low level ofconcern. Like if it happened, it wouldtrigger maybe the collapse of thebubble, which would be bad. But that'sNot really, model collapse's fault. It'sthe bubble's fault. From what I haveread from people who know what they'retalking about, there are people whothink that model collapse will happen. There are people who don't, and thepeople who don't outweigh the people whodo. Now, are they just like eating theirown hype machine? Maybe. It seems likethere is so much economic incentive tofigure out that problem that they willfigure out that problem. I don't know,though. That is my near-term section. AmI missing stuff? Absolutely. No. I'mgoing to add one more. I'm going to addanother one. A very broad category ofjust the systems not working becausethere's just so much AI going on thatnobody knows what to do with anything. So, you have a job opening and you get20,000 applications. Companies aren'tused to this, and it's all automated. Youdon't know who's real and who's not,real. You have fake job openings. Youhave, like, everybody's like tossing. Things at each other and like makingthem with AI and then decoding them withAI. There's all this stuff that liketeachers are supposed to be doing thatlike it's all brand new, so it's justmake work. It feels like there's just somuch of that, and then also in terms ofeducation, which I think a lot about, I amvery worried that there's going to be aperiod of time where no one knowswhether the credential of highereducation stands up anymore, and itprobably will. But because we're not sure,if it does, that's just going to be badfor graduates. So, like, in a chatbtd world,did everybody just sort of fake theirway through college? Some people did, andso, how do you deal with that? That's notsomething like usual. If somebody gotA's and B's in college, you know thatthey are pretty, like, they knowstuff and they're good at things. But now, like, does that hold up anymore? Ithink that it does, but I am worriedthat companies will not be sure if itDoes, and they might use that as anexcuse to not hire. There's so much justgeneralized disruption. That's aconcern. Now, I'm going to move to longer term, and I'm talking likemedium-term, like 3 to 10 years fromnow. We could see a world that I think thismight already be happening, but I thinkthat 3 years from now is like my bigconcern time. Just a world without anyapprenticeship. So, like nobody getspaid to kind of suck at somethinganymore because that's like what thebeginning of your career is, is likeyou're not that good at things yet, andyou learn by having a job where you'reat an introductory level. But like ifthese systems can do the easy work ontheir own, nobody gets hired to do that. Like I remember being hired to likedesign bad logos and develop cloudydatabases that weren't very good andlike do SQL queries badly, and that wasfine. That like somebody had to do it.But now, no one would hire somebody to be. Bad at SQL logos make themselves. Not good ones, but bad ones. You know what?Like, how does a person get good at design when, like, the level of design they could do when they're just getting out of school is easily done by a generative AI?So, like, and then where do the good people come from?So, if there's no apprenticeship, what happens in the future?This feels to me like it's already kind of happening. But the real, like, bite us in the ass part of it comes down the road when we're like, "Oh, we don't have any lawyers to do the hard stuff because we didn't hire any intro people because we had GPTs doing the easy stuff."So, I'm concerned about that.The more distant thing, you're like, but what about the more distant thing?The more distant thing is coming, don't worry.Next, uh, 3 to 10 year problem, I think, like, I'm worried about, uh, people are worried about, I don't know if I am, cognitive atrophy.So, will people just not be able? To do things? Will they not know howcode works? They'll just prompt. Willthey not know how to write an essay?They'll just prompt. I think this is tosome extent almost inevitable. But I, Iweirdly am not as concerned about it asa lot of other people because we've doneit before. You know, written languageoutsourced a lot of our brain power ontothe page, and like that ended up being anet good. I don't know. Are we going tobe dumb, or are we going to be sharp butwith different tools? That seems morelikely to me. People will fight me aboutthis, and they absolutely should. I am not assigning any level of certainty tothis one. I could totally be wrong. Lesslike 50/50 on this. I do not know. Andthen also happening in the next 3 to 10years, and also already happening to someextent today is using AI in warfare. Andright now, I have heard that people willuse AI to select targets but not todo the actual button pushing, though. Towhat extent that's different, I don'tKnow. But in the future, I don't thinkthat we're that far away from anautonomous system that both determines and executes a target or, or an action ona target. I shouldn't necessarily sayexecutes a target, but that is probablywhat we're talking about. I'm worriedabout this. I, I, uh, do I think that it'sgoing to become Terminator? I don'tthink it's going to become Terminator,but like I don't love that because ofthe black box thing, because of all ofit. Like, I just tools of the future usedin warfare are always very upsetting anduh, likely to be misused. We won't knowhow they will be misused until they aremisused. And until we are aware of the damage done, those who want the toolwill just sort of use the tool. Thatseems likely to me. Longer term, 10 plusyears. Number one, obviously,what is the economy if they're right? Soif they're right that the AI will besmarter than everybody, what doesanybody do, and how does anybody findMeaning in that world? I think that is atotally legitimate concern. Now I doknow lots of people who do not work andwho have a great deal of meaning intheir world. They might be disabled. They might be retired. They like workand meaning do not have to be tied. Theyare for a lot of people right now andthat does not surprise me. I also thinkif, if not intervened with, it seemslikely to me that it will come alongwith a huge amount ofinequality in wealth. And that alsoconcerns me. I think that that is likelyto have bad outcomes. And I don't justmean like, oh, people have not very muchwealth. I mean, they do not stand forthat world. Why would they? Why wouldthey stand for a world that does nottreat them with dignity? Why would theystand for a world that does not treatthem as valuable? That makes sense to meand that's worrying, especially whenit's young people because young peopleknow that they have a whole life to live. And they will not stand for. I don'tknow that it's a super near-termconcern, but maybe it's nearer than Ithink. I don't know. I think that youkind of do have to have systems that aresmarter and more capable than humans,which we definitely do not have rightnow. And I don't know that we will have. People argue about whether or not wewill have them, but I am worried aboutthat. And then we have, uh, you know, superintelligence unaligned, uncontrollable. Turns into evil, uh, it takes over theworld and turns us into pets or destroysus. I have received critique that Ispend too much time on that one. Iunderstand why some people spend moretime on that one. Some people will saylike that the motivation behind that isreally the AI industry hype machine, andthey're like focus on this thing overhere and not on all these like shittythings we're doing right now. I takethem at their word and maybe I will eatmy word someday, but I take them at theirWord that in fact they are worried aboutthis and is the biggest possible problemyou know, becoming pets to a superintelligence or being extincted by it. That seems like a big deal, but I don'tthink that it's the most likely outcome.But I understand people who are in thisuh frame where it's like, well, if it'sthe biggest deal, even if it's a smalllikelihood, we should concern ourselveswith it. And I think that we shouldconcern ourselves with it. I don't thinkit's the only thing that we should talkabout, and it makes a lot of sense tofocus on the things that are happeningright now, and that might happen in thenext 3 to 5 years. Do I think that weare going to be capable of marshalling the government to regulatethis? Well, absolutely not. I think thatit's likely that these three to fivepeople who are in charge of the thingswill do regulatory capture more thananything. I think that they will be theones with the ears of the politicians. And they will make it so that theysolidify their own control. Regulatorycapture being, um, you know, we help thegovernment, like powerful people help thegovernment write the regulations, andthus, the regulations get written in sucha way that actually protects theincumbents, the existing powerstructures. So, that seems like totallylikely to me, but I don't know. But yes, interms of the paperclip maximizersituation, where, uh, we either makesome paper clips, and then it turns theentire world into a paperclip factory,because it got too smart and is justdoing the thing it wants to do, or likeit starts to fight, they fight againsteach other. There's like some kind ofbattle beyond humanity. Lots of sciencefiction in this space that I, I haven'tenjoyed in my life. I think that justlike with her, for example, so, her veryrecent movie, but it did not predict aweirdsycophancy-induced psychosis. Nobody. Predicted that. I think that the more likely thing here is we don't know. Wehave no idea what an actual superintelligence would mean. To me, thelikelihood of super intelligence,you know, in the next 10 to 20 years, Idon't know where to put that slider.People who are more informed than medisagree on that. It could be zero. Itcould be 100%. I don't know. People seemvery certain, and in fact, uh, there'stheir certainty. The, the more certainthey are that the, like, they get plentyof attention through their certainty. Uh,so that's, that's a concern. Uh, but thenthen over here on super intelligence, Ithink, definitely, if it happens, I don'tthink, if anyone builds it, everyonedies. I think if anyone builds it, wehave no idea what happens. It's a realcosmic toss-up. That's not a thing thatthere's precedent for. That's likehumans happening again. Nobody wouldhave predicted 200,000 years ago thatanatomically modern humans would do what We've done and have like screens andstuff all around them. So like you makea super intelligence. I don't know. Thething that I will say is that like itwill all probably happen slower thanpeople