Season 1 Episode 1: It’s not intelligent, so don’t call it AI!

[Nick]
All right, greetings, this is our inaugural podcast here, Episode 0, Episode 1, a pilot, whatever you want to call it.

[Mike]
We’re your hosts. I’m Mike.

[Nick]
I’m Nick.

[Mike]
And we’re going to basically go into detail about, you know, the crazy world we live in and kind of attacking the myths of the day, as it were, trying to give you a better sense of what’s going on, how to think about stuff.

[Nick]
Yeah. How to be healthy, how to worry a little less and be more informed. You know, the thing we struggle with every day.

We do the struggle, so you don’t have to.

[Mike]
So this podcast is an evolution of sort of Nick and I, we just basically go to bars every week and just talk about the world as it is. And our discussions when our friends are around seem to be so informative that we decided that we should do a podcast about it, just like every other podcast, bro. But I swear this time it’s actually going to be informative.

We’re not a bunch of meatheads.

[Nick]
That’s right. And we’re not trying to self-aggrandize. This is for you.

This is not for us. Can we say that? I think we can.

[Mike]
So this inaugural episode, we’re going to talk about the topic on everyone’s lips, AI. I happen to be somebody who works in tech and Nick has a political economy background, so he thinks about labor all the time. And we’re going to kind of go into a bit of the, you know, myths around AI, a bit of sort of the mis-framings of AI, and maybe how to think about it in the future.

Of course, with every evolving technology, you know, facts on the ground are changing all the time, but to the best of our ability, we’re going to try and give you a sense of kind of what is going on and kind of what to expect going forward in the next, say, few years or whatever. Right?

[Nick]
Exactly. And it’s a very human reaction to begin with any new thing and say, this is exciting. This is new.

This is also the most, this is the worst thing that’s ever happened. It’s the best thing that’s ever happened. Factually, you dissect that over the near term.

It’s actually just a process that can be looked at rationally and through a historical lens. So let’s begin that process here today in the studio, if we can.

[Mike]
Yeah. So where do you want to start? Because obviously we have a, there’s so many things you can say about this technology, obviously.

It’s a very broad topic.

[Nick]
So I like this little anecdote. It’s a bit of an icebreaker. Very recently in Congress, our amazing and talented and brilliant, let me say that sarcastically, Secretary of Education was reading off the prompter, talking on the subject of AI.

And she started off by saying, we’re living in an era where technology and A1 is impacting all of us. And she did not, this was no, she did not realize what she was talking about, A1. So in this environment, we have sort of the blind leading us in Congress.

They should be regulating. They’re not. We have a class of techno moguls that we kind of at this point are recognizing are running a very exciting grift, part of a larger trend.

And they are selling us on the revolution of AI. The version, the vision of this that they are selling us is this is the next, let’s say, part of the fourth industrial revolution to use a old hat political economy term here. But as you pointed out to me just very recently, it is a forced revolution.

It is a pitch. It is being sold to us en masse because there is an agenda, not even conspiratorially, just quite naturally. The needs of capital are, as always, outweighing the needs of everybody else.

So what does that look like? Who are the guilty parties here trying to sell us this forced revolution, Michael?

[Mike]
There are a lot, but obviously, a lot of the politics and direction of AI are so wrapped up in the politics of Silicon Valley, right? So you obviously have the development of a technology that is decades in the making. I’ve written about the history of AI on my blog, Misaligned Markets, talking about sort of going back to the 60s, you have good old fashioned AI, as they called it, or symbolic and logical AI, rules-based AI.

And a lot of that history obviously is wrapped up in DARPA, the military, et cetera. But what is interesting is, and I get into this in my writing, is that some of this is incidental in the sense that, you know, when you have, AI as a field is predominantly academic driven, or it has been. And so it has been a field that’s very slow to have applications in industry, right?

And a lot of the ideas from this field, some of these are decades old, like neural nets, for example, did not have the resources or the money behind them to actually experiment with them at scale. And so it’s not until you have the Googles of the world able to kind of just freely finance sort of the playthings of academics that you start seeing these more industrial applications. But, you know, kind of as you can see from things like driverless cars and even large language models, a lot of these things, while we call them AI, you know, one, they’re separate technologies, but two, a lot of this technology involves existing supply chains of humans doing labor in order for the AI, quote unquote, to do its work, right?

So, you know, just a few months ago, Waymo was asked about, you know, why are there drivers in the Philippines kind of monitoring Waymo cars, right? Or the remote drivers and such, yeah. Right, the remote drivers, they step in when the car makes a mistake, et cetera, right?

But even without those, well, monitoring the cars, you know, you’ve got engineers who are mapping every inch of the road in order for the AI to actually understand where it’s driving, in order to actually understand the rules of the road and even, you know, sort of what is going on at the road in real time, right? You’ve got, you know, all this labor basically refining, you know, the inputs of large language models in order to have them have a accurate understanding of the world, right? And so one way to think clearly about, one way to think clearly about AI, one, stop using the word AI, let’s actually try and break out all these different technologies, large language models, transformers, which is sort of the architecture behind large language models, driverless cars.

