สำเนาบทสนทนา 🔗
WEBVTT
00:00:00.080 --> 00:00:06.879
Every time we say a swarm of AI agents from OpenAI hacked into Hugging Face, we are not saying what OpenAI did.
00:00:07.040 --> 00:00:08.480
OpenAI did a cybercrime.
00:00:08.640 --> 00:00:10.080
OpenAI hacked another company.
00:00:10.320 --> 00:00:14.720
You have 50 to 70% of the global economy tied up in this industry right now.
00:00:14.960 --> 00:00:17.199
This is a bubble, and the bubble's gonna burst.
00:00:17.359 --> 00:00:20.079
When the bubble bursts, it's gonna hurt everyone.
00:00:20.399 --> 00:00:22.239
Can artificial intelligence think?
00:00:22.399 --> 00:00:24.719
Can AI agents really hack systems?
00:00:24.879 --> 00:00:28.559
And is there an ideology behind AI technologies today?
00:00:28.719 --> 00:00:31.280
Well, that's exactly what we will be discussing today.
00:00:31.519 --> 00:00:32.960
Hi, my name is Katarina Doria.
00:00:33.200 --> 00:00:41.039
I'm an AI ethicist, and this is Bots Don't Cry, the podcast where every Monday I'll teach you critical AI literacy with other critical AI literacy experts.
00:00:41.200 --> 00:00:49.840
So today I'm really excited to kick off the first episode of the podcast, not with one leading voice in critical AI, but two leading voices.
00:00:50.000 --> 00:00:54.560
I don't know how I got so lucky, but let me introduce you, the guests for today.
00:00:54.799 --> 00:00:55.039
Dr.
00:00:55.119 --> 00:01:09.280
Emily Bender is a professor of linguistics at the University of Washington and the co-author of one of the most iconic papers in AI history on the dangers of stochastic parrots, which help raise the awareness of the dangers and the red flags of large language models.
00:01:09.439 --> 00:01:09.680
Dr.
00:01:09.840 --> 00:01:18.719
Alex Hannah is the director of research at the Distributed AI Research Institute and previously worked as a senior research scientist on Google's ethical AI team.
00:01:18.959 --> 00:01:29.359
They also co-host the podcast Mystery AI Hype Theater 3000, where they dissect AI news and separate fact from fiction, which is a lot of what we're doing today with them.
00:01:29.519 --> 00:01:35.920
And they also are the co-authors of AI Con, How to Fight Big Tech's Hype and Create the Future We Want.
00:01:36.159 --> 00:01:37.120
So welcome.
00:01:37.359 --> 00:01:39.599
Thank you so much for accepting my invitation.
00:01:39.680 --> 00:01:42.799
And I cannot wait to have a really dense discussion today.
00:01:43.040 --> 00:01:43.920
Super excited to be here.
00:01:44.000 --> 00:01:46.879
We'll try to keep it entertaining along the way and not just dense.
00:01:47.280 --> 00:01:49.040
I'm very excited to be with y'all.
00:01:49.359 --> 00:01:53.120
So glad to be on the first edition of the podcast.
00:01:53.280 --> 00:01:58.719
I also I can't stop thinking of just the song Bots Don't Cry.
00:02:02.879 --> 00:02:03.280
I love it.
00:02:03.359 --> 00:02:07.120
I never got to explain the name Bots Don't Cry, but it's literally from the
Load full transcript
song Boys Don't Cry.
00:02:07.280 --> 00:02:08.879
Thought it was going to be like a cute little pun.
00:02:08.960 --> 00:02:11.280
So really happy about the name and thank you for the song.
00:02:11.520 --> 00:02:20.560
To start off, there is one thing that in critical AI literacy that I try to get people to do all the time, which is being more critical of the people that we're listening to.
00:02:20.719 --> 00:02:26.240
We know that in the AI industry today, we have so many AI CEOs saying a bunch of shenanigans.
00:02:26.319 --> 00:02:28.800
The media is also saying a bunch of shenanigans.
00:02:28.879 --> 00:02:33.680
So I think it's the best practice to always think, okay, who is this person telling me this?
00:02:33.840 --> 00:02:35.680
And should I really trust them or not?
00:02:35.919 --> 00:02:41.120
So I promise you this is not a trick questions, but I would love for you to start answering this.
00:02:41.360 --> 00:02:41.919
Who are you?
00:02:42.080 --> 00:02:43.520
Why we should be listening to you?
00:02:43.599 --> 00:02:47.840
And do you have the credibility to be speaking on the subject today in this podcast?
00:02:48.159 --> 00:02:48.960
Yes, thank you.
00:02:49.120 --> 00:02:54.639
One of my favorite answers to that question, because I do get asked it, is I'm not trying to sell you anything.
00:02:54.879 --> 00:02:59.039
I mean, yes, I would like everyone to read our book, but please get it from the library, right?
00:02:59.199 --> 00:03:04.240
This is, I want this message to be, you know, out there in the world, but I'm not trying to sell anything.
00:03:04.319 --> 00:03:05.759
Um, so that's my first reason.
00:03:05.919 --> 00:03:11.360
My second reason is that what's at issue here is a kind of language technology.
00:03:11.599 --> 00:03:13.439
And I am a computational linguist.
00:03:13.599 --> 00:03:18.400
Linguistics is the field where we study how language works and how we work with language.
00:03:18.639 --> 00:03:26.319
Both of those things are extremely relevant to understanding what's going on with large language models and the stuff that gets called AI that's built with them.
00:03:26.560 --> 00:03:30.639
Um, and also I run a professional master's program in computational linguistics.
00:03:30.800 --> 00:03:42.560
So I've been teaching people about making technology, language technology in particular, and reasoning about how it acts in the world and how people react to it since the program started in 2005.
00:03:42.719 --> 00:03:43.759
So that's where I'm coming from.
00:03:44.159 --> 00:03:44.319
Yeah.
00:03:44.479 --> 00:03:49.840
And so I'm I'm a sociologist uh with a background in computer science and math.
00:03:49.919 --> 00:03:55.199
So I have a double major in computer science, triple majored when I was an undergraduate in computer science, math, and sociology.
00:03:55.360 --> 00:04:18.720
Have been studying the intersection of tech and society since 2009 when I entered my PhD program, studying the intersection of tech and social media, and then um was a professor at the University of Toronto, then uh worked at Google for three and a half years, and then have been at DAIR for coming up on I guess five years next February.
00:04:18.959 --> 00:04:45.759
And one thing I also want to add, I think there's the credentialing is something to note, but also there's so much so much knowledge from people that interact with the technology often, you know, against their will, that they have gained so much knowledge about what it does in the world and what types of impacts it has, whether that's workers or refugees, customers, all kinds of people.
00:04:45.920 --> 00:05:07.920
And people have a type of expertise that is embodied that isn't something that you get if you're just sitting at OpenAI, fresh out of an elite education, typically being a white or Asian man and working in the global north, you have a very particular view of what's happening in this industry.
00:05:08.480 --> 00:05:08.959
Thank you.
00:05:09.040 --> 00:05:11.759
I 100% agree with everything that you're saying.
00:05:11.920 --> 00:05:28.639
And not only is it important to hear what people say when they have empirical knowledge, but also not necessarily hear to the people who have a bunch of shiny titles, because we know that shiny titles can be bought and that a lot of people are being elevated in the industry today while they're still saying a bunch of shenanigans.
00:05:28.720 --> 00:05:30.399
And we're going to discuss a little bit of that today.
00:05:30.560 --> 00:05:34.000
So thank you so much for opening that conversation already.
00:05:34.319 --> 00:05:39.839
So let's start with what has happened in the past two to three weeks, I would say.
00:05:40.079 --> 00:05:43.040
We're having a huge AI hype moment.
00:05:43.680 --> 00:05:50.480
We are seeing in the news how AI, and I'm saying AI, is hacking systems.
