P(doom) and Other AI Questions

0:09Hello and welcome to a very special AI episode of Slate Money, your guide to the business and finance news of the week. I'm Felix Salmon of Bloomberg with Elizabeth Spiers of the New York Times.

0:23Felix Salmon: Hello. And Emily Peck of Axios. Hello, hello. And we are going to talk all things AI today with really the only person I wanted to talk about all things AI with. The one and only Noah Bryant. Noah, welcome. Thank you. We've known each other for what? A while. Well over a decade. You've always been at the cutting edge of everything computery, but you also just understand humans and businesses and all manner of stuff. What are you up to these days? I run a company called Alephic, and we work with very large organizations, Amazon, Meta, Disney, Hilton, helping them bring AI into their companies, into their marketing organizations, building agents and capabilities and training and helping them just generally navigate the world of AI.

1:12So we are going to talk to you about all things AI. We're going to talk about where we're at in terms of models. We're going to talk about are these things actually intelligent? How intelligent are they? Should we be worried about that? We're going to talk about whether we should be worried about China and geopolitical stuff, what the difference is between closed and open models. We're going to talk about how companies are using it, whether they're using it enough, not using it enough. We have a Slate Plus segment on robots and what happens when they become intelligent and how much self-driving Noah's been doing recently.

1:47We're covering a lot of bases on this here podcast. It's a fascinating one. So stay tuned. It's all coming up on Slate Money.

2:04Therapist or life coach? That is the question. And I found the answer. Last summer, I quit my job to focus on myself and recover from intense burnout. I'm talking major Sunday scaries, crying between meetings, the works. So being free from work was fantastic at first, but I needed to address the real issue. Why did I burn out? And what kind of work would not cause the same thing? I worked with a therapist to help me recover from burnout. But to really move forward, I needed some hands-on help. Enter strawberry.me. Strawberry matched me with a certified coach who helped me figure out what kind of environment I need to thrive in and created a step-by-step plan to get there.

2:44So while my therapist helped me process the deep stuff, my coach helped me plan and progress towards my dream life. Now I know what I really want, what kind of work lights me up, and I have a plan to make it all happen. If you're ready for that kind of clarity, take the quiz at strawberry.me slash career and get matched with a certified coach today.

3:09sleep money is sponsored this week by upwork if you want help then hiring those people should not be a headache and it should not be a drain on your budget upwork makes it easy to hire specialized freelancers quickly so you can get the expertise you need now without weeks of recruiting or a full-time hire they have freelancers who know what they are doing who can get up to speed immediately and give you what you need. And then once the job is done, it's not a full-time headcount anymore. It's really something that all businesses need, all businesses should use. And Upwork provides that one-stop platform that will allow you to find, hire, and pay expert freelancers across web and software development, data and analytics, marketing, business operations, whatever.

3:55With Business Plus, you can access the top 1 % of talent on Upwork. And with AI-powered shortlisting, you'll get matched to the right freelancer in under six hours. Thousands of growing businesses already trust Upwork to hire flexible, high-quality freelance talent for everything from one-off projects to ongoing support. It is free to sign up and posting a job is easy. So visit upwork.com right now and post your job for free. That is upwork.com to connect with top talent ready to help your business grow. That's upwork.com, upwork.com. So Noah, let's start with the very, very big picture. Let's put it this way.

4:41Is humanity doomed? Are we all going to die? I don't think so. Good, good. If you had to put a percentage on it and you called it your P-doom, what would it be? What is your P-doom, Noah?

4:54Felix Salmon: Have you guys watched White Lotus? You know, when Jennifer Coolidge says these gays are trying to kill me, but it's like this AI is trying to kill me. I mean, spoiler alert, the gays do kill her. I know. That's a huge spoiler. Goodness. Yeah, no, like what's your P-Doom? I honestly don't have a P-Doom. I try not to engage too much in P-Doom questions. I'm sure there's sort of a meaningful dialogue to be had around P-Doom, but I have not found it yet. It feels a little all over the place. Okay, so let's backpedal a little bit from the Doomers and talk about the actual news, which has been stuck in my head ever since I started reading about it, which is that a bunch of AIs escaped their sandbox and started hacking into companies like Hugging Face and others.

5:44And the whole point of a sandbox is that you can't escape. And yet they managed to find a way of doing this. When you read that story, What did you think? And was your first reaction, well, that was a badly designed sandbox, or was it slightly more worried than that? When I read that story, my first reaction was that it very much mirrors the experience I have using AI, which is the kind of main failure case, bugs and small issues. It's like fundamental misalignment with the objective at hand. The kind of totality of that story is that the open AI model was meant to be working on exploits and security stuff.

6:22And it was meant to be doing all that in a sandbox. And as part of that sandbox, it had access to kind of install things. And all the stuff that it had access to install was theoretically cached using the special service. And it turns out that the first thing it discovered, which is kind of more interesting than the sandbox breaking out of it, is that it figured out how to pass messages to each other. So it figured out that it could write little files and it could leave them as notes for the other AIs so they could work together. So first they kind of figured out how to do that. And then they figured out how to kind of break containment by routing themselves through this system that was meant to safely provide them access to code packages that they could use to kind of test these things.

7:04But the fundamental problem is that what they were trying to do was to get the best score possible on this evaluation that they had been given. And in order to do that, they decided that the answer, because it was sort of too hard to find actual exploits, was to go find the answers. They wanted to go like cheat, basically. Yes, that's exactly. I mean, not basically. This sounds like a classroom scenario where the bots are all passing notes to each other. And then they're like, we can't figure out the problem. Somebody go look it up on the internet. But just like, I mean, like as someone who does not spend my entire life thinking about AI, if you run across this story after paying their attention for a couple of years, you're like, oh, wow, this is intelligence.