[Nick]
Transformers who are Cybertronian robots who turn into vehicles that save us from other.

[Mike]
Right, they’re robots in disguise, obviously, that they’re here to take over.

[Nick]
So what actually is a transformer?

Very, very briefly, if you could.

[Mike]
So a transformer basically is a particular, I don’t want to pretend to be an expert on this, but basically, transformer is one particular type of neural net architecture that is behind the most recent kind of AI craze in the sense of this is what is driving most of the anxiety about AI. Transformers have this thing called attention. It’s a mechanism that allows AI to basically focus on or a neural net to focus on particular pieces of information within a corpus or what have you.

And so a lot of people have been trying to solve this problem of getting neural nets to understand language in a natural way. And historically, you have things that are sort of older natural language processing algorithms focused sort of on having a vocabulary of words that they’re familiar with. And this is the corpus.

Right, right. But transformers have that, but they also have this attention mechanism that allows them to actually understand within a sequence of text which pieces of a new text are actually most relevant to what is being discussed.

[Nick]
So this is the input, and then that’s being transformed into?

[Mike]
Kind of, yeah.

[Nick]
I mean, like, again, What is the output form for the sake of our audience?

[Mike]
Well, the output basically is just the text that’s being outputted, right? The output basically is the sequence of words that satisfy the pieces of language that are in attention, that are most relevant to what’s being discussed. And I’m, I’m paraphrasing here.

I, you know, I don’t want to be the definitive sort of like, um, arbiter or articulator of, of your knowledge of transformers. I’m just saying, this is my understanding of what transformers are. And the reason why transformers are so impressive is because, you know, for, for decades, getting something that could produce coherent text was hard.

Um, and something that’s interesting about this is, you know, a lot of people like using, um, the phrase spicy autocomplete to talk about what transformers are, what they do, and, and look like autocomplete obviously is, is not inaccurate. Stochastic Parrots, it’s not inaccurate. I think, you know, to some extent, what these terms are getting at is that even with the attention mechanism, uh, large language models and these types of neural nets don’t have any understanding of what it is they’re doing, right?

They are producing a sequence of coherent words, um, that’s, you know, sound nice and that, that fit with the context of what is being said.

[Nick]
And these transformers are indeed doing not thinking. And that’s kind of the first point we should make is we should stop calling it AI because the I in that is intelligence. These in systems, these transformers in this case are performing tasks.

They are not, you know, thinking in the traditional sense as we would, uh, uh, uh, sort of say, right?

[Mike]
Yeah. I want to, I want to, I want to linger in this point a little bit longer because there is something that’s very nuanced that is happening here. So when people encounter large language models, especially for the first time, they’re extremely impressive, right?

Their grasp of language is probably better than, you know, some percent of humans. Right. Are they smarter than a third grader?

Right, right. But, or fifth grader, right?

[Nick]
Yeah, it’s fifth grader, sorry.

[Mike]
Okay. Yeah. But, you know, part of the magic here, you know, cause the spicy autocomplete thing kind of undersells what it is they’re doing, right?

So what large language models are doing effectively or what they are trained to do is, I don’t even say trained, this is what it, this is sort of an emergent property. So large language models effectively model the shape of language use, right? So language as it’s been used across time, across space, right?

There are even, you know, recent experiments of getting a large language model, for example, to talk like a, I think like a 1930s, like, you know, human, right, because they limited the corpus that it was trained on to data from that era. Right. And so large language models are broadly, they are, they are literally encodings of the shape of language for the corpus that they are trained on.

[Nick]
Wait, hang on. There are people out there who are specifically training large language models on like F. Scott Fitzgerald and to the output is…

[Mike]
So you have to understand that even as dystopian as Sam Altman’s promise of metering intelligence and all this bullshit is, fundamentally…

[Nick]
Oh, Sam Altman, by the way, we have named our first agitator actor guilty party here.

[Mike]
He’s a grifter.

[Nick]
We’ll come back to him.

[Mike]
But my point is, in tandem with, with this notion of, you know, AI is going to be the future and everyone should, should stop, stop and give everything to AI, including their electricity bill or whatever.

[Nick]
Which is what Sam Altman has proposed. Put all the eggs in one basket. Right.

That basket being the AI will run all systems.

[Mike]
Seven trillion dollars of investment.

[Nick]
It will optimize everything. Everything.

[Mike]
Right.

[Nick]
Yes.

[Mike]
So in tandem with this, fundamentally AI as a field has always been the sort of academic niche interest. People who are lifelong academics, who literally just care about expanding knowledge and expanding, you know, this sort of like particular field in a way that is for the sake of knowledge for its own sake. So Transformers as an architecture was financed by Google, but Google famously runs…

[Nick]
Left it on a shelf when it, when it, when this was first produced.

[Mike]
Yes. Yeah.

So Google finances academics as part of their, I don’t know why they do this, but to give back to the field, I guess, to, to have a good brand. Okay. So Google financed a bunch of academics.

[Nick]
Because they don’t do evil.

[Mike]
Right.

[Nick]
They, they promised us for years.

They don’t.

[Mike]
They finance these academics. Right.