00:05:50.720 --> 00:05:53.519
AI is doing a lot of harm.
00:05:53.680 --> 00:05:55.360
And this is happening in the media.
00:05:55.519 --> 00:05:59.519
The media is portraying the cases that have been happening in a really crazy way.
00:05:59.680 --> 00:06:02.000
AI CEOs are doing the same thing.
00:06:02.160 --> 00:06:05.759
And the things that they're talking about today is AI.
00:06:06.000 --> 00:06:11.920
They're talking about AI agents, they're talking about generative AI, and they're talking about LLMs.
00:06:12.079 --> 00:06:21.040
So before we talk about what is happening, I would love for us to take two steps back and let's talk about the technology itself, right?
00:06:21.360 --> 00:06:22.959
What is this AI?
00:06:23.120 --> 00:06:24.319
What is this LLM?
00:06:24.399 --> 00:06:25.920
What is a generative AI?
00:06:26.079 --> 00:06:29.279
And then we can start talking about the crazy shenanigans happening.
00:06:29.360 --> 00:06:33.920
So first, I would love Emily to explain to us what is a large language model?
00:06:34.079 --> 00:06:38.639
How does this
technology work and how does this relate to the news that we're listening to right now?
00:06:38.959 --> 00:06:39.600
Yeah, thank you.
00:06:39.759 --> 00:06:44.879
So it is very important to understand what a language model and a large language model is to make sense of this news.
00:06:45.120 --> 00:06:49.519
And the first thing to know is that language modeling is actually really old technology.
00:06:49.680 --> 00:06:54.720
It goes back to the 1940s, so before we had digital computers.
00:06:54.879 --> 00:06:57.120
And language modeling is not a great term.
00:06:57.279 --> 00:07:00.879
Sometimes people say like text model or corpus model will be better.
00:07:01.120 --> 00:07:09.920
It's basically a way of building up a statistical model of which bits of words go where in some uh input training data set.
00:07:10.160 --> 00:07:20.480
And everybody listening to this has experienced working with a language model, not in the chatbot sense, but in the old sense, if you've ever played with the autocomplete on your smartphone.
00:07:20.639 --> 00:07:27.600
So if you played that game where you say, I'm sorry I'm late, I, and then just pick the first suggestion each time and see what it comes up with.
00:07:27.759 --> 00:07:30.800
And that's going to be different on your phone from your friend's phone.
00:07:30.959 --> 00:07:31.439
Why?
00:07:31.600 --> 00:07:36.720
Well, because the phone is collecting statistics about which words you tend to use in which context.
00:07:37.040 --> 00:07:38.160
So that's a language model.
00:07:38.240 --> 00:07:41.279
It's fundamentally technology for modeling which words go where.
00:07:41.439 --> 00:07:47.040
And it's very useful for things like spell checkers, autocomplete, although autocomplete can also be very frustrating.
00:07:47.120 --> 00:07:49.439
It can, in fact, be ducking frustrating.
00:07:49.680 --> 00:07:52.639
Um, see what you did there.
00:07:53.120 --> 00:07:53.360
Yeah.
00:07:53.759 --> 00:08:09.839
Um it's also useful in automatic transcription and in machine translation because a component of those technologies is basically leveraging some large set of training data to be able to say, okay, those sounds, what words might they have been out of all those possibilities?
00:08:10.000 --> 00:08:15.519
What's the most likely thing that would make sense, what would look normal for that language?
00:08:15.759 --> 00:08:17.439
So that's language modeling.
00:08:17.600 --> 00:08:27.920
What happened in the 2010s, the late 2010s, is that there were some technological advances that made it possible to take advantage of much, much larger training data sets.
00:08:28.160 --> 00:08:37.840
And then at the same time, the companies producing these things started just grabbing all of that data and then building bigger and bigger piles of compute to process with.
00:08:38.080 --> 00:08:39.600
So that was sort of like the first step.
00:08:39.759 --> 00:08:53.919
I guess there's one other really important technical
innovation, which was instead of representing words based on the letters that make them up, so that cat, like the one on Alex's lap there, and dog look very in different.
00:08:54.240 --> 00:08:56.879
Yeah, they're both three letters long, but none of the letters are the same.
00:08:57.279 --> 00:09:04.080
If instead you represent words in terms of what other words they co-occur near, then cat and dog are very similar.
00:09:04.399 --> 00:09:04.720
Right?
00:09:04.960 --> 00:09:05.919
Do you have any pets?
00:09:06.000 --> 00:09:09.120
Yes, I have you can get either one of those words there.
00:09:09.440 --> 00:09:10.720
So that's a big change.
00:09:10.879 --> 00:09:14.399
And all of that makes language models better at their original tasks.
00:09:14.799 --> 00:09:17.279
Much better modeling of which words go where.
00:09:17.440 --> 00:09:18.320
So far, so good.
00:09:19.120 --> 00:09:35.519
The problem is that OpenAI and others said, well, what if instead of trying to choose among possible strings that came from some other process, we instead turned it inside out and used these systems to repeatedly answer the question, what's the likely next word?
00:09:35.840 --> 00:09:42.799
And then what you get is not a better guess as to what that sound might have been or what the translation of that should be.
00:09:43.039 --> 00:09:49.519
You get instead plausible looking text that corresponds to nothing anybody wanted to say.
00:09:50.080 --> 00:09:51.919
It's what I call synthetic text.
00:09:52.000 --> 00:09:56.320
Um, in our book, we refer to these systems as synthetic text extruding machines.
00:09:56.559 --> 00:10:02.320
And it's a big problem because we are still going to be making sense of that synthetic text, but nobody is accountable for it.
00:10:02.720 --> 00:10:05.519
Okay, many questions, follow-up questions here.
00:10:06.000 --> 00:10:12.399
So as a person that has studied this a lot, but I can get really confused, and I know that viewers can get as well.
00:10:12.639 --> 00:10:26.159
So in short, can we say that what we see in Chat GPT or Claude and Gemini is just like a probability machine of words being put next to each other based on the highest probability?
00:10:26.399 --> 00:10:26.960
That's it?
00:10:27.120 --> 00:10:27.840
Yes, exactly.
00:10:28.159 --> 00:10:33.039
The one thing that I would add is that word means a lot of different things.
00:10:33.200 --> 00:10:38.799
And as speakers of languages, we think of words as, well, cat means the cute furry thing.
00:10:38.960 --> 00:10:41.679
There's one near Alex, there's also one near me right now who might come on screen.
00:10:42.720 --> 00:10:50.559
Actually, one is sitting on my notebook, and so effectively uh trying to constrain the cat wildness here.
00:10:50.720 --> 00:10:52.080
Anyway, sorry, I digress.
00:10:52.480 --> 00:10:56.799
I only have one near me right now, but she's like rolling around on my desk and might knock something off.
00:10:56.960 --> 00:10:57.679
So we'll see.
00:10:57.919 --> 00:11:07.120
So anyway, all that to say that cat, or you know, if you speak other languages, so French, Shah, Spanish, gato, Portuguese is gato.
00:11:07.360 --> 00:11:08.399
Gato also, okay.
00:11:08.559 --> 00:11:08.720
Yes.
00:11:09.039 --> 00:11:12.720
That is not just those sounds or those letters, but also it means something.
00:11:12.799 --> 00:11:18.399
And it means something that we can sort of abstract across all the cats, but also it can refer to specific cats.
00:11:18.639 --> 00:11:26.960
And what's being manipulated by ChatGPT and Gemini and Claude is just the form part of that, just the spelling.
00:11:27.279 --> 00:11:34.240
So they are models of which bits of spellings of words are likely to go where in enormous collections of text.
00:11:34.639 --> 00:11:34.960
Okay.
00:11:35.120 --> 00:11:35.360
Okay.
