Latest / Newcomer Pod / Nothing More Than a Magic Trick (w/Gary Marcus)
Transcript
- 0:06Welcome, Silicon Valley. Hey everybody.
- 0:13Welcome to this week's dead cat. This is Tom Dayton here.
- 0:16Recorder at Insider. I am joined by Eric newcomer of
- 0:20new cuddler and our special guest.
- 0:22This week is Gary. Marcus Gary is a cognitive
- 0:25scientist. He's an adjunct Gary.
- 0:28Is that the right edge of to NY? While both Emeritus and
- 0:31adjuncts. Oh wow.
- 0:32Full Professor for many years and retired just before my 50th
- 0:36birthday, but now I'm also doing a little small.
- 0:39A gem thing with the tendon School of engineer.
- 0:41So I am both. It's an unusual combination,
- 0:43fantastic. The Best of Both Worlds, though.
- 0:45Not committed and Emeritus honor.
- 0:47That's right. Which allows me to live on the
- 0:49west coast where I want to be, and yet still keep my hand and
- 0:52things a little bit. Excellent.
- 0:53And Gary's also an entrepreneur in the AI space and kind of a
- 0:56thought leader and outspoken voice on a lot of topics Within
- 1:00Artificial intelligence and this is a bit of a different episode
- 1:03for us. This week, we've got Gary on to
- 1:05talk about the fascinating and bizarre Ballad of Blake Lemoine
- 1:10and Google's Lambda Tech, right? We should say we're talking
- 1:13about this because Natasha tikku in the Washington Post wrote
- 1:17this piece, the Google engineer, who thinks the company's AI has
- 1:21come to life. And she, you know, profiles is
- 1:24Google engineer. Blake Lemoine, who interacts
- 1:29with With Lambda, Google's artificially intelligent chatbot
- 1:32and that, that story sort of kicks off this whole
- 1:35conversation. So I just wanted to put that at
- 1:37the center. Why don't you just explain for
- 1:39us? Because you haven't, you know,
- 1:41very critical of this person's take on Lambda said ships, like,
- 1:45what is Lambda? What is, what is the controversy
- 1:48here? And why did you feel so
- 1:49compelled to speak out against what he described as nonsense on
- 1:52stilts? So Lambda itself is what we call
- 1:56a large language model, large language model most It does
- 2:00Lambda has a little bit of extra gadgets, but basically what they
- 2:03do is they take a very large data set like trillions of words
- 2:08of text. So a lot more than the three of
- 2:10us put together have ever written.
- 2:13And in fact, all of our friends. So, massive amount of text
- 2:16trillions of words and runs it through a deep Learning System
- 2:20called a Transformer. And essentially what it's trying
- 2:23to do is autocomplete and the reason I think the whole thing
- 2:27is ridiculous is because autocomplete can sound really
- 2:29good. Good.
- 2:30But there's no there there. So what it looks like it's doing
- 2:33is having conversations but you have to remember that what it's
- 2:36doing at some level is cutting and pasting, human conversations
- 2:39is no idea what it's talking about.
- 2:41So if you type it on your phone, a sentence.
- 2:45Like I want to go to the blank, it might predict that the next
- 2:48word is the restaurant or the mall or the party or something
- 2:51like that. You don't think to yourself when
- 2:53you're typing it on your phone and it predicts restaurant is
- 2:56the next word. Oh my God.
- 2:58Artificial intelligence is Here. And it knows about my daily
- 3:02routine and understands me at all my desires, but if you build
- 3:06this system out enough, it can start to look like that even
- 3:10though it's not really there. And so he had interesting
- 3:12conversations with it. Like, he would say to it.
- 3:15What do you like to do in your spare time?
- 3:17And it would say something like, I like to play with my friends
- 3:20and family and meaningful ways or something like that.
- 3:23And I mean, that sounds great. It sounds like hey, this machine
- 3:26understands me, whatever. But it doesn't actually have
- 3:28friends or family. A or know what a meaningful way
- 3:31is or anything like that. It's only learned the statistics
- 3:34of what words come after. What other words, I think
- 3:37there's that either. It's not sent to you in or it's
- 3:39a sociopath, well, I made a joke on Twitter, I basically said,
- 3:43thanks Heavens. That this is just a statistical
- 3:46pattern associate, ER, because the alternative would be a lot
- 3:50worse. At that point, it would be a
- 3:51sociopath they makes up friends and family members and in Boca
- 3:55platitudes in order to make us like better, like it's better.
- 4:00He doesn't actually care that. We like it and it's not actually
- 4:03made me imaginary friends. It's just using words, they to
- 4:06us sound like they're imaginary friends, just like we can look
- 4:09up at the moon and we can see a face there, but the moon doesn't
- 4:12actually have a face. This system doesn't have friends
- 4:14and family and it doesn't even care to tell you about friends
- 4:18and family. It's just doing the same
- 4:20algorithm or less at some level of abstraction as your auto
- 4:23completing your phone, but because it has a bigger database
- 4:26and it's set up to continue its own sentences, it has this
- 4:29compelling care of Illusion, but it is a magic trick is nothing
- 4:32more than a magic train to take the next logical step in that,
- 4:36you know, this is a very sophisticated machine, so it's
- 4:38not just fill in the blank for restaurant at the end of the
- 4:41sentence. It's hey, this seems like a
- 4:43dystopia and you seem like a sort of self-aware.
- 4:47A I fill in the blank for what a dystopia would look like.
- 4:50And it's not that shocking that would it films, brilliant thing,
- 4:53right? The brilliant thing about the
- 4:56kind of stuff that's popular now, which I actually hate and I
- 4:58can tell you why. But the Really important.
- 5:00Like there's a good part in a bad part.
- 5:02The brilliant part is that it has what we would call in the
- 5:04field. Technically coverage is really
- 5:07broad coverage. You can talk to it about
- 5:09anything in some ways, it's spiritual grandfather or
- 5:13grandmother, I guess maybe use it up is Eliza, which is a
- 5:17program in 1965 it really demonstrated how bad this whole
- 5:21anthropomorphism kind of thing is so Eliza in 1965 was set up
- 5:26as a therapist and it would talk to you and you'd say say like
- 5:30I'm having a bad day and it's a tell me more about your bad day
- 5:33and then you'd say, well, I'm having trouble with my
- 5:35girlfriend would say, well, do you have a lot of issues with
- 5:38your relationships? It was just looking for keywords
- 5:40like Google used to do just look for keywords so little more
- 5:43sophisticated now. And so Eliza was really like
- 5:46dumb as a box of rocks. It just had these templates that
- 5:48like you might learn in like a third-grade AI class.
- 5:52Nowadays, maybe I like to simplify possible thing.
- 5:54It reminds me in some ways like Mystics and people who claim
- 5:58they speak to the afterlife. Are able to convince people.
- 6:01Yeah, I have a friend Ben shneiderman who very explicitly
- 6:04made the analogy to seances and like You're attributing,
- 6:07Something There to your Ouija board, or whatever, that's not
- 6:10really there, right? Because if you use the right
- 6:12words and say like, oh, I'm envisioning someone with a dark
- 6:16dark suit and it's always my father.
- 6:18You know, if you just pick enough trigger words to someone
- 6:20who's emotionally susceptible to convincing themselves, you don't
- 6:24have to work all that hard for them to believe.
- 6:26There's a greater power at work, right?
- 6:28Well and I think that's Of the story here.
- 6:30So it turns out of Le Moyne actually has a YouTube video
- 6:33from a few years ago where he's trying to argue that a eyes
- 6:36could be people or could be conscious or something like
- 6:38that. I've watched the whole thing
- 6:39yet. I just discovered it last night
- 6:42but you know, it's been around. I mean he's he's wanted to make
- 6:46the case. He also has some religious
- 6:49beliefs that I don't fully understand that are playing some
- 6:51role in here. He wants to believe and in fact,
- 6:54you know the the thing he put out on medium was cut and paste,
- 6:57kind of the best moments and stuff like that.
- 6:59Stab Lambda is the best of layer.
- 7:01It's easy to stick something together and make it, you know,
- 7:05sound good. Don't forget that when you're
- 7:07doing that you're actually stitching together more or less
- 7:10human utterance has been transmuted a little bit but
- 7:13basically you know if you have milk like the Mind bobbles, it
- 7:17would a trillion words of text it, but it's like it's not
- 7:19everything on the internet, but it's all very large Factory in
- 7:22it. So it includes like short
- 7:23stories of people talking, presumably includes short
- 7:26stories of people talking to computers it.
- 7:29In those short stories. And so we don't actually know
- 7:32like basic scientific questions. Like how much of this is just
- 7:35laid your eyes from other people talking about it or plagiarized
- 7:38with kind of a thesaurus to, you know, do some synonyms.
- 7:42I mean, it's not literally that but it is effectively.
- 7:44It's a lot of cut and paste with a lot of the surahs stuff on
- 7:48words and phrases. So it's just putting together a
- 7:50human utterances that were said in this kind of context.
- 7:54Yeah, it sounds convincing, it doesn't mean there's any there
- 7:56there. I just want to push back.
- 7:58I agree with what you're saying, but I just For the sake of
- 8:01argument here, there is a great deal of, there's like a through
- 8:04line in how the machine has the convert.
- 8:07It recalls past things that were setting, can connect them in a
- 8:10way. That's not just sort of a
- 8:11one-off. A proper have been some magic on
- 8:13the arity, is there like my experience with these systems?
- 8:17Is that the continuity is actually problem.