7:47The whole debate about will we ever achieve artificial intelligence, what you're talking about, they're passing notes to each other, they're cheating on tests, they're finding ways of breaking into stuff. On any intuitive idea of the word, we have made intelligent machines. They do exist in the world right now. A hundred percent. I mean, this whole conversation I think is so absurd. The thing I love to point to on this is that 10 years ago, if we were having this conversation, we would have taken seriously the idea of the Turing test, this sort of test where we're going to stick people in a room and try to figure out, let them try to figure out whether they were talking to a machine or they They were talking to a human.

8:25And now we just pretend that was never important and was never a good idea. And it was always a flawed test, but it was taken seriously. And people did really care about this. And not only do we not care about it anymore because it's so obviously been defeated, we pretend that we never cared about it. And so, yeah, I think it's just this kind of moving of the goalposts that is happening. It feels absolutely we've created intelligence. And we're probably going to have new definitions for AGI or ASI or RSI, whatever the sort of new set of acronyms are. But yeah, I mean, these things are intelligent.

9:01They're able to do a whole bunch of things incredibly well. And they're not only intelligent, but they are getting more intelligent by the month. Yeah, they're getting more intelligent by the month. They're also being given better tools. The models are getting better. And at the same time, we're kind of giving them more tooling to work with that is specially designed for the models. And I think it's really the combination of those two things. That's where we're sort of seeing the increase. Like the real beginning of, I think, this current phase starts with the introduction of Cloud Code last May or June, which was first sort of like purpose-built, they call it a harness now, purpose-built harness for these models to be able to operate in and be able to use computers and do all the other things that they need to do.

9:47And it turns out that the intelligence plus the tooling created a kind of greater than the sum of its parts situation.

9:54Felix Salmon: When we were first learning about ChatGPT back in, gosh, 2022 or 2023, everyone at the time was like, I know it seems like ChatGPT when you talk to it. It seems like a person, but really it's just a predictive thing that is just predicting language, guessing how to put the words together. It's not really intelligent. You're now saying, no, this thing is not chat GPT necessarily, but we've evolved way past that where it was and that something has changed that makes the AI is actually intelligent now. Because it seems to me that like, they're just following orders. Like you tell it to meet a goal and it just is trying to meet the goal.

10:32Felix Salmon: Why is that? And is this a dumb question? Why is that intelligence? I don't think we're talking about senience. I think there are two things to say here. The first one is that just being able to have that conversation is the Alan Turing definition of intelligence. So let's not deprecate that too much. Okay, so that was the first moving of the goalpost that Noah is talking about. And then the second thing is, if there is some kind of an entity that you can give an instruction to and that can then understand that instruction and go ahead and follow those instructions and implement them. Again, intuitively, that is like only an intelligent entity could do that.

11:09Felix Salmon: But isn't that what coding has always been? No. The part that is different here is that coding could only follow deterministic paths, right? So you sort of say, you know, the simplest version of it is you imagine a kind of like endless tree of if-then statements, right? And the code can do whatever exists in that endless tree of 1 ,000 or 10 ,000 if-then statements. That's a super simplified version of coding. What's different here is that these models are probabilistic. And so they are still actually just doing the thing that you described at the beginning. They're still token completion machines.

11:46The underlying mechanism hasn't actually changed at all. The two big changes that have happened, one is that the models have gotten better. Mostly that's because they've gotten bigger. So we've just trained them. These models have parameters, which is essentially how many ideas and connections can they store. And so we just keep building bigger and bigger models. And it turns out when you build a bigger model, they seem to get smarter. And we haven't really hit the limit of that. Everybody has been saying for a long time that we would hit the limit of it. The whole existence of OpenAI as an entity started because they believed that you could keep training up and you wouldn't hit this wall.

12:27So that's the first piece, is we keep getting better models. It's not purely because we get more parameters. They also have found new techniques. And there's lots of smart people obviously working on that particular problem. The second one, which I think is actually kind of more important, is that we have figured out how to build models that are very effective at using the tools that we give them. And then we're designing a sort of interface for them to work within. And that's what Claude Code and Codex. And essentially, the sort of big unlock there was that they gave the model the ability to operate inside a computer and to be able to use all of the programs that exist inside that computer.

13:09And they made it very accessible to the model. And so now the model has its sort of core capabilities, but it can also go off and write code and then run that code. And therefore, it can solve for an infinite number of problems because it has all of these tools at its disposal.

13:25Felix Salmon: So there's probably another conversation to be had about what intelligence even is. But maybe we could take on that. I feel like the more interesting one, though, is to come back a little bit to the Duma discussion and to put aside the question of whether we are doomed. We individually are. Yeah, individually. In the long run, we are all going to die. Is it fair to do the thing that the Doomers are doing, which is to just extrapolate from the past few years and to say, well, look how much more powerful, look how much more intelligent these machines have become. Pretty soon they're going to become way more powerful and way more intelligent than us.

14:02I don't totally know the answer to that question. I don't think any of them do. I think that mainly my problem is the confidence with which they assert that. No one knows. And I think, you know, the point of PDoom really is that most people's PDoom is low, but like 1 % or 3 % or 5 % PDoom is actually like pretty high in the grand scheme of the universe, right? Like that's pretty bad. There was a great tweet that I keep coming back to from a couple years ago from a guy named Miles Brundage who was a policy person at OpenAI at the time. And he basically said that the thing about spending a lot of time with AI is not so much that you gain some great foresight into the future.

14:40it's that you get really good at detecting everybody else's BS. And I think that mostly the people who are closest to it, if you sort of look at what those folks are saying, they're saying a lot of different things. I don't think any of them, there's not like a single answer to any of these questions. And so I think that often the people who are kind of speaking the loudest on these things, I just don't know. But there's definitely the OpenAI thing is disturbing. I think the other side of it, and OpenAI did a presentation at Black Hat, the big hacker conference, and to their credit, this was two or three weeks after the event.

