How does AI Solve the Problem of "Mouthy" Workers?

Disciplining the Labor Force by threat of Automation, an Interview with Cory Doctorow

Transcript

And AI doesn’t need to do the programmer’s job. It just needs to

allow them to uh discipline the labor force so that you don’t get Google walkouts saying we’re not going to make

tech for um uh drones. so that you don’t get uh mouthy workers who demand higher wages, who hop from employer to

employer. You remember that the Silicon Valley firms about 10 years ago uh were settled a big criminal case where they

had uh gotten together and done a no poaching agreement saying they wouldn’t give raises to each other’s talent to bring them over because their wages were

very high and they had a lot of worker power. And so, you know, an AI that is controlled by a skilled worker doesn’t

solve that problem for bosses and the AI that that is being used in ways that um are really impressive for software

development. That AI is the AI that doesn’t fulfill that goal. the AI that uh you know poops out code faster than

any human being could possibly review it that is then you know has its homework marked by coders who are terrified they’re going to lose their job and are

working as Sergey Brin says a six-year week uh which he says is the target that everyone working with an AI should be uh

shooting for um those workers do fulfill the destiny but those are the workers saying you know I’m a software developer

in aerospace we’re using AI to write code never get on an airplane again because we’re producing the most ghastly tech debt at scale. So, you know, until

you can like cleave apart these subtle issues, you can’t know where AI is affecting the labor market, how it’s affecting the labor market, and how that relates to the investor story.

QUESTIONS:

1. Do you anticipate that greater adoption of AI will discipline the workforce?

2. Should we use this opportunity to renegotiate the social contract? If so, how would you like to change the social contract?

3. How do concerns over AI code affect your view of airplanes and other critical systems?

TRANSCRIPT CONTINUED:

on the on the on the question of hype and it correct me if I’m wrong but you seem to view this you know what’s coming

from the AI companies and the CEOs you know as at at least primarily hype that it would that it’ll the popping of the

bubble will leave based behind something but it’s mostly hype and then if you have a salesman who can go to a company

and say hey this will allow you to replace your mouthy workers that that’s pushing on an open door but that is that

isn’t of course a long-term business plan like that. It has to actually it has to actually work like and so it’s at

some point if if it’s not you’re going to start to see some dialing back which you’re you’re starting to see a a little bit of that from companies now which are

finding that actually uh you know they’re not able to completely replace people at this point. So much of this though seems to be organized around kind

of people’s thoughts and motivations and ideas rather than what’s actually happening. Um we put up F8.

What this is kind of a counter um argument being made by one of the uh one of the AI folks is I think this is an

open AI person. I can’t remember. He says so he says uh so many intelligent people still cling to the idea that everything the AI labs do is downstream

of IPO hope or whatever. Some kinds of intelligensia are uniquely unsuited for the moment because their lens is quote strategic skepticism where normal people

can employ magical thinking. And the latter part of that what he means by the magical thinking is like is observing uh

the AI mechanistically and saying okay we have we have some ideas about how this is working and why it’s working but

we don’t we don’t actually completely understand it and there is something going on here that is currently beyond

our understanding of it. Uh yet it is doing a lot of these very interesting things. is solving math problems that you know humanity has not been able to

to solve which puts it beyond just kind of a a guessing game robot like a a very significant thing.

So while we’re in the realm of just ideas and motivations and everything, what do you think of this challenge to kind of people like yourself people

like me in some respects that we are uniquely unsuited for this moment because our constitutions are constantly

skeptical? But at some point you’re going to run up against something that is real. And how how are we going to tell the difference if we’re constantly

looking for the fakeness behind the the frauds that run these companies? What if actually underneath it there are these

computer scientists who actually have developed something?

Well, I I think it’s a very weird argument that um being discerning makes it hard for you to find the truth. Um you know, approaching things and saying,

“Well, that’s an extraordinary claim. I don’t need any extraordinary evidence because the claim itself is so interesting, I think is a very weird way

to do truth seeking. And I think if you want to talk about magical thinking, the idea that shoveling words into the word guessing program makes it wake up and

turn into God is is extremely magical thinking. I I always say it’s like believing that if you keep breading horses to run faster, eventually one of them gives birth to a locomotive, right?