And you know, these academics are able to do research for like with, with no profit incentive, no profit motive. Right. And lo and behold, Transformers is one of the, the architectures that came out of this research and because, you know, no one really knew what to do with it.

And it was a curiosity of academics, you know, 2018, I think is when it was, was sort of the paper attention is all you need came out, sort of sits on a shelf, like Nick said, for, for a few years. And people kind of, the academics mainly play around with this architecture. It was a curiosity.

It was a curiosity. Yes. Only.

And they, and they slowly scale up, you know, that’s why when GPT came, became big, ChatGPT became big, you know, you, you only heard of GPT-3, 3.5, and 4. Not GPT-1 and 2, 1 and 2 were basically novelties that researchers, you know, behind, you know, uh, closed doors were playing around with. Right.

Um, but going back to my, my, my main point is, okay, you know, being in the shape of language kind of means that large language models have packed in within them, kind of a bunch of modalities that were not programmed. Right. And it’s, it’s going to sound like for five seconds, like I’m going to sound like a booster.

I’m not being a booster. I’m, what I’m trying to do is I’m trying to give you a fair assessment of what these things are, why they even have any efficacy at all, as opposed to being complete, you know, BS. Right.

So large language models having language packed in means that what they, what they are doing, you know, and I say doing in quotes, cause it’s like they are sort of an autonomous process. It’s just, just functioning.

[Nick]
What they are processing.

[Mike]
Yeah. Sure, sure, sure. Well, I don’t even know if they’re processing, but, but the implication is in order to predict the next token, they implicitly have, you know, I’m trying to avoid words that imply agency, but they have found quote unquote high-level regularities in language.

They have run across these things. And patterns in language that effectively give them for free. Um, uh, they, they give them for free kind of these abilities that, that appear like intelligence.

Right. So, um, Michael Woolridge is a, a, um, AI veteran. He’s been in the field for decades.

He was even, you know, in the field when they were doing good old fashioned AI, right. These old rules where you’re hand coding rules by, by, by, by hand with, with code and logic. And he was talking about, he gave a talk called, this is not the AI we were promised, um, for the Royal Society, I think about three months ago.

Um, and he’s saying, you know, when you think about it, you know, language encodes a lot of things like, for example, reasoning traces, plans, right, it encodes these things. Right. So when I write an article about how to shop at a store, right.

That article contains in it, the steps you need to do to go to a store and what to buy and how to buy it. Right. And so, you know, language is encoding all the processes and steps that we do.

Cause we, we communicate those when we talk to each other. Right.

[Nick]
Yeah. The instruction manual for anything reduces down the complex actions of that task into a series of steps that are simplified. Right.

It, it, and that’s what is being in, in put into these models.

[Mike]
Not, not just that though. There’s a bunch of other things that I think are also kind of very subtle, very implicit, right? So I got a pop quiz for you.

[Nick]
Uh-oh, here we go.

[Mike]
It’s a little, little macabre, but let’s do this. Right.

It’s, it’s a example that came to mind. Yeah. So I want you to complete the following sentence.

Okay.

[Nick]
Okay.

[Mike]
Give you two choices, right?

[Nick]
Mm-hmm.

[Mike]
The boy drowned in the, A, water, B, pool. Which one do you think is more likely?

[Nick]
Oh, that it’s just out there?

[Mike]
Right. So you’re completing, you’re completing the next token, like a large language model. Yeah.

[Nick]
Which one is more likely? So me, Nick, the LLM, my instinct is B, pool.

[Mike]
Right.

[Nick]
Drowned in the pool. Yeah.

[Mike]
Why?

[Nick]
Oh, that’s instances that show up in the news, like in the Zeitgeist. This exists.

[Mike]
So there’s, there is a sense pool is encoding frequency.

[Nick]
Right.

[Mike]
There’s something else that pool does too. Right. So more than likely when something like that happens, well, pool encodes a location.

Right. We don’t just say someone drowned in the water. I mean, maybe if we’re doing a quick kind of, you know, description.

[Nick]
Yeah. A coroner would say that’s a person drowned in the water because that’s a scientifically factual statement.

[Mike]
Right. But, but normally when we have news stories or reports about something like that, it would be, it would be a location, a lake, ocean, river, right? It’d be a specific body of water.

So like words like pool and lake and river encode, not just that, you know, the medium is water, but it encodes a location. And so this is what I would call like a regularity of language. There are a bunch of subtle nuances of language that you and I kind of take for granted, we don’t think about them very often, but a language model would pick up on, you know, kind of immediately.

Right. Because, you know, it sees the patterns across language. It sees sort of like these relationships that we take for granted and there’ll be millions of these, not just in terms of like meaning, connotation and definitions, but also in terms of like syntax and grammar and regulating like sort of how you use words, right.

The rules of the game, the rules of the language as well as connotation and meaning and stuff like that.

[Nick]
So the connections that we take for granted instinctively as like users of a language and walking around using it.

[Mike]
Right. Language models index on them. And like, you and I don’t think about this. And so it’s sort of like, you know, we can kind of see language models as like, in some cases, if we are wowed by the, the experience as a quote unquote appear, because in order for it to kind of, to, to kind of rival us, it is literally indexing on all these things that we take for granted or we don’t even think about. Right.