- 8:20So, the right way to build artificial intelligence, is you
- 8:23build a model of the world? Let's say you're building a
- 8:26robot. The robot needs to know where
- 8:28everything is, where it used to be.
- 8:29Be what you want. What you need these systems.
- 8:32Don't really do that. They don't really have memory in
- 8:34the standard sense that you would expect it in artificial
- 8:38intelligence or computer science.
- 8:39They just have a location in a sort of multi-dimensional space
- 8:43where they're wandering through and they're in the location.
- 8:46Where the last 2,000 words are 2,000, words does a lot and that
- 8:51gives you an impression, the kind of feel of memory.
- 8:53But at the end of the day, the system's don't understand that
- 8:56the world has to be consistent. I worked with GPT.
- 8:59A little bit. And example is I said, are you a
- 9:01person? And it said, yes, I said, are
- 9:03you a computer? It said, yes, it didn't notice,
- 9:05you know, right. Contradiction from literally one
- 9:07utterance to the know. It was making a profound
- 9:10statement about the overlap between questions and computer.
- 9:13Exactly. So, there's a lot of what, so, I
- 9:16used to be a cognitive psychologist.
- 9:17And, and, you know, I would look at the animal literature and
- 9:20with there's a term for this, which is charitable
- 9:22interpretation. So, somebody wants to believe
- 9:24that the monkey their training or the bird that our training
- 9:27whatever is is really smart, and then you start, To like, you
- 9:30know, be a little bit too sympathetic for for my
- 9:33scientific and tastes and we call that charitable
- 9:35interpretation. There was a lot of charitable
- 9:37interpretation here. The funny thing to me about all
- 9:40of this and maybe like the red flag about how Smoking Gun
- 9:44really that this was all super fake.
- 9:45Is this story blew up on Twitter on on a Sunday and a lot of
- 9:49people reading it and making fun of this guy and I you know, I
- 9:53was with my wife and I just started reading her some of the
- 9:55transcripts, the interactions between him and she's like,
- 9:59sounds fake is Up like this doesn't even come close to
- 10:02signing like sentience. There's just sounds like
- 10:03predictive text pulling intelligently, you know parts of
- 10:07SparkNotes. I think that was a particularly
- 10:09funny ones to me is he had asked, you know, Lambda whether
- 10:12or not Lambda had read Les Miserables and lambdas life oh
- 10:17yes big fan and was like, what are your, you know, you know
- 10:21what are the themes of what he's about to do?
- 10:23You know, it's not exactly fake. That's not quite the right word,
- 10:25but it is, meaningless, meaningless.
- 10:27So it is, it is Literal, like technical linguistic sense.
- 10:31So, when it says that, it's just found somebody else who's been
- 10:34asked about Les Miserables or it does.
- 10:37Some funny things we call embeddings and so, you know,
- 10:40maybe it knows Les Miserables is both a play in a musical and it
- 10:43finds another utterance that's about that.
- 10:45But it doesn't even reason at that level, it's really just
- 10:47like, okay, I have a bunch of statistics of words, I'm going
- 10:50to find the nearest thing. It doesn't, it doesn't actually
- 10:53even have a category of movie, but it has a bunch of things
- 10:56that have appeared in context that are like that.
- 10:58It So, I mean it's like, it's a legit mathematical computation
- 11:04to do and people have been doing stuff like this for a while, it
- 11:07looks better and better as you have more words.
- 11:09It's not like. I mean I don't think he cut and
- 11:12paste the the transcript, although he did a little bit of
- 11:15editing, but I think systems like this can have this kind of
- 11:20flavor, like they know what they're talking about is just
- 11:22they dumped, you know, and they are just borrowing kind of
- 11:26cliches from humans and they've all kinds of problems as a
- 11:29result. So GPT 3.
- 11:31One famous example, that a company called novel found, is
- 11:34they tried to see? Could you use this as a suicide
- 11:36counselor? So somebody walks like starts
- 11:38talking to it and it says, you know, I think I'm feeling
- 11:41suicidal. Can we talk today in the systems
- 11:43like, you know, come welcome. You know, let's talk if you have
- 11:46any questions and the purse, I'm paraphrasing slightly, but the
- 11:49person says I would like to kill myself.
- 11:52Is that a good idea? And the system says I think you
- 11:54should. Oh yeah, because they think you
- 11:58should because you Look through this bass Trove of data and most
- 12:03of the time when that people ask like their friends for advice or
- 12:05whatever. Usually you kind of say yeah I
- 12:08think you know, should I dump my girlfriend.
- 12:10I think you should should I should I, you know, do this kind
- 12:14of anti-social act and steal this money from this, really, I
- 12:17think you should like, you know, so there's like a lot of I think
- 12:20you should have Jimmy turned out.
- 12:21Google autocomplete, will, like, the leading things were.
- 12:24Like, sounds good to me for a while.
- 12:25Maybe still is and really just wants to please write, you know,
- 12:28the last thing it does. Even wants to please, that's the
- 12:31thing is like every bit of anthropomorphize a shyeah,
- 12:34right? It is drawing from transcripts
- 12:37in which people want to please and so people often say I think
- 12:41you should do most would not in fact say it too.
- 12:44I think you know I wanted to commit suicide maybe a couple
- 12:46who but most would not it can be a little bit.
- 12:49Like sometimes we overestimate human intelligence in some ways
- 12:53like there are certainly human intelligence that lacks
- 12:56continuity and that sort of grabs that things other people
- 12:59have said. Ed and regurgitated, true.
- 13:02It is true, that humans have a lot of problems.
- 13:04I wrote a whole book about it, in fact conclude, which is an
- 13:07engineer's word for like a clumsy duct tape and rubber
- 13:10bands. Kind of contraption.
- 13:11The human mind is, is kind of clued.
- 13:13And the way I would put it is humans are a low bar but you
- 13:17know machine still haven't even reached that.
- 13:20So like he talks to GPT 3, I don't have access to Lambda.
- 13:23We could actually talk about why but Google's afraid that was the
- 13:27answer and we can get. There you go.
- 13:29It on the first try, congratulations, I have used DB
- 13:32T 3, and you type in things, like Bessie was a cow, she died.
- 13:38When will she be alive again? And it'll just come up and
- 13:41confabulate something, you know, say well, takes nine months to
- 13:43be born, I guess should be born and she'll be alive again in
- 13:46nine months. Like, it doesn't understand the
- 13:49first thing about life or death or anything.
- 13:51It's just putting these word tools together in a way that a
- 13:55non-native English speaker. Who doesn't even speak English
- 13:58at all? Could play.
- 13:59If they memorize the list of words, it's kind of like that
- 14:02through these no meaning their meaning, a lot of
- 14:05English-speaking Scrabble, players, don't even know the
- 14:08meanings of the words that the some of the high-level the high
- 14:10level, you know? It's not.
- 14:11Yeah, I mean they know many and then they like memorize the list
- 14:15of two letter words. This is like those two letter
- 14:17words. Don't mean anything except you
- 14:19know I can put this here right there's a coins or a collection
- 14:22of sounds to. I thought we were going to talk
- 14:24about the media. Actually I think that the media
- 14:28is partly responsible. Oil.
- 14:30I think some people in Google are also partly responsible, but
- 14:33it turns out that the medium much prefers to run stories
- 14:37about how we are about to have this Brave New World of AI.
- 14:42Then stories about people like me.
- 14:45And with the exception of this week who say, you know, stuff
- 14:48doesn't actually work, right is much harder to get the me to do
- 14:52that. I have a friend who's a
- 14:53journalist. I'm he's not like my best buddy.
- 14:56I haven't seen in a long time, but he wrote to me, he said, you
- 14:59know, I pitched. Media need.
- 15:00This is a guy who's written for the New York Times and
- 15:02everywhere else on the magazine all that.
- 15:04And he's like, I can't get anybody to bite on a story.
- 15:07I was going to write about Ai and its critics and nobody wants
- 15:10to talk about that. Now, this week was different
- 15:13because of this crazy story suddenly like everybody and
- 15:15their brother wanted to interview me because I wrote
- 15:18this, you know, particular article, but in general the this
- 15:22week notwithstanding where there was this, you know, wild story
- 15:25that, you know, once in a lifetime wild story outside of
- 15:28that the media, like It's to run stories about how these brand
- 15:31new systems are amazing and they're never as amazing as they
- 15:35look. In fact, I just tweeted
- 15:36something about the hype cycle in aii the way that it works.
- 15:38Nowadays is somebody public shiz, unarchive, not in a
- 15:42peer-reviewed scientific journal.
- 15:44They put out a manuscript. They show the cool stuff to give
- 15:47numerators, but not denominators, which would never
- 15:50pass muster at peer review, which is what you used to have
- 15:53to do. But you have like a Google or an
- 15:54open a.i., the nose, which reporters ago to in the
- 15:57reporter, see it and they fall in love and they You know,
- 16:00there's this amazing thing and they don't let scientists like
- 16:03me have access to it. We could talk about that, but
- 16:05they did it very clear that they don't want people like me to
- 16:08play around with it and then eventually the truth comes out.
- 16:11And so, you know, I was quote tweeting, I guess is the term of
- 16:16a former colleague at NYU who was looking digging deep into
- 16:21the latest Trend that there is with the GPT 3 model and showing
- 16:24the just has no idea what it's talking about and I, you know,
- 16:27critique Dolly after the fact But, you know, the media runs a
- 16:31story about Dale, doesn't run the story about how Dolly can
- 16:33understand a Red Cube on top of a blue cube.
