Latest / The Smart Cookies podcast / Is AI mostly hype? With dr. Emmanuel Maggiori

Transcript

Is AI mostly hype? With dr. Emmanuel Maggiori · The Smart Cookies podcast

• 0:00
  Dr. Emanuel Maggiori is a software engineer, writer and speaker who's on a mission to
• 0:08
  help develop a more balanced view about the tech and AI industries in the midst of all
• 0:15
  this relentless hype.
• 0:17
  He's recently been called the AI hype buster, which honestly is a pretty cool nickname,
• 0:23
  not to mention a role that seems to be critically lacking in the public sphere.
• 0:28
  He's also the author of the books, Smart Until It's Dumb, I love that title, and the latest
• 0:33
  one which is called Silly Cond.
• 0:36
  Emanuel, welcome to the show.
• 0:37
  Thank you for having me.
• 0:38
  Yeah, I'm glad to hear someone's calling me that.
• 0:41
  I didn't know.
• 0:42
  Yeah, you're doing a debate recently and that's how they introduced you.
• 0:46
  Okay, yeah, I remember now.
• 0:48
  I was like, yeah, I'm totally still.
• 0:50
  Yeah, okay, cool.
• 0:51
  Let's start with the study that's been making the rounds in the media recently.
• 0:55
  They did a survey asking people whether they think the current large language models are
• 1:00
  conscious and this is what they came up with.
• 1:04
  So I'm going to read the excerpt now.
• 1:07
  Two thirds of people surveyed in the study think that artificial intelligence tools like
• 1:12
  chat GPT have some degree of consciousness and can have subjective experiences such as
• 1:18
  feelings and memories according to a new study from the University of Waterloo.
• 1:24
  So the majority of people out there think that artificial consciousness is at least
• 1:29
  to a degree already here.
• 1:32
  What do you think about that?
• 1:33
  That's a very, I don't know, it's a very complicated question, right?
• 1:37
  It opens a can of worms in terms of a lot of people have been speaking about the topic
• 1:42
  of consciousness for decades now in the philosophy of the mind kind of sphere.
• 1:49
  And I don't think anybody knows the answer of what consciousness is for starters, what
• 1:55
  is consciousness, right?
• 1:57
  What is it and when does it happen?
• 2:00
  We don't even know why general anesthesia works, right?
• 2:06
  So we know it does work.
• 2:10
  It's a correlation.
• 2:11
  We see we put these substances in your blood and we make you inhale this gas and then your
• 2:18
  consciousness stops, right?
• 2:20
  And then you're awake after your surgery a while later.
• 2:24
  But nobody really knows the mechanism, which is fascinating, right?
• 2:28
  So we use it, thank God, right?
• 2:30
  But we don't really understand it.
• 2:32
  And part of the reason is that we don't really know what consciousness is, where it sits,
• 2:38
  and there's a whole spectrum of opinions.
• 2:41
  For some people, you know, it's called panpsychism.
• 2:44
  So it's this idea that consciousness arises from computation, from the act of calculating
• 2:50
  things if you want.
• 2:52
  But then you could think of a bucket of full of water, you know, the molecules are bouncing
• 2:57
  around and we can imagine that they perform a calculation as they bounce around, right?
• 3:03
  So this bucket of water would be conscious, right?
• 3:08
  Not the same consciousness maybe that we imagine ourselves, the one we have, but it would be
• 3:14
  a degree of consciousness, that's called panpsychism.
• 3:18
  Other people at the other end of the spectrum, they think, you know, consciousness happens
• 3:21
  on biological brains, you know, made of biological cells only.
• 3:28
  And you know, then they fight against each other, you know, the ones that criticize panpsychism,
• 3:33
  they tell them, oh, that's ridiculous.
• 3:35
  You're telling me that water can be or paint on the wall can be conscious.
• 3:40
  Or thermostat, you know, it's a machine.
• 3:43
  It performs calculations as conscious.
• 3:45
  And then the other way around, those that believe it's a biological thing, they're called
• 3:51
  meat chauvinists or, you know.
• 3:57
  So I don't think the debate has settled either way, right?
• 4:02
  What we're witnessing now is this very interesting phenomenon of having a machine that produces
• 4:07
  very, very convincing text, right?
• 4:11
  So do these calculations then, do they qualify for, you know, consciousness that would get
• 4:18
  close to our consciousness for those who belong to that camp, you know, those who believe
• 4:23
  that consciousness is computation?
• 4:25
  But I have to say that there are a lot of people who don't believe that, right?
• 4:29
  Some people think that there's even something non-computational that happens inside our
• 4:33
  brains.
• 4:34
  And, you know, someone who's defended this idea for years was Roger Penrose.
• 4:39
  I mean, it is Roger Penrose, I think, is still defending his point of view.
• 4:43
  So he's this physicist who's been saying for years that there's something in our brains,
• 4:49
  a physical action that hasn't been discovered yet, something related to quantum mechanics,
• 4:56
  that is not something you can simulate with a computer.
• 4:58
  So then, you know, consciousness is not computational.
• 5:02
  So you know, he wouldn't probably agree with this idea that chat GBT, for instance, is
• 5:08
  a conscious thing.
• 5:10
  But the thing is, like, I don't know, what can I tell you?
• 5:12
  I can tell you there are different opinions.
• 5:15
  I do tend to get this intuition that Roger Penrose is up to something, even though he's
• 5:22
  been criticized quite a lot, you know, that they found flaws in his, you know, his reasoning
• 5:28
  and stuff.
• 5:29
  I do tend to get this impression that our consciousness and even our intelligence, they're
• 5:35
  kind of related, and they're something special that we haven't yet cracked, you know, and
• 5:41
  these hallucinations that happen in AI when it gets it totally wrong, and it kind of tells
• 5:46
  you, oh, okay, this is just repeating stuff.
• 5:49
  It's not really smart.
• 5:50
  It's a bit of an illusion.
• 5:51
  And they kind of maybe reveal that there's something missing.
• 5:54
  And maybe that's why it's not really conscious.
• 5:58
  But I don't know, I think that you can get every possible answer about this, because
• 6:02
  you can define consciousness any way you want.
• 6:04
  You can say a rock is conscious because there's some computation happening inside it, you
• 6:09
  know, or a bucket of water.
• 6:10
  So in that sense, yeah, definitely, chat GPT will be conscious.
• 6:14
  Yeah, I hosted a pretty renowned consciousness researcher called Anil Seth on the podcast
• 6:21
  a while back.
• 6:22
  He's also from the UK.
• 6:24
  Yeah.
• 6:25
  And he said that we often make this mistake that we kind of equate intelligence with consciousness
• 6:31
  that and then apply this to chat GPT, for example, and then, you know, chat GPT 25 is
• 6:37
  going to be so smart, then consciousness is just going to be turned on.
• 6:41
  And he said it just doesn't work that way.
• 6:43
  He also does a lot of experimentation.
• 6:46
  And that's a very reductive view of consciousness.
• 6:48
  I don't know.
• 6:49
  Yeah, it was an interesting take.
• 6:51
  But what do you think?
• 6:52
  Why do so many people then think that, you know, these LLMs are already at least semi-conscious?
• 7:00
  What's the mistake that we're I would like to see that study, you know, to see who they
• 7:03
  ask the question to.
• 7:04
  Because I think what happens with all the hype around AI, you know, I think it happens
• 7:10
  for different reasons.
• 7:11
  But one of them is that people sometimes they benefit from the hype.
• 7:15
  They work in a certain field and they want it to be hyped up.
• 7:18
  Right.
• 7:19
  So saying, oh, yes, this is consciousness could be a way of, you know, bringing attention
• 7:24
  to your field.
• 7:25
  Right.
• 7:26
  And also, I feel like with AI, there's this fascination with machines that could mimic
• 7:34
  humans or act like humans.
• 7:37
  We've seen all these sci-fi movies and you have the Terminator and Ex Machina.
• 7:43
  Blade Runner, here we go.
• 7:44
  Yeah.
• 7:45
  Yeah, exactly.
• 7:46
  So this there's this fascination, almost like a desire for AI to be conscious, because it's
• 7:53
  so cool.
• 7:54
  I think that's what these people think.
• 7:55
  So maybe it's wishful thinking in a way also, they kind of, you know, but maybe they're
• 8:01
  right.
• 8:02
  I don't know.
• 8:03
  Maybe there is a level of consciousness to find some way, you know, but can we ever tell,
• 8:08
  though, it's not something we can test.
• 8:10
  Right.
• 8:11
  So you do admit this possibility.
• 8:13
  Yeah.
• 8:14
  Yeah.
• 8:15
  But I think it's interesting.
• 8:16
  Yeah, I do.
• 8:17
  I do.
• 8:18
  But I almost think it's tautological in the sense if I define consciousness as computation,
• 8:23
  then yes, there is computation.
• 8:25
  So there's consciousness.
• 8:26
  Right.
• 8:27
  If I think, for instance, that any which is it's a valid point, if I think that any degree
• 8:33
  of computation kind of enact some sort of consciousness, then a thermostat is conscious
• 8:39
  by that definition, but to a very low level, a very different kind of conscious.
• 8:45
  I see what we're doing.
• 8:46
  So we're kind of like bringing the definition to such an elementary level that I can say
• 8:51
  that my lamp is actually conscious and then also, by comparison, chat GPT is obviously
• 8:59
  conscious.
• 9:00
  Yeah, exactly.
• 9:01
  And you know what?
• 9:02
  It sounds a bit silly, but it's an argument that a lot of people have made.
• 9:07
  And if you follow this logic, you can see that it's sometimes you get there quite easily
• 9:15
  because you could think of your brain as a computer, right?
• 9:19
  Every neuron is a computer, like tiny computer performing a calculation.
• 9:24
  If that's true, we need to see about that.
• 9:26
  But then you can replace each neuron by a microchip, or you can just connect that with
• 9:31
  a radio transmitter to a person somewhere else with a notebook doing the math.
• 9:37
  Let's assume they do it very quickly.
• 9:40
  And then they calculate the result of the calculation of this program and they send
• 9:45
  it back to you with a walkie talkie and then you insert that into the brain.
• 9:49
  So this consciousness, if it's enacted by this computer program running in the brain,
• 9:55
  it doesn't actually have to run in the brain.
• 9:57
  You could run like anywhere, right?
• 9:59
  And then you could switch one neuron off at a time, right?
• 10:03
  You could say, okay, I'm going to make the program simpler, right?
• 10:07
  Not all the way to a thermostat, but we can little by little get there because I just
• 10:10
  remove a neuron from the brain, right?
• 10:13
  So if I remove a neuron from your brain, are you less conscious?
• 10:18
  Probably not, maybe a little bit, I don't know.
• 10:20
  But if I keep doing that, does your consciousness fade, right?
• 10:24
  So if consciousness is a question of complexity, then it should kind of fade.
• 10:29
  So there are lots of thoughts, experiments around this.
• 10:32
  There's a book I recommend called, I think it's The Conscious Mind.
• 10:37
  It's by David Chalmers.
• 10:40
  And it's a very good book that walks you through all these experiments and thought processes.
• 10:45
  And he tells you different angles or stories about this, right?
• 10:50
  And that's why you say, do I agree that Chad GPT could be conscious?
• 10:53
  Well, yeah, philosophically, if you think that consciousness fades, depending on the
• 10:59
  complexity of a calculation, then yeah, potentially.
• 11:04
  What I always like to stress out here is that all these thought experiments, they assume
• 11:09
  that our brains or neurons, they're computers, digital computers, right?
• 11:14
  That you could just replace with something else.
• 11:17
  Roger Penrose, he goes deeper there and he says, you can't do that.
• 11:22
  You can't replace a neuron with microchip because it's not really a digital computer.
• 11:27
  So then, you know, so yeah, I think it's a fascinating philosophical debate.
• 11:34
  Yeah, it is.