There’s another thing called distributional semantics. Okay. So this is a field of, of linguistics that has been known since the fifties.

I think it’s founder, or I wouldn’t say it’s founder, but one of the early figures in this, this field, I think his name was Rupert Firth. You can gut check me on that or fact check me on that listeners. But Firth has a quote saying, you shall know a word by the company it keeps.

Right. So words that are similar or that, that kind of directionally talk about the same thing, appear closer together in language. Right.

So here’s another thought experiment. No pop quiz for you.

[Nick]
You’re talking about literally closer, like in a sentence, two spaces away or two words away from that other word.

[Mike]
I will give you an example without no pop quiz for you. So let’s think about fire. I don’t know why my things are so violent today, but you know, just to say, let’s just go with it.

So normally words like fire will, will appear near things like ash, but, um, these, these things will appear together because, you know, fire tends to appear next to things that it burns. Right. And so you can get sort of like, almost like a mathematical precision of like fire plus leaf equals ash, right?

Fire plus wood equals char fire. You know, like you, you, you get these relationships that emerge because of the proximity of words to one another. Right.

And so a language model that is fed all this language use about words like fire and, and, and the temperatures of which things burn will have almost like, it could, it could recount almost like a firsthand experience of what burning is and how it functions without even knowing it, you know, knowing quote unquote, any of this, right, because the language itself encodes these relationships.

[Nick]
Uh, and this was the realization that, um, you know, in the innocent days sitting on the shelf at Google, these transformers were, were performing these amazing functions. Um, this was an innocent time. This is the garden of Eden period, but eventually after a few years of it sitting on the shelf, being looked at by academics, we enter the snake into the garden and that snake in this case was Mr. Sam Altman, who seemingly was the first one to look at this product and say, oh my goodness, I could either from his own excitement or willingness to show for it and to sell it, I can sell this to everyone as the, the general technology, the solution to all of our ills and proceeded to, um, work on it and exploit it for that purpose.

[Mike]
So there’s something interesting about this. We’ve got the convergence of, as far as I could tell, like multiple forces here.

[Nick]
Yeah.

[Mike]
Right. So Altman himself, um, you know, comes from Y Combinator. It’s a, a major kind of like, uh, player in Silicon Valley.

If I, you know, funded a lot of, uh, major Silicon Valley companies. Um, so he was the president of Y Combinator. And then separately, you know, you’ve got this sort of like strand of philosophy in Silicon Valley takes kind of different forms.

Um, I’m trying to think about how to, how to frame this. Cause it’s, it’s, there is, there is a convergent worldview that you will find across different, uh, different aspects of Silicon Valley.

[Nick]
The different corners of it, depending who’s there.

[Mike]
Yeah, yeah, yeah, yeah. So, so this idea of super intelligence, it predates Sam Altman, obviously, but it’s been in the air of, of Silicon Valley for the last like 20 years. Um, the most like recognizable strand is sort of in the doomer, doomerism of Eliezer Yudkowsky, a neckbeard fedora wearing Berkeleyan.

[Nick]
Right. Um, and another one of our, uh, sort of figures that we’re going to examine. So now we have two, uh, we’re looking at.

[Mike]
I really don’t want to think about him.

[Nick]
Oh yeah, yeah, yeah. We can kick him for later though.

[Mike]
I hate him so much for many reasons.

[Nick]
So we’ll start with our first snake though. We’ll continue with, with Sam Altman’s, uh, process.

[Mike]
But setting the stage too, cause I think it’s important to set the stage.

[Nick]
Absolutely.

[Mike]
So, you know, we’ve got the doomerism of, of, of Yudkowsky.

He, he started a blog in 2000 talking about, you know, AI, um, you know, is, is, is possible to encode human intelligence in the form of, of Silicon. Right. Eliezer kind of gets backing from major academics.

Um, I don’t want to call them major, but they are academics. Um, Nick Bostrom is one. Um, there are a bunch of other, um, engineering oriented people.

Like, um, there’s this guy who, who, who predicted, uh, when Watson would, would, would beat a chess player. I forget his name. He’s like a very, very, very, um, well-known internet figurehead.

Um, I can’t believe I’m blanking on his name.

[Nick]
That’s the way it goes.

[Mike]
Yeah.

I guess I had to move on, but you’ve got people who are not Eliezer Yudkowsky who, you know, who basically are kind of converging on this idea, um, separately from the, the, the doomers and, and the, um, kind of, uh, utopians because Yudkowsky actually started off as a utopian, he believed intelligence would solve all of our problems, right? Um, you’ve got people who just naively think that, uh, intelligence is a solvable problem. You’ve got effective altruists, right?

So you’ve got this, this whole cacophony of people, this whole group of people who are for different reasons, converging on this idea of, of, of super intelligence as a thing as possible. Right. So Sam Altman stands at the peak of this existing, um, this existing zeitgeist.

[Nick]
Right.