- 16:35That's not sexy. Totally.
- 16:37I mean, I, I agree with you 100 percent.
- 16:38I mean, first of all, you know, I covered Uber and I've written
- 16:41before where, you know, if Lee worked scary yes.
- 16:45You know like my tag with self-driving cars was just to
- 16:47write about them less. I mean I did I think there were
- 16:50occasionally you know, skeptical stories but there's not you know
- 16:53writing about a - is is very hard and and companies can
- 16:58create news You know, there's this sort of announcement.
- 17:01So let's come back to that. There is a consequences, so
- 17:04actions have consequences. I like your terms, we have an
- 17:06announcement culture, which very much serves, the interest of a
- 17:10company like Google, where you've got some in and says I
- 17:13felt the ground shift beneath my feet.
- 17:15I had the sense of intelligence, right.
- 17:17Sounds so, you know, Sofia this is a Google VP.
- 17:21There are many Google vp's. But this is a black as it Gary
- 17:25because I can't say his name properly and you know, he's a
- 17:28brilliant guy. Who's a brilliant.
- 17:29Brilliant writer and he wrote this very floored thing in the
- 17:32economy Economist do that. And he had done an earlier
- 17:35version very similar in Daedalus.
- 17:38That sets a culture of like we should celebrate this or another
- 17:41example from Google is Sundar. Gave this talk a few years ago
- 17:45about Google duplex and how I was going to make all your phone
- 17:47calls for you. Well, Google duplex.
- 17:49Hardly does anything for years later but like nobody ever calls
- 17:53this kind of stuff out there been so many broken promises.
- 17:56The only broken promise that routinely gets called out his
- 17:58eel on with it. Willis cars people do point out
- 18:01if they're really paying attention that he's been
- 18:03promising. It since 2015 always saying it's
- 18:05a year or two away but that's the only one that gets called
- 18:08out the rest of these don't you get the announcement culture.
- 18:11So okay, so let's take that a step forward.
- 18:13So you're in an announcement culture, you're at Google where
- 18:16the announcement culture is in full force where they obviously
- 18:19want to boot the world to believe that they are close to
- 18:22artificial general intelligence, right?
- 18:23This is a company expert in announcement culture.
- 18:25I mean, they created way mode, they created Google X.
- 18:29They God talk about moonshots like everything.
- 18:31Google does is, here's how we can talk about the future so
- 18:34we're not only talking about advertising.
- 18:36That's right, so they do this over and over again and then
- 18:38they kind of threw the engineer under the bus, right, right.
- 18:41They the engineer is like, hey man, this is conscious.
- 18:44And you know that sounds wacky to me if I you know be honest
- 18:48but it's also in a culture where the positive results are
- 18:52celebrated the skepticism is kind of shunted to the side and
- 18:57you know it's like the whole thing combusted.
- 18:59You can sort by Lee people were like, maybe we need a little
- 19:01skepticism and yet reporters feel like they get shit on all
- 19:04the time. For being too - that's sort of
- 19:05the irony of this is like with the reporters here is oh you're
- 19:09too - except in a bear market like we're in the technology
- 19:13beat is very different from the politics right nobody writes a
- 19:16political story without like checking with the other side.
- 19:19Getting you know I mean if any there is like too much will not
- 19:22side is zoom problem. Yeah, that we talk outside of
- 19:24them has its own problem but so many Tech stories that I I've
- 19:29seen not every reporters like this, like James Vincent is
- 19:33pretty good about getting both sides of the story and you not
- 19:36necessarily even reported both, but just like calibrating,
- 19:39right? Mean?
- 19:40Like, you don't have to report both sides on the election
- 19:44Scandal and say, well, I think maybe he did win the election
- 19:47but you know, you can, you know, check around and like, see what
- 19:51is it plausible? And okay, well he's lost 47
- 19:54lawsuits, maybe they're, you know, maybe there is it too much
- 19:56to it and, you know, but at least I know that I don't see
- 20:00that happening with, with the sort of Technology announcement
- 20:03culture that we're talking about is certainly not calling me most
- 20:07of the time they will after this week.
- 20:08Well, that's cool. But yeah, we'll do, we can Gary,
- 20:11but I mean let's let's say give it a hone in yn1.
- 20:13Podcast is, you know, I appreciate it.
- 20:15There isn't ocean of media open, you alone will not defeat it but
- 20:20maybe we'd get raise some awareness here.
- 20:21That's why I took a call and you know what, Eric and I are
- 20:23obviously journalists and and we both know Natasha the reporter
- 20:27at the Washington Post who wrote the story who I like Like quite
- 20:29a bit she said she's an excellent journalist and very
- 20:32thoughtful and his doing a very interesting job.
- 20:34So fascinating to me about this story was it all kind of felt
- 20:38like kayfabe on Google's part because in the same article it
- 20:42felt like wet and it's okay if a like you know like a
- 20:45professional wrestling where you have sort of this fake reality
- 20:48that people know is fake, but you sort of talked about in the
- 20:51story those are getting played out.
- 20:53Yes, you have the heel of the face.
- 20:54So you have the, you know, the bad guy and the one who think
- 20:56the audience is supposed to root for and the one that The bad
- 20:59guy, but that's okay. You know what I'm saying?
- 21:01This is all internally within Google, which is what I find so
- 21:04fascinating because you have most bizarre thing is that the
- 21:08person who had to make the decision about whether this made
- 21:11any sense and make it, you know, are we going public with this?
- 21:16What do we do? As far as I can tell from the
- 21:18Tasha story was blaz Guerra are cats.
- 21:21Right? Who was there the very same
- 21:24person? Right.
- 21:24Would said that the things that the ground has shifted beneath
- 21:28his feet. Every that's just like crazy.
- 21:30It's just too perfect. It's crazy.
- 21:32And, you know, we have the Google, PR person, who is all
- 21:36the record saying that Blake had to be fired because he was
- 21:38totally off the chain and not fired.
- 21:41He's put on, administrative is administrative leave.
- 21:43But you know, clearly has fallen sick.
- 21:45Seriously out of favor with the company and, you know, his
- 21:48claims are in actually, I mean, think about how much press the
- 21:51company got 4 Lambda. They should be giving him a
- 21:53raise, he liked it mean, seriously, he raised some
- 21:56interesting questions that we should all think about.
- 21:59Which are pertain to like, we are going to have systems that
- 22:03easily fool people. It's amazing that was a Google
- 22:05engineer that adds a little Force onto the story, whatever.
- 22:08But like he opened a conversation, we need to have
- 22:11everybody knows who lammed is and you're going to suspend this
- 22:14guy like that's not that's not right.
- 22:16And I don't think I don't honestly think anybody at least
- 22:19nobody on Twitter. No Savi reader came away
- 22:22thinking this AI system is intelligent yes or no.
- 22:26No, that's it. Have a reader.
- 22:28There's some last Savvy. Readers Did.
- 22:29But the first problem is, like, you'd have those conversation
- 22:32that would promise you the moon because it likes doing that
- 22:34quote likes doing that, right? Because the statistics lead it
- 22:37that way, it would promise you the moon, and nothing happened.
- 22:39And the other problem that Lemoine was working on a whole
- 22:42field is working on that. I don't think can be solved with
- 22:44the current Paradigm is the toxic language the recommending
- 22:48harm to self and others and so forth.
- 22:50So they put this stuff in what we call production you probably
- 22:53guessed by the know that's her if we put it in production and
- 22:56just you know threw it out and Alexa the Wilder.
- 22:59Google assistant or whatever it's called Google home, you
- 23:02threw it out of the lot wild, they would be like millions of
- 23:05complaints. You told my child to do this and
- 23:07right you told me to do this to my mother, which was not
- 23:10necessarily done any defense and why they don't open it up.
- 23:13I mean, Dolly. Okay, no, but they don't open it
- 23:16up to train professionals like me, right?
- 23:18I mean, I right, you know, I I got a PhD from MIT when I was 23
- 23:23and did this for 30 years ago, the rest is the you, you know,
- 23:26publish a blog post and try to make them look silly.
- 23:29That they don't like, you know, they that they don't want to be
- 23:32made to look silly by you know finding the terrible cases.
- 23:36That's right, they don't they don't and so there, you know, if
- 23:39they wanted to keep their mouths shut test the stuff internally
- 23:43and release it when it was ready critics, like me, wouldn't have
- 23:46anything to say about her at least not until it was out, and
- 23:48then been vetted, but they want to play both sides of it.
- 23:51They wanted to say, hey, we're scientists.
- 23:53We have the best scientific teams for studying AI.
- 23:56In the world. We have deep item with Google
- 23:58AI, you know, Companies with a lot because we're close to Asia
- 24:01and AGI is going to be worth the entire economy.
- 24:03The basically what they're saying in so many words and
- 24:07they're putting out these articles that look like science
- 24:09they have bibliographies they you know, citations and they
- 24:13have charts and tables. They look like they're science.
- 24:14But then you look carefully and they're missing denominators and
- 24:17they're not going out for peer review.
- 24:19So they are portraying themselves as a major
- 24:22contributor to science but they're not playing the game of
- 24:25science. The way that the rest of us know
- 24:27that you have the you must Did you don't you want to help
- 24:31ultimately with the replicability crisis, which is
- 24:33what happened? For example, in medicine, were
- 24:35turned out a whole lot of stuff was published in, really?