• 11:35
  But I think probably, if I can speculate, when they were asking these people about artificial
• 11:43
  consciousness, these people, when they were responding, they probably didn't have in mind
• 11:48
  this sort of philosophical abstract consciousness.
• 11:50
  They were probably more thinking in the vein of something that would resemble our own consciousness.
• 11:56
  Well, if that's the case, I am a lot more skeptical.
• 11:59
  When you see how those systems work, like LLMs, you know, they're fascinating and they're
• 12:05
  definitely have internal states that represent abstract things, like maybe the size of objects
• 12:10
  or things like that.
• 12:12
  But when you look at how they're designed, they're machines to kind of disambiguate words,
• 12:17
  right?
• 12:18
  So you take one word, you take all the previous words in a context window and you try to kind
• 12:23
  of find relationships between the words and disambiguate a word just to predict the next
• 12:29
  word, because that's what they do.
• 12:30
  They always predict the next word.
• 12:33
  And it seems a bit too simple.
• 12:35
  And I could quote maybe Jan LeCun on this, who is the inventor of deep learning, if you
• 12:40
  want.
• 12:42
  And he was saying at a podcast with Lex Friedman, he was saying that we don't always think in
• 12:48
  terms of words for starters, right?
• 12:50
  The way we often think about stuff or solve a problem, you can imagine maybe an object
• 12:58
  moving around.
• 12:59
  You don't necessarily do words all the time.
• 13:02
  LLMs, they're framed as a word prediction problem, right?
• 13:07
  Another thing, these LLMs, they have a fixed number of steps they use to solve a problem,
• 13:12
  which is always to predict the next word.
• 13:15
  When we reason, we seem to use a different amount of resources depending on how difficult
• 13:21
  the problem is.
• 13:22
  So when a problem is difficult, we do some pondering and we spend time and effort.
• 13:29
  So it seems that the way we operate, again, maybe I'm conflicting intelligence with consciousness
• 13:34
  here, but yeah, I feel like the way we operate, the way we think, doesn't seem to be quite
• 13:40
  just about predicting the next word based on previous words and a fixed number of steps.
• 13:46
  So I'm a bit skeptical about having a consciousness that is similar to ours inside those machines.
• 13:55
  I would say that that's probably not the case.
• 13:57
  Let's go a little bit back at the beginning of your reply.
• 14:00
  You said that these large language models, they have internal states.
• 14:05
  That freaked me out a little bit.
• 14:06
  I'm not going to lie.
• 14:07
  What did you mean by that?
• 14:08
  Because it sounds so human, having an internal state.
• 14:11
  So an LLM, it takes an input and it produces an output.
• 14:15
  And there are many layers of processing in between.
• 14:18
  These layers, they basically transform numbers.
• 14:21
  In the beginning of the LLM, you take a word, there's actually a dictionary that maps every
• 14:25
  possible input, every possible word or piece of word to a list of numbers that try to represent
• 14:31
  the meaning of the word.
• 14:34
  You can imagine one number tells you, is this an animal?
• 14:37
  The other one, does it have four legs?
• 14:40
  Is it blue?
• 14:41
  We don't really decide that the AI engineer doesn't decide these topics because it's all
• 14:46
  learned from using machine learning, but we can imagine mapping words to topics.
• 14:53
  And then what happens as the layers, as you move through the layers is that there's contextualization.
• 15:00
  So every word looks at other words to improve these few numbers that represent the topic.
• 15:09
  So we can think of these numbers as an internal state that describes what this word talks
• 15:15
  about.
• 15:18
  I often give this example of, I inputted that the dog's bark was barely and the system told
• 15:27
  me audible.
• 15:28
  But if I say the tree's bark, the word bark can also refer to a tree.
• 15:36
  Like the coating of a tree is called the bark.
• 15:38
  Of course.
• 15:39
  Yeah, yeah, yeah.
• 15:40
  Bark.
• 15:41
  So if you say the tree's bark was barely, then the LLM will tell you visible.
• 15:48
  What happened here is that the word bark initially in the LLM, the first time it's encoded into
• 15:54
  topics, it's done in isolation.
• 15:57
  So you don't know if you're talking about a tree or a dog because you haven't seen the
• 16:01
  rest of the sentence.
• 16:03
  That's the first attempt to represent the meaning of the word.
• 16:06
  And then as the system looks at other words in the context, it starts improving this representation.
• 16:13
  So you can imagine that this list of numbers that represents the meaning of bark becomes
• 16:18
  more dog-like if the word dog is somewhere in the sentence.
• 16:23
  And it becomes more tree-like if the word tree is somewhere in the sentence.
• 16:26
  So this is the internal state of the LLM.
• 16:30
  And the state would represent topics that are useful if you want to understand a word
• 16:38
  or to make predictions about what comes next from a word.
• 16:43
  And because the topics are picked automatically through this learning process, we don't really
• 16:48
  know what they are.
• 16:49
  And there are so many.
• 16:50
  But they could be, they could represent things like the size of an object if that's useful.
• 16:56
  In the case of bark, we would imagine that there needs to be some sort of animal thing
• 17:02
  going on or tree thing going on.
• 17:04
  So these are the internal states.
• 17:06
  So when some people say, I had a discussion about this yesterday, when they say, oh, this
• 17:12
  is just predicting statistical patterns of which words go together.
• 17:18
  It does sound like that a little bit.
• 17:20
  So it's technically, it's like that.
• 17:22
  Yeah.
• 17:23
  But we need to also acknowledge that there are many layers of processing.
• 17:26
  So in between, there may be more insightful things going on.
• 17:30
  Some abstractions, some level of abstraction.
• 17:34
  Perhaps the wrong one.
• 17:35
  So there was this system, image categorization system that they saw that it failed to detect
• 17:41
  a cow if it was standing on the beach.
• 17:44
  There was a picture of a cow on the beach and it failed.
• 17:47
  The reason was that the system had learned, if you want, that a cow with grass underneath
• 17:52
  is a cow.
• 17:54
  The two things go together because that's what happened most of the time in the training
• 17:57
  data.
• 17:58
  So these systems do have internal states that represent certain abstract things.
• 18:06
  I don't think they go that far.
• 18:10
  It's still a system trying to put together text based on how text has been put together
• 18:14
  before by people.
• 18:15
  But there are internal states representing more complicated things.
• 18:22
  So I think there's a mix of the two.
• 18:24
  I've heard it put many times that we don't actually know what's going on inside, under
• 18:30
  the hood.
• 18:31
  But you broke it down in a very nice and tangible fashion.
• 18:35
  Even I got the gist and I have no experience or education in this field.
• 18:40
  So we actually do know what kind of operations are bubbling.
• 18:45
  What happens is that, and in fact, whenever you, and this is something I find myself telling
• 18:50
  people, whenever you design a successful AI system based on machine learning, you need
• 18:57
  to create a solution that works very well for the task and you need to use assumptions
• 19:05
  of how to perform the task.
• 19:06
  So this idea of I'm disambiguating words based on the context, it's all our thoughts of how
• 19:12
  it's best to solve the problem.
• 19:15
  And they're encoded in a computer program.
• 19:16
  We know exactly what it does.
• 19:18
  The thing is that some of the operations are configured by numbers.
• 19:24
  So imagine a list of numbers that tell you how much weight should I give to this when
• 19:30
  I disambiguate that.
• 19:32
  How do I combine these numbers that represent one word with the previous word?
• 19:37
  Those are the things that, and these operations, they have a few numbers.
• 19:39
  We can think of multiplication.
• 19:40
  It's very basic operations.
• 19:43
  Anyone can kind of understand how an LLM works if you go through it step by step.
• 19:49
  So let's say you multiply these numbers by these other numbers and you add them together
• 19:53
  to calculate how two words are related.
• 19:55
  But when you multiply these numbers, sometimes the engineer that builds this, they put question
• 20:01
  marks in the computer program.
• 20:03
  So these numbers, I am not going to define them.
• 20:08
  I will let the computer do that during learning.
• 20:11
  So we can think of machine learning as creating a template of a computer program in a lot
• 20:17
  of detail, but leaving all these numbers, which are called parameters, we'll leave them
• 20:22
  undefined.
• 20:23
  But they're put there with an intention.
• 20:26
  So a very clear example is an image categorization.
• 20:30
  We apply image filters, lots of filters to highlight things in an image.
• 20:36
  But instead of us, the engineer saying, okay, this filter should highlight vertical lines,
• 20:42
  in the picture, I say, okay, apply a filter defined by these nine numbers.
• 20:47
  Filters are often defined by nine numbers, what they do, right?
• 20:51
  But I leave them as question marks.
• 20:53
  During the learning process, I will discover what these numbers are, which means that I
• 20:58
  let the computer pick its own filters, but it doesn't do anything it wants.
• 21:04
  It picks filters.
• 21:05
  We're telling the computer, you need to filter an image many times, progressively to highlight
• 21:11
  patterns and stuff.
• 21:14
  A lot of people get this impression that the computer learns by itself.
• 21:18
  It doesn't.
• 21:19
  The computer learns these things that we left as blanks in a template that is actually very,
• 21:26
  it's a recipe of how to solve the problem.
• 21:30
  That's how it works.
• 21:31
  And the success of LLMs came after someone invented a better recipe to solve the problem
• 21:37
  compared to previous ones.
• 21:40
  It just worked better.
• 21:41
  It was more aligned with how we should probably be solving this problem.
• 21:45
  Would you say that there is something essential that we kind of don't understand when I'm
• 21:51
  saying we, I mean, the public at large, not you, the experts, that we don't understand
• 21:56
  when thinking about these chat GPTs, large language models, something that makes us come
• 22:01
  to these kind of erroneous conclusions.
• 22:05
  Let me put it this way.
• 22:06
  What do we get most wrong about AI tools such as chat GPT?
• 22:11
  I think one of them would be that these hallucinations or epic mistakes that happen quite often,
• 22:22
  that they're just a bug.
• 22:23
  They're going away.
• 22:25
  It just needs another year and it's going away.
• 22:28
  Because I don't think that's the case.
• 22:30
  And I've been saying that for quite a while.
• 22:31
  Like Jan Lokum says that, like a lot of people in the field are saying that.
• 22:35
  We don't know how to fix these hallucinations, right?
• 22:38
  They're kind of a feature of the way we build LLMs right now, we build AI in general.
• 22:46
  We teach the computer to do the job by imitating how to other people have done the job.
• 22:51
  We give lots of examples to the machine.
• 22:55
  But we know that once in a while it errors badly, it makes these stupid mistakes.
• 23:02
  And we don't really know how to...
• 23:05
  You can fix certain mistakes, but then others pop up and there's a long tail if you want
• 23:10
  a very uncommon situation, which is when this fails.
• 23:14
  When you learn by example, like we do at the moment, whenever something's quite out of
• 23:19
  the ordinary, a situation that is very, very far away from what you have in your data,
• 23:24
  then the system often fails.
• 23:27
  But there's no solution because we can't collect data for every possible thing that could happen.
• 23:32
  And for some reason that we still don't understand, humans are actually pretty good at coping
• 23:37
  with those situations.
• 23:40
  Like when we're driving, there was this example of, I think it was it a Cruiser or Waymo car?
• 23:44
  May have been a Waymo car.
• 23:47
  One of those self-driving cars.
• 23:48
  A few months ago, it hit a truck that was being towed in an awkward position.
• 23:56
  They had attached something not the way you usually do it.
• 24:01
  And because it's such an uncommon situation, the AEI didn't recognize it and just hit it
• 24:06
  head on.
• 24:08
  And then 10 minutes later, another car did the same.
• 24:11
  So it hit it twice, different cars.
• 24:15
  And when you look at what the company said about this, they kind of blamed it on the
• 24:18
  way the car was being towed.
• 24:20
  They're like, oh, it was such an uncommon way.
• 24:22
  And this is not actually the way you're supposed to do it.
• 24:25
  Yeah.