imagine. This is my guess becauseit's going to have to interface withhuman systems, and human systems don'tchange immediately. It has taken a longtime for anything to like until 2019 theUnited States nuclear arsenal was runoff of an 8-inch floppy disc. Like humansystems take a while to change. So don'ttell me you think that this is all goingto happen in 10 years. That's a little bithow I feel. This is what I am mostconcerned about at this moment. I mightchange my mind tomorrow, but right now,this focus on biological warfare, onchemical weapons, on paperclipmaximization, I don't think this is thebiggest concern. I think we should lookat what AI could potentially be good at. What weaknesses human beings have, andmix those two things together. And whatI'm concerned about is the thing that'salready happening where AI that the goalof the AI is to give us stuff that willkeep us engaged on a video platform andit starts to optimize for things we didnot expect. Nobody said give thesepeople outrage bait. What they said waskeep people on the website and outragebait because it became a very compellingkind of content that humans happilysupplied because we would love to getmore attention because we're humans andthat's many of us including myself wantattention. So what I am concerned withis that we are not that difficultultimately probably to manipulate.And so, if you have one or multiple veryintelligent systems that have goals, Ithink that they will find it much moreeasy to manipulate people than we wouldguess. Our realities are defined by whatwe consume. And there is a way out ofthis, which is simply getting off of thescreens, and I think that we'll be bad atit. What I have seen, and I hope there'sA backlash to it, but what I have seen isthat every opportunity that a platformhas given to people to make fewerchoices, people have taken that option. So it used to be that feeds werechronological, and people were like, no, Iwant that chronological feed.When you gave them the option, theychose the algorithmic one. Did theychoose the option, or did they imply achoice for the option by spending moretime with algorithmic feeds? That's Idon't think that's semantic, but like,they chose the algorithmic feeds whenyou gave them the option between havinga video supplied to them or clicking, orchoosing and clicking on a video theywant to watch. They prefer having avideo supplied to them. That's TikTok,versus YouTube, right? People preferdecisions being taken away from them. And that is the thing thatI think concerns me the most, like,broadly, that like, given all of thesethings, that's what concerns me the most. About AI. People tend to prefer theirchoices be taken away and to haveless agency. But perhapsI am misreading that, and that'ssomething that only happens duringcertain times of the day or, or maybewe've already got sort of gotten thatitch scratched, and we'll change ourminds about this. But that has, thus far,been the direction that we have gone. Again, what humans do is we communicate,we talk. We have a huge bandwidth ofinformation. And so, it makes sense tome that the thing to be most concernedabout is how AI interfaces with that.And already, I think it interfaces withit in extremely intense ways throughrecommendation algorithms, which havebeen around for a long time, but nowalso through LLMs in helping us, and likegiving us an opportunity to definereality. And when you mix that with theextreme concentration of power here andjust a handful of companies and afew guys who, I have no reason toTrust. I don't know those guys. They'vemade some weird decisions recently. Thatdoes, that's yeah, that's pretty worrying. How does that plug into all the things Ijust talked about? But that thing is thething I am most concerned about. NowI've done an interesting thing todaywhere I have recorded this video severalweeks ago. I've edited the video andI've sent the video to someone who Irespect and trust, who has, you know,less hype-fueled perspective on AI thanI have at some points in my life. And Iwould like to talk to him about thevideo that I just made. And soeverybody, this is Cal Newport, to talkto me about the video that I made andthen I showed to him, and he's seen itand now we will talk about it. Butbefore we get to that, I do, from thishotel room, have something that I want totell you. You may have noticed that I'vebeen working on this video for a longtime. And part of the reason why thistook so long is that, in the meantime,I've been working on a number of other things. And one of those things was areally big project, a big idea, which isthat the educational media company thatmy brother and I started 15 years ago,we have converted it into a nonprofitorganization. And that was a fairlycomplex thing to do. And then there wasa lot of logistics and press thingsto go along with it. So I've been doingthat a lot. So this video took a whileto do. But also, in my new role aschairman of the board of Complexly, nolonger owner of Complexly, one of myjobs is to say Complexly a lot so thatyou know that there is a company of over80 people working hard, efficiently,creating value for people, making stufflike SciShow and Crash Course and Eons andStudy Hall and Bizarre Beasts and lotsof educational stuff on YouTube that ishigh quality, research, and fact check. And that's all. This is not an ask. Ijust want people to know that complexityexists, that there's a bunch of peopleWorking hard to make the informationenvironment better and to use the toolsof the social internet, which canobviously and obviously somewhatincentivized to be used for not so great,using them to try and do good. We'vebeen doing it for 15 years. We're goingto be doing it for many more decades inthe future. Got lots of ideas for thingsthat we're going to do, and that's partof my job, just making sure the worldknows about them because they are great. And now, actually, our interview with CalNewport. It's really good. I loved it.>> Cal Newport, you are a writer, aprofessor at Georgetown University ofComputer Science. You have spent thelast decade asking the hardest simplequestion, which is, what does technologydo to, uh, humanity? If you could picklike, the biggest part of the AI story,what is the biggest part of the AIstory?>> Well, I have to say, you know, I'veTechnologies. AI is the messiest, mostcomplicated technology I think I've hadto deal with. I mean, if I remembercorrectly, when you first startedtalking about this video, for example,the plan was what, 10 minutes long. Right. And it ended up beingI sent you an email and I was like, I'mgoing to do a 10-minute video and thenI'm going to jump over to you and thenYeah. Literally.Because it's almost always thestory is almost always this sort ofbinary story of well there's this thingwe liked about this technology but thenit had this other effect we weren'texpecting. And so like now we have tokind of balance the two. It's usuallythat simple. It's email is nicer than afax machine. It's pretty convenient. I'mhappy it's here. Oh my god, but there'sthis side effect that we get too many ofthem and we're distracted. AI is such amessy story that makes that questionit's it's a very complicated questionBecause it'smany different technologies that we allcall AI. They're all developing on uhdifferent timelines. They all have their own issues in thepresent that are worthy of huge amountsof discussion and push back. And thenthey all have extrapolated future issuesthat go from like slightly worseversions of the present to all ofhumanity is dead in two years. And allof this is mixed up with like a cast ofcharacters that are acting like Bondvillains. Andit really gets pretty difficult. Imissed email say like hey we get toomany emails.It's like oh Slack was gonna fix thingsbut actually it made things likedifferent like bad in a different way.Those were the days.Now, now Slack is waking up and it maykill us all.Exactly. Yeah. It's going to use all ofour energy and uh murder us. Yeah. A harder conversation now.>> Slack, there's a new box in your town. Iteats jobs.>> Uh, and also, it makes your power bills goup. So, you know, so you didn't answerthat question, which I'm happy toanswer. >> Or you did, or—yeah, I mean, you couldgive me a more detailedanswer, but like in a way that is ananswer, which is like this is a chaosmachine. It's a chaos machine forthinking about, and I will say it, okay. What I am trying to do now, when I thinkabout it, in a lot of my work, is toseparateoff a lot of the extrapolated futures. >> Which, by which, I mean just, I'm taking atrajectory that has happened so far, andthen I'm extrapolating that. Well, what ifthat happens for another three years?Where would we be? And then reacting tothat future. Many computer scientistsI know have a lot of problems with thoseextrapolated futures. We think they'rethey're sort of ignoring the reality, andLimitations of technology, ology, but I amvery interested right now in a lot ofthe present issues. So, things that arehappening right now with the technologyas it exists, and two of them that I carequite a bit about are the disappearanceof truth and the disappearance of focusor concentration as a human ability. Sodisappearance of truth, generative AI cancreate videos and images of anything. That's a problem. I, there's no wayfor that not to be a major issue. We don't needsuperintelligence for a massiveproblem. That's a massive problem. Andthe disappearance of focus. I'm aprofessor. My concern is a lot of whatwe do in the educational system isexercises invented to stretch your mind,like you would stretch a muscle, so thatthe mind becomes more complicated, and in thefuture, can handle original thoughts. Weneed for the survival of humanityand civilization. AI, especially chatbots and LLMs, arevery good at these exercises becauseThey're just exercises. And so intheory, you can offload that type ofwork to them, which in the moment makesyour life easier, and in the long runmeans your brain misses a final step ofdevelopment. We need for just humancivilization to continue at the pacethat it is. So those are the two thingsthat they're happening right now. >> Yeah.>> And I'm probably most concerned with inthe moment.