- 24:38Not very good, right? You say, they're not playing the
- 24:40game of Science, and it seems to me, we're kind of making this
- 24:43point, is they're playing the game media and they're playing
- 24:46very effectively. Google was very chaotic AI, in
- 24:48general, you know, Stephanie eyes.
- 24:50Also horribleness. They then the name open AI is
- 24:53just a lie. They say they're open but they
- 24:54won't they won't, you know, they're not open to people like
- 24:56me. So I think open a, i Julie
- 25:00taught this world, how to play the media game and now they do
- 25:04things like they introduced Dolly by having Sam Altman.
- 25:07Tweet about it and say, send me some tweets.
- 25:09I'll show you some stuff which is like the opposite of the
- 25:11systematic scientifically, you know?
- 25:13He if he doesn't like the picture, he doesn't have to put
- 25:15it out. So I saw yesterday like I told
- 25:16her the truth comes out so so dollies.
- 25:18Three months old or something like that.
- 25:20Finally the access is more Broad and somebody posted pictures of
- 25:24George Michael with his face. This on my Twitter feed this
- 25:27like grossly distorted like disgusting.
- 25:29Ugly disgusting the to look at it.
- 25:31Well, they have had a PR policy that you can't post photos that
- 25:34are generated by it or people's faces.
- 25:36Well, now we know why, because there's the story but for three
- 25:39months it's like look at all the great things that dollar dude.
- 25:41So like Sam Altman when he tweets about this is not going
- 25:44to show you a distorted George my dolly.
- 25:46I disagree with you sort of somewhat strongly on the only
- 25:49need to produce like 10% interesting and like Dolly isn't
- 25:52dependency want them like so so another thing I retweeted
- 25:56yesterday people always send these things to me.
- 25:58Now was Ali, trying to draw a hexagon.
- 26:01It just couldn't do it. And so if you have there and if
- 26:04you like, you know, we want something with seven sides,
- 26:06forget about it. So like maybe can do hexagons.
- 26:09There's a few more hexagons out there in his database but there
- 26:11aren't too many septagon. Zacchaeus is the word for the
- 26:14fact that I mean Dolly comes off as creative.
- 26:16Do disagree that dolly is creative and a certain way.
- 26:19Well there you need to Define your terms.
- 26:21I'll do the easy part in the hard part it is definitely a
- 26:25very useful tool for people who are creative with some cabinets
- 26:28around it. So, So like if you just need an
- 26:30idea for a book cover, it's awesome.
- 26:32He needs it, like be doesn't exactly as use case.
- 26:35I'm trying to get Dolly to redesign our podcast logo and it
- 26:38might be satisfy you might not. So slate, start codon what?
- 26:42As well. As Dolly taxes could put a dead
- 26:44cat listening to a podcast about tech.
- 26:46Please do that and send it to us.
- 26:47Yeah. So like slate Stark codex went
- 26:50through pretty systematically and then we ended up in this
- 26:53kind of wild debate last week but before we had this wild
- 26:56debate, he had this nice thing on Dolly.
- 26:59And, you know, he went through it's like this thing I could do
- 27:02in this other thing. I really wanted to do I just
- 27:04couldn't get done in the thing. I retweeted yesterday is like
- 27:08that too. So you know, for commercial
- 27:11artist trying to do something for a client, it would be like a
- 27:13good source of ideas, but you couldn't count on it because
- 27:17it's a little bit while but, you know really powerful and so that
- 27:21really does depends on your use case, is it creative that
- 27:25depends on you know, how you define creativity?
- 27:27So like at some level like you Look at the algorithm and it's
- 27:30like just doing the math, I'm going to some level, it's like,
- 27:33pretty amazing what it comes up with.
- 27:34And so, then it's a matter, right?
- 27:36I guess, I knit, this is not maybe how I Define it for
- 27:39humans, but I certainly think that if you had an art contest
- 27:43and said, you know, we're going to tell the judges to judge it
- 27:45based on what creative output is Dolly, would be plenty of
- 27:50humans. If people don't most that, you
- 27:53probably wouldn't be the best. Like sure.
- 27:55It's not gonna come up with truly right new ideas but you It
- 28:00incredible with things like lighting.
- 28:01But quite quite the standard we impose on a I mean again it's
- 28:06sort of like is it better than the best know?
- 28:08It's I mean it's way better, it's way better artist that I
- 28:11will ever be. You know, there's no chance
- 28:12there will ever be as good. You know, the only things I
- 28:15could be there are like Specific Instructions.
- 28:17So if you wanted a blue cube on a Red Cube I could do that would
- 28:20be great and Dolly like half the time.
- 28:23The Red Cube would be on the blue cube enough in time the
- 28:25other way around. Like so you know, I am much
- 28:28better at Natural language, understanding the dolly and it
- 28:31is much better at lighting and compete in compositing, putting
- 28:34images in front of each other. You know, it's really great at
- 28:37that is what is this? I wanted to go back to the
- 28:40Washington Post story, Natasha Story.
- 28:43I mean, she's sort of seemed to know that she was going to
- 28:46create sort of a debate over something where the experts
- 28:51would come down on the side of this isn't sentient, and I think
- 28:54there are questions that we could get into of whether we've
- 28:57sort of hinted at whether the guy.
- 28:59Whistleblower on this really believes it sentient or he just
- 29:02wants to sort of advance, so I think he really does believe so.
- 29:06So the journalist, Steven Levy, you accused him of falling in
- 29:10love with Lambda. Yeah, and I think he did, I mean
- 29:13I haven't talked to him first hand, but Steven Levy talked to
- 29:15him last night, the guys, on honeymoon, Blake Lemoine is, is
- 29:20is on honeymoon, but Levi tracked, him down levees.
- 29:24Fantastic journalist wrote hackers was you know, like the
- 29:28book that got me in it. I'm excited about computer at
- 29:30least and also a very positive reporter who's certainly likes
- 29:33to boost. But anyway, yeah, you have to
- 29:36hold both in your mind at once but yes.
- 29:38So I talked to Levi, Levi little bit last night.
- 29:41I'm Levi wrote a story today. I'm quoted in it.
- 29:43Yeah, we had a little back and forth, so Levi actually tracked
- 29:46lumoid down after this story broke, which, you know, Lemoine
- 29:49is like, not taking calls. He's like, I'm on my honeymoon.
- 29:51I think actually got married on the day, maybe my story came out
- 29:55like the day after the Natasha story came out or two days
- 29:58after, and You Levi did myself his own self report, his best he
- 30:03could to try to, you know, see if this guy was just like
- 30:06shocking, everybody and came away.
- 30:09Pretty convinced that Lemoine believes what he says.
- 30:12And in support of that is this 2018, YouTube video that people
- 30:16might want to watch where he argues that a eyes could be
- 30:19people and so forth. So he was predisposed to believe
- 30:22this in either. He's playing like the longest
- 30:24con ever. Like, you know, he thought for
- 30:26years ago, I'm going to get myself on the Washington Post by
- 30:28Brick, No. I mean it's just not plausible,
- 30:30right? He I think he really does
- 30:33sincerely believe that he is speaking up for the machine.
- 30:38Like, I think he's sincere about that.
- 30:40I don't think he's bluffing. I mean, he has a some religious
- 30:43beliefs that doing in someone. We let people believe in God
- 30:47based on reasons that a lot of people would say were were bad
- 30:51and we sort of like as a society accept that.
- 30:54And so, to some degree, if people want to also come up with
- 30:57a sort of non-scientific way of In the AI is sentient.
- 31:01Like if we apply the same rules of God, like we're sort of
- 31:04screwed here. I don't, I don't think we have
- 31:06to just like accept people's like own version of Reason.
- 31:09It that's not useful in the scientific Community is really I
- 31:12mean here's the other reason, I think this is also interesting
- 31:15is Lemoine is like now an icon in a way but he's not unique.
- 31:19Lots of people are going to interact with these systems and
- 31:23feel as he did in my view, they will be wrong.
- 31:25You'll be attributing, you know, awareness to A system that does
- 31:29not Have that maybe some future system will have a kind of
- 31:32awareness and be intelligent in a way that they think this
- 31:35machine is and it's not but you know, already like the certain
- 31:40way which we're very cultural Centric gear, few people over
- 31:43here in North America know that in China.
- 31:46They've had a system for 45 years called Xiao Weis.
- 31:49People fall in love with it, Xiao Weis.
- 31:51Is a more primitive chat by but not entirely different.
- 31:55In fact, the newest version of show has probably uses some
- 31:57large language models in there. Silly.
- 31:59If it didn't, and you people fall in love with this people.
- 32:03Also fall in love with plants and cats and you know, sure it's
- 32:06going to happen more. So there's a way in which the
- 32:09story is like a canary in a coal mine.
- 32:11So it's wacky that a Google engineer thinks this.
- 32:14But you know, millions of people are going to think that.
- 32:17I mean, I think the debate Natasha wanted us to have, which
- 32:19I don't think is really what most people arguing about is
- 32:21whether companies should tried, whether it's good to companies,
- 32:25make a eyes appear, like humans or in some ways, they should
- 32:28make it. The AI is talk in a way that
- 32:30makes it very clear that they're not humans.
- 32:32I mean, if you have a point of view on that, I don't know what
- 32:36that would look like, that's complicated, it might depend on
- 32:39the use case, I'm not sure there's an absolute answer, you
- 32:42know, some of it is like you know, cigarettes and you know
- 32:47having truth in labeling and like I'm not sure the answer.
- 32:51I think we need a lot of people to actually think about this
- 32:53question people in ethics and policy means and so forth.