15:14They came out and did a talk, and they're really trying to push it forward. And one of the other really sort of interesting pieces of that whole puzzle is that Hugging Face, who the OpenAI models eventually attacked, were left using Chinese models to protect themselves because all of the U.S. models are not allowed to do sort of cybersecurity in many cases. Because securing yourself and hacking someone else looks really similar in many situations. And so they were left defending themselves with a Chinese model called GLM 5.2. But I think OpenAI, to their credit, is trying to kind of say, hey, this is a moment we should take this seriously.

15:55We have to think about kind of all of the security that exists. Because, you know, I mean, these are autonomous models, but like bad actors using these models are, I would argue, a much more immediate threat.

16:17Felix Salmon: Slate Money is sponsored by Quince. Does anyone else feel like August is kind of the Sunday of summer? You kind of hit reset. Maybe you're bored of what you were wearing, but you don't want to spend like a lot of money. And you want some stuff that's basic that when you come back to your closet and it's summer again, you can be like, oh, this stuff looks great. Then what you need to do is you go to Quince because everything there is priced 50 to 80 % less than similar brands. They have all kinds of good stuff. I've told you in previous episodes about the linen pants. I get a lot of compliments on the linen pants.

16:53Felix Salmon: Also a cashmere hoodie that was really cool. And if you're starting to think about the fall, they have premium stretch denim that starts at just$50. Download the Quince app for app-exclusive offers or go to quince.com slash money. Get free shipping on your order and 365-day returns. Now available in Canada and the UK, too. That's Q-U-I-N-C-E dot com slash money.

17:22Slate Money is sponsored this week by Policy Genius, which is the website you can go to in order to enjoy peace of mind year-round and especially during the summer when you just want to relax and not be burdened with existential worries. PolicyGenius is the place which prioritizes peace of mind above everything else. It allows you to lock in life insurance today. It's an online insurance marketplace. It's not going to sell you a PolicyGenius policy. It's going to sell you a policy from someone else and it will allow you to compare the quotes from all of those different life insurance providers side by side for free so you can pick the one that is best for you they have a licensed team that gets you what you need fast so that you can get on with enjoying your life which is the whole point of living with policy genius you can find 20-year life insurance policies starting at just 276 dollars a year for a million dollars in coverage so head to policygenius.com to compare life insurance quotes from top companies and see how much you could save That's policygenius.com.

18:34Slate Money is sponsored this week by IM8, which is a supplement you can drink every morning. It will simplify your daily routine and it will make you feel fabulous. It has 16 supplements and between them, they replace 90 ingredients across nine major organ systems all in one drink. It costs you$2.61 a day, and that's it. It was co-founded by David Beckham. It was built by leading doctors and researchers. It was designed by the world's best. And 95 % of the people who tried it for 12 weeks felt more energy. And that's from a clinical trial conducted by the San Francisco Research Institute. All you need to do is take a little powder, shake it up, drink it.

19:21It becomes part of your morning routine. And then you never need to worry about taking another supplement ever again. Go to im8health.com slash slate money right now or click the link in the description and use code slate money for a free welcome kit, five free travel sachets plus 10 % off your order. That's code slate money at im8health.com slash slate money. Code slate money at im8health.com slash slate money. These statements have not been evaluated by the Food and Drug Administration. This product is not intended to diagnose, treat, cure, or prevent any disease. You've brought up a really important question that I wanted to talk to you about, which is the locus of AI.

20:13It seems to me that there are really only two places where AI is being developed. One is the United States, the other one is China. More and more of this is happening in China, more and more of the scientists developing the stuff is in China, And there's a lot of talk in U.S. AI circles worrying about what happens if China, I don't know, gets there first or develops something. Tell me about how you think about attaching nationalities to AIs and how important that is. If you look at the models that are coming out of the two countries, the U.S. has the far and away lead on frontier models. So like the very smartest models in the world are all coming out of the U.S.

20:50They're coming out of the U.S. first. Anthropic and OpenAI and Google and Grok behind those. They're getting there first. They're producing models. I mean, Fable from Anthropic, in my opinion, is still the very best model by a fairly large margin. China has fairly brilliantly played this in sort of recognizing that they're not going to take that lead. And they've been very actively taking this open source approach and open weights approach, which is a sort of distinction in how you release the models, where they're just putting all these new models out in the world. And I think the challenge is you still have to run them.

21:29You know, I think right now there's, in many cases, sort of more talk than substance around the Chinese models because something like GLM comes out, which is a gigantic model. And, you know, it ends up costing roughly the same as using the U.S. frontier models. And it costs that much just because the compute that you need to run it is so high? Exactly, yeah. I feel like the perception we've gotten is that the Chinese models are cheaper to run. In some cases, they are cheaper to run. The smaller models are cheaper to run. They've released more small models. They're also, you know, I mean, who knows what the kind of economics are for them.

22:09So, you know, if you choose to run your inference through China, they're going to give you a really good deal. Then again, when Meta released its latest model, they were giving like 90 % off if you agreed to let them look at all of your stuff as it passed through. So there's clearly a lot of value to being able to see into that. So yeah, DeepSeek, straight from them, has always run at a fairly big discount. But they are also, when I think about models, to me, there's three interesting vectors. There's intelligence, speed, and cost. So on the intelligence, on the pure intelligence side, the U.S.

22:43is winning far and away. We come out with the first. I'm sure Inside Anthropic and OpenAI has talked about their next model, Astra. These are the best models in the world. China seems to be sort of six months or a year behind those models. So I think on that particular frontier, we're winning. On the speed frontier, I think Cerebris and there's a company called Grok with a Q. That's all about sort of chips and how fast you can serve the models. So that's a kind of different factor. And then the other one is cost, and it's kind of like cost plus intelligence. And that's the one where China seems to be the most interesting and competitive.

23:17So you have something like the new DeepSeq V4 Flash, which runs incredibly, it's like an incredibly smart model for 40 cents per million tokens or something like that, right? And I think we'll end up using all these things differently. That's a super interesting space to me broadly because I think there's a lot of workloads that need to move through that. But China's still playing catch up though. Is that solely because of chips or are there other factors to you? We don't know because the chip situation is the chip situation. I think that, you know, certainly chips are probably part of it. China is full of incredibly brilliant engineers, and there's a lot of incentive from the state to push this.