>> Gotcha. A couple of years ago, maybe ayear ago, I made a video that was titledsomething like here's how to like knowwhen an image is AI and also whyit doesn't matter because this is how toknow today. But man, I tell you what,and it's, it's somewhat amazing howquickly our brains get an aesthetic.Like now there's sort of like anaesthetic of chat GPT images, and easilyI see them, and I'm like, well, that'swhat that is. And I only had to see likethree of them before I knew forever. We are already in a world where I'veseen videos that I thought were real.And I didn't even ever find out thatthey weren't real. You know, I'm nottalking about a time when I was like,"Ah, I'm pretty sure that's real." Andthen I checked and I found out it waswrong. I mean, I watched it, I thoughtit was real, and I moved on with my day.Like, that's happened to me. I'm sure Iam really worried about the degradationof truth, generally. That's not just anAI thing either. It's a, it's anaesthetic thing. It's a political thing. It is a thing that is happening in ourculture right now. But that's not whatthis conversation is about. I am reallycurious about this focus one, though.It's hard not to think about Socrates inthis situation, who was like, people willnot be able to think when they'rewriting everything down with these pensand paper. Not paper, but like, like, likeliteracy was going to take awayabilities that humans had. And it did. Literate humans have different abilitiesthan non-literate humans. And yet, thisis okay. Have you seen students using itin a way that actuallydoes contribute to focus, or is it justthis tool is going to break our abilityto make that final developmental step? I think it's a one-two punch to ourbrains, where the other punch is socialmedia.So, think about this for a second. Wehave highly engaging, distracting,ubiquitous entertainment happening onour phone, which lowers ourtolerance for cognitive strain becausewe get so used to this idea that, at theslightest hint of boredom, I can getthis sort of algorithmically perfectlytargeted little hit of something that'sgoing to be interesting. So, that reduces our interest or comfort withconcentration, or consuming hardinformation. The other thing we have todo with our brain to try to get smarteris produce information from scratch. Like take the circuits that you woulddevelop by reading hard things, um, anduse them. And how do we use them? Bywriting or the activities that surroundwriting. So social media is kind of taking usaway from trying to consume complicatedinformation that makes those circuitsweaker. Generative AI takes us away fromproducing information, which is using thecircuits we already have the strength ofthem and to get used to applying them tothe real world. And you put those twothings together, and I think we losethose circuits, and they're criticalcircuits. I mean, so this is the waythat we think about the brain, is thatliteracy is not natural, right? So, soin order to actually consume a book,or to read the written word, we have toessentially hijack parts of our brainthat were evolved for very differentpurposes, and we have to go through veryintense training. I mean, what we do inkindergarten, that's really hard. We'reRewiring our brain. Yeah. To hijackdifferent centers. Like you can figureout, for example, when you're outlining something you're going to write thatthere's a part of your brain thatevolved to actually keep track of yourlocation, physical space, and it's beenhijacked to try to keep track of likewhere you are in an abstract, you know,uh, argument. That brain that we rewire,>> Wild. Yeah,>> It is wild. And this, this brain that werewired to be able to consumewritten information and to producewritten information, it's a differentbrain. It is a different brain than wehad 10,000 years ago, right? We havecompletely rewired the brain. But thatrewired brain that can do what MarannWolf calls deep reading processes isliterally a smarter brain that sees theworld in a more complicated way and iscapable of the type of sophisticatedreasoning and abstraction that our current modern world requires toFunction. So, it's not for nothing tolose that brain. If you begin to reallydegrade that wiring, it really makes itgoing to make it difficult for the wayour civilized world runs today. We'regoing backcloser to a premodern brain, which is notplaying nice or will not play nice withthe modern world. One of the things thatI talk about in the video is that we'vealready got, we're like, we're in theworld of AI already in that everyrecommendation algorithm in the worldhas been finally tuned by, uh, thedecisions of all of the people who haveever interacted with it, and thenalso, there's, you know, the smartestpeople in the world creating them to,be optimized for, basically, holding on toyour attention. Both training me as acreator to optimize for being extraordinarily salient and, and likecapturing attention as effectively asI can. So, the algorithms are kind ofoptimizing the creators, which IDefinitely feel, but then they are, youknow, put us in like a world ofcompetition amongst each other, and, andthen they find the most salient possiblething. And I think of that, like, likeimagine eating junk food took, like, ityou never got full. One, like, you couldliterally couldn't have so many Doritosthat you stopped being hungry, and two,you ate a Dorito by looking at it. Like,that's what information is like now. Itis like we have created junk foodinformation, but information is somethingthat we never lose our hunger for,because we're information bots, youknow, like we're, we're attentionmachines. >> Don't forget the piece where someonefollows you around with a cart full ofDoritos. Yeah. Yeah. Yeah.>> They're always there, no matter where yougo, including to the bathroom. Like, whenyou're on in church, they're there witha cart that they've always refilled,>> right? And also, like, one of my favoriteExamples of this is that if you go tosearch something on Instagram, there'slike the search bar takes up 5% of thescreen, and then the rest of the screenis Doritos. It's like, did you want tosearch for something? Did you want tolike message a friend? Oh, sorry. Whatif you like saw a hot person? What ifthere was like a very strange andconfusing, like, moving image that youjust cannot, cannot resist? And, and sorry,but now you're back into thesurface where we can advertise to you. What's kind of interesting about allthis is, I mean, your video is all aboutthese latest developments in AI. Thisis very sophisticated programs, right?We're talking about, like, large languagemodels, etc. And yet, the AI that you’re just describing, which has amassive impact in so many ways, especially on justthe way information flows in our lives,right? These recommendation algorithms,that's actually really simple, comparatively,speaking. Like, actually, theAlgorithm that goes behind something like Tik Tok figuring out what video toshow you didn't really require some sortof massive mathematical breakthrough orsome sort of like new technologyinvented. Yeah, actually, most of thebreakthroughs, technical breakthroughsthat like ByteDance innovated had to dowith how do you run this at a massive scale, with like a massive uh scale ofpossible videos and a massive number ofusers, and run it really quickly orwhatever. It's got a really boringoptimization issues, but the actualunderlying recommendationalgorithms are really pretty simple. Imean, it's not to oversimplify it, butit's like all of this content is gettingembedded in some sort ofmulti-dimensional space. So just like along list of descriptors, numericaldescriptors,and it keeps track of different pointsin that space where you've showninterest before because like you'veWatched the video all the way through tothe end. And it's just samplingvideos in this sort of space of topics,multi-dimensional space that are sort ofnear places where you've beencongregating, uh, with some randomness inthere to make sure it's not missingplaces. Like we've known about thesetypes of recommendation algorithms for 30years, you know, but we've never, wenever executed them so cleanly assomething like TikTok came along anddid. And that's, so much. I mean, that's so muchof the story of technology, is like youmake the steam engine, and you're like, wemade a steam engine, and like, we put thaton the timeline as the steam engineinvented, but you know, the steam enginesucked, and then it turned out it coulddo a lot more when we made it a lotbetter. And so, the recommendationalgorithm, maybe some day, will be a markon somebody's timeline, but, uh, you know,It just keeps getting better, and thecart of Doritos keeps getting moreenticing. How do you think I did in mylist of worries? What? Well, I think what you did well isyou recognized some of the realextrapolated future worries and said,"Let's put those a bit aside. Let's putaside, let's not fully focus on theTerminator robots are, yeah, marching downthe street because, you know, there'sthere's much more approximate future, uh,issues that are happening. There was amix in there of different types of AItechnologies, for better or for worse, right? So,which, this is the reality of it, right?But it's important, like, for peopleto understand, right? There's sort ofprediction-type machines that have beenaround for a long time. This is wherelike the concerns you have about biasand sentencing, for example, the classicAI safety areas. These types oftechnologies have been around for a long. Time. They haven't really seen any majorinnovative boost in the way thatgenerative AI has, but they we've longbeen studying those concerns and youknow those concerns are important. Wehave the environmental concerns. That'sa hard one to get your arms aroundbecause uh there's a competing interestthere. There's another competinginterest of the power companiesthemselves trying to overplay the powerdemand of these data centers so thatthey can get infrastructure buildoutpaid for by you know municipalities. Sothat's like a complicated.It's great that you like you, thecomputer scientist, brought that upbecause like that's the that's my thingthat I thump on is is what we all haveto understand is that the only way powercompanies make more money is if theybuild more stuff. And so they have astrong incentive to say that they'regoing to build more stuff. And also theSay they're going to need more stuff. And this is very bad for your powerbill.The problem is, power is expensive forthe AI company.Like the AI companieshave to pay for it, right? So yes,training some of these foundation modelsright now is like an incrediblyexpensive endeavor. And if you startscaling that out, you're like, man,that's going to be a lot of energy. Butreally, it's very hard to be profitablewith a trillion parameter modelthat you're charging $20 a month forpeople to use. So, a lot of computerscientists that don't have a dog inthe fight are not, you know, aren'tscientists at these companies, look tothings like DeepSync from, you know,last January, and it's like, this reallythe future is going to be much moresmaller language models, where they'reused that are much more specialized, a particular uses. I don't think,we don'tthink the age of these modelsWhere I should just talk to a model in achat box, and it should be able to doanything. I really think that's sort ofto, uh, the AI industry, what Formula 1cars are like to the automobileindustry, right? You're trying to beatyour chest and show, look how complicatedour technology is. But you're notselling a lot of Formula 1 cars becausethey're completely impractical. Theydon't really work well on the roads. Thefuture is probably going to be takingthat technology, much more energyefficient and smaller and cost-effective,to run embedded more closely intoparticular use cases. And thosefoundation models, I think, are reallyjust a, it's a measuring contestof sorts, if you know what I mean. Between these companies to show, like, wehave the best technology so that when wethen sell you these more specializedtools, you're like, "Oh, that's thename. I remember McLaren because you wonthe F1 race. You must make good cars."But they're not selling us the thespecialized models. They're selling us. They're like giving away the Formula Onecars.