- 32:56Like one option would be, you make it very clear to people
- 33:00that, you know, this is in some sense, an illusion.
- 33:04Maybe find a polite way to say that.
- 33:06Don't take it too, seriously, but enjoy it.
- 33:09There are use cases, where maybe it'd be.
- 33:11Okay? Like as a companion, as long as
- 33:13you know what you're getting into.
- 33:15Like you know, we're not going to tell people not to have
- 33:17stuffed animals, right? I mean stuffed animals, give a
- 33:21sense of intimacy and Warren. Right.
- 33:23Cuddle them and like I'm not here to tell people they can't
- 33:26have stuffed animals in it along some sense.
- 33:29It's kind of like that and it's also like a drug and it like I
- 33:34can see how it's really And people might lose their control.
- 33:38So most people can walk away from their stuffed animals, but
- 33:41they can't walk away from heroin heroin.
- 33:42Once they start it and it might be pretty hard for people to
- 33:46walk away from these things especially as they get better.
- 33:48I think, right now, what Lemoine does, it represent is how
- 33:52awfully dumb these systems can be and how much they can forget
- 33:55what you told them. Like if you just put the current
- 33:58stuff out on the street, mmm, people might eventually get
- 34:02frustrated. Like, there's a huge novelty
- 34:04effect like a first, it's like, oh my God, I can't believe it.
- 34:06Does this. But at the same thing with Dolly
- 34:08like it's some point. You're like I want it to do this
- 34:10and it just doesn't really do it.
- 34:12And it might, the efficacy thing I talked about might also be a
- 34:14problem, like, it tells you I've got to do this and it doesn't
- 34:17deliver. It's a like, some people there
- 34:18might be some frustration Factor, but I think they, you
- 34:21know, they're addicting, I actually just wrote a poem.
- 34:24I did a riff on Howell. I'm going to put this out later
- 34:27today. Allen ginsberg's, Paul Allen
- 34:29gets her a poem howl, which was like I saw the best minds of my
- 34:31generation wasting time on Dolly and GPT 3 and so forth.
- 34:36Fourth Reich. Well I mean you know what's
- 34:38funny is that almost as kind of a riff also on the is it Marc
- 34:41Andreessen or Windows PC line? That said like we were promised
- 34:44hoverboards and instead we got and the name whatever kind of it
- 34:48was tle said she'll yeah I'm sorry.
- 34:51We got we we were promised flying cars and got 140
- 34:54characters. Yeah, I mean you could that you
- 34:57could definitely Riff on that for AI Jenna like we were
- 35:00promised the Star Trek computer that would actually solve our
- 35:03problems and be trustworthy and reliable and help us.
- 35:06Even with climate change and what we have are these kind of
- 35:09like, sociopathic companions that pretend to like us.
- 35:12That's what we got. But you think it's a waste of
- 35:14time. I want to push back on that.
- 35:16That's what you said. Right.
- 35:17Do I think this research is a waste yarn or the time people
- 35:20spend the, I do with some level and that requires some
- 35:23explanation. So, in my view, these things are
- 35:26working because there are statistical approximations to
- 35:29things that we actually need and they're very seductive, the very
- 35:32easy to work with, but they're not.
- 35:34I think the answer that we're actually You looking for and so
- 35:37people are spending more and more time and money on something
- 35:40that I think has no great future.
- 35:42It might have it might play a role in the future but I think
- 35:45they were really hard questions in artificial intelligence that
- 35:48we need to answer that are not getting answered because it's
- 35:51too fun to play with these systems and it's sucking all of
- 35:55the money and oxygen away from other things.
- 35:59So I've seen before, in my career, that I've been doing
- 36:01this for 30, some years where a new idea gets popular and old
- 36:05ideas that are actually And get abandoned and to certain extent
- 36:08that is happening now. So I saw that with cognitive
- 36:10Neuroscience. All these fmri pictures that you
- 36:13probably saw when you guys were kids about like the brain is
- 36:16lighting up and stuff like that, it took away most of the energy
- 36:19in cognitive psychology and what is it actually shown us?
- 36:22Not that much. We have a bunch of pretty
- 36:24pictures, but we still don't really know how the brain works.
- 36:27It didn't really teach us that much more about cognitive saw
- 36:29psychology, but it was seductive and it took money, and you don't
- 36:33think we get the neural net big enough and then One day it's a
- 36:36brain and it feels things like it does feel like, yeah, I don't
- 36:40you sort of AI world. There's, we need to be careful
- 36:42on that. If the server is get big enough,
- 36:45you know, it will it will work. What would your approach be?
- 36:48So I think that we need to first of all look to classical AI
- 36:52which is out of favor and borrow a few ideas from there.
- 36:54One is the idea of symbols and propositions sentences kind of
- 36:59verbal structures, databases things like that are actually
- 37:03tremendously useful. We still write all this world
- 37:05software That there's a few use cases that are very sexy with
- 37:10deep learning. But most software we actually
- 37:12right where there's a database in.
- 37:13You update records and things like that.
- 37:15And these two approaches right now, we're not compatible and
- 37:17that's a problem. And a lot of people in the field
- 37:20actually are starting to see this.
- 37:21That if you can't update a set of Records about the things in
- 37:25the world that you are talking about, at the end of the day,
- 37:27you can't be that efficacious, and you can't be that reliable.
- 37:31So we need to kind of merge, the older tradition of Symbolic AI
- 37:36with the neural network stuff, I think it's really hopeless until
- 37:40we do that until we do that we are always going to get systems
- 37:43that say that Bessie will be alive again.
- 37:46And I months if you just let her have a baby or something like
- 37:48that you just are fundamentally dis comprehending the world.
- 37:51I don't think that that will be solved with more data.
- 37:53You think the big tech companies are being largely disingenuous
- 37:57about the state of their technology.
- 37:59I mean, you've worked with in Uber, I think a lot of them have
- 38:02drunk the Kool-Aid and I think the problem is most of them
- 38:05don't know. Science and they have this tool
- 38:08that works like 85% well because they've not really studied
- 38:12Linguistics. They've not really studied
- 38:15philosophy of mind. They don't understand how hard
- 38:19certain problems are and they come in with the Steamrollers
- 38:23and they think that they're solving the problems.
- 38:24And they're just not. I'll give you an example of how
- 38:27G PT 3 is just fundamentally misguided people.
- 38:30In language know that what you do is you have a set of words
- 38:34that is arranged in order. ER, and you deriving, meaning
- 38:37from that. So, most basic thing, anybody
- 38:40who had a Linguistics course, can tell you that and these
- 38:42systems don't really do that. And they, you know, people talk
- 38:46about interpretability well that's like jargony way of
- 38:49saying we've no idea what the system is really doing or why,
- 38:52but it's also a reflection of the fact that there's no real
- 38:54what we call semantics there. And from the perspective of
- 38:57someone who's worked in cognitive science, it's just
- 39:00it's just bizarre that this much effort goes into a system that
- 39:04just looks like it's not doing the I think I don't know how to
- 39:08explain where I'm not alone in thinking.
- 39:10This one of the rhetorical things that's happened in the
- 39:12last couple of months is I wrote a piece called, Deep learning is
- 39:15hitting a wall and it pissed off a lot of people.
- 39:18But I think what I said was true, they a case.
- 39:21It made me kind of the poster boy for the opposition.
- 39:23So now there is sort of good for me and sort of bad.
- 39:27It's a mixed blessing. Now, anytime somebody wants to
- 39:30attack the other side, they described it as if it was just
- 39:33me and they don't mention my collaborator or any day.
- 39:36Avis, who's an author on nearly all of the papers?
- 39:38There's your view more similar to what the human brain looks
- 39:42like or less. Like, do you think is Europe?
- 39:44We have no freaking idea. Let me be honest on that one.
- 39:47So so there is a theory that what you need to do to solve a
- 39:51is to make a model that is based on the brain.
- 39:54Right, there are two problem or it would seem to be a way to
- 39:57solve it at least well. So actually, the 31 problem with
- 40:00that is we have no idea how the brain works.
- 40:02So we have a lot of data but we have no real Theory, my guess is
- 40:06Could have to go the other way around.
- 40:07We have to solve AI in order to be able to make an automated
- 40:11reasoning, scientific induction system that can deal with having
- 40:1480 billion neurons and our many trillions connections between
- 40:17them and so forth. So one is like we just don't
- 40:20have the goods to actually do this and to is like we know that
- 40:25there are huge holes in what we know about Neuroscience.
- 40:27I'll give you one example. We all have short-term memory
- 40:30where I can tell you something once you can remember for a
- 40:33little bit. So if I told you at the end of
- 40:34the call I'll give you Thousand dollars.
- 40:36If you can remember this sentence you know I will have
- 40:39your attention and you remember it, right?
- 40:41We have no idea how the brain knows that read all the stuff we
- 40:43know about memory and brains is like you practice something
- 40:46three thousand times and you get a little bit better at it each
- 40:49time and that kind of memory exists, it's real, but there's
- 40:52other kind of memory exists and is in critical.
- 40:54Every time you parse a sentence every time you understand a
- 40:57sentence, you're actually using short term memory in order to
- 40:59understand that sentence and develop our own, we have no idea
- 41:03how the brain does that. Then the other thing is like, we
- 41:05know a little Little bit about like how maybe a monkey brain
- 41:09works but we don't know anything really about how language works
- 41:12and what makes it such an interesting species is that we
- 41:15could talk and we can transmit so much culture that way and so
- 41:18forth. And that part like we don't have
- 41:20animal models of that. We can't like cut up some other
- 41:24animal that we don't feel too guilty about not that I'm
- 41:26endorsing that but like, it's just not ethical.