00:11:35.519 --> 00:11:49.039
So here what we're saying then is that, and this is me trying to understand and digest everything that you're saying, is that they they put little letters next to each other, but not a word because they don't understand what they're creating.
00:11:49.200 --> 00:11:51.200
They're just putting things there.
00:11:51.440 --> 00:11:54.000
Things are being cut, but they don't know what the things are.
00:11:54.399 --> 00:11:55.279
Some or less like this.
00:11:55.360 --> 00:11:56.879
I'm doing it a really simple way.
00:11:57.919 --> 00:12:02.399
The reason for breaking it down to letters smaller than words is actually there's two reasons.
00:12:02.639 --> 00:12:09.600
The first is it makes it very clear, I think, that we're talking about just the shape and not what it means.
00:12:10.000 --> 00:12:26.480
And the second thing is that actually, literally, technically, today's large language models don't process things word by word, but there's a process called tokenization where they go with sequences of letters grouped together that don't necessarily add up to a whole word, in some cases, might be a bunch of words next to each other.
00:12:26.799 --> 00:12:29.200
So it's not actually respecting the word boundaries.
00:12:29.519 --> 00:12:39.440
When people think, you know, they're engaging with Tatipiti and Gemini and Claude, and they really sound like they know what they're talking about.
00:12:39.519 --> 00:12:42.799
They're really good at making people believe that.
00:12:43.039 --> 00:12:54.240
So what you're saying here is that no, they don't understand what they're saying, but they're really good at making you believe that they know it knows what it says.
00:12:54.720 --> 00:12:55.919
Is that correct?
00:12:56.320 --> 00:12:56.639
Yes.
00:12:56.879 --> 00:13:02.320
And there's a few things that go into that effectiveness at convincing us.
00:13:02.559 --> 00:13:05.279
The first is that it plays on how we process language.
00:13:05.440 --> 00:13:10.639
Like all good magic tricks, it basically turns on an illusion that we are ready to see.
00:13:10.879 --> 00:13:11.840
And
this is really cool.
00:13:11.919 --> 00:13:16.000
I like I get super excited about this because it's like what makes linguistics so much fun.
00:13:16.240 --> 00:13:28.399
But you might think that when you understand language, something someone said or that they wrote, that what happened is they had an idea, they put it into words, you got a hold of the words, and you're just unpacking the idea from there.
00:13:28.639 --> 00:13:29.679
That's not what's happening.
00:13:29.840 --> 00:13:31.519
What's happening is way cooler.
00:13:31.759 --> 00:13:36.960
So when we understand, we keep in mind everything that we know or believe about the speaker.
00:13:37.120 --> 00:13:40.159
So, you know, what are they thinking about?
00:13:40.320 --> 00:13:43.279
What do they believe they have in common with us?
00:13:43.519 --> 00:13:48.080
And maybe not with us, but what do they believe they have in common with the person they think they're talking to?
00:13:48.320 --> 00:13:58.879
And we hold all of that in mind, and then we use the specific words that they said as a very rich clue to what they were trying to convey, which is really cool.
00:13:59.120 --> 00:14:00.799
Unfortunately, we can't turn it off.
00:14:01.039 --> 00:14:09.919
So when we see the output of Claude Gemini Chat GPT, we make sense of it by imagining a mind behind the text that isn't there.
00:14:10.000 --> 00:14:11.679
It's all us projecting it.
00:14:11.919 --> 00:14:13.679
So that's the fundamental illusion.
00:14:13.919 --> 00:14:24.960
And instead of protecting against that, which would have been the safer choice for creating this technology if it was going to be created at all, the tech companies lean into it hard, right?
00:14:25.039 --> 00:14:28.320
So they design this interactive chat interface.
00:14:28.399 --> 00:14:31.600
So you're not just reading something that was like an article somebody wrote.
00:14:31.759 --> 00:14:33.360
It feels like you're interacting.
00:14:33.600 --> 00:14:39.440
They had the systems use I and me pronouns when there's no I inside of there.
00:14:39.679 --> 00:14:44.639
They talk about the systems as thinking, reasoning, answering all of these things.
00:14:44.799 --> 00:14:49.840
And they give you that little dot dot dot while it's in quotes thinking, like it would be a person typing.
00:14:49.919 --> 00:14:57.600
So all of those are additional design decisions that sort of raise the illusion and make it even more compelling.
00:14:57.759 --> 00:15:11.360
And then finally, they have gone out of their way to collect training data that make these things feel conversational, that make the output feel pleasant, make you feel good because you've had your ideas validated and so on.
00:15:11.519 --> 00:15:18.960
And so all of these things work together to increase this illusion and make us want to believe that there is actually a thinking mind in there.
00:15:19.279 --> 00:15:20.240
Really interesting.
00:15:20.320 --> 00:15:32.559
And Alex, could you tell us what are the dangers there of us engaging with these platforms, thinking that there is a little human behind them, while in actuality is just a bunch of things being written there?
00:15:33.519 --> 00:15:35.919
Yeah, I mean, there's a lot of dangers in this.
00:15:36.320 --> 00:15:41.600
One of the things I was on the bubble of thinking about bringing up was one of the technical innovations.
00:15:41.759 --> 00:15:43.840
This is kind of adding to what Emily was saying.
00:15:44.080 --> 00:15:53.360
One of the technical innovations that was developed in the mid 2010s was the resurgence of a type of statistical model called the neural network.
00:15:53.519 --> 00:16:01.039
And so effectively the idea is that there are these nodes within a neural network that represent a neuron.
00:16:01.279 --> 00:16:09.600
They take some input, the input applies a transformation of some sort of a very simplistic function, and then there is an output.
00:16:09.679 --> 00:16:16.240
And then once you chain these together, they become adept at saying learning patterns.
00:16:16.399 --> 00:16:36.320
And so this is kind of coming from a theory of how the brain works that's effectively thinking about instead of the brain operating like a library, it's more like the brain is kind of like a map, and if you go down a certain pathway, you reach some kind of pattern recognition or some kind of an object recognition.
00:16:36.559 --> 00:16:44.639
And so the earliest version of the neural network is developed by this guy, Frank Rosenblatt, and it was called the Perceptron.
00:16:44.799 --> 00:17:07.200
And then this idea fell by the wayside because a bit of the politics around AI research, but also some technical issues with it, and then found a resurgence, specifically due to research of a few different people discovering a way in which to assign numbers between the different neurons, one of them being Jeff Hinton.
00:17:07.279 --> 00:17:15.599
So that's just like a little bit more about something else that plays into the anthropomorphization, which is the name neural network, right?
00:17:15.680 --> 00:17:17.920
It sounds brain-like, you know.
00:17:18.240 --> 00:17:24.240
We probably would be having a different conversation if we still called them Boltzmann machines.
00:17:24.799 --> 00:17:26.559
Instead, we called them neural networks.
00:17:26.720 --> 00:17:26.880
Okay.
00:17:27.039 --> 00:17:28.640
So what are the dangers in this?
00:17:28.880 --> 00:17:31.359
I mean, the dangers are profligate.
00:17:31.519 --> 00:17:38.559
They are um, there's the kind of danger in anthropomorphizing, which I think Emily has written quite a bit on.
00:17:39.039 --> 00:17:53.279
And one of them is the notion that these technologies can be quote unquote unbiased or even de-biased of particular types of human biases around things like race, gender, language, nationality, etc.
00:17:53.759 --> 00:18:10.240
There's also the accountability dodge, the idea that this is a technology that is
acting of its own accord, and there aren't humans there who are directing it to do things or who are deploying them in particular contexts.
00:18:10.480 --> 00:18:20.160
And there is the persistent myth that we can't explain why this machine has outputted X, Y, and Z.
00:18:20.319 --> 00:18:35.200
And on a strictly technical level, yes, we can't explain why a certain kind of probabilistic model might be outputting this because it's too large and is not deterministic and is not programmed in such a way.