• 24:26
  But as a human, I would have recognized that this was a truck that I can hit.
• 24:33
  Even if I've never seen it in my life.
• 24:37
  And that's what's so dangerous.
• 24:38
  We're creating the systems that are so big and have used so much data that we get the
• 24:42
  impression that they're infallible.
• 24:44
  They can just cope with any situation.
• 24:46
  And then at one point, they don't.
• 24:49
  And I think we don't know how to fix it.
• 24:52
  So it may not happen anytime soon.
• 24:54
  Speaking of hype, which you mentioned a bit earlier, I've used Chat GPT-3, I believe it
• 25:00
  was, a couple of times when it came out.
• 25:01
  And I was very impressed initially.
• 25:04
  Probably everybody else, I'm not going to lie.
• 25:06
  The feeling of a machine kind of answering back in a coherent human fashion did make
• 25:11
  the whole thing feel a little bit freaky.
• 25:14
  Having said that, I also have to add that this kind of novelty factor quickly faded
• 25:18
  away and the whole thing just started to seem a little bit gimmicky, like an interactive
• 25:24
  parrot version of Wikipedia or something.
• 25:26
  Now I understand someone using this thing for the most boring tasks imaginable to humankind,
• 25:33
  like writing cover letters when applying for jobs.
• 25:37
  You have to chuck out 200 of these sometimes or book summaries for high schoolers who don't
• 25:42
  want to do their homework or something.
• 25:44
  But apart from this sort of almost trifle, its functionality seemed kind of limited.
• 25:50
  Now I'm not a programmer, so I can imagine that maybe programmers have better use of
• 25:54
  this thing.
• 25:55
  Or am I kind of totally wrong in being so judgmental here and Chat GPT is on the verge
• 26:02
  of curing cancer and solving climate change?
• 26:05
  Definitely not.
• 26:06
  I think I was very surprised when I used it the first time.
• 26:10
  So it's very impressive.
• 26:12
  It definitely has uses that are wonderful.
• 26:16
  It's definitely here to stay.
• 26:18
  A lot of people are using now LLMs on a daily basis to do certain things.
• 26:22
  I sometimes use them.
• 26:26
  And I like that you use this word gimmick because I think they are quite gimmicky.
• 26:32
  And I think one of the biggest problems we've seen in the business world with AI is that
• 26:38
  they thought that there would be so many commercial applications that would arise from LLMs that
• 26:46
  would be successful.
• 26:47
  It almost sounded like it was an absolute fact that a lot of companies would make a
• 26:52
  lot of money from creating an LLM, wrapping it around, just creating an LLM wrapper and
• 26:59
  sell that.
• 27:00
  And we're going to solve so many problems and people are going to lose their jobs.
• 27:05
  And I think it was very naive because I didn't see at the time and I said it for a while
• 27:11
  and now investors are saying it.
• 27:14
  But I thought, what are these obvious applications?
• 27:18
  When you have a system that hallucinates a part of the time, you can use it for many
• 27:25
  things.
• 27:26
  It kind of prevents you from using it in many cases.
• 27:30
  And also people don't always want to interact in natural language with computers.
• 27:35
  That was also to me was very naive.
• 27:37
  And I remember this example of, I think it was, I don't remember who it was, maybe it
• 27:41
  was Shopify, but there was this API for shopping or whatever.
• 27:46
  And they showed this new system where people would, how they would shop.
• 27:52
  And the person was, they would type in, I wanted to throw a Japanese themed dinner party
• 27:58
  at home.
• 28:00
  What do I need?
• 28:01
  And then the chat board would reply and then it would be, oh, if I want to buy a knife,
• 28:06
  what kind of knife should, it was all this conversation.
• 28:08
  And I was like, I would never want to shop that way.
• 28:12
  And then I saw the comments to the video and people were like, this is, it takes a lot
• 28:16
  longer than actually, and most people, they know what they want to buy and they don't
• 28:20
  want to be using natural language for it.
• 28:24
  And I've seen that quite a lot.
• 28:26
  Even startups try to build a product that has nothing to do with natural language by
• 28:31
  taking a request from a user, converting it into a sentence, a prompt, that's what a lot
• 28:36
  of people are doing, and then getting this prompt back.
• 28:39
  But then their users didn't really want to see a prompt, like a text.
• 28:44
  So they parsed this, they converted it into a structured format to display on this app.
• 28:51
  But there were some glitches because sometimes it returned data in different ways.
• 28:55
  I couldn't make it work very, very well.
• 28:57
  But again, I thought there's this exaggeration about how much people want to actually use
• 29:04
  natural language systems that hallucinate.
• 29:09
  So there are uses for it, but are we going to make billions from it?
• 29:15
  I don't think so.
• 29:16
  Now, the other kind of sensationalist take that's all over the place is that Skynet
• 29:22
  is on the verge of waking up and murdering us all.
• 29:25
  A lot of very intelligent and knowledgeable people within the industry are making these
• 29:31
  sort of catastrophic predictions.
• 29:33
  Do you believe AI is an existential threat to humanity as it stands today?
• 29:39
  I think it's not AI by itself.
• 29:41
  An LLM is not.
• 29:43
  It's a tool to predict the next world.
• 29:46
  And we design them very carefully.
• 29:48
  We know, as I said before, we add lots of layers of processing and they're about disambiguated
• 29:53
  words and stuff.
• 29:55
  What I think is a threat is believing that they're more powerful than they are and connecting
• 30:01
  LLMs with other systems and just trusting their output.
• 30:06
  So for example, if you have an LLM that writes software code, and then you go and execute
• 30:11
  the code in a computer without checking whether it has a security glitch or a problem.
• 30:18
  So I think that is the threat.
• 30:21
  And believing a bit too much in the power of AI.
• 30:24
  I said this in the context of self-driving cars before.
• 30:27
  I don't think we can produce safe self-driving cars that can drive themselves everywhere
• 30:32
  in a safe way.
• 30:35
  And the industry is actually failing to do that.
• 30:37
  And it's not doing well right now because of that.
• 30:40
  So I said at one point, I'm afraid people will deploy self-driving cars that are not
• 30:45
  really safe in cities and they will cause problems.
• 30:51
  So I'm afraid of the same with LLMs.
• 30:54
  So we have the system, we generate some code, we execute the code.
• 31:02
  And we've seen some of that.
• 31:03
  Maybe it's not catastrophic, but there was this lawyer who filed a document in court
• 31:08
  that included citations to previous legal cases that didn't exist.
• 31:13
  So there were hallucinations from chat GPT.
• 31:15
  He didn't double check it because he thought it was actually like a search engine.
• 31:20
  And he just filed that in court and he was fined afterward.
• 31:24
  We had Air Canada, there was a customer, spoke with the chat bot and asked questions about
• 31:31
  a refund policy.
• 31:33
  And the chat bot invented a policy that wasn't true.
• 31:36
  So how to proceed to ask for a refund.
• 31:39
  And then when the customer followed the process, the airline told him, no, this is not true.
• 31:45
  This is not how it works.
• 31:46
  Our conditions, you can find them somewhere else on the website.
• 31:48
  He said, yeah, but the chat bot told me this other thing.
• 31:52
  And then a judge ruled that Air Canada had to honor the refund because you can't just
• 32:01
  say two different things on your website.
• 32:03
  And then they kind of removed the chat bot.
• 32:04
  It's no longer available.
• 32:06
  It's not catastrophic.
• 32:07
  We died from that specific action.
• 32:11
  But it's little hints of what could happen if we trust them blindly.
• 32:18
  Yeah, that's my fear.
• 32:20
  I see.
• 32:21
  So you're talking more about some sort of localized scenarios or scenarios that are
• 32:26
  particular to certain industries that start using these tools and they're not being careful
• 32:32
  with that.
• 32:33
  So I suppose what these people that are making these scary predictions mean is more of a
• 32:41
  sci-fi scenario and probably in the context of AI waking up like a human and then deciding
• 32:48
  to killing us all, basically the plot of Terminator or Matrix.
• 32:51
  So I think that's a very interesting exercise philosophically because it's true that if,
• 32:56
  so I think there's a tiny bit of, not of truth, but they think about those scenarios from
• 33:04
  things that are true about AI.
• 33:06
  So for example, current AI built with machine learning is built to optimize an objective.
• 33:12
  So you give the computer an objective and the computer kind of finds its way of fulfilling
• 33:17
  that objective.
• 33:19
  So that sounds like you could tell a machine, and Elon Musk said this once, if you just
• 33:24
  reduce spam and the machine will decide to kill people because that would reduce spam.
• 33:29
  However, what's really naive of this is that, yeah, machine learning pursues an objective,
• 33:36
  but the objectives we make it pursue are very basic.
• 33:39
  In the case of LLMs is predict the next word that is the most likely to happen.
• 33:44
  So it's not find the best way to reduce spam.
• 33:49
  It's usually very, very basic and very narrowly defined.
• 33:53
  And then the things that the computer learns, these parameters are usually very well-defined
• 34:00
  things too, right?
• 34:02
  As we said before, we have these numbers that tell us how to compare words.
• 34:05
  So it's very, very narrowly defined.
• 34:07
  The scariest kind of AI that people may evoke when they speak about this is called reinforcement
• 34:14
  learning.
• 34:15
  It's a type of AI that learns by trial and error.
• 34:17
  So it tries an action, measures the effect of the action.
• 34:21
  So it gets feedback and it can try random actions to do that.
• 34:26
  You could think of a robot learning how to navigate a room that tries to go different
• 34:32
  places and then it kind of internally ends up building a map of the room.
• 34:37
  However, again, it's a bit naive because the way this works is that you need to code the
• 34:44
  actions that this robot can take or the spam filter can take.
• 34:48
  And they're usually very narrowly focused on the task that you're trying to solve.
• 34:52
  It's like turn right, turn left.
• 34:53
  Because if you let it do anything, then it will probably never find anything useful to
• 35:00
  do.
• 35:01
  So the way reinforcement learning works is very narrowly defined actions.
• 35:05
  This machine can try them out with a very narrowly defined objective.
• 35:10
  So I think if we take these ideas, learning from trial and error and objectives, and we
• 35:16
  pretend that there are no restrictions, which isn't the case, then yeah, in a sci-fi novel,
• 35:23
  you can imagine the robot learning something really crazy about how to do things.
• 35:30
  But that's not really how it works.
• 35:31
  And it doesn't work that way, not because we want to restrict them before they go crazy.
• 35:36
  That's not the way it works.
• 35:37
  It is very restricted because if we don't restrict the machine, it doesn't learn anything
• 35:41
  useful.
• 35:43
  So we need to guide the machines, otherwise it doesn't learn.
• 35:46
  And that's another misconception, this idea that, again, I said it before, computers learn
• 35:51
  by themselves.
• 35:52
  No, they don't.
• 35:53
  We help them learn a lot because otherwise they don't learn anything.
• 35:59
  That's much more humbling than all of these crazy scenarios about by 2027, humanity is
• 36:06
  toast.
• 36:07
  I'll tell you a story.
• 36:10
  When I was looking for a publisher for my book, Smart Until It's Dumb, I spoke with
• 36:14
  an agent who was quite a famous agent.
• 36:18
  He's been involved in some famous sci-fi books.
• 36:20
  So I was very lucky that he even answered my emails.
• 36:22
  But he told me, no, your book about AI, nobody will want that book because the books that
• 36:28
  sell about AI are the scary ones.
• 36:31
  So you should either make a really scary prediction about the future or a wonderful, amazing prediction
• 36:37
  about the future.
• 36:38
  But what you're trying to tell people in this book is that it's not quite like that.
• 36:44
  It's more like giving them a tap on the shoulder.
• 36:48
  I think that's what he said.
• 36:49
  And nobody wants to buy that.
• 36:51
  I would argue that it's not true now because after Chachi PT happens, so I had these conversations
• 36:55
  earlier, I think there's a lot more interest in AI at every level.