The the thing that we don’t know what consciousness is. Uh we have people who make AI who say things like this is 51%

of the way to consciousness. And when you challenge them and you say you don’t know what consciousness is, they say, “Well, how do you know this isn’t consciousness if you don’t know what

consciousness is?” Which is again a very strange little rhetorical slight. Uh you know, the last time someone said that to me, I said, “I don’t know what the weather in Paris is and I don’t know

what the weather in Lisbon is. That doesn’t mean they have the same weather.” The fact that two things that we don’t understand exist doesn’t mean that they’re the same thing. uh and

while we don’t know exactly how AI works and why while AI does exceed many of the technical uh uh guesses that we’ve made

about how making small changes in the techniques that we apply to machine learning would produce uh outputs that

would be commensurately small and in fact and instead they’re larger which is always exciting you know the formal term for that is a breakthrough breakthroughs

are cool and interesting um that uh even though that’s true the world is full of systems we don’t fully understand and

whose inputs and outputs we don’t um control uh precisely that are instead in a kind of brittle equilibrium that that

makes all kinds of sways and and changes all the time including just you know firms and uh even small relatively well

understood markets do things that are chaotic and hard to anticipate so I think saying oh well you know you’re uniquely unsuited because you’ve never dealt with a thing we don’t fully

understand means that you failed to grasp ask how many things we don’t fully understand which is most of them.

So uh go ahead Ryan.

Uh the one I guess two counterarguments that you hear from um defenders of of the virtues of the or the power of the

technology are that calling it calling it like a word guessing game is you know minimizes it to a to a a degree that

just misunderstands what’s going on here. And one and the two two examples they often give are, you know, a a couple that has been together for

decades that is able to complete each other’s sentences. It’s not just a kind of random mathematical guessing game that they’re doing. It’s like they actually deeply understand each other.

And it’s not they’re not just predicting from the past. There’s there’s something there’s reasoning and understanding behind that. And then the other example

they give is let’s say you you’ve you’ve [clears throat] read a mystery novel or sci-fi novel but a let’s say it’s a mystery novel and in the last page they

say uh and the detective says I understand I figured out who the killer is and the killer is and then it stops

and then the reader if you ask the reader what’s the word that’s going to come after this the reader who can

accurately guess the next word can do so because of reasoning and understanding like They they understood the book. They understood the story.

And so it’s not just mathematical computations and guessing. And so that’s true when people I’m sorry. Go ahead. I I cut you off.

No. So that so so they’ll say so there is some level of reasoning and understanding going on here that we

don’t quite understand ourselves yet how it mechanistically comes together, but it’s there’s something there. So it is

true that when people do it, it’s because we’re doing a thing called understanding and reasoning, which we only uh have a a

a cursory understanding of our are what’s going on both mechanically within the brain and then you know uh uh sort

of chemically and and so on cognitively what’s going on is not perfectly understood. But we actually have a much better idea of what the computer is doing. And if we’re discovering

something about what computers are doing, it’s that we are discovering that word guessing is actually very very powerful. Uh which does not make it

reasoning. So it turns out that there are like linguistic corelates latent in the structure of a mystery novel that

make it possible to if you count all the words, which is like so the idea that um counting words and then guessing based on the count is not what they’re doing

is just technically wrong. That is in fact what they’re doing. You that’s what model training is, right? You’re measuring the frequency distribution of

words in their relationship to other words and phrases. That’s what it is.

Now, saying that’s all it is doesn’t mean that it’s trivial. It means that there’s a lot of stuff that you can figure out by making those kinds of

guesses. But those guesses are not the same as comprehension, which is where we get this funny term that we use when we describe the defects in AI products,

which is hallucination. It’s not a hallucination. It’s just a a limit to statistical guessing. M now one of the

things that AI did for us this current generation of AI is it showed us that statistical guessing is more powerful than we thought it would be when you

scale it up that uh you the returns to scale from adding more training data and measuring more and more subtle

relationships are are larger than we thought they’d be. That is a legitimately really interesting thing to

discover about computer science and about frankly the nature of reality. And it’s not the same thing as reasoning.