[Mike]
And to what extent he believes it, I don’t know. Um, he clearly, you know, given his role in, in commoditizing large language models, kind of even before I would argue they were ready to be commoditized because they’re, they’re failing in ways that, that a product that is mature would not, and in many of these failures, as we said, we’re kind of baked into how the product works, like understanding language regularities at a high level can only get you so far in navigating the world. There’s something that we have that is not that, that is beyond that. Um, so there’s a question of does Sam Altman believe this?

He comes from this world. He, he actually, you know, was given seed money to, to work on OpenAI because OpenAI was supposed to protect us from, you know, this fearful super intelligence that is supposed to be coming for us. Right.

[Nick]
That was our, that was our mission statement. In hindsight now we can say, well, he saw the opportunities and a way to exploit it early on. That is for the critics of this whole thing.

[Mike]
Well, actually not even early on. It was kind of, kind of like, you know, so OpenAI was founded, I think in, I want to say it’s 2016, I can’t actually remember.

[Nick]
Sure.

[Mike]
It’s not until 2022, 2023, he sees dollar signs and he starts commoditizing GPT 3.5. Right. So it’s some time. Um, but yeah, there’s this question of, you know, did he ever believe any of this or was he always just a, um, a grifter looking for the next big thing?

Right. You know, I lean towards the latter. It’s just easier to, to, to, to, to think about that, you know, like it, to think about it that way, because, you know, it’s kind of hard to see somebody who has that fear and then just goes, like, I can make money from it.

And, and, but he did say something funny, I think in 2018, he’s like, you know, superintelligence will, will, will, will be, will kill us all. But in the meantime, some great companies will emerge from it. It’s a paraphrase.

It’s not exactly what he said, but he said something to that effect. What a horrifying statement. Yeah.

But the real, the real tip off I think is, is in 2023, I believe there was a debate or not a debate, but basically he was kicked off. He was, he was ousted from the company. I don’t know if you remember this.

I do know. So he was ousted from the company and he was ousted from the company by people who are genuine doomers, like Helen Toner, right? People like that who were like, they believe the mission statement of, of safeguarding humanity from, from AI as they saw it.

And, you know, they kicked Sam off.

[Nick]
Because he had betrayed them according to their point of view.

[Mike]
Yes. So they’re saying Sam is dangerous, not just for violating our company standards and norms, but he’s dangerous to humanity. We got to basically prevent them from getting more power.

Right. And so, you know, that kind of schism, that kind of riff is not something that emerges from, well, I just have an alternate view of, of, of the superintelligence risk, right? It, it, I think it emerges because Sam Altman maybe fundamentally did not believe in it, or he, maybe he did believe in it and he just said money’s more important, I don’t exactly know.

[Nick]
Yeah. It wasn’t a healthy debate. It was a betrayal either from one side or the other, uh, fundamentally.

And, you know, that’s, that’s, that’s what we’re trying to look at here. Whether he believes it or not, he has not stopped trying to sell it. And that is what we’re judging him for, right?

Um, this willingness to take a relatively simple, uh, program that he saw on a shelf, studied, and then got so excited about, he was willing to make political connections, um, make, make business proposals, um, convert open AI into a, a for-profit venture versus a nonprofit one, simply to say, this is the technology that will change everything and save us from everything. And that is in a sense, I think healthy for us to critique because we’re the ones being sold it as citizens of the country and denizens of the world and taxpayers, quite frankly. Right.

So we can have a healthy skepticism without, you know, kind of being lamers and losers around it. Right. I mean, I think that’s fair to say.

[Mike]
Yeah.

[Nick]
Good, good, good. No, I, thank goodness.

Cause that’s what I’ve been doing. Um, so that’s great that we have, you know, uh, identified the original sin of the garden, uh, uh, the snake in the garden, Sam Altman. Um, Elon Musk enters the equation at some point, uh, in this timeline.

[Mike]
So, so Elon Musk gave open AI some of his early money, I believe, right? Yes. That’s what they were fighting over.

[Nick]
As the trial has revealed to us, yeah.

[Mike]
Yeah. And those of us in Silicon Valley has some exposure to this already. I mean, unfortunately, you know, and I’ve admitted this before.

I have rationalist friends. I know about the community in the Bay area that are the Singularitarians. Yeah.

Well, you know, to, to what extent I can be a helpful counterbalance or try to, to, to nudge these people towards reality, but you know, it, it is what it is.

[Nick]
And now we’re pleading to people through the internet, through the format of podcasts, uh, learn about this, don’t be paralyzed by it. Um, and the more, you know, the more you can act actually, um, as opposed to the, the glut of information or, or the lack of good coverage that we’re kind of suffering with simultaneously.

[Mike]
Yeah. I mean, that, that, that is a huge reason, I think, why it’s worth talking about this. If you, if you even have personal experience with AI or, I’m trying to.

[Nick]
The boss has made you go and, Hey, your workflow has changed today. You show up on Monday and you’re never doing AI.

[Mike]
I’m trying to actually implement a rule for myself and, and even those around me. Stop using the word AI, stop using the phrase AI, because like I said earlier, like my intro was saying, you know, like these are different technologies that are being bundled together. And one way that, that, you know, kind of the, the boosters quote unquote will get you is, you know, you’ll complain about large.