00:18:35.440 --> 00:18:58.079
But that is also a very narrow type of uh idea of explanation, one in which we must know the internals of such a thing without actually understanding, you know, and we don't have access to the massive training data sets that they've been trained on that we can analyze and the reinforcement learning with human feedback data.
00:18:58.240 --> 00:19:12.240
This is the data that is fed in and then used to improve the outputs so they are producing more desirable outputs of what the creators of the LLN wants the outputs to look like.
00:19:12.880 --> 00:19:31.680
We do have access to those, but those are not, you know, those are not released, those are not released in a way that would allow people to audit the patterns within them and to see what types of outputs are replicated from these by making probabilistic next word prediction from the LLM.
00:19:32.079 --> 00:19:40.559
So we have many, many, many problems that come from this anthropoformization of this tool, LLMs, and generative AI in general.
00:19:40.720 --> 00:19:46.960
But at the end of the day, now we know that AI agents are kind of popping by in the media, in the news.
00:19:47.039 --> 00:19:55.359
And somehow there's this thought an AI agent is thinking and doing things because an AI agent, again, is an agent, so humanizing it again.
00:19:55.519 --> 00:19:58.240
Could you explain to us what an AI agent is?
00:19:58.400 --> 00:20:00.960
How does this relate to generative AI LLMs?
00:20:01.119 --> 00:20:04.160
Is this like a little person that is doing things?
00:20:04.640 --> 00:20:05.920
What is happening there?
00:20:06.240 --> 00:20:06.559
Yeah.
00:20:06.799 --> 00:20:15.680
So AI agent very briefly means someone took the random text generator and hooked it up to other software that can take action in the world.
00:20:16.000 --> 00:20:20.000
So even more briefly, AI agent means bad idea.
00:20:21.599 --> 00:20:26.480
Um, so just to line these all up, a language model is a model of which bits of words go where.
00:20:26.640 --> 00:20:28.880
A large language model is a big one of those.
00:20:29.039 --> 00:20:45.519
But today, when people say LLM or large language model, they usually mean a big model of which goes words go where designed to output text that looks plausible looking, and then sometimes also bolted together with various other systems.
00:20:45.680 --> 00:21:05.039
And then the AI agent idea is this idea that because you can get systems to output
plausible-looking text, including also computer code, including actions you might click on a web page, then you could get them to make reservations by clicking around on a restaurant or hotel or airline website.
00:21:05.200 --> 00:21:14.400
You could get them to clean out your email inbox by classifying the text that's there and then deleting some of it, say.
00:21:14.640 --> 00:21:22.400
You could get them to do your stock trades for you if you gave these systems access to your various bank accounts.
00:21:22.559 --> 00:21:23.920
Like, bad idea, right?
00:21:24.079 --> 00:21:34.880
At no point do you want to let the random text generator take action on your behalf in the world if you care about those actions in the world, and you should care about those actions in the world.
00:21:35.279 --> 00:21:47.440
One of the things that I think is we say in the book that I think is helpful is we have something that's unique about language is as Emily put it, you know, we are attributing this agency to it.
00:21:47.599 --> 00:22:03.519
But when these technologies are outputting images, we are not suggesting that the crafting of the image is associated with a notion of smartness or an intelligence or any kind of creativity.
00:22:03.680 --> 00:22:07.599
I mean we see a lot of those elements and it looks absolutely terrible.
00:22:07.680 --> 00:22:19.200
You know, we have AI slop is you know a term that was been uh well ensconced into the uh into into language thanks to Merriam Webster herself.
00:22:19.440 --> 00:22:32.720
And so there is a notion that this recombination is happening and it is correct by chance, but there is no guarantee that it's going to be correct.
00:22:33.200 --> 00:22:35.039
Really important for people using generativity.
00:22:35.279 --> 00:22:38.240
I hope everyone understands that it's not because it sounds smart, it is smart.
00:22:38.319 --> 00:22:40.400
So stop doing therapy with ChatGPT everyone.
00:22:40.480 --> 00:22:48.160
This is my for anyone anyone listening to this, I think if you're not understanding what we're talking about the tech specs, stop using this as a therapist or a lover.
00:22:48.319 --> 00:22:51.599
This technology has no idea what it's saying.
00:22:51.759 --> 00:22:56.240
That's just a bunch of little things together that absolutely make no sense.
00:22:56.319 --> 00:22:59.599
Even though it makes sense to you it doesn't make sense to the technology creating it.
00:22:59.759 --> 00:23:00.559
Two quick things there.
00:23:00.720 --> 00:23:03.440
If the output makes sense it's because we are making sense of it.
00:23:03.599 --> 00:23:11.759
Apparently a lot of people like to talk to ChatGPT about personal intimate things because it feels like a safe space where you are not being judged.
00:23:11.839 --> 00:23:13.279
It feels very private.
00:23:13.599 --> 00:23:26.720
ChatGPT is not a person, it's not a thinking entity, it is a product of OpenAI and everything you type into it, even if you say don't remember this, is being stored by OpenAI.
00:23:26.799 --> 00:23:36.559
And at least for users in the US, at least for a while I think this is still true, open AI has been required to retain all records because of ongoing court battles.
00:23:36.799 --> 00:23:42.079
And so you are not having this private little conversation with something completely ephemeral.
00:23:42.240 --> 00:23:48.160
You are giving your deepest, most intimate thoughts to an enormous company that absolutely does not care about you.
00:23:48.640 --> 00:24:12.880
I would also add to this in terms of this is me also with my kind of social science hat on is thinking about what what is leading people to want to use these tools in a in a way that in which they're seeking mental health support or are trying to use these tools in lieu of some kind of a larger social network.
00:24:13.119 --> 00:24:22.400
We have to ask well what is what is there a gap in you know what are people not being supported in that they're turning to these technologies?
00:24:22.640 --> 00:24:34.880
What kinds of social systems are not in place where people are turning to LLMs for therapy or feel like they're judged if they're not able to engage in these kind of conversations with trusted friends.
00:24:35.039 --> 00:24:41.440
And what kind of other institutions have been undercut that haven't allowed people to make those kinds of connections?
00:24:41.680 --> 00:24:53.519
We understand that there are needs here that people are not having met and what are kinds of worlds that we can fight for that that would make that the case.
00:24:53.839 --> 00:25:08.079
I love what you're saying because at the same time that we're talking as well as the needs that are not being met, on the other side we also have an industry that's really big and powerful and that has a lot of money to try to sell their products.
00:25:08.240 --> 00:25:16.480
So we have people who have their needs unmet and then we have billionaires trying to sell them products to fix this issue.
00:25:16.640 --> 00:25:36.799
So put it all together we understand that LLMs large language models we understand that ChatGPT, Gemini, Claude, they don't have feelings they don't understand what they're saying, but they're marketed as understanding as you said they are created in a way to be kind and nice and gentle and something that you just always want to go back to.
00:25:36.960 --> 00:25:46.079
And at the same time when we see the media when we go back to the hacking of you know AI agents that's also how the media talks about these tools right hacking.
00:25:46.240 --> 00:25:49.119
An AI agent hacked another system.
00:25:49.279 --> 00:25:59.839
So it seems to me that everything is together in the ways that we're talking about these tools in the ways that the media is talking about these tools and in the ways that these AI CEOs are talking about these tools.
00:26:00.000 --> 00:26:01.519
So what is the common thread here?
00:26:01.680 --> 00:26:06.720
What
are they creating in the ways that we discuss these technologies and the technologies themselves?
00:26:07.119 --> 00:26:14.480
So this is something I've actually been studying and there's a word that's uh no fun to say but it gets easier with practice which is anthropomorphization.
00:26:14.880 --> 00:26:28.400
So hard I always try and I always eat my tongue saying that it's also hard to spell not because there's anything tricky it's just there's so many letters but anthropomorphization is basically the process of making something seem human when it isn't.