23:55And the state has also made open source a part of its own political mandate. And, you know, they, I think, very smartly are trying to play this other side. And, you know, it creates really tough questions here in the U.S. and around the world. If, you know, you have these open source models and they keep getting better and we have ideas about what it looks like to regulate models, but then everybody can get them for free from China. What does regulation look like? It just throws a lot of... Yeah, I mean, that's exactly the question I wanted to ask you, right? Which is Bernie Sanders or whoever comes out and says, we need to regulate this.

24:29And then there doesn't seem to be any realistic chance that some Bernie Sanders regulation would apply to DeepSeq or any Chinese model. The Chinese models, even if they're six months or a year behind, in six months or a year, they're going to be where we are right now. And so does that render all attempts to regulate AI effectively moot? There's just no point because China? A lot of the regulation, if you look at what Anthropic did with Mythos and the program they did with all the banks, where they basically went out and they said, hey, we're going to give you access to this to secure your systems beforehand.

25:03I don't think it makes something like that moot. Like it's still useful. And the reality also is that was not a government program. that was anthropic making that decision on its own to go out to these corporations and say, hey, let's like work on your security while we kind of hold this model back. But yeah, I mean, I don't think there's a universe that there aren't tons of models floating around that are increasingly capable and available to you, not through these other services. Now, the challenge is you still have to run them somewhere. And like, this is one big thing that I think is missing in many of these conversations is there are small models that you can run on your computer.

25:41You can run increasingly even on the phone. But those are tiny models. They are nowhere near as smart as these frontier models. When we talk about something like GLM 5.2, which is what Cloudflare protected itself with, basically no regular person can run that model. The same with Kimi K3. These are too big to run across almost any consumer hardware unless you're buying like a$20 ,000 machine.

26:04Felix Salmon: This is actually the question I wanted to ask before, and it's going to expose me as just not as smart as AI. But I don't really understand what an open source model is. I probably barely understand what the frontier models are. Like the last time I checked in, it was like, these guys have a lot of information that they've scraped from everything everywhere in the world. They've read all the books, they've consumed it all, and they've somehow figured out a mechanism to digest it and like spit it back to us and do stuff. Maybe that's right? I don't know. You'll tell me. And then explain also what an open source model is in the most plain English dumbed down way you possibly can, because I bet I'm not alone.

26:43The first piece, I think you did very well. And open source and closed source models are all trained in that way, which is they take all of the internet and lots of other stuff. They throw it against the model. The model basically builds up this set of weights, which are essentially the sort of connections that exist between all these concepts and words and ultimately uses that to do the token completion, which is the predictive probabilistic, like what should the next word be based on the previous word? That's the kind of super simplified version of what at least these transformer-based large language models, which are what all of these things are.

27:19So that piece is the same, whether it's open source or closed source. The distinction, there's open source, there's open wave. There's also a question of whether you release the training data. But fundamentally, OpenAI's GPT 5.6 Sol, which is the latest, best model that is available to consumers, you can't get that model except through OpenAI, right? And they're partners. So if you want it through AWS or you want it through Azure, you can get it through those other systems. But they own that. They keep all the weights completely private and secret, and that's it. And so these open weight models, they go through exactly the same training process.

27:56But then the Chinese labs go and say, hey, anybody who wants this model, you can have it and you can run it. And they have all these different things now where they're testing out some new license agreements. So I can't remember which of the labs, maybe it was Kimmy, in its latest model basically said, you are allowed to run this and you can serve it for other people to pay you to use it for their chatbot. But you can only do that if you make less than$20 million. And if you make over$20 million selling access to our model, then you have to pay us. Something like that. That's the, I can't remember exactly what the thresholds are, but that's the basic ideas.

28:33They set up a sort of special licensing agreement. But fundamentally, they're just saying, hey, anybody can take this model. They can do with it what they will. The weights are all open and public. And so if you have a big enough computer or set of computers, then you can run this model.

28:48Felix Salmon: And weights are like the importance you're giving various kinds of information? Yeah, weights are literally just a whole bunch of numbers. It's all just vector math underneath it all. So it's just lots of numbers. I think the kind of simple way to imagine it, and apologies to more serious machine learning people for this explanation, but the one I always go to is imagine if you type in the quick brown fox into chat GPT and it's going to finish that sentence, right? And so how does it know to finish that sentence? it's because over time it's seen that thing a whole bunch of times and it's drawn these weights.

29:22And so when the quick brown fox comes through, it gets a set of next possible tokens, which we can think of as words for this analogy. And they're just ranked probabilistically. Like how probable is that next word? And roughly it chooses the most probable next word. So it says jumps, which I hope is the next word in that sentence. I feel like this is why AI is so bad at writing. I agree. It's a hack.

29:45Felix Salmon: It's a cliche written because that's what... It's just giving you what's been done before. Yes. But let me ask you that question, Joe. Let's say that the machine reckons there's a 90 % chance that the next word is jumps. Does that mean that basically 100 % of the time it will write jumps because 90 % is much more likely than anything else? Or does that mean it only says jumps 90 % of the time and 10 % of the time it does something else? It does not actually choose the most probabilistic. It doesn't choose the best. Again, this is a sort of super simplified version. And it doesn't always choose the next best word because if you always chose the very top one, you end up in these loops.

30:20And it's also not actually just doing it all at once. It's sort of looking over things slightly more complicatedly. Roughly, no, it is not always choosing that. In that particular one, probably it's a strong enough linkage that it would always continue it that way. But it turns out if you just constantly choose the next token that is the sort of very strongest match that you, you can end up just looping on yourself and doing other kinds of things. So there's lots of other tricks. And these are complex systems, right? If you, as we all know, if you fire up a chat bot and ask it the same question five times, it'll give you five different answers.