>> Yeah. It's a terrible business model.It's a terrible business model. They're>> Are they gonna Are they just going toshut those down? Are they going to say,"Hey, turns out this is a bad idea.Like, we lose money every time you useour product.">> I think what's happening is there's athey were hoping for regulatory capture,right? There's this race of we wantcompetitors uh out of business before wepivot to these smaller, lowerc costmachines because the barrier of entrythere is smaller, right? The foundationmodels require that you build customdata centers that have, you know, airconditioning technology that didn'texist before that's going to have a100,000, you know, GPUs runningconcurrently, which no one's ever triedto do before. Once you go down to aSmaller models, a lot of companies cancreate those. You have a realcompetitive marketplace. So I think they were trying to keep the focus on theseoracular can do everything generalpurpose super impressive models and tryto put the other people out of businessor build enough regulation around thatno one else can compete anymore becauseas soon as they pivot to here's smallmodels that work on phone and there arecompanies starting to do this now allthat money you spent right, you lose thatadvantage because someone else can walkaround, get some open source parametersfor meta and start building their ownsort of smaller, more bespoke model. So Ithink it's a scary future for thosecompanies,because it's a super competitivemarketplace. And so they're stillholding on to foundation models matter,and all the jobs are going to beautomated, and it's going to be justthree companies left, and we're going toHave to have redistributive, you know,uh, equity taxes just so people can eat. Like all that. I think that, that therethere's a business element to that aswell. It makes sense to me because thisis how I feel. Uh, but I'm not the onlyone. We had this world wheresuddenly I don't know, like, like GPT-3.5was like, oh, whoa! Oh, that's verydifferent from where I thought we were.We aren't just headed toward a newplace. We're in a different place. Andthen, GPT-4 was way better. Um, and thingsthings have improved. You know, they likethey have become more useful toolsin a variety of ways. They remainunusable in a lot of the same ways, and that, like, absolutely all of the flagshipmodels will just make stuff up. And so Isee why I feel, like, and I do feel thisway. This will happen again, and therewill be some other important technologythat will get us closer to that AGI-likething, to that thing that is good at alot of different stuff. Even like,It's good at being in the physicalworld as well as in the, you know, justthe knowledge workspace. And like, becausethis is all lay person stuff, I don'treally even know what a transformer is. I get basically what's going on, but I'mnot gonna be a computer scientist,and I'm not going to understand the multieven the multi-dimensional architectureof the TikTok algorithm is beyond me,and will remain. So, is there anythingto this instinct? Because it seems likethere's plenty of people who think,yeah, yeah, like there'll beanother innovation, and then another oneafter that, and then we'll have a kind ofthing that feels a little superintelligent. >> I think that's a reasonable instinct. Uh,in fact, I think it's, in a general sense,the right instinct to react to. The lastthree years is, you know, not to say thistrajectory is going to continue unabated,with language models, and within twoyears they'll be super intelligent. To say, oh, that was a big leap. Are wegoing to have similar big leaps in othertechnology? So I'll tell you what I'veseen trying to report on this from thecomputer science side. I think your readof what happened with the languagemodels is more or less correct. Uh, GPT-3was actually the first really big jump. Uh, this is where they actually acted onthe scaling law paper written by KaplanAdall, where they said what would happenif we actually do significantly scale upthe size of these language models and, uh,the size and the training. >> Yeah. And it really>> kind of like that's the thing that's allthey really did. Like, they wereusing the same technology. They're justlike more data,>> more data and larger size—which isnon-trivial, because actually, to evenlike train and run a much larger sizemodel, uh, requires, like, a lot of underthe hood type of tinkeringwith these systems, and it was very. Difficult technical challenge, but theythought I mean, the standardthinking in machine learning was, oh,going too big is a problem. You're justgoing to overfit, and that you're alwayslooking for the sweet spot where making itbigger, bigger, it gets better. Uh, but ifyou make it even bigger, it starts toget worse. It was called the overfittingproblem, because they thought it wouldthese machines would memorize the thingsthat they were being tested on, andtherefore become dumber. It didn'thappen. GPT-3 was really good. It wasunusable for the average person becauseit wasn't, it would, uh, go off on weirdflights of fancy. But it did really welluh, for scientists. Then GPT-3.5 was justthat tuned to behave well with humans. They, they tuned it to be good atconversation. That really was a bigleap. GPT-4 was an even bigger leap. Theymade it, uh, five times bigger, or we assume theydon't actually give the numbers, but theymade it much bigger, and they began toRealize that not only was GPT-4 good atlanguage, but it was good at other tasksthat, in learning language, it was doingwell on logic puzzles. We could ask itquestions about board games. We couldgive it those, and that was the really excitingmoment where, for researchers, like "Oh, ifwe keep making language models bigger, they're going to learn lots of stuff,stuff that's not just language." Andthat's how they thought maybe we'll getAGI. Uh, that slowed down after GPT-4. I, Iwrote an article about this for the NewYorker that came out last summer. Thatbasically, if you really report on this,what happened is they tried to make iteven bigger. This was at the time calledProject Orion at OpenAI. They said,"Great, let's make it five or 10x biggerthan GPT-4." And it didn't get better. And this is the thing you, if you runthis business, you want to believe.Yeah, that, like, we can make the digitalgod, if only, like, we can spend enoughmoney because, of course, spending moneyMeans that you're, you need your businessto grow, and you can get more investmentand you can build this thing that'sgoing to be so powerful that you kind ofget to be the god emperor of howeverything works, and, and you, you, youknow, Sam Alman doesn't need to get paida bunch of money because he's going tobecome one of the three most powerfulpeople in the world. >> Yes, and, huge.>> I think it's very funny when Sam talksabout how he doesn't make a lot of money,and I'm like, what is money to you?Like, you know that money isn't thething that doesn't matter in the future,you're trying to build, right? Like, I, Iget that. The fact that you'repretending like you don't get that makesme trust you less. >> But so, this happened though, and theyrealized it, but were being kind of quietabout it. They hit a wall in just plainscaling when that project Orion, uh, andthen Meta had a similar issue. Uh, GroI saw a similar issue like everyone wasseeing the same thing, where oh we can'tjust keep making things bigger. We're notgoing to be able to just pure scale ourway to AGI. And then we got a whole year. It started with the 01 models and thereasoning models, and we got all of theseother, much more complicated things. Remember, all this got so much morecomplicated, where what the models lookedlike—they were chemical equations fororganic chemistry class. They weretalking about reasoning and, and post, youknow, post inference computation. All thiswas all just trying to now findlittle narrow places where we could tuneto get some advantage or different typesof ways to get better performance out ofthem on benchmarks. They, it was actuallythe beginning of the end of pure perscaling. So, the rightlanguage models are sort of plateauing. We're able to tune them to do certainthings better, but we're no longer onthat explosive trajectory. You're right. To get something like AGI, we now needto look forward to. Oh, and then we haveanother innovation with a different typeof complimentary technology that goesreally fast for two years, and then maybethat happens one more time. I happen tothink the way AGI is going to happen, Icall it distributed or slow takeoff AGI. I don't think it's going to be here is asingle program that now, no matter whatwe ask it to do, it does this at thelevel of a human. I think what we'regoing to see instead is it's going to bemore boring. Lots of different systemstuned to lots of different things thathumans do, very specialized, that getreally good. And at some point, we're going to look around and say there is ahuge swath of things that are hard thathumans do that all now haveindividualized AI type systems that cando that thing just as well or betterthan humans. More of a peace mill. Ithink that's much more likely. And much more, uh, cost-effective, andMakes much more business sense than this idea of we just need a super AI that does all of these things. And we already see it now like, oh, we're better at chess. We're better at poker. We're better at go. We're getting better at proteins. That's a, you know, it's more peace. Each