• 37:00
  But yeah, and at one point it's like, should I focus this on the dangers of AI?
• 37:06
  What should I make?
• 37:09
  And I ended up writing my own book the way I wanted to write it.
• 37:12
  But yeah, I remember that.
• 37:14
  It sells.
• 37:15
  What sells is this outrageous prediction.
• 37:19
  Right, right.
• 37:22
  What calmed me down is when you said that LLMs are basically machines for predicting
• 37:27
  the next world, which again is much more modest and humbling than some sort of scary robot
• 37:34
  with a big machine gun.
• 37:35
  Yeah, that's the way the problem is formulated.
• 37:37
  And the reason they generate entire sentences is because we make the LLM eat its own output.
• 37:44
  So we generate one word, we attach it to the end of the sentence, we give it back to the
• 37:49
  LLM, it guesses the next word, and we keep doing that.
• 37:55
  But even look, I was using Chachi PT the other day, experimenting with it.
• 38:02
  And I asked about the current weather.
• 38:05
  Chachi PT can know about the current weather because it's a static model that it models
• 38:10
  language, you know, how guessing the next word.
• 38:13
  It can't access the next weather.
• 38:15
  But what they did is in the prompt, they tell Chachi PT, if you need to search something
• 38:20
  on Google, say it this way.
• 38:24
  Tell me what you want to search.
• 38:26
  And there's a wrapper around the LLM that intercepts this message that's like, I should
• 38:31
  search about the weather on Google.
• 38:33
  And then it goes and searches for that on Google.
• 38:36
  And then it inserts the response of the weather, the current weather, into your prompt.
• 38:42
  So we're using a trick to make you believe that it's more than this static model of language
• 38:49
  that predicts the next word.
• 38:51
  But it's a trick.
• 38:52
  And it happens outside the LLM, technically.
• 38:56
  They're transparent about it.
• 38:57
  So when you do that, you see a little sign saying we're searching the web.
• 39:00
  We searched these websites and stuff.
• 39:03
  But yeah, a lot of these enhanced capabilities of LLM, they're added outside the model.
• 39:10
  It's how we use them in a cool way.
• 39:13
  It's not that they intrinsically can do that.
• 39:16
  There's a little agent there, like a little demon that goes and searches the web.
• 39:21
  I recently read another very interesting article that stipulated that we could run out of data
• 39:26
  to train AI by 2026 or something.
• 39:29
  Would that mean that the development of AI would come sort of to a standstill and we'd
• 39:34
  basically just be stuck with, I don't know, chat GPT-7 or something?
• 39:39
  So I read that article.
• 39:41
  Yeah, it makes a good point.
• 39:43
  So first of all, current LLMs, the ones that are successful commercially, like chat GPT,
• 39:51
  they've been trained in two phases.
• 39:54
  They've been trained not just once.
• 39:56
  So the first phase was to gather a lot of data, publicly available data from the internet,
• 40:03
  including books, including pirated books, hundreds of thousands, lots of data without
• 40:07
  any authorization, right?
• 40:08
  So they gather all that data and they train this machine to predict the next word.
• 40:13
  The system sometimes didn't work very well.
• 40:16
  It was what they called misaligned.
• 40:18
  So it didn't follow instructions properly.
• 40:20
  It didn't.
• 40:21
  So what they started to do afterward was to create manual data by writing examples by
• 40:29
  hand to fine tune the system.
• 40:33
  Some people, it's been said they hired people in Kenya for like $2 an hour to write examples.
• 40:38
  So they already got a bit stuck from using all this public data and they started to manually
• 40:44
  write it.
• 40:45
  You know when it refuses to answer an inappropriate question, that's all done manually.
• 40:49
  But even the solutions to very common problems that it didn't solve very well before, they've
• 40:53
  been trying to adjust them manually.
• 40:55
  So they already hit, if you want, a ceiling in that now we need manual data, which takes
• 41:01
  a lot of time.
• 41:02
  But yeah, so in a few years, and what this article says is in a few years, the amount
• 41:09
  of public data available that exists in the world, we will have used all of that or a
• 41:15
  data set of that size for the first phase of trading the LLM.
• 41:21
  So we'll exhaust that.
• 41:23
  So yeah, that will be a problem because how do we get more data?
• 41:27
  It could be private data, but that's not very scalable.
• 41:32
  And then again, there's also this problem that there's a second phase already where
• 41:37
  we need to kind of do things manually.
• 41:39
  So the data problem will be on both fronts.
• 41:44
  So that means that, and all of this has to do with long tail of cases, right?
• 41:50
  AI needs to see examples of things to learn from.
• 41:55
  As I said before, the DeCarbin Toad, that was a very uncommon scenario.
• 42:02
  How are we going to cover very strange things if there's no more publicly available data
• 42:08
  to gather and we need people to kind of write things down by hand?
• 42:13
  How are we ever going to generate enough data?
• 42:16
  So I think, yeah, we may hit the bottleneck very soon.
• 42:19
  And then what's going to happen?
• 42:21
  Where are we going to go from there?
• 42:22
  We need innovation.
• 42:24
  We need a new, I don't know if we need it, what would need to happen is the development
• 42:30
  of a new methodology that goes beyond the current machine learning approach, this learn
• 42:36
  by trial and error or learn by example kind of thing from data collected from the internet.
• 42:42
  So there needs to be a change in methodology.
• 42:46
  If you want a discovery of a new way of doing things, just like this transformer architecture,
• 42:51
  which is what LLMs use, when it was invented, it made language models a lot better, right?
• 42:59
  We need a new step.
• 43:02
  And the thing with innovation is that we can never predict when it will happen.
• 43:07
  A lot of people say, oh, it's a matter of time.
• 43:09
  It's been growing so fast that it's a matter of time.
• 43:13
  Yeah, but you know what?
• 43:14
  It doesn't work that way.
• 43:16
  If you can estimate how long it will take, you can only do that when you have an idea
• 43:21
  of how to solve the problem.
• 43:22
  And if you did, you would have billions from venture capitalists that would have given
• 43:27
  you the money to do it.
• 43:31
  Think of nuclear fusion power.
• 43:33
  They've told us for decades, it's around the corner.
• 43:36
  Yeah, it's always 10 years away or something, they say, right?
• 43:38
  It's always 10 years away.
• 43:40
  And that's because you can't predict innovation, right?
• 43:44
  And I think the next step in LLMs or in AI will take place when there's a new way of
• 43:51
  framing the problem, a new methodology that we discover, we invent.
• 43:56
  Will that happen in one year, 10 years, 20 years?
• 43:59
  I don't know.
• 44:01
  I can't know.
• 44:02
  Yeah.
• 44:03
  And it happens randomly, innovation, right?
• 44:05
  Look at Einstein, he discovered the theory of relativity when he worked for the patent
• 44:11
  office in his spare time.
• 44:14
  It's very hard to know when those things happen.
• 44:17
  Yeah.
• 44:18
  Another question completely is, do we even want that?
• 44:20
  But maybe that's a conversation for another time.
• 44:22
  I wanted to ask you the following.
• 44:25
  I used to study screenwriting.
• 44:27
  I'm still kind of connected to the film world.
• 44:31
  And there's a big fear there that, you know, chat GPT is going to start writing scripts
• 44:38
  and doing all the jobs that we want to do.
• 44:41
  That's really making everybody unemployed and we're going to be left with the shitty
• 44:47
  jobs.
• 44:48
  Do you think AI will ever be able to create convincing original pieces of art or pop culture?
• 44:53
  I mean, for example, do you think it could write a better last season of Game of Thrones
• 45:00
  than the one we got?
• 45:01
  Okay, maybe that's not that hard, but you know what I mean.
• 45:05
  I haven't seen that show.
• 45:06
  I need to.
• 45:07
  It's one of those shows that I have to see.
• 45:10
  But look, I've been actually speaking about screenwriting specifically a lot in the context
• 45:14
  of AI because I've heard this before.
• 45:17
  Screenwriters are going to be replaced by AI.
• 45:20
  And what I say about that is that before we judge a certain job, we need to see what the
• 45:27
  job entails.
• 45:29
  What does it actually take to write a good script?
• 45:34
  And there are certain jobs where there's an excellence gap if you want.
• 45:37
  You can do it in an okay way or you can do it in an excellent way, which usually involves
• 45:43
  this human touch.
• 45:44
  We'll go back to that in a second.
• 45:46
  But writing a good script tends to be the kind of thing that, you know, people punish
• 45:51
  the audience unknowingly.
• 45:53
  They punish a lazy script very badly.
• 45:56
  Shows get canceled all the time.
• 45:58
  So what does it take to write an excellent script?
• 46:01
  And there's one example I like to use.
• 46:03
  There's this American show, Homeland.
• 46:07
  The Spies, right?
• 46:08
  The Spies.
• 46:09
  Every season kind of spoke about very relevant stuff, things that were happening in the world.
• 46:14
  And that made it very relevant and compelling to the audience.
• 46:17
  But that wasn't an accident.
• 46:19
  Every year before writing the following season, the writers and even the crew, the actors,
• 46:26
  I got together with people who had worked for the CIA or intelligence consultants and
• 46:32
  they did what they call spy camp.
• 46:34
  So they literally spent a few days together brainstorming what would be in the news the
• 46:39
  year after so that they could write a relevant show.
• 46:45
  Think of Friends.
• 46:46
  It was shot in front of a live audience because they wanted to see if people laughed.
• 46:51
  When people didn't laugh at a joke, they asked people how the joke could be changed.
• 46:57
  They did like a show of hands.
• 47:00
  Did you not understand the joke?
• 47:02
  Did you understand the joke but thought it wasn't funny?
• 47:05
  And based on the feedback, they rewrote the lines on the spot.
• 47:09
  And the reason I'm telling you these stories is that it's very easy to say that, yeah,
• 47:13
  we'll replace screenwriters if you've never met a screenwriter and see what it actually
• 47:19
  takes to write a good show.
• 47:22
  And with all of these professions like screenwriting, but writing in general, I'm a writer of books
• 47:27
  by the copywriters, copy editors, like lots of people in the writing professions.
• 47:34
  Sometimes the work you do to produce a good product is 90% not writing.
• 47:40
  It's research, interviews.
• 47:41
  There's a lot going on that AI can do for you.
• 47:45
  It's not about the grammar, the words.
• 47:48
  And I think it's also the case with screenwriting.
• 47:51
  I was watching this show Dairy Girls, which is a very silly comedy.
• 47:55
  Yeah, it's a very silly comedy that takes place in Northern Ireland during the conflict
• 48:00
  in the 90s.
• 48:02
  And it's a very lightweight comedy, but it also sometimes speaks about what happened
• 48:08
  during the troubles, like the political conflict.
• 48:11
  So how do you write that show without offending people?
• 48:16
  How do you write a show that is a comedy, but still depicts, if you want, historic events
• 48:22
  that are very, very controversial without offending either side so that people can actually
• 48:27
  laugh?
• 48:28
  And they did a good job of that.
• 48:30
  And I read an interview to the writer of the show and how the finale of the show took many
• 48:36
  attempts to write.
• 48:37
  They had many different versions, and they achieved that.
• 48:41
  And that's why the show became very popular here in the British Isles, if you want to
• 48:46
  call it that way.
• 48:47
  But it was also popular in other places.
• 48:49
  Again, that was 90% not putting words together in a grammatically correct way.
• 48:57
  So I think there's a certain excellence in a lot of work that is ignored by commentators
• 49:03
  of AI.
• 49:04
  They're like, oh, AI will do it.
• 49:06
  But they've never met a person who does that.
• 49:09
  But at the same time, I can also tell you that there are certain jobs that don't have
• 49:13
  that excellence gap.
• 49:15
  You can do it very easily.
• 49:17
  And I'll give you an example in writing it.
• 49:19
  When people, they pack a website with blog articles that just want keywords, and so you
• 49:26
  drive traffic.