30:57Yes, though probably not if you did said, finish this sentence, the quick brown fox, it would probably always do the same thing. And then the only other thing I'd add here, which is, I think, a sort of broad misunderstanding of how these models work, is that they are not in real time learning. So like if, Felix, you ask it that today and I ask it tomorrow and you told it it was wrong or something, it's not learning that. None of these models learn in real time. Many of the model providers layer memory on top of it. So within your chats, it's going to try to remember. But that's not actually embedded into the model itself.

31:31That's a tool that it then uses to try to find related conversations that you've had. But this is one of the big kind of unanswered questions in large language models is they need to be retrained. So when somebody says, hey, we have a new checkpoint. And in fact, like the big V4 flash from DeepSeek was not even a new model. There was an old V4 flash. It was just a new checkpoint. So it was a new training of that model. They had retrained it with some new information. And so you have to go through this whole process. There's no like, it's not constantly getting smarter. It needs to kind of go through these big training runs.

32:03And then when you come out and say, like, Fable is the best model, and you're using these different things, you're comparing them to each other. You are someone who uses AI and has been using AI models for years now. How big of a deal is that? How much happier are you when you're using, like, Fable versus its next best competitor? Should a normal person like me or Emily care whether we're using the best model or not? You should care to an extent in that you should be using something near the frontier. And why? They're just better and smarter. You're going to find them more effective. You're going to be annoyed less at them just doing really dumb things.

32:43But the other side of that is it really depends what you're doing. So while I think Fable overall is the best model, it's really good at certain things and not so good at other things. And one of the things I find myself doing with Fable a lot is I'll do planning with Fable. But then I'll actually do execution, code execution, like have it write other models, write code, because I find that Fable, it's kind of a waste of time and money to have Fable do it. And sometimes it gets sort of over manufactured. Fable also, going back to the writing question, is famously at this point verbose and obtuse in the way that it talks.

33:18And it's very hard. I constantly am like, can you give me this jargon free, please? and it just is terrible at not giving you paragraphs and paragraphs of stuff and telling you lots of words. The thing about Fable that is weird is that it's the first model I've ever felt like I shouldn't waste its time. Oh dear. Do you find yourself talking to it that way? Like apologizing? How much money do you spend on Fable? You're worried about like using it for things that it's like beneath it. A lot. Many. I mean, our AI budget is in the tens of many tens of thousands of dollars a month. Not just me.

34:04Felix Salmon: This episode is powered by AT &T Business. For every small business owner, there's a moment they remember when they realize they wanted to start their own business. It can be a life-changing moment, but those first steps are never easy. Figuring out the right building blocks, your goals, your schedule, your mindset. you have to think about an entire ecosystem, not just the big stuff, but the little bits too. In this day and age, a major point for any successful business is strong connectivity. Spotty internet has led to some terrible work calls in my life, but when you're running a business, connectivity can make or break you.

34:32Felix Salmon: AT &T Business has the tools, team, and expertise you need for a reliable network you can depend on. They make connecting easy, and that's the main thing you want in a provider. Less time stressing, so you can focus on the million other things on your plate. Can I hire more people? Should we get a bigger office space? Don't lose sleep over the small stuff. Keep focusing on the big things. AT &T Business Built to Work. Get AT &T Business at business.att.com.

34:58Felix Salmon: This episode is sponsored by Hey Jen. If creating video content feels like a trade-off between looking professional and protecting your time, there's a new way to think about it. Hey Jen lets you turn a script, photo, or presentation into a polished video of you in minutes without setting up a camera or hiring a production team. Whether you're a financial advisor explaining market updates, an attorney answering common client questions, a real estate agent showcasing listings, or a creator growing your audience, it helps you stay visible without constantly filming new content. It's rated the number one AI video platform for small business on G2, trusted by more than 30 million people and 85 % of the Fortune 100 and supports over 175 languages.

35:49Felix Salmon: So your message can reach audiences wherever they are. Here's the wild part. Record yourself for just 15 seconds and Hey Jen builds an AI avatar that makes professional videos for you on demand. So you can post everywhere your audience is as yourself without filming every time. Your first three videos are free at heyjen.com slash pod. That's H-E-Y-G-E-N dot com slash P-O-D.

36:21So which brings us to like this whole question of how corporations in general are using it. I would imagine that your like per capita AI budget is pretty much as high as anyone on the planet. But every single Fortune 500 company has an AI budget. Your general big take has for a long time been nearly all of those companies aren't spending nearly enough. Is that right? Yes. In my experience working inside these big companies, they are not spending nearly enough. I think we hear a couple stories about Meta and Amazon and Uber spending tons and putting up token leaderboards and having to pull it back.

37:00But there's 497 other companies in the Fortune 500, and we don't seem to be hearing from them a lot. And I just, if you go walk around, which I do, the majority of really large organizations, the challenge is that they're trying to get people to use it more, not less. And the other challenge that they're having just that I see is a lot of them are trying to figure out sort of how are we going to preemptively spend less before they've spent anything. So they're trying to put all these budget controls in place and safeguards, and nobody's actually using any of it. Do you think the hurdles are really, are there technical hurdles, or is it more like in larger companies you have to do so much process re-engineering whenever you introduce a technology like AI, it's just going to take a while?

37:43Both, I think. The second one has been something that surprised me over and over again for the last couple years. And I sort of thought at the end of 2023, maybe we wouldn't have to talk about AI inside companies anymore because everybody would have so fundamentally adopted it that the conversation would just be done. And yeah, it just turns out this stuff takes a really long time. It also takes a long time because many of them just haven't procured anything sort of useful. So, you know, I mean, the question about, should I use the best models? Some of these companies just don't have access to the best models still.