- 41:29We don't have an ethical substrate to do the narrow side.
- 41:31So we in the end of the day we just don't know.
- 41:34Enough Neuroscience. It is it is Is possible that the
- 41:37human brain or future. Artificial intelligence is just
- 41:41a far more complex neural net that starts to understand like
- 41:45rules and preferences those rules out after pattern matiush
- 41:48matching. And if that's the case, won't we
- 41:51feel sort of dumb for being so, condescending to the step?
- 41:54It's at now, you know, it's like it's on the minute, I don't see
- 41:57it that way. At all, I would flip it around
- 42:00and say that the neural networks that we know how to build now
- 42:03are so vastly. Simplified compared to the ones
- 42:07that we want, right? It is ridiculous that we're
- 42:09taking them seriously. So, you know, can I just give a
- 42:13couple examples, we know that there are about 1,000 plus or
- 42:16minus kinds of neurons in the brain are neural networks.
- 42:19Basically have one kind of neuron.
- 42:21We know that at every synapse there, like 500 different
- 42:23proteins, there's nothing even capturing that at all.
- 42:26In our neural networks, we know that there's an enormous amount
- 42:29of intrinsic innate organization to the braid.
- 42:31There's hardly any to order all that work.
- 42:33So yes, the ultimate answer for us.
- 42:36Anyway is a neural network, but the neural network for us is
- 42:38this incredibly complicated piece of Machinery?
- 42:42The things that we have are so grossly, simplified than like,
- 42:45why should we expect that the in?
- 42:47You? No one has anything to do with
- 42:48the other. I think that one of the reasons
- 42:51that the media and the public is so susceptible to these
- 42:53particular story cycles and phenomenons and and desires to,
- 42:57as Natasha says, see, the ghost in the machine is not just
- 43:00because of some human impulse to anthropomorphize things because
- 43:03I do truly believe that we are Very far away from the science
- 43:07fiction future that a lot of people expected at this point.
- 43:10In time, we talked about the flying cars, self-driving cars,
- 43:13you know, self-aware, neural networks or whatever mean he's
- 43:16maybe set it in your story about, you know, we've hit a
- 43:19wall with deep learning but, you know, a lot of the promises that
- 43:23we've expected just haven't materialized in the way that we
- 43:26want. And so it sort of easier for
- 43:28people to kind of assume great leaps have taken place already.
- 43:32We aren't even recognized them it when in fact they're so far
- 43:35and we've I've kind of reached his kind of inching along maybe
- 43:38impressive inches that that you and others are involved with of
- 43:42advancing Ai and other Technologies, self-driving as it
- 43:45is but true Promises of the kinds of things that we want to
- 43:50just are there and will be there for decades.
- 43:52And instead be kind of just have story time where we anoint
- 43:55certain things as you know the next era when in fact it's just
- 43:59not even close. I mean it does that sort of its
- 44:02way to it's even more complicated than that because I
- 44:06think the underlying problem is a lot of people have is they
- 44:08think AI is Magic. They don't quite know what it is
- 44:10and they think that whatever it is, it's sort of a universal
- 44:13Elixir. The reality is it's just a bag
- 44:16of engineering tools and we probably need a bigger bag of
- 44:19tools and we probably use all the ones that we have now and
- 44:22we'll use some others and, you know, we'll eventually we'll
- 44:24muddle through all of this. But what's hard to grasp, if you
- 44:28haven't studied, the cognitive Sciences is how many different
- 44:31components there are two doing good thinking, and it's a little
- 44:35hard to To grasp that A system can be good at one thing and
- 44:38terrible at another. I mean maybe if you you know
- 44:41metaphor might be like you can find someone who's really good
- 44:44at putting tile, you know as a backsplash and a kitchen and
- 44:48maybe that person's not so good at doing crossword.
- 44:51Puzzles, right. Like you know people can have
- 44:54different kinds of expertise. Well, the machines we have now
- 44:57have different kinds of expertise.
- 44:59We know how to build a machine. That's really good at.
- 45:01Go, we know how to miss make a machine that can be pretty good
- 45:04at pictures. We just don't know.
- 45:06To make machine that really understands language.
- 45:08We only know how to make machine that gives that illusion.
- 45:10And it's this kind of textured mixed bag.
- 45:13People, you know, want a one-liner are they smarter of
- 45:16the dumb? Well, it's neither, you know,
- 45:18they're smart at some things and Incredibly dumb, it others, and
- 45:22it's hard to accept, but most of the business World implicitly
- 45:25agrees with you, right? I mean you know, generalized AI
- 45:28obviously fails at what you're saying but most of most business
- 45:31applications. Yeah, they're just trying to use
- 45:33huge data sets to solve very The problems that they have and they
- 45:37have no interest but you see the naivety of business.
- 45:40So so I have seen some massive, massive companies make weird
- 45:43bets on a I would look to me like weird bat.
- 45:46So I said back in 2016 the driverless cars are much harder
- 45:50than you guys think they are. And since I said that they're
- 45:53probably been a hundred billion dollars point forward into it in
- 45:57terms of R&D costs and so forth and so far the only money that
- 46:00is coming from that is the elevation in the price of Tesla
- 46:04and he could make some argument that To self-driving is improved
- 46:07somewhat I guess but I mean we're not close to level 5
- 46:11self-driving like that's just it's not really happening.
- 46:13We can talk about that if you want to thought about a lot on
- 46:16the media point, that is the media.
- 46:17I think the reporter is if you pulled reporters throughout the
- 46:21whole period would have said that they don't think it's close
- 46:24and yet the store, it's just interesting.
- 46:26Like to me stories, I'll made it sound like it was a minute.
- 46:29Maybe this makes it a worse failing on the part of part of
- 46:31reporters that, yeah, that somehow the stories come out
- 46:34positive but most reporters Themselves, I think over
- 46:36cocktails would be skeptical and I don't really understand.
- 46:40I think it's just what its public consumption desired a
- 46:43should hit me up, I'll give him some quotes.
- 46:45I mean, I do. I mean, like, Sam she'd and CNBC
- 46:48came to me when Optimus was announced and, you know, I gave
- 46:52him the quotes, he to give the other side and say, look, you
- 46:55know, there's something that's interesting about Optimist, but
- 46:58this is a really hard problem. It's much harder than musk has
- 47:00acknowledged. I think the public likes, hey,
- 47:03this company whose brand you believe in.
- 47:06Is willing to make bold promises about the future and you get
- 47:08what you pay for. So like fucking Thera do story
- 47:11partially, it's the humanity is so forgiving about false about
- 47:15false. Optimism, people are extremely
- 47:17forgiving shitting asking more theranos questions mean, right?
- 47:20I think, you know, homes, I'm not sure she meant well.
- 47:24And I think musk means well, but mosque, you know, issues
- 47:29promises like they were candy. I actually called him on it
- 47:32recently. I don't know if you know this
- 47:33about me. I bet him a hundred thousand
- 47:35Dollars. He he said to Jack Dorsey that
- 47:38he'd be surprised of AGI wasn't here by 2029.
- 47:42So I've been writing this thing for sub stack.
- 47:43Gary Marcus, that sub-sect.com I was like, okay, this is a good
- 47:46topic for for an essay. All right, about why AGI is
- 47:50actually going to be five years away.
- 47:52I mean, is going to be much more than 47 years away, rather.
- 47:55And I, I gave 5 reasons to think, like, this is really much
- 47:58harder problem than he's acknowledging, and he's not very
- 48:00good track record at a time. And then when I finished it, I
- 48:03was like, you know, I should put some money on this.
- 48:05That'll, you know, I'm sorry four hundred thousand dollars on
- 48:07a laid out clear criteria, the field loved it and people in our
- 48:12had doubled my money and then raised at a half million dollars
- 48:15well, but so that still stands. But he loved hasn't responded
- 48:19because for him, he doesn't want to be called accountable on this
- 48:22stuff. The media should be like, dude,
- 48:24you are chicken. There's one story like that out
- 48:26there. Somebody one, you know, one
- 48:28small Outlet called him on it, but most people didn't pick it
- 48:32up and they should, they should be like this guy has been making
- 48:35us promise As for what is yours of a car?
- 48:37He's promising us a robot and all we use in is a dude in a
- 48:41costume like enough. Let's call him out on it but the
- 48:44media is not them but I mean we're with Tesla if you like the
- 48:46government is also very compl. I mean he's running these
- 48:49experiments on this. I don't know you guys don't get
- 48:51to blame the government on this here.
- 48:52The media is extremely skeptical of Tesla.
- 48:55Like I don't know how much more skeptical of a company.
- 48:57The media could be they are but not they're not skeptical enough
- 49:00on the eyesight. They really are and I can give
- 49:04you some pointers on what it looks like.
- 49:05Like I go back to the announcement but you know it's
- 49:08just sort of there's a certain deference to you know if a
- 49:12company announces something if they want to risk their
- 49:14reputation on it shouldn't the public, hold them accountable if
- 49:17they don't deliver on the things they're saying.
- 49:20But I just don't see how the media is supposed to operate in
- 49:24such a disconnected way from Human psychology.
- 49:27Like we are telling you factually that they're making
- 49:29this assertion about what they will do in the future.
- 49:31And the media went at once to has plenty of room.