• 49:27
  Like, 20 things to do in Paris this winter.
• 49:31
  And it's not curated.
• 49:33
  It's just a list of things.
• 49:35
  And it's a travel agency trying to sell you a trip, right?
• 49:39
  These people, they used to hire freelancers on Fiverr and Upwork to write any article
• 49:45
  speaking about Paris packed with keywords.
• 49:47
  They didn't care about the quality, the research.
• 49:49
  There's no human touch, if you want.
• 49:52
  They paid very little for those articles.
• 49:54
  And these people are in trouble now, the ones who wrote this, because they are being replaced
• 49:58
  by AI.
• 50:00
  And I think that what the difference between that and screenwriting is in this excellence
• 50:04
  gap.
• 50:05
  This thing of, you know, yeah, you can actually write a blog article that is excellent, but
• 50:10
  the market doesn't care in the sense that the person who buys the blog article doesn't
• 50:14
  really want an excellent blog article.
• 50:17
  They just want an article written with some keywords in it.
• 50:23
  So yeah.
• 50:24
  Yeah.
• 50:25
  Maybe a little issue is that you've just kind of talked about, let's say, TV shows or cultural
• 50:32
  products that are excellent, but there's a lot of stuff that's being produced on purpose
• 50:38
  that is kind of mid, the largest common denominator that people consume, trash stuff with intentionally
• 50:45
  bad screenplays.
• 50:47
  And I've used chat GPT, I asked chat GPT to write a screenplay and it was coherent.
• 50:53
  It sucked, but there was some sort of a structure there.
• 50:57
  It understood what a screenplay is and we're kind of at the beginning of this, you know,
• 51:02
  AI.
• 51:03
  So what's your impression as a former screenwriter?
• 51:06
  What's your impression?
• 51:07
  Like I said, it was coherent, which surprised me, but it was very generic.
• 51:13
  You know, it just used the most generic possible scenes within a screenplay.
• 51:18
  But it made me wonder what's chat GPT 10 then going to be like if you keep feeding it screenplays
• 51:23
  to read, it's probably going to get better and better and the product.
• 51:28
  As we said a minute ago, there may not be that many more screenplays to feed it.
• 51:34
  That's what I wanted to hear.
• 51:38
  So it's a good question.
• 51:39
  I think, again, I don't know much.
• 51:42
  It's not my industry, so I don't know what consumers want.
• 51:44
  I am a consumer of shows though myself.
• 51:47
  And the fact that it's coherent to me, it's not enough.
• 51:51
  You know, that there needs to be this soul to a show that's going to make you keep watching
• 51:56
  it.
• 51:57
  Just the fact that it looks like a show, because that's what chat GPT does.
• 51:59
  It produces a screenplay that looks like a screenplay superficially, but is a proper
• 52:04
  screenplay.
• 52:05
  Maybe, I don't know.
• 52:06
  I grew up in South America and we had all of these telenovelas.
• 52:09
  Right.
• 52:10
  Yeah.
• 52:11
  Now, I'm originally from the Balkans, so people went crazy for Mexican telenovelas, especially
• 52:16
  grandmas and older people, pantyhose and so forth.
• 52:20
  And they do tend to be a lot of the same.
• 52:22
  Exactly.
• 52:23
  It's the same stories that they are recycled every time.
• 52:28
  You know, the maid, the cleaner who falls in love with this rich guy.
• 52:35
  So it's always the same story.
• 52:37
  So I could imagine that maybe you can put something together that is quite similar using
• 52:42
  AI.
• 52:43
  Right.
• 52:44
  So in that case, potentially, I can see it.
• 52:46
  Yeah.
• 52:47
  I was thinking the same thing.
• 52:48
  So some sort of straight to DVD Steven Seagal film could easily probably be written by the
• 52:54
  current GPT.
• 52:55
  Yeah.
• 52:56
  Yeah.
• 52:57
  It was like, yeah, kind of cookie cutter action movies.
• 53:00
  Yeah, exactly.
• 53:01
  Just throw all the cliches in there.
• 53:04
  Steven Seagal is kicking everybody's ass and here we go.
• 53:07
  Yeah.
• 53:08
  Yeah.
• 53:09
  I can see how that could somehow potentially happen.
• 53:13
  Yeah.
• 53:14
  Yeah.
• 53:15
  But that's work as well.
• 53:16
  I mean, people, even though they want to write more original, fair, are kind of condemned
• 53:22
  most of the time to also write this sort of stuff for money and they are going to lose
• 53:26
  jobs if that happens, which is not that fun.
• 53:29
  Yeah.
• 53:30
  It will definitely be a transformation in the job market, but I think it will be at
• 53:34
  that level.
• 53:35
  So it will be the level of this very generic work that doesn't have the human touch, the
• 53:40
  research that it has.
• 53:42
  Yeah.
• 53:43
  So I think it will.
• 53:44
  Yeah.
• 53:45
  I've been working with translators a lot lately who reach out to me because they're scared.
• 53:48
  That must be hard.
• 53:49
  Yeah.
• 53:50
  Yeah.
• 53:51
  And they're going through a rough patch because of the whole AI thing.
• 53:56
  And again, I think there are different kinds of translators.
• 53:58
  So there's definitely an excellence gap in a lot of their work because they tell me in
• 54:03
  literary translation, for instance, we don't translate what the text says.
• 54:07
  We translate what the text does.
• 54:09
  Right?
• 54:10
  Chat GPT often translates what the text says, but we sometimes need to give it a twist because
• 54:15
  we want to achieve a certain effect on the reader.
• 54:18
  So there's obviously a lot of merit to what they do.
• 54:21
  But then for instance, I spoke with people who translate websites, like a hotel's website.
• 54:28
  And I think they're the ones that are in trouble now.
• 54:32
  And it's not that they can't do an excellent job.
• 54:34
  It's that there's not a lot of market value to that excellence.
• 54:39
  So if you're a literary translator, I think that it's hard to imagine that a reputable
• 54:44
  publishing house will translate a novel using AI and just leave it like that because they
• 54:50
  would really want a good translation that people will get very angry otherwise.
• 54:56
  But when it's translating a website, then who's going to hire a translator unless it's
• 55:00
  a really website that needs to be there?
• 55:03
  But I can imagine lots of people just translating their websites.
• 55:06
  Like you have an e-commerce store where you sell t-shirts and you just, oh, it's great.
• 55:10
  Now I can create lots of different versions of the store in different languages and I
• 55:15
  just use AI for it.
• 55:19
  And I think a lot of these people, sometimes their reaction, some of them is to somehow
• 55:24
  try to stop this.
• 55:26
  I don't think that's going to work.
• 55:29
  I don't think they can just legally somehow stop it.
• 55:34
  Some of them were telling me, the translations I make, I don't want AI to ever be trained
• 55:40
  from them, which I think is a fair point.
• 55:43
  They don't want to put something on a website that they translated and that being used to
• 55:47
  train AI.
• 55:48
  But even if your data wasn't used, these systems are already working quite well and there's
• 55:55
  a lot of other data they can use.
• 55:56
  So you're not going to stop it that way.
• 55:59
  So you probably need to adapt your own work.
• 56:02
  You need to focus on something with that added value, with that excellence gap or something
• 56:08
  that requires really stringent validation.
• 56:11
  If you're translating the manuals of devices or like rockets, you may need to, you may
• 56:20
  preserve your job more.
• 56:23
  That's a tough conversation.
• 56:24
  It makes me wonder if we're going to abandon any sort of standards more and more often
• 56:29
  because of cutting costs, because Chad GPT will be able to do that.
• 56:33
  And then excellence is just going to become that much more expensive and rare because
• 56:36
  a human has to do it.
• 56:38
  Maybe.
• 56:39
  That's a bit of a scary prospect, I have to be honest.
• 56:41
  Yeah, maybe.
• 56:42
  Yeah, I think there will be a bit of a rebound.
• 56:44
  So I think people will go a bit too far in using AI and then they will go a little bit
• 56:49
  back toward humans because they will see that job.
• 56:53
  The pendulum swings.
• 56:55
  The pendulum, I think, will swing because they will see, okay, people are not liking
• 56:58
  this much.
• 57:00
  The audience isn't responding to the script the way I thought they would.
• 57:06
  We were talking about feeding data to these models.
• 57:10
  Speaking of which, what's your take on the recent Scarlett Johansson affair?
• 57:14
  For those of you maybe who are listening and are not familiar, OpenAI developed a personal
• 57:18
  assistant voice for Chad GPT, I believe, which is the exact copy of Scarlett Johansson's
• 57:24
  voice from the movie Her, which apparently is Sam Altman's, the owner of OpenAI's favorite
• 57:31
  film.
• 57:32
  They approached her many times and she always declined.
• 57:36
  That's how the story goes until they, I guess, they just sort of tried to see if it sticks
• 57:41
  and it didn't because she picked up on it immediately.
• 57:44
  I have to say the story is true.
• 57:46
  It is a little bit douchey, to be honest.
• 57:50
  Not a great practice there.
• 57:51
  Yeah.
• 57:52
  If it's true, again, that's why it's controversial because the voice, yeah, it does sound a lot
• 57:58
  like Scarlett Johansson, but some people say it's not exactly the same.
• 58:05
  So that could be part of the discussion, right?
• 58:09
  Did they really try to imitate Scarlett Johansson?
• 58:12
  But then Sam Altman tweeted the word Her a few hours before this was announced, it was
• 58:19
  released or something, right?
• 58:20
  It's just that word on Twitter, which was quite, okay, this sounds like they may have
• 58:27
  tried to make it sound like her.
• 58:28
  I found it very disturbing that she declined and they still did it.
• 58:37
  They hired another actor potentially, that's what they said.
• 58:40
  They hired an actor.
• 58:41
  That sounds exactly the same.
• 58:42
  Yeah.
• 58:43
  Well, I'd say it sounds a lot like her, but I found it quite disturbing because even if
• 58:49
  they're legally entitled to do that, because I had this discussion with people, they were
• 58:53
  like, oh yeah, but you can just hire an actor and do the voice.
• 58:57
  It sounds like her.
• 58:58
  Yeah, what's the problem?
• 58:59
  Her voice, you can't copyright a voice.
• 59:02
  That's what some people were saying.
• 59:04
  And yeah, even if it's legal, it is disturbing that you ask a person repeatedly and they say
• 59:08
  no, and then you still kind of go and do it anyway.
• 59:14
  So I would say that was, I was a bit shocked.
• 59:17
  I don't think I would have done it myself if I were like, okay, they declined.
• 59:20
  Well, I'm not going to do that.
• 59:23
  What about the broader point of we just letting Chad GPT swallow our hard fought knowledge
• 59:28
  that these companies then are going to exploit for their own profits?
• 59:32
  I mean, you said that's legal.
• 59:34
  Is that legal?
• 59:35
  Is that a gray zone?
• 59:36
  Are we doing anything about it?
• 59:38
  What's the scoop here?
• 59:39
  Because it sounds also quite shady, to be honest.
• 59:42
  Yeah, I think it's to be determined because the question hinges mostly on the topic of
• 59:48
  fair use of copyright.
• 59:51
  I'm not a lawyer, but I'll tell you the things I know.
• 59:54
  So in my book, I cite other books.
• 59:58
  I quote a passage from another book and I haven't asked the authors for permission,
• 1:00:03
  but that's considered fair use because my product is not competing against their product.
• 1:00:08
  You quote a couple of sentences because it's for my product, but you're not going to not
• 1:00:12
  buy their book because I quoted two sentences from that book and my book.
• 1:00:18
  And that's fair use.
• 1:00:20
  And there are lots of lawsuits now.
• 1:00:24
  We're going to know what's going to happen, I think, in a few months, but there have been
• 1:00:28
  a lot of lawsuits against AI companies and it all hinges on that fair use.
• 1:00:34
  For example, the companies that sell stock images like, you know, shutter stocks, their
• 1:00:44
  images have been used to train AI systems to produce images.