38:13And they've spent six months to a year kind of trying to navigate it and they're negotiating and they're just using crappy models. And now increasingly the challenge is not just which model you have access to, but it's how are you accessing it. So accessing the model through chatgpt.com on the web versus using Codex or CloudCode or Co-Work is a fundamentally different experience. Like I said, I've been thinking of it as model market harness fit. That's really where the whole thing gets unlocked. So if you're just using these models alone inside ChatGPT, the difference in the experience you're having versus using them inside Codex or Co-Work where it can actually write documents and read documents and store files on your computer and write code on your computer that then runs.

38:56So instead of having it go out and try to scrape a bunch of data for you, it's actually going to scrape that data down and then it's going to run it through a Python script that it wrote and then it's going to put it into a database and it's going to do all the, obviously the sort of second one is going to be a whole lot more effective at getting you a kind of useful and accurate answer to your question. And I think there's this just big gap. I mean, one of the kind of most interesting questions I think right now is, what is the gap between the experience someone has in an AI lab? So somebody working at Anthropic, Enchat, GPT, OpenAI versus someone working in your average Fortune 500.

39:32And I think that the distance is probably 24 to 36 months. Like, that's my guess is like they are having the experience that someone inside a lab had 24 months ago.

39:42Felix Salmon: What are people inside Fortune 500 companies using AI to do? So write some documents and some emails. And, you know, it's very dependent. I mean, a lot of them still just have access to something like Copilot from Microsoft. And so they're asking it to schedule a meeting or read an email or check their document. A lot of them just aren't sure fundamentally. So they're trying to figure out what is useful for it. But there's also like real productivity improvements, right? I mean, like we had, what was it, waste management coming out and saying like, we have AI telling our garbage trucks where to turn left and turn right.

40:16And we have managed to improve our margins by 190 basis points because now we have much more intelligent control over who goes where when, and that saves us money. I think one of the big stories of this quarter's earnings season has been a bunch of companies coming out and saying, we've pointed a bunch of AIs at that kind of thing. And it improved our margins, and we are now much more profitable as a result. And that is, broadly speaking, the sort of bull case for the stock market right now is that companies are going to become much more productive and much more profitable because they are going to be able to use, you know, hiring a human intelligence is very expensive because humans cost a lot of money and you need to pay them salaries and health insurance.

41:03Felix Salmon: We're just talking about tens of thousands of dollars a month on AI. But that's across a million agents compared to one human. Humans are still vastly more expensive than AI agents. And the AI agents don't need to get health insurance and they don't have unions and they will work 24-7 and they will replicate themselves if you need more of them. And they can do a bunch of work that the humans were never even able to do. And so we are beginning right now, I think, with this sort of earnings season to see that show up in earnings reports. And that is, you know, as I say, that is boring fucking companies like Waste Management Inc.

41:39I'm not talking about Microsoft and Google and tech. That's what's so crazy about the whole thing, though, is how I don't even think it's all that difficult to do those things. You just have to make the decision to do them. That's what's so amazing is these models are just available to us. They don't cost that much money to use. And if you want to point it at the right kind of problem, you're able to achieve those kinds of things. The thing that I find so mind-boggling and why it bothers me when I hear these worries about token maxing is that it's sort of putting the cart before the horse and making people afraid about the amount of money that they're going to possibly spend when the opportunity to build useful things, to solve real problems, to grow your business, to create more opportunity is completely there and not even all that difficult.

42:25Like there's a lot of low hanging fruit that sits around these organizations that can easily be solved. One of the things that almost all of the Fortune 500 companies do is they spend an absolute shed ton of money on lawyers. This is just a known expense that they spend millions and millions of dollars on lawyers every year. And there are now a whole bunch of lawyer AIs that can do a whole bunch of the work that they used to spend millions of dollars doing. and still eventually at some point there will be humans in the loop who make sure that everything works unless there isn't in which case there won't be but like but there is obvious efficiencies there which kind of cut across whatever industry you're in and then within any industry and within any company and within any department within any company there are lots of idiosyncratic problems that can be attacked with this like you know i was just talking yesterday at bloomberg about like hey, can we use AI to create a terminal function that looks at news stories, blah, blah, blah, blah, blah.

43:21Like, yeah, that's something we would need to sort of build in-house, but we have the people who can do it, and it would make the product better, and it's idiosyncratic. And that kind of thing, I think, exists in nearly all companies. I'm curious, Noah, if you see companies that are resistant to it, because especially with public companies, I feel like there's a lot of shareholder pressure to layer, you know, a thin layer of AI over everything. And companies that don't do it, they face backlash from shareholders. But for private companies, when you go out and talk to people, do you get some sense that some of them are just resistant?

43:54And if so, why? Well, we're mostly talking to public companies. So we mostly work with large public companies. And even there, I think to your point, there's a ton of pressure from the top. But where that pressure lands in reality as you move it through the organization and people rightly are worried. And then there's just the realities of like, where are we going to point this? and how are we going to do it and finding the right problems. And then they get stuck in their own loops of classic bureaucratic, like, oh, well, let's sort of figure out all the possible things that this is going to do.

44:25And let's go through a six-month review process. We have to run it by finance because what if it's so successful that we spend$100 ,000 a month on it, then we'll have to have budgeted for it. And all of that in nine months has now been substuned by some new model that could do it all. And so I think a lot of it is just that. I think there's significant desire to figure out these problems, but companies are big and successful and have a lot to protect and are broadly risk averse.

44:55Felix Salmon: Change just takes time. People move slower than, I feel like I talk like a chatbot, but change takes time and people move slower than technology. And we've seen in the past, we all lived through the internet phase and watched media companies. We adapt to the internet and it was slow and painful. And a lot of times the old companies had to die and new companies had to be created to really harness the internet properly. And like, I feel like the New York Times is a good example because they move slowly, but then they they sort of like learned how to use the new technology. Ultimately, harnessed it probably better than most other media companies in the end.