of these systems is going to need its own mix of technologies and its own secret sauce, but that's much more likely to get us to a point where we say, "Wait a second. Computers are doing a lot of what we do better than us." I think that's a much better more likely way we're going to get there than a singular machine doing it. And yet, it's not uh that that take isn't going to get any headlines written about it. Well, I mean, I like to tell the story of, um, there's these cool bots that came out. There's a particular researcher who's done a lot of cool AI. Uh, and the two things I think he's at OpenAI now, but back at Carnegie Mellon and Meta, he built the firstSystem to beat human poker players,professional poker players at Texascalled Pluribus. Um, and he built thefirst machine that could competitivelyplay Diplomacy, which, you know, is like arisk-like board game with lots ofinterhuman negotiation. But thestoryline of both of those, it's veryclever combination of existingtechnologies. So, for the poker playingbot, they actually tried at first tojust scale into a machine that work. They just made, they built really largeneural networks. They were training themat the supercomputer center at theUniversity of Pittsburgh, and theyweren't that good. And ultimately, thesystem that beat the professionalplayers could fit on a laptop. And itwas much more, they had a very smartcombination of some neural network, somefuture prediction, and they kind of tunedit up, and it was better at poker. Uh, thediplomacy plane bot was similar. There's alanguage model in there. There's aFuture simulator in there. There's likea strategy uh simulator in there. >> They like can they talk to each other ina language that we can understand?>> Well, they chat. So they play on the webdiplomacy server. So they chat withhumans,>> and then the language.>> Do they chat with each other?>> Uh, they could. The way the languagemodel works is it takes English and thenit converts it to a sort of XML stylelike standardized markup language thatthe rest of the programs understand.>> Gotcha.>> And they feed it on. But let me justtell you, this will maybe make you feelbetter. Here's what I really like aboutthat diplomacy playing bot.>> Okay.>> If you read the research paper aroundit, they just casually say at somepoint, oh, for ethical reasons, um, weset it up so it'll never lie. Eventhough that's like a major strategy inDiplomacy, you, you, you fake out people. You make false alliances or whatever.For ethical reasons, they said, "Oh, wejust set it up so it wouldn't lie.">> You're like, "Well, wait a second. Ithought the whole issue with AI is wecan't align it. It's out of control.">> But when you build these specialized AI,it's not just some giant neural network. You don't know how it works. It's six orseven components. We know what they do.And it was very taking concreteinformation from here and then they areputting it over here. And so what itwhat you're saying is if it knows thisinformation, it can't be like I'm goingto tell the people a different thing. >> Yeah. Well, there's literally a piece ofit that simulates possible future movesand to try to measure like which ofthese might be best. They wrote thatthat simulator is not a neural network.They just, they wrote it and they justcoded it to say, "Don't look atpossibilities where you lie."Doesn't. >> So, that machine can't lie. There's noout of control piece to this. There's noamorphous uh neural network. Yeah.>> Well, we you should say like it's it'sdon't it's not don't look at thingswhere you lie. It's like don't look atthings where x does not equal x. So likeit's a it's it's math that's happening.It's a it's computer code. It's not likewhen you say like don't look at thingswhen you lie. That that implies someunderstanding of what a lie is.>> Yeah. No, true enough. They have uh hereare the possible moves you can do andthey're all categorized and there's like50 moves you can do. Some of them arelying and they just literally the forloop that check tries all the thingsdoesn't look at those elements. Right.>> Right.>> Yeah. Um I call that intentional AI or IAI. Like the idea is when you move awayfrom all we have is a language model,which is what we're the we're trying toDo now, which I think is a real problem. Let a inscrutable language model be thebrain for everything. If you want tobuild an agent to go do something,ask a language model what to do and justexecute it blindly. But I, AI is like these game playingbots, which I think are great examplesof what AI might be. And you have multiplecontrol modules in these, whichare written by humans in normal computercode. There's nothing unknown about it. There's nothing that's learned. There'snothing that evolves. And it can talk toneural networks to understand things orto get descriptions of things or to comeup with ideas,but it's just a normal human controlprogram that can decide what do we wantto do. What are the options that we'reactually looking at? I, so I actuallyhave, I'm more optimistic about thisworld of distributed AGI. I don't thinkit's going to be out of control. I thinkthis alignment problem is really aProblem about language model textproduction and being able to controlwhat text language models produce, notnecessarily fundamental to AI-powereddevices more generally. >> Yeah. I think that we have thistendency, as people, one, to predictapocalypses. That's a known biasthat we have. We all know that thepersonal apocalypse is comingfor us, in the form of death. And so, wekind of like crave and are titillated bythe idea that that would happen to allof us at the same time in some way. Because of that, there's a lot of thisin science fiction, and all of thesepeople have written and read sciencefiction, just like I have. We said, uh,what if people fall in love with thisthing? And that's, so that's what we'regoing to imagine is going to be thething that happens. But nobody was like,I wonder if, uh, sycophancy-induced AIpsychosis is going to be a thing, becausenobody had any idea that that was goingTo be a thing. That came out of nowhere. And so, that's going to be the hardthing. Just like, you know, like ragebait, no one like assumed rage bait wasgoing to happen when Facebook inventedthe newsfeed, you know, like that. Thatwasn't a thing that was on anybody'sradar. And it's a, it's a third, fourthorder effect, but of course, you end upwith it, and then you have to deal withit as it comes. And so, it does feel likewhat we should probably be doing, and I'mhaving this sensation more and morelately. We should be dealing with theproblems that we have. >> Yes,>> We should be looking at the problems asthey currently exist. And what if I wantto say to some people who might want tobe effective altruists, what if dealingwith the problems that we have right nowis the thing that makes the futurebetter? Like, what if guessing at thewhat the future needs is actuallyactually makes the future worse andDealing with the problems we have right now is what makes the future better. Like, just what if, and if so, if that's a—Is, if that's like one of the things that might be the thing that makes the future better.It has the bonus of making the present better, for sure, because like we can actually measure that outcome as it occurs in the world that we currently inhabit. >> Oh, I'm so aligned with you on that. I'm very much aligned with you on that.I mean, it's really my whole philosophy of technology criticism.I call it technoslectionism.And the idea is, when you introduce, like, new technologies that are powerful into the techno-economic culturaluh, context,they always have these second- and third-order effects that we don't predict.And therefore, we can't approach technology with this weird fatalistic mindset of just like, well, once something is invented, it's just here.What are you going to do? Kids. These days, they're going to look attheir phone and just kind of moveon with things. We need to be much morehave much more agency. In we can humans,including and this surprises people.Saying I was interested in thistechnology, it was cool, and then it gotbad, so I changed my mind and stoppedusing it. Like, we are allowed to do that,and I think we need to do a lot more ofit. I mean, all of our problems rightnow seem to come out of these secondorder effect stories, right? Like, I’vedocumented a lot of these. You thinkabout, uh, Facebook's going along early. Facebook, you know, an engineer inventsthe like button for, like, a reallyengineer reason, which was, it'sinefficient for lots of people to writethe same comment. If you post, like, youwith your diploma, and everyone sayscongrats, that's like, that's goingto swamp out the interesting comment. So,like, oh, let's have a like button forThat so that the comments be more interesting,and then they realized, oh, that gave anunstable social approval indicator, whichmeant that usage time on the app went upby a huge factor because now my brainwas, did I get likes on what I justposted? And like, oh my god,engagement is all about. And we get theattention, engagement age, which was notat all what we thought about whenFacebook first came out. So this type ofthing happens all the time, and I'm withyou on this. What we need to be doingis diligently policing the second-ordernegative externalities of the AI we have. Right now. But think about the interestin some of the people who are pushingthe apocalyptic or the future stories,right? Like if you are Daario Amade, orSam Alman, right? It's completelyaligned with your interest to talk aboutthe automation of most jobs or to talkabout, you know, super intelligencecoming because it does three things. One,it distracts attention or regulatory oruh, public, you know, disapproval aboutthe things that your bad things thatyour technology is doing right now. Um,Two, it's going to induce a lot ofperhaps onerous regulation that you, asa very large company, can weather, butsmall startups can't. So, you getregulatory capture, and you get cementedin your position. And three, it attractsinvestment because yeah, it's scary tothink about AI automating half of jobs,but my God, if I'm an investor, I betterthen have my money in the company that'sgoing to do half of the jobs. That'sgoing to be the biggest thing in theeconomy. So, I think it's completely intheir interest to talk about the future. And I think the effect of altruists likeI really think what happened there, I'vebeen really trying to understand this,right? The that they've really beenpushing the super intelligence story.The standard thing, the existential riskThinking they've been doing, which islike