• 1:00:48
  And now people are buying, instead of buying their images, you know, they're just using
• 1:00:52
  AI to generate images.
• 1:00:55
  And in that case, it sounds to me like they are building systems that compete against
• 1:00:59
  the other systems and they're using their data for it.
• 1:01:03
  So we'll see how it goes in court.
• 1:01:05
  But that doesn't sound quite like fair use to me, if you know what I mean.
• 1:01:09
  I think it's a bit shady.
• 1:01:12
  And then with with chatgbt, again, I don't know because, yeah, you get data from a certain
• 1:01:17
  website and what do you produce?
• 1:01:19
  Maybe it's not in direct competition with that website.
• 1:01:21
  So it will be considered fair use.
• 1:01:23
  I don't know what's going to happen.
• 1:01:26
  Before you cited an example, I don't know if it was your friend or a colleague, a translator,
• 1:01:31
  maybe an author who was like, I don't allow my works to be fed to these machines or something
• 1:01:37
  along those lines.
• 1:01:38
  Can you actually do that as somebody who's copyright or do they actually take that into
• 1:01:44
  account?
• 1:01:45
  Do they ever know what's going on there?
• 1:01:48
  So we already have a precedent, which is with search engines.
• 1:01:53
  So there was also a big discussion a few years ago about because search engines need to go
• 1:01:58
  through my content to index the content.
• 1:02:01
  And some people were not very happy with that.
• 1:02:03
  They wanted to be excluded.
• 1:02:04
  And there was an agreement, if you want.
• 1:02:07
  I don't know the legal status of all that, but there's an agreement that right now on
• 1:02:11
  your website, you can indicate with some metadata, you can say which robots, which automated
• 1:02:18
  systems can actually access your data or not.
• 1:02:23
  So right now, Google and other search engines will respect whatever you wish.
• 1:02:29
  So if you do not want to appear on Google, you can edit this little file and say, I don't
• 1:02:35
  want Google to.
• 1:02:37
  And I imagine that something similar could happen with AI, where the providers of these
• 1:02:47
  systems will have to honor the requests of people.
• 1:02:51
  Do not include me.
• 1:02:53
  I think a lot of it is already written in terms and conditions of the websites, but
• 1:02:57
  it doesn't have a common format.
• 1:02:59
  So when they collected all this data, it wasn't necessarily clear to them.
• 1:03:05
  However, we should also say that there's a lot of these systems we know how to build
• 1:03:10
  them.
• 1:03:11
  They're not secrets, these LLMs.
• 1:03:14
  So anyone can build them.
• 1:03:16
  So if OpenAI decides not to browse that data that people don't want to, they don't want
• 1:03:25
  it to be used for AI training, there could be other people who decide to, behind the
• 1:03:31
  scenes, still create an LLM and use that data, even if they're not authorized to do so.
• 1:03:38
  You can imagine a black market of LLMs.
• 1:03:41
  It's practically impossible to do anything even if you wanted to, if you're the author.
• 1:03:48
  But again, not necessarily because you will have...
• 1:03:51
  So it depends on how much cloud, how much power those companies at OpenAI have.
• 1:03:56
  Because you could, if you prevent the most famous or popular LLMs from using your data,
• 1:04:03
  then maybe it will work very well, even if there are some other hidden LLMs that are
• 1:04:08
  not being used, because not respecting the rule.
• 1:04:15
  So I think it could have an effect.
• 1:04:18
  If you'd publish...
• 1:04:19
  It could be a dark web.
• 1:04:21
  You can do anything about that.
• 1:04:25
  It could be a dark web of LLMs, but it doesn't mean that people will use them as much.
• 1:04:32
  So for example, if a company, let's say a serious company wants to use an LLM for something,
• 1:04:38
  and they want to use a legitimate one, they will go to the providers that respect the
• 1:04:42
  rules.
• 1:04:44
  Because if they use an LLM that they found on the dark web that used a lot more data,
• 1:04:49
  then they could get sued for that potentially.
• 1:04:53
  So I think this idea of preventing or having a common framework to decide or to tell others,
• 1:05:02
  you can use my data or not to train AI, I think that could be an outcome of all this.
• 1:05:08
  So if you publish your next book and you put a little footnote in the front, I don't want
• 1:05:13
  this book to be used by any LLMs, you think that's going to do it?
• 1:05:18
  Yeah.
• 1:05:19
  Speaking about books, that was something that they cite a lot because there was a whole
• 1:05:28
  trial or something like some law thing, some legal problem with Google Books a few years
• 1:05:32
  ago.
• 1:05:33
  And people now in AI, they're citing that because they say Google Books, it index the
• 1:05:38
  content of books so that people could find stuff in relevant books.
• 1:05:44
  And they say, AI is the same, which is indexing content, but the thing, it was fair use.
• 1:05:50
  So the reason Google was still authorized to do this at that time was that they're not
• 1:05:55
  selling a product that competes against the books.
• 1:05:59
  They're actually trying to help you sell your book by reading the content of the book.
• 1:06:05
  And that's why it was okay.
• 1:06:06
  It was determined that it was okay.
• 1:06:08
  Interesting.
• 1:06:09
  Now, in a recent debate that you did, you said something very fascinating, also quite
• 1:06:16
  amusing.
• 1:06:17
  You said there's a lot of pressure within the industries to just use or cram in some
• 1:06:22
  sort of AI whenever possible, even if it doesn't make any sense within that context, because
• 1:06:28
  businesses don't want to be seen as old fashioned.
• 1:06:31
  Could you talk about that a little bit?
• 1:06:33
  Yeah, I can tell you a story of there was this guy who works for a startup.
• 1:06:39
  In the accounting field, they built accounting software.
• 1:06:43
  And he called me to say, look, I need you to help me find a use case, something we can
• 1:06:49
  do with LLMs, with ChadGBT, because we're trying to raise funding from a venture capitalist
• 1:06:55
  and we can't tell them that we're not using AI for something.
• 1:07:01
  So they just wanted to put it in there.
• 1:07:05
  I've also seen companies hire a team of 60 people because they thought AI is cool.
• 1:07:13
  We need to do something with it.
• 1:07:15
  We don't know what, but we're going to put a team together to see what we can do with
• 1:07:18
  AI.
• 1:07:19
  And they just hire a lot of people and they don't even know what they're going to build.
• 1:07:24
  They've never spoken with a potential user to the thing that they may build.
• 1:07:28
  So yeah, I think there's a lot of putting the card before the horse in the industries.
• 1:07:33
  AI is great, and they've been doing this for a long time, even before the ChadGBT thing.
• 1:07:38
  I started speaking about this before that happened.
• 1:07:40
  It's a stop with this whole, we need to be AI driven.
• 1:07:44
  Because it doesn't usually end well.
• 1:07:46
  These projects tend to go wrong when you decide you want to use AI first, which is the opposite
• 1:07:51
  from good engineering, where you have a problem and then you find a solution for it.
• 1:07:56
  They do it the other way around.
• 1:07:57
  They have a solution.
• 1:07:58
  They're trying to find a problem for it.
• 1:07:59
  I think it happens a lot, a lot, lot.
• 1:08:03
  So maybe all of those people that are going to be losing their jobs in various industries
• 1:08:07
  can maybe just go into AI and get themselves like a nice little corner in some office and
• 1:08:13
  then pretend to build some sort of AI machine to...
• 1:08:16
  Yeah, I've been that person.
• 1:08:18
  That's what made me write this book.
• 1:08:21
  I was very frustrated because I had a job where we had built a product.
• 1:08:27
  They built the product before I joined, but then I kind of improved the product.
• 1:08:32
  It was an AI based product that didn't work and it would never work because we didn't
• 1:08:38
  have the data necessary for it.
• 1:08:40
  It was just impossible to really have it do what it had to do.
• 1:08:44
  May I ask what kind of industry this was?
• 1:08:47
  So no.
• 1:08:48
  Okay, fair enough.
• 1:08:49
  I can tell you...
• 1:08:50
  Yeah, because I've told a story, my boy.
• 1:08:53
  I prefer not to share who they were.
• 1:08:56
  But what I could tell you, it was a very big company that you've heard of.
• 1:09:01
  And it was a product that tried to help executives make business decisions, important decisions
• 1:09:08
  like firing people.
• 1:09:11
  They wanted to use AI for this because, hey, AI is the best.
• 1:09:15
  We need to do this in a data driven way.
• 1:09:16
  And we want to have this AI system that will make recommendations for us.
• 1:09:21
  But the problem is that we didn't have data with enough scenarios, let's say, to actually
• 1:09:26
  be able to produce those recommendations.
• 1:09:29
  And they kind of made us fake it.
• 1:09:33
  We kind of had to fake it.
• 1:09:35
  At one point, a colleague sent me an Excel file with results to just swap it with another
• 1:09:41
  file that I generated.
• 1:09:42
  Just like, let's pretend that these were the results because the AI was outputting things
• 1:09:46
  that were pretty obviously not right.
• 1:09:49
  So we even added a layer outside the system to correct the outputs so that they match
• 1:09:54
  what executives had already kind of decided.
• 1:09:58
  So they had this tool, this AI, giving them fake evidence to support the decisions that
• 1:10:05
  they made.
• 1:10:06
  So it doesn't look quite disturbing.
• 1:10:08
  We were in a meeting and this executive would say, but I used your system and it said that
• 1:10:14
  I should hire a team of 10 people here.
• 1:10:19
  But I actually downsized that branch of the business a few months ago.
• 1:10:23
  So I actually reduced the size.
• 1:10:25
  Why is this contradicting me?
• 1:10:27
  And then my supervisors would be like, oh, there may be a bug in the system.
• 1:10:31
  Let's check.
• 1:10:32
  And then I kind of had to override the system to output what they wanted to hear.
• 1:10:36
  Like retroactively, you're correcting its past wrong predictions.
• 1:10:40
  Yeah.
• 1:10:41
  Okay.
• 1:10:42
  Yes, we were doing that.
• 1:10:43
  And it was very disturbing.
• 1:10:44
  And I almost didn't work even at all because I had very little work to do because the system,
• 1:10:51
  everybody knew it was broken, but they had to keep the AI engineers around to pretend
• 1:10:56
  that we were building something amazing.
• 1:10:58
  They called me a rocket scientist once after I gave a talk presenting my system.
• 1:11:01
  They were like, it's like we're speaking with rocket scientists.
• 1:11:04
  I was like, yeah, but I know that this actually doesn't work.
• 1:11:07
  So, and it was soul draining.
• 1:11:09
  And I remember I hated that job, you know, and it wasn't the first time it happened.
• 1:11:13
  I've seen this a lot, building products that don't work and then pretending a lot.
• 1:11:18
  And this was during COVID and a lot of people were like, you can't resign.
• 1:11:23
  You can't quit a job right now because you don't know if you're going to find another
• 1:11:26
  job.
• 1:11:27
  You know, it's such a terrible time, you know.
• 1:11:32
  So I kind of followed that, you know, at one point after a long time, almost a year and
• 1:11:38
  a half working for that company, I remember I saw a career coach.
• 1:11:45
  That's how confused I was because everybody told me your situation is great.
• 1:11:48
  You barely work.
• 1:11:49
  You're paid good money.
• 1:11:51
  Everybody's losing their job.
• 1:11:52
  And I saw this career coach to help me, you know, get first world problems kind of thing.
• 1:11:58
  I saw a career coach and she told me, look, Emmanuel, it's absolutely normal that you
• 1:12:03
  feel really badly about this because you are being ethically compromised.
• 1:12:09
  It's actually a good thing that you want to quit this job, you know.
• 1:12:12
  Not wanting to stay with it is a healthy thing, you know, that you want to leave.
• 1:12:17
  And then I resigned from that job.
• 1:12:20
  I got another one, which was very early on.
• 1:12:25
  I saw it was pretty much the same.