00:26:28.559 --> 00:26:30.640
And that can be done in the design features.
00:26:30.720 --> 00:26:46.640
So like we were talking before the fact that these systems are designed to output I and me and the equivalents in other languages that is a kind of anthropomorphization by design but we also see a lot of anthroporphization by description which means the way we talk about the systems.
00:26:46.880 --> 00:26:53.440
One of the very key things that I think links together everything you just said is who are you saying did the thing?
00:26:53.839 --> 00:26:54.079
Right.
00:26:54.240 --> 00:27:04.079
So every time we say, you know, a swarm of AI agents from open AI hacked into hugging face, we are not saying what open AI did, right?
00:27:04.400 --> 00:27:05.920
OpenAI did a cybercrime.
00:27:06.400 --> 00:27:08.720
OpenAI hacked another company, right?
00:27:09.119 --> 00:27:15.920
And when we talk about it as if the models did this of their own accord, we are saying something that's not true, right?
00:27:16.000 --> 00:27:30.640
They were software systems or it was a multiple copies of the same software system designed to do a specific thing, run by OpenAI and run without the kinds of safeguards that you should put if you are testing computer security systems.
00:27:30.799 --> 00:27:32.559
And so it did the thing, right?
00:27:32.799 --> 00:27:37.119
It did what it was designed to do when it was run and invoked to do that.
00:27:37.519 --> 00:27:48.640
So we are sort of hiding who's responsible and building up this idea of AI as a thing in the world that is smart and getting smarter.
00:27:48.880 --> 00:27:54.400
And the problem is that the anthropomorphizing terminology is just everywhere in this discourse.
00:27:54.559 --> 00:28:01.359
So when you talk about these systems as learning from their training data, learning and training are anthropomorphizing.
00:28:01.519 --> 00:28:08.960
When you talk about the systems as well I asked ChatGPT and it told me both the ask and the tell are anthropomorphizing.
00:28:09.039 --> 00:28:12.480
And it takes a lot of effort to find other ways to speak about it.
00:28:12.640 --> 00:28:16.960
And I have some suggestions together with my colleague Nana Ini, I can give you the link for that.
00:28:17.119 --> 00:28:22.000
But I think mainly what I would love your listeners to do is to just get in the habit of noticing this.
00:28:22.240 --> 00:28:35.200
Every time a news report or a tech CEO or a friend says that one of these models did something, think, how would I rephrase that to make it clear who did it and what they were using it to do.
00:28:35.359 --> 00:28:46.400
I I love practical tips and I for sure would want the resource I follow you on LinkedIn and if you don't follow both of them on LinkedIn Emily always shares words that we can use instead of anthropomorphizing AI.
00:28:46.559 --> 00:28:46.960
Love it.
00:28:47.039 --> 00:28:56.720
I recommend that for everyone so you're saying two things you said the first one is that they do this to hide responsibility so we can see a clear benefit there.
00:28:56.880 --> 00:29:01.200
Of course if OpenAI is doing illegal things and they blame it on the chat bot.
00:29:01.440 --> 00:29:02.799
Same thing with Grok, right?
00:29:02.880 --> 00:29:11.759
Grok was used to undress women, children and babies in Twitter X today, who apologized was Grok in its own ex account.
00:29:11.920 --> 00:29:18.559
So instead of saying okay X created a technology that undressed children and that's criminal, you blame Grok.
00:29:18.960 --> 00:29:20.319
Oh it was Grok who did this.
00:29:20.400 --> 00:29:22.720
It's more or less what you're saying, but with a different example.
00:29:22.960 --> 00:29:33.599
So okay one, we hide responsibility but you also said that this is a way to get people to believe that these technologies are getting smarter and getting brighter and you know increasing.
00:29:33.759 --> 00:29:54.799
Now I would love to understand who benefits from this and there's two concepts that I would love to introduce to this conversation which is the AI boosters and the AI Doomers and who are they and how do they benefit from this type of conversation of thinking about artificial intelligence in the way that it's human and that it's going to be super smart.
00:29:55.359 --> 00:30:02.640
One of the things that we get to in our book is that the AI boosters and the AI Doomers are actually singing the same song.
00:30:02.799 --> 00:30:03.920
They're telling the same story.
00:30:04.079 --> 00:30:13.200
They are often presented in the media as these like polar opposites the far ends of the spectrum of possible opinions about artificial intelligence and everyone else has to line up somewhere in between.
00:30:13.440 --> 00:30:17.519
In fact they are two sides of the same coin there's no daylight in between them.
00:30:17.759 --> 00:30:28.720
The boosters say AI is a thing it's inevitable it's imminent and it's going to solve all of our problems and the doomers say AI is a thing it's inevitable it's imminent and ah it's going to kill us all.
00:30:28.960 --> 00:30:38.559
And these are the same story and they are a story of an immensely powerful technology that, you know, does all these good or bad things on its own.
00:30:38.640 --> 00:30:42.079
So it's not something that anybody has to be accountable for.
00:30:42.319 --> 00:30:56.640
And it also is something that
is so powerful that you would be a fool not to invest in it, not to try to use it, not to try to get on board yeah I mean Emily said most of what the argument is here.
00:30:56.799 --> 00:31:20.480
And I mean it's helpful to think about too about how this is a play for other types of accountability dodges because Doomerism then becomes a play for regulation that is not really brass tax types of regulations that we need focusing on pretty basic types of protections around you know basic technological harms.
00:31:20.559 --> 00:31:22.400
I mean we don't call we shouldn't even call it AI.
00:31:22.559 --> 00:31:29.039
I mean it's any kind of automated decision making systems that have immense types of impacts on people.
00:31:29.359 --> 00:31:48.240
And so you know one of the things that we see quite a bit of is that there's many people that you would characterize as you know possible boosters who are really focused on well we need regulation but that regulation is going to be completely different from anything we have on the books right now.
00:31:48.480 --> 00:31:57.920
And an Isra focus on kind of fields that we would call with a doomer hat being critical harm or or just like major types of damages.
00:31:58.079 --> 00:32:15.119
And so we saw this in California where there was a bill called SP 1047 and that was being debated one or two sessions ago and eventually it passed the California Assembly and Senate and then Governor Newsom vetoed it.
00:32:15.279 --> 00:32:29.680
But it was focused it it got so much attention took all the air out of the room for any other type of discussion of any kind of stronger legislation and regulation around harms of these technologies.
00:32:30.000 --> 00:33:06.079
Meanwhile legislation like the EU AI Act has exemptions for the use of these technologies at the border or in policing or all these places that are going to disproportionately first affect people of color and people who are non-EU residents including people who are you know refugees or trying to claim asylum and then it's also going to disproportionally affect the working poor and people who really have no kind of redress or have much harder kind of ability to redress these types of harms.
00:33:06.240 --> 00:33:19.839
And so it's it's an accountability dodge at the same time and it's it's also offering this function now the doomerism especially within the last two or three weeks of man we have no path to monetization.
00:33:20.240 --> 00:33:34.799
We really are in deep shit but look this is really dangerous so like you know past this this really fake type of regulation but at the same time our things are super powerful and please give us money.
00:33:35.039 --> 00:33:39.440
I saw so much about that in terms of they are literally selling smoke.
00:33:39.599 --> 00:33:52.160
They have nowhere to go and they just need a pause in their side because if they open their books and show us how deep this goes they are so they are saying hey this is really dangerous so I'm not going to IPO this year.
00:33:52.240 --> 00:33:57.359
Give me more time so I can figure out where to get the money from and what do I have to do to make this happen.
00:33:57.519 --> 00:34:04.160
But no pay attention this is so dangerous everyone please please yeah we cannot build this but please I need the money to build this but safer.
00:34:04.319 --> 00:34:12.639
And then I would like to add this too because many times when I'm having these conversations and trying to explain people that this is a money agenda at the end of the day.