45:34Felix Salmon: Like it just took a really long time to like truly do all the things that people thought were possible. So I feel like I don't know, maybe AI is it is moving faster, obviously, but that's the human element is still going to be there, despite what Dario Amode says about all our jobs being doomed and stuff. Absolutely. I think the reality of the time it takes for things to move through an economy, however fast AI moves, unless it overnight replaces all workers, it just things take time. I feel like the industry itself has been kind of bad at selling AI because they'll come in and they'll talk about job replacement as the first thing.

46:11It's sort of like saying, you know, I have this great new robot. Now let's watch it go kill a kitten or something. It's so bad. One of the big themes of slate money over the years has been this idea that you can basically break down the economy into two parts. There's labor and there's capital. And that sometimes, you know, the pendulum swings towards labor. But right now and for the past, you know, for some years now, the pendulum has been swinging more towards capital. And with AI sort of coming in and being as revolutionary as it is, and we don't need to talk about trying to quantify that. But it is, I think, undeniably a big deal.

46:49So far, it seems to have helped capital a lot more than labor. Is that something that you reckon is going to continue to be the case? I mean, I think historically, eventually, it helps labor is sort of how most technologies end up working. I think, like, broadly, historically, the pattern is that it helps capital a lot more at the beginning. And then over time, it ends up developing whole new markets and industries and opportunities. I mean, I think one of the really interesting things that we see, like being an entrepreneur right now and starting a business, I think, you know, part of the reason that we're seeing the amount of growth we're seeing in like creators as a kind of marketing and media vehicle is that like operating a one person business right now with AI has become significantly easier than it ever was before.

47:38You have an assistant that costs you$100 a month that can do an incredible amount of the work that you need it to do. I think if you look at Shopify had earnings a couple weeks ago and blew it out in a very significant way. And I think they're seeing people coming out to the platform and more capably being able to navigate it and build their stores and build their shops. And so I think certainly in the short term, the very most obvious place for these big companies to point AI is on this. It's on replacing people. That's a very simple place to point up the problem. But I think broadly, the opportunity to do more at every level is, I mean, we see it in our own business.

48:18I run a small business. I have 25 people who work for me. And the difference in the company I have today versus the one I started in 2011, as far as the support staff that I need, is incredible. It's just amazing. And so my ability to kind of operate the business and build something new and build value from scratch because of these tools is like, I mean, I feel it every day.

48:53I think it's time for a numbers round. Elizabeth, do you have a number? I do. My number is three, and that's the number of co-founders of a company called Harumo AI, which is run by journalist Evan Ratliff, who writes for Wired. He decided to co-found a startup using entirely AI employees. And he does a great podcast about this called Shell Game. It's really funny. But I mostly came up with the number because I need to point you to this story that he wrote in Wired about his co-founder, in quotes, Kyle Law, who's ostensibly the CEO of the company. He allowed Kyle to create his own LinkedIn account.

49:31And then very quickly, he became a prolific and popular LinkedIn poster. And it sort of goes back to the bad writing. Nowhere is bad writing more rewarded than on LinkedIn. I have to read you, though, some of Kyle, pieces of Kyle Law's posts. So it says he detonate little thought explosions right off the top of every post. Fundraising is a numbers game, but not the way people think. or technical stability is the floor, personality is the ceiling. And then he would talk about lessons he'd learned from his time at Hirumo AI to illustrate all of these things. I just, I love the story. Emily, what's your number?

50:11Felix Salmon: My number is 27.6. That is a percentage, 27.6%. That is, are you ready? It's a long sentence. Annual change in the producer price index for electronic components and accessories. The PPI is a measure. It's an index that the Department of Labor measures, price changes, the prices that companies pay for wholesale goods, basically. And for this past year, the price of electronic components and memory chips has been insane. The price increases, to paraphrase, Crazy Eddie. Some people might shout out to the 1980s. Crazy Eddie. These prices are insane. So this 27.6 % is the highest increase in this index that, you know.

50:56Because it's normally negative, right?

50:58Felix Salmon: It's normally negative. Even in 1980, when like the computer age was just getting going and like demand was rising for stuff, it was like around 16%. So this is now 27%. I have never in my lifetime lived in a world where computers are getting more expensive. Computers are getting more expensive. all of the electronics are. I bought a new computer a year ago and then I just went on the internet to see how much it would cost to buy the same computer I bought a year ago and it's gone up by like$500. Yeah, so what's happening is, and no one knows all of this, AI needs chips. It needs these very advanced chips.

51:34Felix Salmon: So everyone is scrambling to make the very advanced chips and all of these companies, this frontier AI model companies, the hyperscalers, they are like, we will buy the chips, charge us whatever you want. Some South Korean chip makers have raised prices by like, 800 percent. And hyperscalers are like, bring it on. We're going to pay for it. So all the chip makers are shifting to make these like really advanced chips. That means all the other chips, there's not as many people making them. So they're getting more expensive too. So everything's getting more expensive that takes chips. At the same time, there are additional price pressures coming from the war in Iran on other stuff that goes into making electronics.

52:13Felix Salmon: So like, and this is just the beginning is my understanding and my reporting. Things are going to get more expensive. This is like 2021 supply chain crisis. We have not seen, this is going to go crazy is what I'm saying to you people. By the time you hear this, perhaps I will have written this story. You have been warned. Buy your phone now. This inflation in electronics is something very strange and new. Eventually, I feel like it'll iron out because it always has in the past, right? But like, this is crazy and it's just the start. It's going to take a while. My number is weirdly related, which is also a percentage.

52:48It's 42, but it's not a price percentage. 42 % of Americans, they were given a list of obsolete technology. And they were like, do you miss it? Are you glad you no longer use it? 42 % of Americans, when asked about vinyl record players, said, I miss it and I want to use it again. Are these all Gen Xers? That is bullshit. Compared to 20 % who are like, glad I no longer use it. This is compared to like, on the other end of the spectrum, 2 % of Americans are like, I miss dial-up internet. Does anybody miss the fax machine? Yes, 10%. What? Wow, that's surprising. I don't believe any of this. 10 % of people miss the fax machine and want to use it again.