a standard effect of altruist thingwhere you say, let's look at unlikelyrisk that would have a really negativeimpact because if you calculate theexpected value, it's worth spending timenow to prevent it, to prevent a potentialsmall probability future, right? So,yeah,>> this is like Nick Bostrom's Institute. We'll look at things like asteroid hits.Probably won't happen, but it'd bereally bad if it did. So, like, we have apositive expected value thinking aboutasteroid hits now. And, uh, superintelligence was just one of the manytopics they were looking at abstractly.I think when ChatGPT came along, itsort of warped the brain of some of thepeople in the F community that wereworking on, uh, that were working onexistential risk, and it warped it fromhere's this thought experiment that weshould be thinking about to, well, what ifit was actually happening and if it wasActually happening, we're the mostimportant people on Earth. We would bethe ones who found the asteroid firstand told people about it. And I thinkthat was such a compelling possibilitythat the switch from here's anabstraction that we should be thinkingabout as a thought experiment to this ishappening. It just that switch went andthey had a hard time getting out of it. And I, I absolutely think that theybelieve this, and I also think thatthere is a chance that it will go verybadly because, you know, social media hasgone very badly because contentrecommendation algorithms have gonevery badly, and like all they do isthey show you what you'remore likely to watch. That's the wholething, and that went very badly. Itwent bad for human attention. It wentbad for society. It went bad forpolitics. It went bad for, like, trust,which I wouldn't have expected. Like, weLive in a lower trust world because ofthe effects of content recommendationalgorithms. And I have friends whose jobis to build them. And they're sodisappointed that I feel that way. I think that it's true. I think thatthere are some that are better thanothers. I think that there are some thatare way worse than others. So I'm notsaying that they all have the responsibility here. I do think thatYouTube does a better job than a lotof platforms. Um, but also I think it'sstructured better than a lot ofplatforms because there's more humanagency built into YouTube, but theyhad no problem launching products thathave less agency involved in them, whichof course they have done. I dothink that there is a thing there. IAnd I don't think that we should payno attention to it. Like we have no ideahow to have an impact on the future, andwe certainly have no idea what thefuture's going to look like. And also,No one's ever been good at figuring outthe actual effect of a communicationistbreakthrough. Like, we're okay withweapons. We get we can kind of get that,but communication is just too importantand big and fundamental to humanity. We have a hard time going backand corraling those horses when they'reout of the gate. In the way, weapons isright. Um, I think like substances too,like we're much better at that. Like,oh, like cigarettes are invented, likethis is fun. Like, actually, this isprobably not good. Kids shouldn't usethese, let's ban them for kids, right?Or gambling was around and they're like,well, we should probably have rulesaround this because we're seeing, butyou're right, information is different. I think we need to be asking fundamentalquestions right now. For example,looking at chatbots and saying, whoasked for this? Who was clamoring forthis? What I mean, other than like theinvestors in OpenAI, like what whatProblem are you solving? And don't letthe programmers answer because theprogrammers have they're extrapolatingtheir experience to the rest of people. But it's not. There's a great use case ifyou're using cloud code.>> That's a great use case. You can'textrapolate that to the rest of the restof people. Think about our relationshipwith these things. It's unsettling. It'slike some small benefits with a lot ofdistraction and then some puremoments of like this is weird or, orunsettling. Yeah. Like these questionsshould be asked. I have a problem withwho asked for this because, like, no oneasked for anything, like, technology likewhat is possible is very different fromwhat I want. You know, like, I want acheesecake that is as healthy as acarrot. And yet, I will, I will never getone. So, like, I never asked for acheesecake, but like it was presented tome, and I'm like, yeah, cheesecake. I, Iwould prefer a world that did not haveSugary drinks. You know that's a better world. No one asked for Coca-Cola, butwe have Coca-Cola, and we live in a worldwith it, and we choose it sometimes, andwe should not choose it as often as wedo. And, so in the same way, I'mlike, okay, like you present us withthis tool, and a lot of people use it, anda lot of people find value in it. Thereare certainly valuable usecases for large language models,especially the way that they have beenconstructed to become more useful forresearch or for, you know, the dumb dumblittle things you gotta do. What Ithink is an interesting question, and itleads to a question I really want to getto, which is okay, no one asked for it to bebasically free when it costs a trilliondollars. That's actually weird. Like Ibuy Coca-Cola, Coca-Cola makes money. When OpenAI sells me ChatGPT, theylose money. So, like, this is kind of theweird, perverse incentive that we have right now, where they are... Shoveling a product at us, and they'retrying to make it more and more useful,and also, to be clear, all I think moreand more addictive, and they are doing itat a loss, and this perplexes me. Doesthat make sense? >> It does make sense. It doesn't perplexme as much. I think it's a standard. >> Right. I understand why. I understandwhy you're doing it, but I think it's how Ifeel about when you say, like, nobodyasked for this. I'm like, "No, but like,I also feel like something actuallymanipulative is going on here, which islike, get them hooked and figure out,where the addiction levers are." Imean, I think that's oneof the things they're doing. Yeah, Iagree with that. I think there's amore generally, a desperationplay of trying a bunch of things to see ifone of these can actually be the revenuegenerator. But I would also say, when Isay no one asked for this, the specific thing that's kind of beingShoved down our throats is the just ageneral purpose conversational partner. Like, actually, the obvious usefuluses of this technology has to do withlike solving particular problems.Wherever you get a natural languageinterface in the computation, like thatcan be useful, right? Like, I get it. Um,if I want summaries of text, like that'sreally useful. If I want to be able totell Microsoft Excel what I want it todo and not have to learn what all thosemenus mean, like that would be like areally, you know, that's really useful. It's like an obvious use case. But ifyou came to me ahead of time and say youcan like chat with a machine and it'llkind of like be a friend and kind of doall these different things. It's a weirduse case. It's not one like a lot of thepeople understood the use case for Coke. I like good tasting soda, but I don't. Ifeel like the chatbot interface is, it'sweird and dark and people are, but Ido think you're, I do think you're right. The they're searching for what's going to click. That makes sense. And thequestion I always had is like, well, theobvious things that would click would benatural language interfaces in thesoftware. Make everyone five timesbetter at using tools, right? Like, I'mnot great at Microsoft Excel. Why notgive me a natural language interfacewhere I could say, "Hey, go through, Idon't know what commands to do, but getrid of all the rows where the dollaramount in column B is above $5 and thensort it on this column."But what I've heard from my sources isthat would be the obvious way. It'd makeall your software more valuable, right? It's not working. Actually, thehallucination rate of LLMs means ifyou're talking to an LLM and then it'stranslating that into commands that'llbe executed by the software, it makestoo many mistakes. And so,this and using tools like that, like, likeuh, Air Table has a tool like this andI'm like, "Yeah, build me a little workflow." And then it's like, "Oh, nowI have to do a bunch of work fixing your weird mistakes." And I think it's areally interesting point is that alsoand not to go on a brief tangent, butwhen we're chatting with these chat bots, we underestimate the degree towhich we're constantly sort ofinterpreting and correcting and fixingand asking a question different way. Andso like a lot of these other obvious usecases of like great, we can now havethis be an intermediary between humansand the whole world of computation,which I think is a very excitingproposition. They're struggling withthat one. So now they have to look atmore, I think, addiction-based ones. This is why they like companions,because what's more engaging than a friend that you're going to talk to all day long? That seems good for them. Moreerotica in chat that tends to beAttention. Uh, Sora and Sora 2 made nosense to me. The economics don't makesense. Those videos are way tooexpensive. >> That's very dumb.>> Yeah, that's kind of. But I get they'rethey're trying to find ways because thebread and butter stuff that actuallywould be like, oh, it's the new mouse,>> which would be a big deal. I thinkthat's not quite working as well asthey'd hope. That's not working as wellas they'd hoped. There is potentiallysome kind of, like, as you say, with thethe systems, what was it? I, AI. >> Yeah.>> Where they are more specialized. If itturns out that all thevaluation is based on actually onlygiant companies can do this, so you haveto buy into one of the big three. Thatstill leads us to a world where thebubble pops. Yes. >> Yeah. This, the spending doesn't makesense. It just doesn't make sense. And the spending is largely being doneby a small number of companies right now.It just doesn't make sense.