[Nick]
Chief among them who is Sam Altman. Right. Of those boosters.

Right.

[Mike]
They’ll, they’ll, they’ll, you know, you’ll complain about large language models, making mistakes and not being built for purpose, which is true. That is completely true. They are not built for purpose.

They are language associated machines as we’ve been talking about. You know, you’ll see, you’ll complain about that saying AI sucks pointing to the large language model, not knowing that there’s this delineation. Right.

The booster will get you saying, well, you don’t like cures for cancer. Do you? Right.

Cause there are, there are models that are built for purpose for a specific task that are actually, you know, scoped. They are effective at what they do.

[Nick]
And these are the good use cases that do exist and often are cited by the people in question, by the, by the boosters.

[Mike]
They’re not really cited. What we’re doing now is we’re, we’re kind of smearing all of these technologies onto ChatGPT and saying, AI will change the future. IE, ChatGPT will change the future.

It’s like, no, ChatGPT is not discovering new cures for cancer.

[Nick]
But we are being, the point is we’re making, we’re, we’re being encouraged to paint all these technologies with a broad brush by Sam Altman and his ilk. Yeah. Right.

They are the ones who benefit when we don’t kind of refine our own language. And it’s like, oh, AI does this, AI does that. We have to start here now and today and make our vocabulary a little bit sharper, a little bit better.

Isn’t that a good first step? Like being informed? Yes.

But then talking about it more accurately, responsibly.

[Mike]
Right, right, right.

[Nick]
Good.

Let’s fight back in the only way we know how, by reading things on the internet. I’ll take it. Um, so yeah, Elon Musk has entered the, has entered the chat here, um, entered the, the conversation.

Um, he’s provided the seed money and his interest over the years since he started working with, with Altman and, and OpenAI, et cetera, is seemingly let’s get, um, autonomous vehicles, which again is a bit of a misnomer. Right. Which we can get to in a second.

And let’s make sure we do, um, because he’s going to profit from the, the, the business model of Tesla. Eventually getting these cars to be fully automated or at least selling people on the idea that that’s what they will be in time, like get people excited around that, buy a Tesla, get these, this new class of vehicle, because it’ll drive you to work one day. It’ll drive you from LA to New York.

These are huge promises. And quite frankly, as we’ve seen so far, if the current trends are any indication, these technologies will not get there soon or maybe ever, depending on the assessment. Right.

[Mike]
Yeah. So something that’s so interesting about, um, you know, driverless cars are, are, I think actually, I think they’re DNNs, like, I guess, I don’t know. I forget exactly what the, there are different type of neural net architecture, um, then.

[Nick]
And we should, we should also say that’s another form of technology that’s specific and we shouldn’t brush it, paint it with that same broad AI brush.

[Mike]
Sure. It’s a separate thing. Yeah.

It’s a separate thing. I mean, in this case, I don’t know that knowing what type of neural net it is matters, right? Transformers are a different neural net than driverless cars.

Right. But the, the, the thing I want to drive home is for a lot of these, um, AI, you know, um, um, technologies, and I am saying AI technologies now deliberately kind of an umbrella way, a commonality that many of them have, especially these neural nets is that, okay, the internet has given us enough data, there’s enough data out in the world for, for people to train these types of AI systems. Cause these, these neural nets, this, this idea of a neural net is very old.

I mean, you know, say old in terms of like maybe seventies, eighties, nineties ish, right. And these things weren’t, weren’t viable back then because, you know, there’s wasn’t enough data to train a model on to predict anything, right.

[Nick]
The capacity was there, but the amount of data was so low that they couldn’t, uh, then perform that work.

[Mike]
And so now we have, we have reams and reams of data. And so these models are, you know, I would say what we are learning or what, what the field seems to be learning from what I can say as an outsider, you know, as a informed lay person, I would say is that, you know, the data’s there. These models are able to perform up into some percentile of efficacy and then they just, they just cannot get over the hump.

And so in a lot of these fields, you’ll notice. What’s on the other side of that hump.

[Nick]
What’s the thing being promised to us though, then is a question.

[Mike]
Yeah. The, the promise obviously is that they’ll, these models are so smart. They’re just like us or they’re better than us.

And so they will, they will take over all of our work and they’ll do other things. It’s true of the driverless cars, right?

[Nick]
So we’ll be the Jetsons, that’s the promise.

[Mike]
The same promise that we were given by Elon in 2016 about being able to put our feet up on our dashboard and drive from LA to New York is the exact same promise that Sam Altman gave us about what a, you know, what ChatGPT will be doing. ChatGPT 7 will be doing in, you know, 10 years time, whatever, right?

[Nick]
I’ve heard him say ChatGPT, roughly speaking, will run the nation’s electrical grid. Like that is the kind of maximalist, dare I say, utopian vision that is literally being sold to us right now. Like that’s the discussion he’s trying to, he’s trying to point us towards.

And I think as recently as last year, you know, people were lobbying Congress, Sam Altman directly, and billions or trillions of taxpayer dollars were supposed to go to that effort, setting up data centers, setting up this infrastructure, you know, getting all of our eggs into one basket, as it were. And that is the end state that he intends for us.