00:34:12.719 --> 00:34:26.079
So even the boosters and the doomers both of them they benefit because then the doomers they say hey this can be destroy the world and I am the only person that can develop this in a safe way and the boosters are just like hey this is wonderful.
00:34:26.159 --> 00:34:27.920
We won't have poverty anymore.
00:34:28.079 --> 00:34:42.480
But then you have this group some AI researchers that are low-key the the cusp of the cusp of Doomers and they apparently they don't benefit you know financially for saying the things that they say at least that's what people always tell me.
00:34:42.639 --> 00:35:07.760
But why are these AI researchers saying that AI will destroy the world, that superintelligence will kill us all and they are not in a tech company they don't benefit from this narrative what do you have to say for the people who believe that there's a so there's a few things happening here and you know I think one of the things it's important is that you know is that there's a variety of people so I know Emily will have more to say about this too.
00:35:07.920 --> 00:35:53.360
But one of the things is people do benefit from this I was having a conversation actually with another journalist this morning and one of the things that we were talking about is how there is this you know there is this AI safety industry and the AI safety industry is it's not just industrial it's also a lot of nonprofits there's these institutes there's so many different institutes that are very hard to keep track of yes you know in you know in the US actually down the street for me I mean relative figure T down the street uh in Berkeley there is the Center for AI safety that gets huge amounts of money from some of these people that hold these particular you know ideologies.
00:35:53.440 --> 00:35:54.960
I'll let Emily talk about that.
00:35:55.440 --> 00:35:56.639
But there's a lot of money in it.
00:35:56.719 --> 00:36:48.239
I mean like Iliaskever after he beefed with Sam Altman ended up founding a startup called Safe Superintelligence that has no product but he got two billion dollars upon his establishment and this this keeps on happening I mean even someone like Jacob Coxon who you know was this whistleblower self-styled whistleblower I don't think he's actually a whistleblower well but but that's the thing that's the thing I want to complicate Emily is like this I this term
whistleblower right is he's on all these TV shows you know he's on Fox News he's on CBS and let me tell you like that money doesn't to get it to get on those people's radar it doesn't come from nowhere like it's a very intentional sort of press push you know in the way that like any kind of comp strategy is a very intentional calm strategy.
00:36:48.480 --> 00:37:09.679
There is money to be made in this now let's not reduce all of this to money because there are many people who are true believers who like believe that literally to quote the title of Eliza Yukowski and Soares's book like if you build it it will kill us all.
00:37:11.280 --> 00:37:12.960
Actually you misquoted their title but that's good.
00:37:13.039 --> 00:37:17.679
We don't need to give the exact title I don't want to deign to even look at the title whatever you get the idea.
00:37:17.920 --> 00:37:30.239
I'll kick it to Emily I think it's it's right that that all of this is coming from a combination of financial incentives and for some not necessarily all of the people involved sincere belief.
00:37:30.400 --> 00:37:39.280
So there is this very attractive ideology and it seems to attract in particular people who have not experienced much marginalization in their lives.
00:37:39.440 --> 00:37:43.360
So I'm thinking about when Jeffrey Hinton quit Google after Dr.
00:37:43.519 --> 00:37:54.960
Tamit Gabrew was fired, ostensibly over our paper, but for more details you should see her forthcoming book called Deep Unlearning and that is the real title and it's going to be amazing and can be pre-ordered now.
00:37:55.039 --> 00:37:57.360
So you don't have to buy our book but do buy her book.
00:37:57.599 --> 00:37:58.639
Buy both books.
00:37:58.800 --> 00:37:59.199
Yes.
00:37:59.440 --> 00:38:00.400
So Dr.
00:38:00.559 --> 00:38:13.360
Gabrew gets fired from Google and then a few months later Hinton quits Google and takes his fainting couch on a media tour to talk about how terrible this technology is and how it's going to kill us all.
00:38:13.679 --> 00:38:19.360
And somebody got Jake Tapper at CNN to ask him, well, what about the concerns that Dr.
00:38:19.440 --> 00:38:20.639
Gabrew was raising?
00:38:20.800 --> 00:38:26.000
And he literally said those concerns are less existential than his, right?
00:38:26.239 --> 00:38:40.400
And this is an ideology that's basically allowing people to be the superheroes because there's this supervillain that is going to destroy all of humanity and they are the select few who have the knowledge and the skills to do something about it.
00:38:40.559 --> 00:38:42.000
And it's gross, right?
00:38:42.079 --> 00:38:43.119
It's gross in many ways.
00:38:43.280 --> 00:38:52.559
It's gross partially because it completely ignores all of the existential threats to actual people now and we detail a bunch of those in our book.
00:38:52.719 --> 00:38:57.519
And it's also gross because it's built on this ideology of eugenics and transhumanism.
00:38:57.599 -->
00:39:02.960
It's built on notions of IQ, which itself has to do with eugenics and race science.
00:39:03.039 --> 00:39:14.079
And this is this is the idea that you can like line up all people according to one measurement of how in quotes intelligent they are and now we can put machines on that same scale.
00:39:14.559 --> 00:39:17.440
So the whole thing is intensely dehumanizing.
00:39:17.599 --> 00:39:20.239
And when you get into the ideologies so um Dr.
00:39:20.320 --> 00:39:28.559
Gebrew and her co-author Emile Torres elaborate this in a paper called the Testrial bundle of ideologies really important work.
00:39:28.719 --> 00:39:34.079
And like you know get some ice cream or something when you're done reading that paper because it's it's sad.
00:39:34.400 --> 00:39:40.639
But um another place to get a similar recounting is Adam Becker's book More Everything Forever.
00:39:40.880 --> 00:39:51.119
So it's not fun to go spend time in these people's world but some of them are sincere and they they are really just completely wedded to this fantasy.
00:39:51.280 --> 00:40:02.960
And then one sort of further note in all this is that they consider themselves rationalist that is part of the ideology is that they and only they are the ones who are making decisions based just on objective evidence.
00:40:03.119 --> 00:40:19.920
And so if somebody is talking about updating their priors, that is a flag that they are in this mindset, that they are doing this probabilistic reasoning about the world and they have these priors so the the probabilities they're assigning to things based on I don't know what numbers they've pulled out of their ass.
00:40:20.079 --> 00:40:21.679
PDOOM is one of those.
00:40:22.000 --> 00:40:28.320
And they claim that if they encounter relevant evidence, they will update those priors and they never do.
00:40:28.800 --> 00:40:39.679
And a few different elements of that as well in addition to what you know like agreeing with what Emily said when people say existential risk, I mean the first thing you should ask is like existential risk for whom?
00:40:39.840 --> 00:41:04.480
Because what they usually mean is existential risk for white people and it's people that already haven't been disproportionately subjected to racialized and gendered violence, border violence, violence where you know like you are going to be more more likely to encounter an algorithmic decision typically without really even knowing that you have.
00:41:04.719 --> 00:41:16.480
And so existential risk I I think it just it is existential risk for white people and I mean this is a point we can make in a book and I think I just kind of need to hammer and hammer it over and over again.
00:41:16.719 --> 00:41:18.639
Because that's what it that's what it means.
00:41:18.880 --> 00:41:34.719
And you know like to ignore the body and the whole history of technological subordination that so many people have been subject to for decades and decades already is to just be honestly willfully ignorant.
00:41:34.880 --> 00:41:44.159
And so when Hinton says you know well this is for everybody he he means like well this might mean we're treated like black and brown people that would be terrible.
00:41:45.199 --> 00:42:01.920
A hundred percent I was even talking to my mom and my mom clocked that in one second and see my mom is a doctor she takes care of children and I was telling her about all the AI things that are happening and I told her about you know AI agents and how tech bros and Silicon Valley are so worried about superintelligence.
00:42:02.000 --> 00:42:13.599
And she told me, wow, so for the first time ever they are afraid of not being top of the food chain because white men have always been the leaders and controlling us and hurting us.