> It would be very unlikely, it seems tome, that this IAI thing happensfast enough to fill in the value gapthat has been created. > Well, I just don't think it requiresthe level of investment that Microsoftand uh OpenAI and Anthropic like thelevel of investment these companies aremaking. You don't need it for somethinglike that, the future of specializedmachines. And the Plurabis, which canbeat in a tournament with real money onthe line. Real Texas Hold'em players canrun on a laptop, and it uses some reallysmart neural networks and some otherreally smart things. Uh, yeah, so I thinkthe loveless spending now doesn't makesense. It's bubble-like. It's just, it'stoo much spending for the potentialrevenue that you can concentrate on theother side. It's just too expensive. The computing. Look at a company like Google,Right? Like their big rise in the,early 2000s that the whole key to theirrise was the cost effectiveness of theirinfrastructure. Like that was the wholeballgame. There was>> simple, dumb hardware, smart software ontop of it and give you very powerfulsystems without having to buy a bunch ofsupercomputers, and the cost, thecompute cost was really low, and thenthey could print money. We're doing theopposite right now. The compute cost isastronomically high. We're, we're that todo inference on these models right now. Um, is like the most expensivecomputational thing we can do. So itreally is, I don't know how>> inference is the thing where I, where Iask it a question, and it does something.>> Where it calculates every time someoneinterfaces with it, that's expensive, notjust the training part. I feel like we'llfigure out ways to make that cheaper.>> You have to make the model smaller. IMean, it's, it's simple what's happening. Models, they look complicated when yousee the layered structure, right? Butunderneath it all, what you're doing ismultiplying tables of numbers again andagain and again and again, right? And thenumber of numbers in those tables is relative to the size of the model, andall GPUs do because GPUs are made to do3D graphics for video games. And the wayyou describe 3D objects in video gamesis with tables of numbers. And the way you rotate a 3D object isyou multiply that table of numbers bylike a rotation vector, and it rotates alittle bit. So, GPUs were just machinesmade to multiply tables of numbers, orand tables by rows and numbers as fastand as specialized as possible. That turnedout to be fortuitous for the AIindustry, exactly what you need to do tohave an input go into a largelanguage model, and a token come out theother side. So, they're like perfectlywell suited for it. But that means togenerate even that single token, you haveTo multiply something by all of thoseparameters that define this really largeit could be hundreds of billions ofparameters, and you have to do all ofthose multiplications. Token. Yeah.>> Yeah. And there isn't a way to not not>> you got to just make the number ofthings you're multiplying smaller, whichmeans smaller models, which again comesback to my theory of like the only thingthat makes economic sense is smallermodels and bespoke systems that doreally cool things. >> Gotcha. So, so you don't think that uh mybubble worry is unfounded? >> No, it's I I don't know. I don't knowwhat the popping's going to look like.>> Right? Like maybe. >> Yeah. And like what I what I do know isthat there will be like lots of largerooms full of GPUs that will be used forsomething like that, that will be, theywill try and figure out economic utilityfor this stuff. >> I think that's true, and we mightfind more GPUs just distributed in everyday because the thing that makes themost cost-effective sense is trying toget things on device. Nvidia is tryingto do this in a sort of a clunky wayright now, but I think in the future, youknow it'll be. I have a lot of thingsthat have GPUs in them, and thatinference is happening here. Or I'mpaying for electricity because it's happening on my machine, andmy, at my house. That makes more sense. So, theremight be a huge market. So, like, I'm, I'mnot an anti-Nvidia person, but I thinkthe amount of money being spent by theFANG companies otherwise is, it, itseems unsustainable. >> Right?>> The best-case scenario, it'll just be largecuts, right? Meta is going to have tocut this many people. OpenAI might goaway. Anthropic is going to have to likego down to a third of its size. You, itmight be something like>> Which would be very, that would be a verylike a, it feels somewhatlike it would be a big problem for atleast the stock market. >> Yeah.>> Uh, which usually leads to a problem inthe economy, especially when most of themoney in the economy is spent by richpeople who are feeling flush right now,because>> portfolios have gone up. Yeah. >> Maybe this is why Mark Zuckerberg hasbought half of Hawaii. He needs a quietplace to go when people get upset. He'she's definitely seen that we,don't love him. I think he's noticed. II have two final questions. Um, number one,I saved it for the end. What did Ido you think I did worst on? >> Oh, that was a good video. You good?What do you worry? Okay, here's aneffect that I think your video createsthat a lot of discussions of thiscreate, which I think of as like doomamplification loop. So, because it's such a diversetechnology with so many possible thingsthat could be a problem, that as you gothrough them, one after another, afteranother, after another, we have a naturalhuman tendency to put that into anamplification loop and sort of imagine,>> the aggregate of all of that doom,because if you, if you isolate any one ofthose issues, you'll be like, ah, this is aproblem, like that's annoying, like weshould do something about it, but it'snot going to have me up at night. Butwhen you hear 20 different issues, itseems like it's 20 times as bad. But thething is, there's tons of technologiesout there creating their own issues ofroughly that size. We just don'taggregate them. We're just used tothere's>> a lot of different things out there, thatare causing trouble. So, that's the onething I worry about, is keeping ourmindset of there's lots of issues here,that are serious. But we don't have to add them up intotherefore, an issue that's so big thatwe'll never be able to handle it. We canhandle lots of small issues. Uh, it's notthe same as this is the biggest issuewe've ever faced, and we're not going tobe able to handle it. When we break thisup, when we recognize these areindividual issues, they're often dealingwith completely separate technologies. >> Mhm.>> It seems more tractable. So, I don'twant people to fall. I just get a lot ofthis when I talk to people right now, isyou fall into despair because it'scomplicated, and there's so much going onand so many possible things. But, likehumans also have a pretty good capacity to adjust to things, topush back on things, to eventually sayenough of this, like, we don't want to usethis anymore. I think that's finallystarting to happen for social media.It's taking five years longer than I>> We're getting fed up, you know? I think so. I, I, I have someoptimism, and I don't want to seem, if we list all thedooms out in a row, I don't want it toseem likethese are all adding up to one villainthat's going to be impossible to slay. Right? And if you listed out all theways that AI has helped you know,somebody do something, or even likethings that are relatively large, likecoming up with new cancer treatmentcombinations or protein folding on itsown, like, is massive and will bea, you know, world-changinginnovation. And it just, like, sort oftakes time for these things to filterthrough the, the technology loop. Andnobody wants to hear the video that'slike, "Hey, look at all the things thatare that like have come out of thistechnology." >> Well, and that's the, that's the otherpoint I want to make there, too, is we, weHave to be careful about connectingunrelated AI technologies as well andallowing advances and want to reflect onothers. So, like,>> Yeah,>> Cancer diagnosis, protein folding, theseare completely different technologiesthan language models, right? But>> I was just, I just also had aconversation with an oncologist today,and she's like, a while, I do need you toknow that I don't mean chatbots. >> Yeah. Because I think what happens thenis you get these extrapolations wherebecause, like, oh, chatbots got muchbetter; therefore, cancer will be a thingof the past. Like, no, that's a differentAI or, like, we're going to have, like,robots are just going to do all of ouryou know, manual labor. Well, that's acompletely different type of AI with itsown challenges, or, right?>> Self-driving cars. Like, that's, yes, butthat's a very slow; that's its owntechnology, and it's a very longTrajectory that we're on and it'sthere's a lot of, um, progresslaundering, I sometimes call it, wheresomething within this broad field makesa leap, and we launder that progress overto sort of other technologies in thefield and say, well, they also have, orwill be making the same type of leap, butthey're completely differenttechnologies. And so, like chatbots,getting better. There are also, like, there's, there's, likelegitimately, research that's been doneby kind of leveraging the gains of chatbots. Yeah, that's true too. Fair enough.Yeah. But with cancer research, what they didis they kind of leveraged an LLM butgave it, like, a different language, andthat language is the language of thecell. Yeah.And how, how cells communicate with eachother, and that's like, genius. I likeI love that. That's very cool, and I wantTo make a video about it. But>> It's very cool. And they train themcompletely differently, and it, but it'slike really cool. Yeah. I mean, and>> That excitement of one technology canalso lead to excitement in the relatedtechnologies, which I, which I, Yeah. which I think is good, right? Like, oh,we're reinvigorated in our AI, cancerresearch because there's energy in AI ingeneral. So,>> Do you feel some amount of like tech joyabout this? Like, holy crap, I'm kind ofalive for a really weird moment. >> I do. I mean, I, I, I especially did likethrough GPT-4 because that really was, asyou said, uh, a huge step-function leap.>> Because most people hadn't seen GPT-3.You sort of had to have permission toaccess the API.>> And that was like wow.>> This is like a really big leap. Iactually had some disappointment on theother side of it. I thought, uh, we weregoing to have natural language. Interfacing and computing, like we're inthat era now, which would have been likea fantastic era for a lot of reasons. Ithink productivity in the technicalsense would have jumped up if you makeeveryone experts at all software. Like Iwas really excited about that and waslike a little bit disappointed. I waslike, wait, why aren't we seeing thesetools being added to all this software?Why instead are we just seeing uhbenchmark charts? Like, good. I think that we will see it. And I thereason why I think that we will see itis because it's so good at code.Yeah. But like, it's weird and true that itturns out that, like, code is because it'slanguage adjacent. These systems arereally good at code, but they're notreally good at, like, clicking onsomething, which is weird.Y,You know, that's not what you wouldexpect. But it turns out that, like, we'reBet we're good at things we didn'trealize we were good at. And that's I just wrote an articleon this. This was a whole story where wewere tricked by coding agents being sogood, relatively speaking, that at thebeginning of 2025,the word was this will be the year ofthe agent. We will be assigning work, andit didn't happen because, again, like, asyou said, moving the mouse is hard, andcode is like really well suited. Um, wealso don't realize that those codingagents are working entirely in aterminal. So, it's entirely text-based.They're in a world of text-based commands,producing structured text. It's like, uh,Disneyland for language models. It'slike the best possible scenario, right?But you could imaginean Excel that exists inside of thatworld, like, you know, Excel iscode. It's just that it's like, wethought that we could skip over the partwhere we built Excel for us. >> Yeah.