Otherwise, these models wouldn’t make the the the characteristic errors that you get when you’re not reasoning, but just guessing uh guessing words,

right? So, I guess some of what I would say to that is is there any limit though to the that power of scale? Because

there have been a lot of predictions of, oh, they’re going to hit a wall, they’re not going to get better, they’re not going to improve much from here. And that hasn’t been the case. They’ve continued to improve, you know, an open

AI model solved this 80-year-old mathematics problem that humans had been stumped on. There was a situation that unfolded. We could put F7 up on the

screen. I’m sure you probably followed this. There’s a whole controversy over whether a chatbot wrote the this prize-winning story. Um, you know, for

this this famous Commonwealth short story prize. There were a lot of people who looked at this and oh, this is AI writing. They dispute that it was. You

know, it seems like the analysis at this point is that it wasn’t actually AI writing, but I think it indicates how advanced the uh AI even on in

creative endeavors that require, you know, the sorts of things that you would want to imagine humans were sort of uniquely good at that AI has gotten to that place where it’s indistinguishable.

I’ve seen other analyses where experts are judging short stories and they’re rating the, you know, the AI stories just as highly or if not higher than

some of the the human um creative product. So uh you know if if these laws of scale that just adding more and more

and more genuinely do produce these sorts of profound improvements what is ultimately the limit to that?

So the scale is breaking in the sense that they it it is it used to be that linear inputs to scale produce linear outputs to scale or even better than

linear. So you spend an extra dollar, you got an extra units worth of of goodness out the other side, whatever you’re using to benchmark that that number is going down. The amount that

you get out, which means that we’re having to put much more in. That’s why each model, you know, each new foundation model is is less profitable than the previous ones is that they they

are hitting limits to scale. You know, they I I I have heard people say that they have seen AI art that uh they can’t

distinguish from uh um human art of of quality. And [clears throat] I read that short story and it was ghastly. I mean,

it’s a it’s a bad short story. I think that one of the re, you know, and it’s not the first time that there’s been a

literary prize awarded to uh a piece of fiction that I looked at and was like, “This is unreadable, terrible stuff.”

And maybe that’s my bias because I’m a science fiction writer. We write pulp literature. Uh, and you know, within the science fiction field, which is one of

the last fields that actually pays for short fiction, um, science fiction editors are being drowned in AI stories and they don’t have any trouble discerning them because

they are each of them less publishable publishable than the last. They are dreadful, dreadful work. Um, you know, I

I’m not a mathematician. I’m the son of a mathematician. I don’t know enough about math to assess the problems that have been solved. It seems really

impressive. And you know the fact that um mathematicians have figured out how to use computers to solve problems that they couldn’t solve before is both

wonderful and normal. Uh mathematicians have solved new problems using computers for a really long time. And you know a mathematician and my acquaintance was

talking about iterating through uh several different ways of of um uh approaching a proof uh which would be

labor intensive to do personally and which you can automate through a computer. That’s great. Uh, you know, look, there’s a lot of things that that this software does that I think in the

absence of a bubble, we would just call a plug-in. And there have been some really important plugins for the tools I use over my life. I I word processors have really made some leaps and bounds.

My first word processor was a program printed in a magazine I bought at a newsstand and typed into an Apple 2 Plus. And in in the years since, word

processors have gotten all kinds of features, some of which are very valuable to me. Some of which I think are very frivolous, some of which I think when other people use them, it makes their work worse. I’ve never been

in a circumstance where I said, “Well, this plugin is so interesting. Let’s bet the economy on it.” And also, it’s probably time to feed the writers into a wood chip.