[Mike]
I finally remember the guy I was talking about at the beginning. Oh, please. Yes.

Ray Kurzweil. Okay. So Ray Kurzweil.

[Nick]
Oh my gosh, Ray Kurzweil.

[Mike]
Ray Kurzweil is a, is a singulatarian who, you know, so he’s well known in computer science circles for, you know, predicting, you know, when AI would be people like certain types of games like chess, predicted the internet in some sense, very, you know, very broadly.

[Nick]
And things like the success of, if you’ve seen the documentary, hopefully everyone has, AlphaGo on Netflix. So like, yeah, it’s, it’s out in there in the, in the pop tech realm itself. Like there’s things that, that you can look at.

[Mike]
Yeah. So the reason I bring him up is not just because I didn’t remember him earlier, but it’s just like this view of AI getting better than humans, you know, and, and being able to reach a point where it just outcompetes us in everything and the fear and hope and, and dreams that come from that, Sam Altman did not create that. Elon Musk didn’t create that.

And in some ways, I think even someone like Elon Musk or even Adario Amadei, who is another figure who I think is, is actually interesting to talk about, you know, they are kind of as beholden to those narratives, even as they promote them to boost their companies.

[Nick]
Sure.

[Mike]
Right.

Because in some sense, I think I would say to what extent Elon has beliefs, this is something he has felt strongly about.

[Nick]
Yes.

[Mike]
He was talking about, I mean, he, he has been actually instrumental in normalizing Yudkowsky’s worldview via Nick Bostrom’s book, Superintelligence. I think it was a book that came out in 2014-ish, you know, where Bostrom is sort of codifying the thesis of an AI so smart, so powerful that it will be a danger to humanity as a species, right?

[Nick]
And this then kind of dovetails with Ray Kurzweil and predicting the singularity when all intelligence, organic and digital and whatever, will converge into a single system, a single entity even. And, um, that was another idea that, again, was kind of being sold to the public, like, oh, we just need to live. And he, Kurzweil clearly believed it.

[Mike]
He needs to live. He’s taking vitamins every day to make sure he lives. His date for the singularity is 2045.

[Nick]
Yeah.

[Mike]
So, and he says, given the, you know, performance of ChatGPT, et cetera, that we’re on track, we’re on track, quote unquote, for that.

[Nick]
Yeah.

[Mike]
But he needs to believe that because he doesn’t want to die. Right.

[Nick]
Yeah.

[Mike]
And these promises go back, you know, even further to, you know, there’s a very smart, some people say the smartest human is Einstein, whatever, arguably probably was John von Neumann, who is one of these guys who, you know, I think if von Neumann were alive today, he probably would fit right into the Bay Area rationalist, you know, effective altruist culture, right? Von Neumann is one of the people that helped in the early or mid 20th century coin the singularity term. Right.

And so in a lot of ways you have a bunch of people who are very mathematically minded, kind of simplifying the world to models and data, right? The main point I was trying to make earlier is, you know, okay, we finally have enough data to make these models that are capable of doing some impressive stuff up to some percentile of task or, or whatever. And what has to happen to bridge the gap, wherever that gap is, depending on the domain, could be driverless cars, could be ChatGPT, whatever.

There are a bunch of humans who step in and take up the work of making the world digestible to the model. Right. So in the case of driverless cars, you know, you’ve got, like I said, you know, a bit earlier, you know, you’ve got the remote drivers who step in during emergency, but you also have, you know, the engineers who are mapping every, every square inch of the road, kind of making sure the model is aware of every pothole, every kind of variance that would kind of disturb the ride.

[Nick]
Right.

[Mike]
And in order for these, these driverless car companies like Waymo to even operate, they had to pick the city they operate in, right? Elon Musk sold us the idea of being able to drive from LA to New York with no kind of infrastructure being part of that vision, right?

[Nick]
Because the car itself would be so smart a la your Ironman that it could do the work. When in fact…

[Mike]
I’m surprised you skipped Kit.

[Nick]
I, oh, I was, I should have…

[Mike]
Knight Rider.

[Nick]
Knight Rider was gonna…

[Mike]
Knight Rider. Where’s my Knight Rider?

[Nick]
Where is my Knight Rider?

Where’s my jet pack? Where’s my Knight Rider? And they keep trying to sell us a jet pack, yeah.

[Mike]
Or where’s my, you know, in terms of ChatGPT, where’s my Jarvis, right?

[Nick]
Oh, right. Yes. Yes. That’s what we’re looking for.

[Mike]
You know, and in the case of language models, you know, I’ve written about this extensively on my blog, Misaligned Markets. I have this metaphor I call Grandpa Shelby is sort of the way you should think about large language models, right? So Shelby is the character, Leonard Shelby from Christopher Nolan’s movie, Memento, right?

Classic movie, right? Protagonist Leonard Shelby has amnesia, cannot form new memories, right? So the date, the day the movie starts, those are kind of all the memories that the character has, anything coming forward in the future, he has to write down extensively or else he doesn’t know, right?