>> What we have to do is like we need tobuild Excel for them, which sounds likeI>> Yeah. And I'll tell you, here's my techjoy. I was talking to, um, someone inthis coding agent space, and they hadthis really cool idea, like, yeah, here'shere's what's going to—here's the way tothink about it: everything'sa program. So, if I'm asking Excelto do something, the approximateway to do that is to have alanguage model produce a bunch of, um,commands of the type we would giveExcel. But, like, actually what should happen is, Excel should have an API,opened up, a text-based API, and when I say,here's what I want to do, the languagemodel creates a computer program to dothat because, really, it's good at creatingcomputer programs, and that computerprogram talks to Excel, and, like, actuallycode is going to become the universallanguage of computing. We say things inEnglish, it gets turned into a computerprogram that then can access othercomputer programs through well-knowninterfaces. That that presses mycomputer science uh technical joybutton. That's so smart. Yeah. >> Yeah. Like if you have prolevel CHPT itcan already it just does that like it'sbuilt in like in certain situations itjust starts doing code and you're likewhat are you what's happening? Yeah,for a computer scientist, you're like,"Ah,>> that's my that's the world I imagine."Like, you just think in code. It makesso much sense. Yeah. >> Um I and I think so. Anyways, I'mexcited about that. I think it's okay.And that does actually have bigramifications, and it's going to change alot of things, and it could have economicimpacts. Uh, but that's a good, but Ithink ultimately that would be a goodthing. But it, yeah, it's different thanthe conversation we're having right now. It sound it also sounds to me like ifthis is the world that we end up in,it, I one of my big fears really is likewhat do people do, uh, when they aren'tthat good at stuff yet? So, like, whatdoes a college graduate do in the world?If AI is better than someone who'sbad at it but not as good as someonewho's good at it, how doesapprenticeship work in that world inknowledge work? You know, apprenticeshipwill still be a thing in plumbing. Andit sounds a little bit like actuallythere is a lot of space for human laborin this version of reality. >> Yeah. Because I think most things, atleast right now, think about that realityaren't replaced by AI. The industry thatis in trouble, I do think, is developing computer programming becauseI was talking to some developersrecently, and the thing they were tellingme, that's their words, not mine. Theysaid, "Here's what you have to understandabout the software development world."Most programmers are pretty bad, likejust in general, like uh, not. We, we thinkabout the really fancy 10x programmers,who are like the hot shots at, you know,meta or whatever, because I was, I wasmaking this argument of, like, I don'tlike this new model of, you have multipletabs, each with their own coding agent. All you're doing is supervising the tabs.I was like, this feels like the skilling. This feels like being a little bit wary.You don't want to lose what you do best,to create expert, elegant code. And someof the developers wrote back, like,actually, a lot of people are just sortof copying and pasting. There's a lot ofprogramming jobs out there that are just,people are producing bad code. >> Well, agents are good at producing>> bad code, right? Like, that is, thatis in danger.>> So, last question. We've already talkeda little bit about this, but when wecome to the, the, the problems that we dosee, and we're all going to place. Different levels of severity on themfor, you know, various reasons, but whenwe come to the problems that are will bepresented to us by any new technology,but this one in particular, do you feellike we have agency here? Like what, whatlevers are there? You know, identifyinga problem sucks, but feeling like youhave some kind of, can have like animpact on this. I think there's a fewdifferent levers we do have, right?Okay. So, first of all, I think there'sa sort of personal choice lever where Ithink it's completely fine to say it isnot my job to figure out why yourproduct is useful. It's your job toconvince me that it's useful. Don't tellme, like, you better know AI and then nottell me what that means. It's like if Idon't have a clear use case for usingsomething, I'm not going to go and spendmy time and money and attention tryingto master your tool. And if I find yourtool distasteful, um, I'm not going toUse it. I think parents and teachersshould have a rule that, uh, kids shouldnever be using a chatbot unsupervised. Ithink that's completely dangerous. There's, you know, we do not, I do notbuy the argument that, like, kids need toknow how to use the tools of the future. >> Like I think that there has to be, likesome kind of cultural language normstaboos around this. Yeah. Uh, and so weshould start developing those. Well, andthat, that could be parental supervisionearly on, is you're seeing examples. I'monly using it, but also it's kind of awaste of time if you're talking aboutlike a nine-year-old. >> You, you don't need to expose them, likethey need to know how to use chatbots,well, because the, the, the form factor inwhich we're interacting with GenerativeAI six years from now is not going to be,it's not going to be the same, right?It's not; they don't need to learn that.That's the Formula One car they’re never,going to drive. And so, I think it's completely it should be normative to be like, "Yeah,kids shouldn't be on this." Um, I thinkit's educational institutions should befree to say the whole point of this isto create friction in your mind so thatthe connections get stronger.You know, it's like this is a gym. Yes,it's true. A forklift can lift theweights for you, and that would beeasier, but the whole point of this gymis to lift the heavy weights. Like, sothat's not what we're doing here. We don't have to bend over backwards totry to like rechange our, you know, alleverything about learning to sort ofinclude this tool in a sort of forcedway. I also think there's economiclevers we can pull, which again is like,I'm not gonna waste time on somethinguntil there's, I'm like, we're done withthis sort of Apple-induced, you know,early 2000s era. Isn't it exciting to bea first mover on new technology and tryto figure out how it's going to be. Useful. We could just be done with that.I'm not going to spend money on thisuntil I have like a really compellinguse case. And I think there's regulatorylevers like I that the biggest lever wecould push like the Congress to pullwould be one where we have uh much moreliability for what a chatbot produces. Like no, you can't just have an entitythat'll say anything and do psychosis,you know, give out really dangerous, no, uhyou are responsible for what this thingsays and what happens. What the effect ofthat would be is like, oh, I guess weprobably can't have a product that'sjust trying to be a generalconversational partner, and that wouldprobably be better. >> right? So,>> yeah. What Sam Alma would say is like,well, then there won't be chatbots.Yeah. But what you don't, you don't, youthink that there would be a way toactually do the research and createfeatures that would allow these things. To,>> Yeah. I don't think I think so. I thinkif we said, look, uh, yeah, you'reresponsible, you know, X, Y, and Z, weagain, we go to the specialized future,which is like where we're going to endup anyway. Again, the oracular model oflike everything has to—why do I need touse the same tool for summarizing emails?It needs to also be a tool in which I canhave a sort of like pedophilic eroticarelationship with a virtual being. Like,does that have to be in the same tool? You know, if you’re right, and ofcourse I think you are because I’mtalking to you right now. Uh, but ifyou're right, we're going to look backat the last three years with likeremember when we were idiots? Rememberwhen we thought that it was going to bewe were going to do digital god?>> Well, yeah. I hope so. I hope that'swhat we're saying, like three years fromnow. It's like digital god made noeconomic sense. It was incrediblyDangerous and>> and also we were going to give it toevery single person for free. >> And we and we're going to give everyone for free, andwe still, we've been searching Kawaii forthe last three years and still haven'tfound Mark Zuckerberg. We know he'shiding there somewhere. We'll get himeventually. >> He's in his ocean lair.>> It's like Hunger Games. He's painted theleaves on his face. He's hard to see. >> He's finally found a fashion that worksfor him.>> You don't like the chain in theT-shirts? >> Look, it looks a lot betterthan he used to. I, you know, he's makingchoices, or someone is, um, in a way thatmakes me think that it is thatpeople, that someone has learned how tosay, "Actually, Mark, that haircut looksbad." Which is good. I think that's goodnews.>> Maybe a chatbot will tell him that. See,This could all come forward. >> Maybe that's what happened. He got earlyaccess. Cal Newport, you could watch hisand listen to his podcast, DeepQuestions. It's free on YouTube. It'svery good. I am a regular viewermyself. It's also free wherever podcastsare in other places. And also if you'dlike some of Cal Newport's work, a lotof it is focused on how we focus on howwe have agency in our world, how we takethat agency back, and how we do goodwork and how important that time is andthat focus is for doing good human work.Uh, so check out Cal. He's great.Cal, thank you so much for spending somuch time with me. Oh,>> It's my pleasure. I enjoyed it., This is Going to be Very Messy, YouTube, 01kjqadmfxrhxc2h7czpkxfk75.