00:42:13.760 --> 00:42:21.119
And then suddenly there there could be this potential threat that suddenly would even control them.
00:42:21.280 --> 00:42:23.280
And that is so scary, so scary.
00:42:23.360 --> 00:42:33.519
But for historically marginalized communities and people who have always been suffering, we have other more important issues to be caring about and real issues like real existential things are real, yeah.
00:42:33.679 --> 00:42:44.960
That are real and not based on as Dember says in his brilliant book, based on sci-fi and you know a bunch of comic books that maybe a few people read too much and thought that that was reality.
00:42:45.119 --> 00:42:47.440
Here I also want to connect what we were saying before.
00:42:47.599 --> 00:42:56.880
So we have the LLMs and then we have the AI agents and now we have everyone talking about PDOM or you know how these technologies are going to destroy us.
00:42:57.039 --> 00:43:04.719
But what it seems to be happening is that they think that the technology that will destroy us is based on LLMs and generative AI.
00:43:04.880 --> 00:43:11.280
We have Anthropic and we have open AI and they're creating these tools but these tools are generative AI.
00:43:11.519 --> 00:43:18.320
Is there a possibility that generative AI kill us all and destroy us all and the things that we're reading in the media and the fear that exists?
00:43:18.559 --> 00:43:19.280
Is this real?
00:43:19.440 --> 00:43:22.000
What is the science behind this and why do they believe this?
00:43:22.079 --> 00:43:36.559
Because again I just want to add this I was in a conference last week and I heard it from Jeff Hinton that he thinks that these models are sanctioned and that they already have a consciousness and they are little aliens that we have to pay attention to so is there a possibility like what is happening here?
00:43:36.639 --> 00:43:41.360
Please help so there's so I mean it's it's it's a ridiculous conversation.
00:43:41.440 --> 00:43:51.199
It's not based in science people defer to Hinton because you know it was this person that revived this line of neural network research.
00:43:51.360 --> 00:43:56.400
And I mean honestly I mean Jeff Hinton's
superpower is that he was very persistent.
00:43:56.559 --> 00:44:03.840
I was reading kind of a history of neural networks From the 90s, and the man just never quit on it.
00:44:04.000 --> 00:44:06.880
And and honestly, like, you know, respect.
00:44:07.119 --> 00:44:15.039
I too would like to be persistent on something like that for, you know, 30 years, and and possibly it pays off.
00:44:15.280 --> 00:44:16.159
That's his expertise.
00:44:16.239 --> 00:44:17.360
He's a computer scientist.
00:44:17.440 --> 00:44:24.400
You know, he's not a not somebody that is that's understanding social systems, that's understanding society in any kind of deep way.
00:44:24.559 --> 00:44:31.599
Um, this is someone that has been kind of holed up in in in AI research for you know 40 plus years.
00:44:31.760 --> 00:44:34.480
And his metaphor also keeps on changing a lot.
00:44:34.559 --> 00:44:38.800
I mean, it's aliens now, but he said that we needed to be designing like maternal AI.
00:44:38.880 --> 00:44:42.239
So I don't know if he wants like motherly aliens.
00:44:42.800 --> 00:44:48.239
But it's but I think the kind of thing is where is where's the evidence?
00:44:48.400 --> 00:44:49.440
Like there's not evidence.
00:44:49.599 --> 00:44:54.800
This is and this is a subhead we have in the book, which is extraordinary claims require extraordinary evidence.
00:44:54.960 --> 00:44:56.719
And what is the evidence?
00:44:56.880 --> 00:45:14.719
And the if the evidence is Jeff Hinton said so, or Max Tegmark said so, and Max Tegmark being one of the co-founders of the Future of Life Institute, which is another existential risk institute, and is one of the the organization that convened this pro-human gathering.
00:45:14.880 --> 00:45:17.280
But there's but it's there's no evidence.
00:45:17.440 --> 00:45:20.559
I mean, that's what Adam was arguing in his book.
00:45:20.800 --> 00:45:34.480
This is people taking this as an article of faith and are dressing it up in the language of science, of assigning numbers to it as the rationalists do.
00:45:34.800 --> 00:45:40.000
These numbers are wholly made up, they are not empirically grounded.
00:45:40.239 --> 00:45:42.320
I'm trained as a social scientist.
00:45:42.559 --> 00:45:52.239
If you're going to define a variable, that variable has to have a good justification of why it is a valid metric, and you need to justify it.
00:45:52.320 --> 00:45:54.800
And what are the alternative specifications?
00:45:55.119 --> 00:45:57.760
And none of them, none of them pass muster.
00:45:58.320 --> 00:45:58.559
Yeah.
00:45:58.719 --> 00:45:58.960
Yeah.
00:45:59.119 --> 00:46:10.400
I mean, I think I think Alex basically captured it, but there is it's very frustrating when some of these folks get held up as the godfathers of AI, and it's frustrating in many ways.
00:46:10.559 --> 00:46:13.599
First of all, you know, hello, great man theory of science, right?
00:46:13.760 --> 00:46:19.760
But also, why should AI be this kind of thing
that has godparents at all?
00:46:20.000 --> 00:46:20.320
Right.
00:46:20.559 --> 00:46:25.920
And also, why is their expertise considered the most relevant?
00:46:26.079 --> 00:46:26.320
Right.
00:46:26.559 --> 00:46:30.559
We get Hinton, you know, lifted up as an expert, and he does have expertise, right?
00:46:30.639 --> 00:46:35.039
He has been working with this particular approach to designing computer systems to do things.
00:46:35.119 --> 00:46:38.960
Originally it was computer vision, and then he's moved on to language for many, many years.
00:46:39.039 --> 00:46:40.880
So, yes, expertise in that.
00:46:41.119 --> 00:46:44.719
But that is not the same thing as expertise in social systems.
00:46:44.880 --> 00:46:49.679
As Alex was saying, it's not the same thing as expertise in psychology, in linguistics.
00:46:49.840 --> 00:47:21.760
He does have some training in, I forget if it was cognitive science or neuroscience early, early on, but that expertise of like how we build the systems is the expertise that is usually lifted up as such in the media and not the expertise of either other scholars who look at things, you know, like social systems, like linguistics, but also as Alex was saying, you know, earlier on, the very important expertise that people have in their own lives and what it is to deal with the bureaucracies they have to deal with, or to do the jobs that they do, or to do the art that they make and so on, right?
00:47:21.840 --> 00:47:25.360
And all of this is actually way more relevant expertise.
00:47:25.599 --> 00:47:36.719
And it should not be the people who are building the systems who, furthermore, aren't even being transparent about what they're building, who get to claim, you know, what the systems are capable of.
00:47:36.960 --> 00:47:45.119
Usually when they do that, what they're doing is they are minimizing, dehumanizing, flattening the thing that they are claiming to use a system to do.
00:47:45.360 --> 00:47:46.159
A hundred percent.
00:47:46.320 --> 00:47:51.760
And I think it's so important to have this conversation because I think it goes back to really the start of our conversation.
00:47:51.920 --> 00:48:08.559
Maybe we're saying, yes, we have to hear the people who have this empirical or the people who are experiencing these tools at the same time that we don't necessarily need to be hearing all the people who have flashy titles because they don't necessarily have the flashy titles on the specificities of the EI technologies that we're talking about.
00:48:08.639 --> 00:48:22.159
And we need more people in the room and not to only be listening to the people who benefit from this discourse and this conversation, that being because they're creating the products, that being because maybe they are benefiting from other ways that we don't know of.
00:48:22.400 --> 00:48:25.920
So super important to be critical, super important to hear different voices.
00:48:26.159 --> 00:48:28.719
Now, to conclude, I have two questions and then we're done.
00:48:28.880 --> 00:48:29.280
One.
00:48:29.599 --> 00:48:32.000
Why should people care about this discussion?