Yeah. Corey, my last question for you is do you worry at all that if you’re wrong

that if the technology is maybe not as you know we’re going to replace all human beings but if it is significantly you know more powerful and continues to

improve and there is some large chunk of the labor force that is displaced. Do you worry that if you’re wrong and downplaying its capabilities now that

you’ll suck some of the energy out of the the capability of people to to organize in ways that are are needed to deal with that you know very scary potential future?

I actually think that uh you know organizing about AI is a useful endeavor and that we should be organizing about it. I think the local movements that are

talking about data center justice, you know, no city council ever said, “Oh, we approved this planning permission and exercised eminent domain on your farm

without talking to you about it because we knew that you’d be really pleasantly surprised and we didn’t want to wreck it.” And so, the fact that we’re getting these local movements that are building

up this capacity to demand equitable treatment and to demand co-determination with technology, I think, is great. I I

I I mean, if it were the case that tomorrow AI could do a lot of human jobs that uh currently are being done by

humans, a I a functional economy would find work for those people because we have full employment for the next 300 years. We’re going to have to move like coastal cities 10 kilometers inland.

We’re going to have to deal with hundreds of millions of climate refugees. We’re going to have more zoonotic plagues as we get more habitat loss. like you, me, and everyone we know

and everyone who will live for the next three centuries is going to have their hands full. Like we don’t have a technological unemployment problem on our hands. If if we have an unemployment

problem, it’s because we have a social and economic problem uh that is exacerbating our environmental problem.

Well, that sounds like a grim future, just not one where the central problem is unemployment. [laughter] [gasps] Um Corey, thank you so much. I really appreciate um you know, your time

this morning and I think your ideas are really fascinating. I appreciate you spending the time to engage with us on them. Um, the book is The Reverse Centaur’s Guide to Life After AI, and I recommend everybody pick it up.

And if you want to get involved on these issues, visit eff.org. The Electronic Frontier Foundation has been doing this work for 35 years, and I’ve been doing it with them for 25 years.

Amazing. Thank you for that plug.

[snorts] That was an interesting conversation, Ryan. I like that one.

Yeah, I think your question is an important one. What if he’s wrong? Uh, and sort to me, there’s sort of a Pascal’s wager element to it. Um, maybe

it’s all hype. If it is, then it’s another bubble that comes and goes. Yeah.

But if it’s not, you got to be planning for it now. Yeah.

You got to be we have to be thinking and organizing about what the world looks like in the event that it’s not hype.

And I don’t think it’s I don’t think it’s hype. So, and I I’m in the minority. You and I think we’re both in the minority on this on the left. Um,

but it’s something that, you know, it’s we’re going to hopefully live long enough to find out. Yeah. Well, the other thing is, let’s say it is hype.

The very fact that you have all of these oligarchs who are like, “We need to change the social contract.” That also seems like an opening because I would like to change the social I don’t like

the social contract that we have now. It leaves many millions of people needlessly immiserated. So, if they’re giving us that opening, even if it is to

hype their, you know, shitty products, um, which I agree with you, I I don’t think that that is I do think that there’s a lot of there there for the technology. I do think it’s going to be

genuinely transformational. I also think it’s a bubble. I think it’s both of those things. Um, but in any case, if they’re giving us the opening to uh

restart negotiations over the social contract, I think we should take them up on that. Yes, indeed.

All right, guys, that does it for us. We got a great Friday show planned for you tomorrow. Um, I think Roana is joining to talk about his fight with Elon and

reflect on the week’s activities with regard to the the DSA wins. He endorsed, I think, Claire and Brad Lander, but not Daria Lisa, which I’m also interested to

to ask him about his endorsement process. So, should be a good one. Um, Ryan, thank you. Great to have you. Great to see you. Good to see you, too.

Yep. Safe travels and I will see you guys tomorrow.

Hey, if you like that video, hit the like button or leave a comment below. It really helps get the show to more people.

And if you’d like to get the full show ad free and in your inbox every morning, you can sign up at breakingpoints.com.

That’s right. Get the full show. Help support the future of independent media at breakingpoints.com.