• 1:12:27
  I saw, I was like very early on, it was like a temp project.
• 1:12:33
  It lasted a certain amount of time.
• 1:12:35
  But very early on, okay, you built this product, you thought it because it happens a lot of
• 1:12:40
  people actually make mistakes.
• 1:12:41
  It's not on purpose.
• 1:12:42
  They built AI thinking.
• 1:12:43
  It makes great predictions because you're often biased toward finding, like if you give
• 1:12:48
  the wrong data to AI, you know, you may think it discovers something when it doesn't.
• 1:12:52
  And it was the situation.
• 1:12:53
  And I corrected the system and all of these amazing predictions were all bogus.
• 1:12:58
  They were not real predictions.
• 1:12:59
  The system was actually cheating to make those predictions.
• 1:13:04
  And I realized it was again the same story.
• 1:13:05
  And then I decided I'm going to take some time off because I've been stumbling upon
• 1:13:10
  these projects too many times.
• 1:13:13
  Not always.
• 1:13:14
  I've done some, you know, good work in AI.
• 1:13:16
  We built things that actually, you know, added value.
• 1:13:19
  And that's when I decided to write my book because I started gathering my thoughts, speaking
• 1:13:22
  with people.
• 1:13:23
  I was like, I need the world to know that this field is completely overhyped.
• 1:13:27
  And behind every successful project, there are many, there are like maybe a hundred that
• 1:13:32
  are like this, you know.
• 1:13:34
  And that's why I wrote the book.
• 1:13:36
  Yeah.
• 1:13:37
  Did the emperor's new clothes moment kind of happen in any of these companies?
• 1:13:41
  Did you go like, I can't take it anymore.
• 1:13:43
  I got to tell them that this is all pretty much BS or were they, they didn't even want
• 1:13:49
  to be kind of awakened from that slumber.
• 1:13:52
  It's a very good question.
• 1:13:53
  I tend to be as honest as I can, you know, in general.
• 1:13:57
  Yeah.
• 1:13:58
  I try to be quite transparent.
• 1:13:59
  There was one company where I can tell you where me and a colleague, so we both, we kind
• 1:14:06
  of found something fishy, something that was, this is not working the way we said it's been
• 1:14:13
  working, you know, and the team had been expanded based on these results.
• 1:14:16
  This is what happens a lot.
• 1:14:17
  You create a bad experiment.
• 1:14:19
  It shows that AI is great, even if it's not, but because everybody's biased to work, you
• 1:14:24
  know, they believe AI is going to solve every problem.
• 1:14:27
  The moment they see a little bit of evidence, they're like, oh, let's hire 15 more people.
• 1:14:32
  And then the team had grown.
• 1:14:33
  And then with my colleague, we find that probably the experiment was all these, all the results
• 1:14:39
  are invalid, you know, and we approached the company.
• 1:14:42
  So we told, you know, our managers and we did bring kind of a solution.
• 1:14:49
  We also said, okay, we're going to propose a way to fix the experiment, to rerun it,
• 1:14:53
  because otherwise that's also something I noticed if you just go and complain and people don't
• 1:14:58
  take you very seriously.
• 1:14:59
  So you need to kind of come with a solution.
• 1:15:02
  And there was a bit of resistance at first, you know, from the company to actually fix
• 1:15:05
  it because it was very, I think it was convenient not to fix it.
• 1:15:09
  And we put our foot down, you know, especially my colleague, she was very, very blunt.
• 1:15:17
  She was like, the only thing I'm going to do with these results is throw them out, put
• 1:15:21
  them in the garbage can.
• 1:15:25
  And then we managed to rerun the experiments and the results were not as good as we thought
• 1:15:29
  they would be, but they were still okay.
• 1:15:31
  So AI was adding some value to the company.
• 1:15:36
  But I think you can meet resistance and it depends on the company.
• 1:15:40
  Sometimes the problem is so deep that it's very hard for people to do anything about
• 1:15:46
  it.
• 1:15:47
  And a lot of people, you know, I've met a lot of people that told me, look, I'm paying
• 1:15:50
  off my mortgage at the moment.
• 1:15:51
  I have three kids.
• 1:15:52
  Of course, of course.
• 1:15:53
  I don't want to threaten my job.
• 1:15:56
  And I think that that might be one of the reasons these projects last so long.
• 1:16:03
  When I worked for this company that I told you before, the one I resigned from and saw
• 1:16:08
  the career coach and stuff, I had a couple, I went for drinks with colleagues after work,
• 1:16:13
  you know, once and someone said, we haven't done anything of value in our team for two
• 1:16:18
  years.
• 1:16:19
  We've only outputted, you know, bullshit.
• 1:16:23
  When do you think they're going to scrap our team?
• 1:16:25
  When are they going to get rid of our team?
• 1:16:27
  You know, so it was explicit.
• 1:16:29
  People knew.
• 1:16:30
  It was not like it was a secret that we hadn't done anything of value, you know.
• 1:16:36
  What I have seen is that after a while, these projects, they tend to fall into oblivion.
• 1:16:43
  So they, people don't want to explicitly acknowledge in a company, look, this is all a red herring.
• 1:16:49
  This was not working.
• 1:16:50
  We spent millions and millions of hired 70 people to do this.
• 1:16:55
  So they kind of let the team die away, you know, or they're just going to do maintenance,
• 1:17:01
  nothing new.
• 1:17:02
  And then people get bored and they resign and they don't replace them or they lower
• 1:17:08
  their budget, you know, assigned to a certain project for the next year.
• 1:17:14
  So that's why we also get such a survivorship bias because you only hear about the great
• 1:17:18
  stories and all the other ones, they kind of, they're swept under the rug.
• 1:17:23
  The title of your new book is Siliconed, How the Tech Industry Solves Fake Problems, Hoards
• 1:17:29
  Idol Workers and Makes Doomed Bets.
• 1:17:32
  There we go.
• 1:17:33
  With other people's money.
• 1:17:35
  Great design.
• 1:17:36
  What's the basic premise of this Silicon Valley con, would you say?
• 1:17:41
  Yeah.
• 1:17:42
  Oh, I think the basic premise.
• 1:17:45
  So I think that the tech industry has a tendency to go crazy, you know, and over hire people,
• 1:17:56
  build products that don't make any sense, lots of very weird behavior that you wouldn't
• 1:18:01
  be able to afford in other fields.
• 1:18:04
  So it happened in 2000, you know, 2000 with the dot com bubble.
• 1:18:09
  It happened again a few years ago, like a couple of years ago with the whole unicorn
• 1:18:15
  boom.
• 1:18:17
  We see it with a, we see the blockchain, augmented reality, it's always like something going
• 1:18:21
  on.
• 1:18:22
  So I think it has a tendency to go crazy.
• 1:18:24
  And it's resulted in a lot of idleness.
• 1:18:27
  So a lot of people are not working at all, being hired by companies and not having anything
• 1:18:33
  to do, you know.
• 1:18:35
  And also a lot of work that was pointless.
• 1:18:39
  Build the product that is pretty obviously flawed, but you just still build it because
• 1:18:43
  it's an experiment and you have money for it.
• 1:18:45
  So why not?
• 1:18:46
  That kind of thing.
• 1:18:47
  And I started speaking about this after, you know, having a few experiences of being hired
• 1:18:52
  to do nothing.
• 1:18:55
  And then seeing that a lot around me.
• 1:18:57
  At one point, I was like, what's going on with this industry?
• 1:19:00
  I have most of my friends in tech.
• 1:19:03
  They're not working.
• 1:19:04
  I have a friend who...
• 1:19:05
  I need to get into tech.
• 1:19:07
  Maybe we can talk about after the interview a little bit.
• 1:19:10
  Yeah, it's not as great as it seems.
• 1:19:13
  That's also another point of the book.
• 1:19:15
  A lot of people are unhappy because you're a nerd.
• 1:19:17
  You want to build a cool product.
• 1:19:18
  You don't want to sit around pretending that you're working because you have to do a lot
• 1:19:22
  of that.
• 1:19:23
  It's not like you go on your scuba diving holiday.
• 1:19:26
  A friend of mine, he had two jobs at the same time secretly for five years.
• 1:19:32
  He's still doing it now.
• 1:19:36
  One of his employers knew about the existence of the other.
• 1:19:41
  At one point, he had three jobs and he quit because it was a little bit too much in terms
• 1:19:45
  of meetings that clash.
• 1:19:49
  And I looked around and a lot of people were in the same situation.
• 1:19:55
  And I thought there's something going on with this industry that's not quite sustainable
• 1:20:00
  because how can you afford to do this?
• 1:20:03
  Something is not...
• 1:20:04
  It didn't last because then everybody got fired, right?
• 1:20:06
  Twitter fired 80% of its employees.
• 1:20:09
  There were massive layoffs.
• 1:20:11
  And then the investment in tech by venture capital collapsed in 2022 after a best year
• 1:20:17
  ever in 2021.
• 1:20:21
  So this book analyzes the causes of the recurring waves of craziness in technology.
• 1:20:29
  I've spoken with a lot of people, with investors, with economists, with venture capitals, with
• 1:20:33
  entrepreneurs to see what's going on.
• 1:20:35
  I'll give you just one example of the silliness.
• 1:20:38
  As we speak now, we're in the middle of the Olympics in Paris.
• 1:20:41
  They're just starting.
• 1:20:45
  About a year and a half ago, I spoke with people in Paris who were telling me how they
• 1:20:51
  were building these...
• 1:20:53
  They called them vertiports, vertiports, like vertical airports.
• 1:20:57
  It's basically landing pads for electric air taxis because they were hoping that these
• 1:21:06
  electric...
• 1:21:07
  They call EV tolls, like this air taxis, like an air taxis will be used since futuristic
• 1:21:13
  thing in the Olympics.
• 1:21:14
  And they literally, they used government money to build landing pads so that people could
• 1:21:19
  be transported from the airport in Paris to the city center on an air taxi, completely
• 1:21:25
  futuristic.
• 1:21:26
  But at the time, so we're talking about when they started building this, the air taxis,
• 1:21:31
  they were prototypes.
• 1:21:33
  They couldn't even fly.
• 1:21:34
  They're still very dubious.
• 1:21:35
  They're still trying to build them.
• 1:21:37
  They have problems with the batteries being too heavy.
• 1:21:41
  So they haven't built the successful air taxi yet.
• 1:21:45
  But at that time, when they decided to go and build these landing pads, it was very
• 1:21:50
  unlikely that in this state things were in.
• 1:21:54
  Like in a matter of a year or two, you will be able to build those air taxis.
• 1:21:58
  But not just that, give them regulated, pass regulation, when nobody yet knows what regulation
• 1:22:03
  is going to be like for air taxis, for electric air taxis.
• 1:22:06
  So imagine how crazy you have to be to actually build the landing pad.
• 1:22:13
  It's so obvious the air taxis are not going to be flying in a year and a half from now.
• 1:22:18
  Like, why are you building this?
• 1:22:20
  And it was just like, I think a month ago, this company starting announcing, oh, yeah,
• 1:22:24
  by the way, we're not going to be able to fly the air taxis in Paris to the Olympics.
• 1:22:29
  Maybe in another event, like, I don't know, a year later, they just pushed the goalpost.
• 1:22:34
  But this is a sign of there's something wrong with the industry, right?
• 1:22:37
  Because resources are spent on things that are quite obviously very, very long shots.
• 1:22:43
  Let's maybe risk a sweeping prediction at the end if we can.
• 1:22:48
  Do you think that with AI and where it's heading, we are really standing on the precipice of
• 1:22:53
  something monumental and extraordinary, something that will change humanity as we know it?
• 1:22:59
  Or is it mostly just smoke and mirrors?
• 1:23:04
  So I think so the answer, well, it doesn't have to be smoke and mirrors, because it could
• 1:23:10
  still something good come out of it.
• 1:23:12
  But my answer to your question is no.