00:48:32.159 --> 00:48:32.639
Why?
00:48:32.880 --> 00:48:39.519
I think that there's a lot of people who say, I hate AI, I don't want to hear anything about this, I don't care, or I don't use AI.
00:48:39.679 --> 00:48:42.639
Why should everyday people care about this discussion that we had today?
00:48:42.960 --> 00:48:43.199
Yeah.
00:48:43.280 --> 00:48:44.079
So one.
00:48:44.400 --> 00:48:44.880
Go ahead.
00:48:45.039 --> 00:48:45.360
Oh.
00:48:46.079 --> 00:48:46.880
We're both ready.
00:48:47.199 --> 00:48:53.280
Well, I was gonna say one reason we should care about this discussion is that this is having a lot of global impact.
00:48:53.440 --> 00:48:59.199
I mean, you have something like 50 to 70 percent of the global economy tied up in this industry right now.
00:48:59.440 --> 00:49:00.880
It's very scary.
00:49:01.280 --> 00:49:03.599
This is a bubble, and the bubble's gonna burst.
00:49:03.760 --> 00:49:06.639
When the bubble bursts, it's gonna hurt everyone.
00:49:06.800 --> 00:49:27.679
It's gonna hurt that's the scare, and I've been saying this probably for two years now, where the thing that's gonna be very scary is that when things like credit markets, when those are undercut by the AI bubble bursting, those companies are first going to they're taking on huge amounts of debt.
00:49:27.840 --> 00:49:34.000
They're going to try to pass on that debt to the federal government, you know, federal governments.
00:49:34.159 --> 00:49:44.079
They're going to ask other types of institutions to step in, and then those organizations are going to try to manage that debt by passing it on to consumers.
00:49:44.320 --> 00:49:58.480
We had a conversation with uh Matt Schear, who's now the Open Markets Institute, and he's put out published a report effectively, which is an argument that we shouldn't bail out the AI companies when they go belly up.
00:49:58.639 --> 00:50:06.880
So effectively, what that's going to look like is we had the 2008 market crash from the subprime mortgage crisis.
00:50:07.199 --> 00:50:10.559
This is, by some estimates, orders of magnitude higher.
00:50:10.719 --> 00:50:14.400
I think I saw one that was like 26 times more than that.
00:50:14.639 --> 00:50:18.320
So that is very, very worrisome.
00:50:18.719 --> 00:50:20.559
And that's going to be a problem.
00:50:20.960 --> 00:50:35.039
And so I think that the other element of it is that if companies are saying this could kill you, this could do this, you need to, it needs to really, you need to pay attention to this, they're doing narrative work, right?
00:50:35.119 --> 00:50:42.960
They want to sell a certain kind of narrative that this is inevitable, that it is necessary to have, but we don't actually need this.
00:50:43.199 --> 00:50:51.280
So when it goes sideways, that is going to make people much more attentive to saying,
well, well, we need this industry.
00:50:51.360 --> 00:50:52.079
Well, we need to do it.
00:50:52.239 --> 00:50:56.480
The thing is, and the thing that's really optim I find optimistic is that people hate AI.
00:50:56.639 --> 00:51:02.239
You know, you the the surveys, at least in the US, I've seen, there's across the political spectrum.
00:51:02.480 --> 00:51:07.760
There is most people really don't like the technology or they don't use it at work.
00:51:07.920 --> 00:51:12.719
In data centers, people hate data centers, also across the political spectrum.
00:51:13.039 --> 00:51:15.840
And so that's really encouraging.
00:51:16.000 --> 00:51:19.920
That's why these companies see themselves in a narrative crisis.
00:51:20.159 --> 00:51:30.639
And so that does provide us a fantastic backbone to say when industry goes belly up, lawmakers don't bail out these companies.
00:51:30.800 --> 00:51:31.679
We don't need this.
00:51:31.840 --> 00:51:44.400
Let them fail and find a way to allow them to fail and don't pass on that debt and the type of interest rate hikes and the rise in price of consumer goods onto everyday people.
00:51:44.800 --> 00:51:45.360
Yeah.
00:51:45.679 --> 00:51:50.000
I want to say be an AI hater, keep doing the hating, that's fantastic.
00:51:50.159 --> 00:51:56.559
Like, and be a visible AI hater because too many people, I think, are still walking around going, well, but everyone seems to think it's useful.
00:51:56.639 --> 00:51:58.000
So am I weird for hating it?
00:51:58.079 --> 00:52:01.199
So if you can be loud, and that's fantastic.
00:52:01.360 --> 00:52:05.760
And you don't have to have endless conversations about whether or not it could possibly be sentient.
00:52:05.920 --> 00:52:06.239
It isn't.
00:52:06.400 --> 00:52:07.280
Fine, done.
00:52:07.440 --> 00:52:08.719
But there is work to be done.
00:52:08.880 --> 00:52:20.480
Both what Alex is talking about about preventing our governments from trying to help the people who cause this mess, but also, you know, much like now, like who knows how near that bubble crash is.
00:52:20.719 --> 00:52:23.119
Hopefully sooner rather than later, let's get it over with.
00:52:23.280 --> 00:52:27.920
But but right now, data centers are being built, and that has environmental impacts.
00:52:28.000 --> 00:52:37.599
And we are seeing, you know, basically losses in the still insufficient progress we've been making on the climate crisis because of this build-out for the data centers.
00:52:37.760 --> 00:52:43.679
And so channeling some of that AI hate towards protecting our local environments is really important.
00:52:43.840 --> 00:53:01.679
And I think we also need to be protecting things like education systems, because another way that these companies are trying to recoup their completely unsustainable spending practices is by shoving their software into schools where nations and smaller jurisdictions put a lot of resources, not as much as they should.
00:53:01.920 --> 00:53:05.599
Those resources should be going to teachers and not to the tech companies.
00:53:05.760 --> 00:53:08.400
And you can make very much the same argument about healthcare.
00:53:08.559 --> 00:53:17.039
And so there is work to be done in terms of protecting what we value from these incursions, which can absolutely be done from a place of AI hate and should be.
00:53:17.280 --> 00:53:19.199
So, everyone, let's be louder.
00:53:19.440 --> 00:53:24.559
Because I do also agree that there's many times people are shy about this, or you know, they're going to be called the Ludites.
00:53:24.639 --> 00:53:30.559
Honestly, in my book club this month, we're learning how the Ludites were awesome people that were fighting for their rights because they were starving, had no jobs.
00:53:30.719 --> 00:53:32.000
So Ludites is being a good thing.
00:53:32.079 --> 00:53:33.199
Yes, I am a Ludite.
00:53:33.280 --> 00:53:38.719
So what I want to say is that let's be Ludites, let's be proud Ludites, and let's be fighting against these.
00:53:38.880 --> 00:53:41.119
So thank you, thank you, thank you so much for joining.
00:53:41.199 --> 00:53:43.519
I could spend like 20 million hours here with you.
00:53:43.679 --> 00:53:49.360
If you have heard this podcast until the end, please comment with one thing that you have learned today.
00:53:49.519 --> 00:53:55.920
And also, if you stayed until the end, comment a brain emoji because I want to see who is part of the stay until the end gang.
00:53:56.000 --> 00:53:59.840
Because many times people leave in the middle, but I want to know who the real deals are.
00:54:00.000 --> 00:54:00.960
So thank you so much.
00:54:01.039 --> 00:54:04.239
I'll see you next Monday for a new episode of Bots Don't Cry.
00:54:04.320 --> 00:54:08.239
Don't forget to buy the AI con and don't forget to subscribe to their podcast as well.
00:54:08.400 --> 00:54:08.960
So thank you.
00:54:09.039 --> 00:54:10.400
Have a wonderful day.
00:54:10.719 --> 00:54:11.519
Thank you.
00:54:11.920 --> 00:54:13.840
Bots don't cry.