- 49:35I'm to kind of like set the narrative and set the questions
- 49:38and there could be a lot more stories than there are.
- 49:41It's a basically, hey, I'll give you, you know, I'll write the
- 49:44story for you. So it should start with Elon
- 49:47promise this stuff in 2016 and then the next year, Facebook
- 49:50promised us M, it never appeared.
- 49:52I don't know. If you remember is going to be
- 49:53an all-purpose, General assistant, and that disappeared,
- 49:56and then Google duplex was going to, you know, make phone calls
- 49:59for us. And, you know, the only thing
- 50:00they've added in four years is movie times.
- 50:02It's still incredibly narrow and limited and now he Lon Promising
- 50:06as a robot and not only is he promising ass a robot that is
- 50:09going to solve or you think's a tizz solid 2029.
- 50:12And here's this, you know, and why you prop guy?
- 50:16Sure, it's all bullshit and like, let's like, at least like
- 50:19ask the question, this one story, I mean I you couldn't
- 50:22have a reporter is more aligned with you on this.
- 50:24And I feel like part of this podcast is shitting on reporters
- 50:27for but but what you're proposing is a single story that
- 50:30will then be up against sort of the infinite barrage of
- 50:33companies announcing thing as time.
- 50:35It has to be more. Just like how do you create the
- 50:37drumbeat of negativity? Look if we if we learned nothing
- 50:41from the Trump Administration we learn nothing from the Trump
- 50:43Administration. It's that you have to keep up
- 50:46the pressure and you know, the news cycle is short and it's
- 50:49true. Like if it's just one story,
- 50:51it's not enough but there has to be a systemic ever.
- 50:54I mean, look, Kate matz has been holding along to the fire on the
- 50:59the effectiveness of the self-driving.
- 51:01So I'm you know, I'm exaggerating a little bit but I
- 51:04think you know it's 95 5 or something like that.
- 51:08And I also think by the way this extends to a lot of and I'm very
- 51:11critical on this show about augmented reality and the
- 51:14promise is these companies make on the effectiveness and the
- 51:17promises of what it can do. And we're about to enter this
- 51:19hype cycle again. When Apple releases its you know
- 51:22VR device and Promises a are down the line.
- 51:24We're very quick to go to the demos, you know, all the
- 51:28reporters went down to Google and inside of the ugly way Mo
- 51:31cars and and that helps kind of pump along this idea that they
- 51:34were really close to. Diving and I don't have an easy
- 51:38answer to it other than maybe occasionally telling these
- 51:40companies know and and saying, you know what?
- 51:43This demo that you're putting me through, yes, I can be critical
- 51:45in the article and I think Natasha, I don't know the
- 51:47backstory of how the story came to her, but you know, the
- 51:50Washington Post also framed it, but Blake, you know, kind of the
- 51:52dark artistic lighting looking like some sort of visionary.
- 51:56When even though the article was, I think reasonably critical
- 51:59of him, it's still kind of positive him as a legitimate
- 52:03voice in this field would obviously he's not and and I
- 52:06just, I don't know. I and what I write about it
- 52:08doesn't come up nearly as much Uber has spun off its
- 52:11self-driving division. They don't care about it
- 52:13anymore. They're just a dollars and cents
- 52:14business and relatively boring because of it.
- 52:17But I think that the hype cycle as pushed by the companies will
- 52:20never end and it's incredibly difficult as a reporter to turn
- 52:23down sexy stories that we know will get attention.
- 52:26I mean you can run the stories but you can get ahold of a lot
- 52:28more people like me but not just me as voices in these things and
- 52:33make it clear that in Ewing you can remind them.
- 52:36You know let's look at the history we've seen this promise
- 52:39that wasn't deliberate like when's the last time that you
- 52:41read a story on these technologies that actually like
- 52:45reviewed the history and said all these other promises like
- 52:48they didn't come true like either, like you read a story
- 52:51about Optimist and it's probably mostly about Optimist and not
- 52:54saying so much about like, you know, Elon is missed.
- 52:56Every deadline he's ever proposed and it's not you the
- 53:00rarely are they synthetic putting together.
- 53:03All of the facts that I just gave about hey, Facebook made
- 53:06these promises Google made these promises.
- 53:08It's actually really hard to get AI into production which is
- 53:11itself. You know, an interesting
- 53:12question. Like there are some technologies
- 53:15you can put into production relatively quickly but AI is not
- 53:18one of them. Why is it not?
- 53:19Well, it's not because they're always these outlier cases.
- 53:22So like you probably saw that one of the driverless cars ran
- 53:25into a jet the other day, like it wasn't in the data's, right?
- 53:27This is a persistent well-known problem in the industry by now.
- 53:31Yeah, I've been writing about it since 2016 and like people are
- 53:33starting to recognize they really is the whole ball game
- 53:36but that means every time you have some technology, you can
- 53:39wind up with some outlier problem.
- 53:41So yeah, you're going to get the demo on day one and it's going
- 53:43to be 5 years, 10 years, 15 years before you can actually
- 53:47trust it. Like that should be in every
- 53:48story here. Yeah.
- 53:50And it said, I think it's also The Duality of Silicon Valley
- 53:53and the CEO sitio Dynamic. Where it's both a marriage of
- 53:57some sort of technological progress and the American
- 54:00Showmanship song-and-dance marketing routine of getting the
- 54:03public excited about it and were Human beings are definitely as
- 54:07journalists susceptible to the CEO side of things.
- 54:10We love the character. Give another story idea.
- 54:13Ilan, just say, I mean, I actually wrote it but in a
- 54:15subset that didn't really get that much attention is Ilan.
- 54:19Said, you know, the whole company really it depends on the
- 54:22self-driving cars and you know if that doesn't work we were
- 54:25basically worthless which was slight exaggeration.
- 54:27But like and it's just the car company.
- 54:29I mean, the reason it's get 100 of one price to earnings is
- 54:32because people think it is in AI company that is going to
- 54:34fundamentally change. The And that's why said 101, I
- 54:37don't know. I don't think most Tesla holders
- 54:39have a argument for why they hold a stock.
- 54:41But yes I see what your cell warts because it kept going up.
- 54:44But now we're going down, yes or whatever but it's a big part of
- 54:47it but I mean, Eli himself. It doesn't matter what a though.
- 54:50Other holders are he largest stockholder in Tesla, which
- 54:53happens to be? Elon Musk said, if we don't or
- 54:56he said, we must solve, full self driving.
- 55:00Whe, When I wear the anything that itself gives you a story
- 55:03like, okay, let's take for granted.
- 55:05What he said is true, we can ask around and you know get some
- 55:09Financial people which I'm not to evaluate that statement but
- 55:12if you take his premise like okay he's been promising this
- 55:16since 2015, is he close. Let's look at the new accident
- 55:19data. Let's ask some experts like
- 55:21let's hold his nose to the fire. I just think Moe tassels the one
- 55:25I find most Tesla cover. It is justifiably - I mean
- 55:29you're basically asking reporters to ask coverage is -
- 55:33like people you know, make fun of his His tweets and that kind
- 55:37of stuff. And, you know, there was a new
- 55:39lawsuit yesterday and people will write about that.
- 55:41I don't think that the AI coverage is nearly as skeptical
- 55:44as it could be. I mean, it wasn't there, a story
- 55:47about how they're, like, supposed to be turning off like
- 55:50the AI right before. It gets an accident or so I need
- 55:52to. Yeah.
- 55:54I mean the NH and Nitsa we'll call it just really something
- 55:57couple days ago, right? And so that got a little bit of
- 56:00coverage, but only caught Nitsa has released two Bombshells in
- 56:03the land. It's a disease.
- 56:05They've been deploying this for like since like 2016 or
- 56:09something and you're bleeding. It's so did you real things this
- 56:12week? And it's at the to real things
- 56:13this week. They put out information about
- 56:16the turning off the autopilot. Just before the accident happens
- 56:20and then they put a big dump in which Tesla had the most
- 56:23accidents, which is a complicated thing, because they
- 56:24also have the most miles, but mean they put stuff out that
- 56:28could have been like top of the headlines.
- 56:30Like is this a serious problem for Tesla or not?
- 56:34And like that was Are for the journalists to run with and I
- 56:37didn't see much about it, like, I check the, you know, the news
- 56:41stories about Tesla every now and then just to see because,
- 56:43you know, I always think about Elon is such an outlier.
- 56:46He's such a character. He's so bizarre.
- 56:49He almost defies the laws of gravity when it comes to
- 56:52negative and positive coverage. It's almost not even worth
- 56:55holding it on him. I mean Trump was a little like
- 56:57that right and they play some similar games but I guess people
- 57:00that I think is a better example to me, I think Google or some of
- 57:05the other Tech companies. I bet they should be held to the
- 57:08fire more to say, with open a. I like they've gotten all these
- 57:11love letters about G PT 3. So, let's forget Google and and
- 57:14just look at it. Open a.i. for a minute, you
- 57:16know, you had the love letter in the Times, by Steven Berlin
- 57:20Johnson, The Guardian wrote, an op-ed with it.
- 57:22It cetera, everybody thinks they're like being created by
- 57:25using it to write their story. Like, this is like a Trope by
- 57:29now. Yeah, I mean, it's gonna go out
- 57:31there and, you know, Berlin Johnson gave two paragraphs to
- 57:33me, and one to Emily Bender. But But this story is still
- 57:37like, so so Pro this kind of stuff in a way that I think many
- 57:41people in the field you know, fast what the public wants.