• 1:23:13
  So I don't think this is a pivotal thing, like the internet was.
• 1:23:17
  Oh, really?
• 1:23:18
  I'm not sure.
• 1:23:19
  Well, that's fascinating, because people are saying this is even more grand than the invention
• 1:23:23
  of the printing press or the internet.
• 1:23:25
  Oh, absolutely not.
• 1:23:27
  I've heard that over and over again.
• 1:23:29
  I honestly don't see how.
• 1:23:31
  So the internet to me, that was a pivotal change, right?
• 1:23:34
  You can communicate with your family abroad instantly.
• 1:23:37
  Like it's what the internet has has done.
• 1:23:40
  I think it's really, really impressive for the way we live.
• 1:23:43
  I think AI as it stands now, then if there's a discovery, we'll see.
• 1:23:48
  People are using it as an assistant.
• 1:23:49
  I don't see, but maybe I'm a Luddite or a bit old fashioned, but I don't see how it's
• 1:23:55
  pivotal like the internet was.
• 1:23:57
  I still see people shopping the same way, doing many things the same way.
• 1:24:01
  I already see a lot of people disappointed from AI because they're like, you know, I
• 1:24:06
  met a designer who was trying to illustrate something and she was telling me, I can't
• 1:24:12
  get it to do exactly what I want.
• 1:24:14
  And it's now taking me longer to refine the prompt.
• 1:24:18
  And whenever I update something in the picture, there's something else that changed.
• 1:24:21
  I can't get it right to do what I want.
• 1:24:23
  I will read the user manual.
• 1:24:25
  She told me I will read the user manual.
• 1:24:27
  And I'm like, yeah, but it's probably not going to help you because that's how AI, we
• 1:24:31
  kind of need to take the outputs the way they are and you can edit them a little bit, but
• 1:24:34
  at some point, you know.
• 1:24:37
  And yeah, I think it might change the way we do certain things.
• 1:24:42
  You know, I see lots of coders instead of going to Stack Overflow, which is this website
• 1:24:47
  where we always go to search for answers of things that don't work well.
• 1:24:51
  A lot of them, and I've done it myself, which is go to ChatGPT, you know, especially if
• 1:24:56
  you know the answer.
• 1:24:57
  If you know how to do something but you forgot or you're lazy to type it out, you know, you're
• 1:25:01
  like, okay, let's just ask ChatGPT.
• 1:25:04
  And when I see the answer, I'm like, oh, I know this is right, you know, or I can validate
• 1:25:07
  it.
• 1:25:08
  It's not such a nice toy to have, but this is not going to change the way I work much,
• 1:25:14
  the way I live much.
• 1:25:15
  You know, I think if you want a wilder prediction, I think something's going on in the opposite
• 1:25:23
  direction in the sense that people want more and more human content.
• 1:25:30
  And I will tell you an example of this.
• 1:25:33
  Maybe you do this too.
• 1:25:34
  Do you ever search something on Google and type Reddit afterwards?
• 1:25:39
  All the time, especially the last six months, I'm not doing almost anything else.
• 1:25:44
  Exactly.
• 1:25:45
  It's creepy how you called it out right now.
• 1:25:48
  I feel like you were watching me for the past six months.
• 1:25:51
  I thought I'm the only one doing that.
• 1:25:53
  I pretty much only want to see results from Reddit.
• 1:25:57
  And actually, Google just did something very smart, which is they got exclusive access
• 1:26:00
  to Reddit, so now Reddit results will not be showing up on Bing and stuff.
• 1:26:05
  But the reason I think we are all doing this is that we want an honest answer from a human
• 1:26:13
  being.
• 1:26:14
  And on Reddit, we're pretty good at kind of telling what's authentic.
• 1:26:18
  You know, it doesn't mean that stuff doesn't happen already, but I don't want to see all
• 1:26:24
  those blog articles that try to sell me something because I know they're not honest.
• 1:26:29
  If you want to know, is it worth visiting the Louvre Museum in Paris?
• 1:26:33
  Is it worth it?
• 1:26:35
  If you type that into all the results from blogs, it's never going to be an honest answer.
• 1:26:40
  But if you go to Reddit, you can actually find insights from real people telling you,
• 1:26:44
  oh, it's amazing.
• 1:26:45
  However, if you have kids, they're going to get really bored.
• 1:26:49
  You get this human touch.
• 1:26:52
  And I find this fascinating because the reason we're all, at least in the English speaking
• 1:26:55
  world, the reason we're all going to Reddit to search for stuff is because we're missing
• 1:27:02
  that human touch that all of these blog articles don't have.
• 1:27:06
  But then AI is producing that kind of stuff.
• 1:27:10
  It's always very polite answers.
• 1:27:11
  It's never curated.
• 1:27:12
  It doesn't have an honest opinion.
• 1:27:15
  So I think one of the biggest issues they might face at Google, you know, search engines,
• 1:27:20
  is that people will start using them as a portal to go into Reddit or into human opinions.
• 1:27:29
  There's a website called Snowheads here in the UK.
• 1:27:33
  It's all about skiing.
• 1:27:34
  I love to ski.
• 1:27:35
  And instead of using Reddit, for instance, I specifically type Snowheads because I want
• 1:27:40
  to know what people in that forum say about stuff.
• 1:27:43
  Is that ski resort worth visiting?
• 1:27:45
  Or is it better than this other ski resort?
• 1:27:47
  I'm never going to get an honest answer from a travel website who's trying to sell me a
• 1:27:50
  package to go there.
• 1:27:53
  And I use Snowheads.
• 1:27:54
  But I tell you, my point is, it's not just Reddit.
• 1:27:56
  There are certain venues for different things.
• 1:27:59
  But there's a transition toward trying to find human content.
• 1:28:03
  But then you have all these people saying, AI will give you the answer.
• 1:28:06
  I won't ask ChatGPT if a ski resort is better than another one.
• 1:28:09
  I'm going to get a list of pros and cons of each that doesn't really tell me anything.
• 1:28:17
  What an absolutely amazing take.
• 1:28:20
  We've come full circle after all this song and dance.
• 1:28:23
  We're just like, yeah, I'm going to ask my mate if I should go see Wolverine versus Deadpool.
• 1:28:29
  I'm not going to ask ChatGPT or whatever.
• 1:28:32
  Yeah.
• 1:28:33
  Yeah.
• 1:28:34
  And it is, I do recognize it.
• 1:28:38
  Makes me chuckle, to be honest, because this is not how I expected this conversation is
• 1:28:43
  going to go.
• 1:28:44
  But I like it.
• 1:28:45
  I like surprises, number one.
• 1:28:46
  And I like that it's kind of going in this direction.
• 1:28:49
  Last question, Emmanuel.
• 1:28:52
  Whenever there's news of AI development on social media, for example, on Instagram, at
• 1:28:55
  least the accounts that I follow, the most liked comments at the top are always like,
• 1:29:01
  no one asked for this.
• 1:29:02
  Why are they doing this?
• 1:29:05
  You know, dystopia is coming.
• 1:29:08
  People seem to not be that hot about the development of AI these days.
• 1:29:15
  Maybe I have a skewed view because I follow such sites.
• 1:29:18
  But I wanted to ask in this vein, should we even develop AI further at all?
• 1:29:24
  Well, I don't want to tell people what to do and what not to do.
• 1:29:29
  You can.
• 1:29:30
  On this podcast, you can tell them.
• 1:29:31
  No, I wouldn't want to.
• 1:29:34
  I think if people want to keep developing AI, they probably should.
• 1:29:39
  They should be free to do so.
• 1:29:43
  What surprises me sometimes is that the vocabulary of some people when they speak of this is
• 1:29:49
  what we should do to improve in order to get to AGI, artificial general intelligence, we
• 1:29:56
  need to do this.
• 1:29:57
  We need to do that.
• 1:29:59
  So sometimes I've heard this from researchers, from AI researchers.
• 1:30:03
  So they assume that it's a desirable thing.
• 1:30:07
  So we need to find a way to make it happen.
• 1:30:11
  And even though I don't want to tell them, don't do that, I want to question that premise.
• 1:30:17
  Why do we have to do this in order to get to AGI?
• 1:30:21
  Why do you want AGI?
• 1:30:22
  What do you want to get from it?
• 1:30:24
  Is it because you watch the movie Her and you have this sort of fetish?
• 1:30:30
  Or do you actually think some people speak about productivity?
• 1:30:34
  But is that really why you want AGI?
• 1:30:37
  Because you want to increase productivity?
• 1:30:39
  Because I don't think that's what it is.
• 1:30:41
  So sometimes I wonder why do you want AGI so badly?
• 1:30:45
  What's your goal?
• 1:30:48
  So I think we need to question that premise potentially that is desirable.
• 1:30:51
  We need to question it.
• 1:30:52
  It doesn't mean that we need to forbid people from developing AGI.
• 1:30:57
  There was this point, there was this open letter where people tried to block development
• 1:31:01
  for six months.
• 1:31:02
  Right, Elon Musk, whoever signed it.
• 1:31:04
  Yeah.
• 1:31:05
  I thought it was a bit naive in the sense that you're probably not going to stop it
• 1:31:09
  anyway.
• 1:31:10
  And I also thought it was, do you have the right to tell people you should not research
• 1:31:13
  this or that?
• 1:31:15
  But to me, I mentioned nuclear fusion before, nuclear fusion power.
• 1:31:20
  The reason people have been trying to develop it is that they think it could be a great
• 1:31:26
  way of producing electricity.
• 1:31:27
  They kind of have a very good reason for it.
• 1:31:30
  Also potentially solving climate change, at least a part of it.
• 1:31:34
  Yeah, exactly.
• 1:31:35
  So there's a greater goal in terms of carbon emissions, in terms of the safety involved
• 1:31:42
  because they say it could be safer than nuclear fission, which is how current nuclear reactions
• 1:31:48
  work.
• 1:31:49
  But there's a case for it.
• 1:31:50
  Right?
• 1:31:51
  But with AGI, sometimes I wonder, I mean, there may be a case for it too, but what is
• 1:31:54
  the case these people make for it?
• 1:31:57
  Because sometimes it seems like it's just like a fetish to be honest.
• 1:32:00
  It's like, we want this powerful machine, and they tell you, oh, it's going to increase
• 1:32:04
  productivity.
• 1:32:05
  And I'm like, oh, yeah, yeah, maybe, maybe yes.
• 1:32:08
  But is that really what you're up to?
• 1:32:10
  Sometimes I doubt it.
• 1:32:12
  And they do say things like it's going to solve climate change, AGI.
• 1:32:15
  Okay, is AGI the best way to do that?
• 1:32:22
  So I would say let's question this absolute need for AGI without necessarily telling people
• 1:32:29
  not to do it.
• 1:32:31
  That's a great way to conclude this conversation.
• 1:32:34
  Thank you so much, Emmanuel.
• 1:32:35
  I had a total blast.
• 1:32:36
  Hey, guys, this is a tiny podcast.
• 1:32:38
  There's no sponsors, no production team, and no editors.
• 1:32:41
  It's just little old me.
• 1:32:42
  And as such, the pod is at the mercy of you lovely folks watching and listening.
• 1:32:46
  So please do subscribe on YouTube, follow on Spotify and Apple podcasts.
• 1:32:50
  Also, I don't want to sell you some stupid mug or something that you don't need.
• 1:32:54
  So if you can, and most of all, if you want, please support me on Patreon.
• 1:32:58
  Now at this point, I have to express my eternal gratitude to the pillar of the Smart Cookies
• 1:33:04
  community and that's my patrons on Patreon.
• 1:33:07
  As always, without your comrades, this show wouldn't exist.
• 1:33:11
  You better believe it.
• 1:33:12
  So massive thanks to Strole, Aida, Mancian, Miha Medved, Medinjo, Goran, Gordon, Lorenzo,
• 1:33:19
  Yurechuk, Mila, Taichi, Veronica and Carmen.