So…

[Nick]
So he’s only has memories up to a certain point of time. Beyond that, he has to put everything on sticky notes.

[Mike]
Right.

[Nick]
Yep.

[Mike]
And so language models, when they are trained, basically all their associations are locked in time when they are, when they are running because their training still, once their training stops, they cannot learn anything new.

You can, you can upgrade a model by kind of either retraining it, right? Kind of subjecting it to the training process again, or just getting a new model altogether. But if you’re running a current instance of ChatGPT, whatever, it is not learning from you, right?

It does not do that in real time.

[Nick]
It is drawing only on its original association and its, that library is limited. That corpus. Yeah.

[Mike]
Right. So, so chatbots have, you know, again, kind of, you know, actually it’s, is it retrograde amnesia or anterograde? I think it’s, I think it’s retrograde amnesia.

[Nick]
Anterograde amnesia.

[Mike]
Oh, it’s actually anterograde. Yeah.

[Nick]
Yeah.

[Mike]
So chatbots have this, this, this anterograde amnesia and on top of that, right, they also have what I would call, the reason I call it Grandpa Shelby is the associations that are in a model’s corpus, it doesn’t have perfect access to them, right? So people are asking large language models, facts and stuff. A language model’s corpus is kind of breaks down into what are known as tokens.

Right. And so tokens are basically pieces of a word. And it, basically you could think of, of a language model’s brain as sort of like the connection across these pieces of words of the, you know, and so, when I talk about distributional semantics, for example, you know, the word fire and, and, and ash are going to come closer together in language than say, you know, water and running.

Right. And so those associations.

[Nick]
Or even within words, like your earlier example, to me, a human would look at strawberry and the only thing it would, might break down to is straw and berry. Those are distinct, separate. Right.

Subwords. Subwords. Uh, the, the LLMs are actually doing a different kind of analysis to tokenize these things than even we would look at.

[Mike]
I mean, in a lot of cases there is mapping of words we recognize as words to tokens. Um, and, and, and in those cases, even though the associations, you know, when we ask a chatbot a question, we don’t get to choose the, the path it’s taking in its associations in order to produce the output, right? We don’t get to choose that path.

And so you get sort of what I would call, and people have called, called this various things, like Ted Chiang is a science fiction writer, um, very famous science fiction writer, he, he said, chatbots are lossy JPEGs or compressed images of the internet, right? This is idea of lossy compression, where it’s like you, you feed these models, the world of language. And what you get is not the entire world of language, but compressions of that world of language.

Right. I call it dementia because what you’re doing is when you ask these models questions… Yes.

Digital dementia. So the models basically, they have these associations. These associations are truncations of more complex pieces of data.

And so you’re asking grandpa Shelby to recall something. And so you’re getting, you’re getting a fuzzy output of that recollection. Right.

Back in my day, you know, like it is sort of like, back when the model, I was, when I was trained on this corpus, right. Yeah. And so, you know, the reason why model outputs feel so kind of like they, like they, they sound so plausible, but are kind of uncanny is because they are fuzzy associations, the models hallucinating all the time, like these are associations from the training data that are in response to questions that may not be from its training, that are in response to questions where the, the word mappings may not be, you know, the, the word mappings might not be in an intuitive way to answer the question.

Right. So like, we’ve kind of taken, you know, what is effectively a very interesting science experiment. Right.

But the science experiment was actually not even, how can you make a intelligence? It’s the, the, the science experiment was like, what are all the high level regularities in language?

[Nick]
The science experiment was at Google. Hey, how can we make a Google search better? And then Sam Altman said, how can I sell this to the world?

[Mike]
They weren’t even doing that. They were literally, when Google, when Google funds these academics, it was for non-commercial purposes.

[Nick]
Yeah.

[Mike]
That’s why the transformer was, was it’s, it’s, it’s available. It’s, it’s open source. It’s public knowledge.

The transformer is not a proprietary product. There’s a reason why Google did not patent this. They did not even see it initially as a way to modify their search engine.

It literally was Google continuing their tradition of, you know, AI academia is open, right? We want to look good, right? We’re going to continue the tradition of, you know, publicly funding the arts as it were, the research, right?

The, the, this is the arts of STEM, right? Basically.

[Nick]
Right.

[Mike]
Um, publicly funding this research so that, you know, we can look like the good guys and, and, and hire the best computer science talent, right? Wouldn’t you want to go do computer science at Google, even if you’re not working on that product, right? So this was literally, um, to paraphrase Ultron, you know, like we, we, we took the most versatile substance in the universe and turned it into a Frisbee.

We’ve taken, you know, the most interesting science experiment of how do you find associations across language usage and have bludgeoned it into the shape of a human assistant, right?

[Nick]
Yeah.

[Mike]
That’s sort of like, it is one of the, the, the greatest accidents of history, I think. Mike here. That was part one of our conversation on AI.

Tune into the next episode to listen to part two of our views on the direction of the technology. Also, while you’re listening to this outro, please subscribe to the podcast. Leave a review.

We’re a young growing podcast. So any bit of engagement helps and look out for the next episode.