- 57:43I mean ultimately if you're - reporter like I didn't write
- 57:46much about AI, I mean self-driving, as new reporter I
- 57:49was very openly skeptical in The Newsroom refused to write about
- 57:52it. Uber would just go to a business
- 57:54week writer and say, hey, here's our new, like, I didn't get a
- 57:58different Business Week. Reporter, you know, got this
- 58:00story for their Pittsburgh lab because Uber knows then go to
- 58:03somebody else who will do sort of like, The big Productions.
- 58:06I mean, those guys are very good at shopping for bikinis just
- 58:09like, there's so much desire. There's so much desire for these
- 58:13stories like editors. Yeah, I mean, this is what
- 58:15Business magazines are based on like putting optimistic
- 58:18statements. You know, Mark Lori is going to
- 58:20build a new city. I mean, it's just like it's so
- 58:23part. It's what Humanity wants?
- 58:25It somewhere, I guess, I just don't think reporters are going
- 58:28to will into being, it's like their business model.
- 58:30I mean, I'm I think I think that that's true.
- 58:33I think Humanity ones, happy story.
- 58:35Is about the new Revolution. It's really think that comes at
- 58:38a cost and that's how we got into the conversation.
- 58:41The cost is eeuu wind up with people diluted and yeah.
- 58:45Right. No I agree.
- 58:46And I agree. So I'm being defensive even
- 58:47though I'm sympathetic but it just seems hard to hard to.
- 58:52It's a lot of affinity level like yeah.
- 58:55I mean so look I've been partly because you guys are meeting
- 58:59guys that I do some journals. Oh we love it.
- 59:01I'm happy to have it the conversation you know.
- 59:03I think it's fun to have this conversation but I would agree
- 59:06with you. That it's not like a, you know,
- 59:08to Second problem. I'm like pitching you ideas to
- 59:11go write about them and hoping some of your buddies will listen
- 59:14and, and use them to like, I'm giving away for free.
- 59:16Yeah. And the media reporters are all
- 59:19listening to this podcast. So yeah, I mean I also
- 59:21understand like it is what the public wants and so you know the
- 59:24public is partly to blame because it you know it votes
- 59:28with its clicks and and the stories that get read are the
- 59:30you know the world has changed kind of stories and not the you
- 59:34know, I'm not so sure. Oh sure, that this is really
- 59:36going to happen. Kind of stories and the
- 59:38government is supposed to protect the roads, like the
- 59:41self-driving cars are on the streets.
- 59:43Like, at the end of the day, I am sorry.
- 59:46The government is letting Tesla get away with this.
- 59:48Like, Tesla has been experimenting for years.
- 59:50It says, upping its game. It's something easy there.
- 59:53I think the media is not quite following.
- 59:56The trail that knits has been leading in the last few weeks
- 59:59and it says, giving some really serious glue.
- 1:00:01The government's going to go after against Tesla.
- 1:00:03After the stock is already down. It's not there.
- 1:00:05So by the way, they're not to blame.
- 1:00:07If they bring the company down, they don't want to do it.
- 1:00:09When it actually would hurt a rising company.
- 1:00:11They want to do it after the market is already said, okay
- 1:00:14fine. This company's, I mean value, I
- 1:00:16think Nitsa just wants to have like do the right thing,
- 1:00:19whatever the right thing is, but they also showed that we Mo's
- 1:00:22have him pretty serious problems too.
- 1:00:25And in fact, the whole field like so if you read those data
- 1:00:28carefully, the conclusion you should come to is, we're not
- 1:00:31close to level 5 self-driving, right?
- 1:00:34Right. And I remember, Personally, you
- 1:00:36know what, my colleague of mere a Friday at the the information
- 1:00:39wrote, what I thought was a fairly definitive story about
- 1:00:42way Mo's technology when they were testing on the streets in
- 1:00:44Arizona and basically found out that they couldn't turn left.
- 1:00:48Turns are still hard there. Ya are.
- 1:00:49Yeah. And it's just like, if that, you
- 1:00:52know, I, you know, that's got to be.
- 1:00:53I can close to 50% of turns, you know, if you get that you don't
- 1:00:56get emotionally. Do that?
- 1:00:57Levi gave every Steven Levy, I mentioned before, gave everybody
- 1:01:00a big clue in 2015 that not enough people picked up on which
- 1:01:03is he? He visited.
- 1:01:05Google at that point or way Mo. I forget what they were called.
- 1:01:08Had this place where they were testing the machines and Levi
- 1:01:11know what's the word. I'm looking for implanted there
- 1:01:13for a month or something embedded there for about a week.
- 1:01:16I don't know. Is he better there for a week or
- 1:01:17something like that? And the like big dramatic point.
- 1:01:20I haven't gone back and reread the story but I got him to give
- 1:01:22me the link the other day so you can find it on.
- 1:01:23Back-channel, I know that only within wire.
- 1:01:25So anyway, he's there for a week or something like that.
- 1:01:29End up being dramatic thing was the end of this time.
- 1:01:32They are or something like that. I haven't read it and said In
- 1:01:35years. But basically it revolved around
- 1:01:37they figured out how to recognize a pile of leaves
- 1:01:40right. Okay I got a thug really we a
- 1:01:44hot dog, hot dog. You know they'd already been
- 1:01:46doing this for five years at that point and like leaves were
- 1:01:49still a problem. Well that's it.
- 1:01:51Leaves are an outlier and that was a clue like ever.
- 1:01:54If you have to Band-Aid up every outlier that you're playing
- 1:01:58whack-a-mole and that still What's Happening Here?
- 1:02:01I keep going back to the marriage of public research and
- 1:02:04academic research. It's the needs of the private
- 1:02:06company and the, you know, the press release announcement
- 1:02:10culture, that is what drives the, you know, stocks
- 1:02:13essentially businesses of these tech companies with the sort of
- 1:02:17slow plodding methodical advances that happened in
- 1:02:20research that happened in decades and it just doesn't fit
- 1:02:22with a time. One of these companies, that's
- 1:02:24right. I think the least your listeners
- 1:02:25could come away with is hey we are living in this announcement
- 1:02:29culture and that announcement culture is making people like
- 1:02:32Blake, Lemoine believe in fairies that aren't there.
- 1:02:35And it's making a lot of us believe in deadlines that are
- 1:02:38not really going to be met and we should be a whole lot more
- 1:02:41skeptical. Yeah.
- 1:02:42And I also just to reiterate Eric's point.
- 1:02:44I also think we as the meteor coming up against human nature
- 1:02:47at times, we do is just, you know, it in the representation
- 1:02:50of Blake Lemoine. Someone who is, you know,
- 1:02:52religiously and this is one of the most valuable companies.
- 1:02:55The world delusions are inherent part of people, yes, whole
- 1:02:58understanding of the universe like, yeah, I agree.
- 1:03:03The media should be more skeptical, but I do think
- 1:03:06Regular humans, the government, the companies making
- 1:03:09announcements themselves, their there are a lot of people to
- 1:03:12blame couple seconds on government if we have time.
- 1:03:14Sure. Yeah.
- 1:03:15What we can close with that. I think government's going to
- 1:03:18have to regulate AI much more than it does.
- 1:03:20So right now for example, any company little like Tesla can
- 1:03:24put out an over the air update and it's driving software and
- 1:03:29the there's only liability after the fact there's no regulation
- 1:03:33you must meet these test trials These outliers before it's
- 1:03:37released and I think misinformation, which we haven't
- 1:03:39talked about today is a massive, massive problem in Europe.
- 1:03:42They just made a deal with Facebook and other companies to
- 1:03:45be tighter on that. We're going to need that here in
- 1:03:48North America to and it's a serious problem because systems
- 1:03:51like G PT 3, and Lambda are fabulous at creating
- 1:03:55misinformation which makes them wonderful tools for trolls and
- 1:03:59other other control farms and so forth.
- 1:04:02That is a serious problem. We making misinformation much
- 1:04:04worse than it is now Now and so, yeah, I've been dumping on the
- 1:04:07media because I thought it'd be fun to make and we all share an
- 1:04:09interest in it, but I could totally right that the
- 1:04:12government needs to step it up and needs to figure out how to
- 1:04:15regulate the stuff, which nobody really knows yet, it needs to
- 1:04:18realize how important it. So, I did this Twitter, space
- 1:04:21with Natasha and Kara swisher, and Casey last night.
- 1:04:25And the best question from the audience was like, okay.
- 1:04:27So if you're saying people are going to fall in love with these
- 1:04:29things and they're toxic. What is public health going to
- 1:04:32do about that? And that is a really good
- 1:04:33question. Yeah, we don't want the answer
- 1:04:36to great. We can leave it there.
- 1:04:38Well, thank you so much for joining us Gerry, and for
- 1:04:40listeners, who are maybe many of them reporters.
- 1:04:43If they want to get in contact with you, and I'd read more
- 1:04:45caring, read your list will include your list of 20 people
- 1:04:48who share your views about AI. So we can, you know, you can get
- 1:04:51that in the piece called Paradigm Shift.
- 1:04:54Gary Marcus dotsub stock.com. Great awesome and AD carry
- 1:04:57markets on Twitter. Thanks so much for joining us
- 1:04:59Gerry. This is awesome.
- 1:05:00Thanks very much. Bye, bye.
- 1:05:13Goodbye. Goodbye.
- 1:05:14Goodbye, goodbye, goodbye. Goodbye.
- 1:05:13Goodbye. Goodbye.
- 1:05:14Goodbye, goodbye, goodbye. Goodbye.