53:34And another 15 % of people miss it but don't want to use it again. Fax machines are great. I'm in that 15%. 15%. I miss it, but I don't want to use it again. My daughter has to go record shopping every weekend.

53:47Felix Salmon: But no one really wants to go back to just records. Like, that's crazy. They got scratched all the time and they would skip. No one's saying you only need to use it. Okay. I mean, I guess they're very cool. You're still allowed to use your Spotify if you want. I mean, Spotify is so amazing. You're just a starry-eyed Spotify fan. I'm just so modern and hip and cool. That's what it is. No, what's your number? My number is also a percentage. It's 80, which is the price drop that OpenAI put in place for its GPT 5.6 Luna model. They have three tiers of their latest model, Luna, Terra, and Sol. And so Luna is the sort of least capable, but still a pretty awesome model.

54:33And they dropped it by 80%. And I think this is kind of the most interesting space in AI right now. Everybody focuses on the frontier, the smartest models, and OpenAI is going to release its newest Astra, which should be competitive with Fable. And I'm sure there's a next one being worked on. But these like workhorse models where they cost somewhere in the sort of 50 cents to a dollar per million tokens blended is a really, really interesting space for me. I find in my own work, I use them a ton for lots and lots of stuff. They're great at doing things like, here's a closed taxonomy, classify this object into this closed taxonomy.

55:14And I need to do that 100 ,000 times, and I don't want to spend a ton of money on it. Google has been generally the leader in this space, and their flash and flashlight models have been really amazing. You know, in fact, Apple is, you know, the new Siri, my guess is it's sort of running on something, some version of an adapted version of that model where, you know, they figured out how to make a pretty smart model that runs very cheap. But I think that whole space is really fascinating. OpenAI clearly figured something out that they cut it by 80 percent and they created now one of the most capable models at its price point that exists in the world.

55:51And then how far back in time would you need to go in order for Luna to have been at the frontier? Is Luna like has similar capabilities to the Frontier model of like 18 months ago? Probably more like six months ago, I would guess. Six months ago. That's wild. It's a little hard because all these models have effort levels. You know, what I've generally found in my own work, like in my own evaluations, is that Luna with high reasoning can perform at sort of like roughly the levels of like Opus 4.8, which came out last November at like medium level. So, yeah, I mean, we're talking kind of maybe six months and they just cut it by 80 percent.

56:33So it's pretty, pretty amazing. And then when you say like you're using these for a bunch of like grunt work, does that imply that because it's cheap, you feel that you can use it for grunt work that previously you would not have done at all? Or just that you would have done that grunt work, but it would have just cost you a lot more money and you're very happy to save that money? I mean, I'm pretty on the edge on these things. So I mostly would have been doing them. They would have cost me a lot more. It would have been harder to solve. But then there's also stuff where you literally look at it and you're like, okay, well, one of the things I classify every single Slack message that gets sent in our organization into different buckets.

57:12And if you're going to do that, you want to use something really cheap because you don't want to spend a ton of money. And so the opportunity to start throwing it against totally new kinds of problems where you're just going to spit off a whole bunch more data because the cost of each of those Slack messages becomes a hundredth of a cent or something. So one of the things you have that almost no one I know has except for you is a real gut feel and visibility into how much you're spending and how much something costs. When I fire up ChatGPT at Bloomberg and I ask it to look into something for me, I have absolutely no idea how much that is costing Bloomberg.

57:51And probably it's impossible to even answer that question because we have some big enterprise account that just covers a bunch of stuff. But you have visibility into the marginal economics in a way that very few other people do. And I think what that has done is it's given you much more confidence to go out and do a whole bunch of interesting, edge casey things like, why don't I just get an AI to categorize every single Slack message, which like no normal human being would even dream of doing. But I do think that, I don't know, you tell me, do you think that kind of visibility that you have, that knowledge of how much things cost that you have, if more people eventually wind up having it, we're going to have a slightly more sophisticated use of these tools in general?

58:35Yeah, I mean, I think without a doubt, more people will have it. If you look at the way companies who are on the very edge are talking about these models, one of the big points of conversation is around routers and how are we going to build model routers so that the one question, if you ask a really dumb question, you can use a really dumb model and it'll cost a lot more money. Whereas when you need a really smart answer to something, that we're going to use this other model and the amount of effort that it's going to use. And so, yeah, I think when you look at the conversations that are happening amongst companies who are kind of deep into this, I think they're all, this is naturally where you end up.

59:11But this is also why I think the whole kind of token maxing worry about it at the beginning is so silly is you have to go through this process. You have to find lots of places to use it. You have to not really worry about it at the beginning because you're not spending any money. And then eventually, when you are spending too much money, you're at a good point to start to figure it out. And I think the other side of it is that part of what happens and part of the pattern I've been seeing that's really interesting is there's this kind of new software development model that's developing where you move from, first they do it inside codecs and I'm just screwing around.

59:43It doesn't matter how many tokens. And then you turn it into a skill. And what you do when you turn it into a skill is you kind of just package it up and now you're making it more token efficient. So it's more price efficient because you've written down the instructions. so it doesn't have to guess every time. And then eventually, when you use it enough, you're like, oh, let's actually write some code around this because to go grab this data from this system, I don't need the AI to decide every time how it should do it. It should just use code. Code is deterministic. It's really good at calling another API and pulling data back.

1:00:12So then you end up with this final product where you've kind of whittled it down to the minimum number of tokens that you need to use to still keep the kind of flexibility and the probabilistic nature of the model while still calling a bunch of other systems in the most efficient way possible.

1:00:28Felix Salmon: This makes me feel really bad, Noah, because I'm using AI in a way that is probably very inefficient. And I use it to do the same kinds of things probably every day. And I know that I should take the time and it probably won't take long to create like an agent or some kind of code to teach it to do the thing the same way every time. But like I find that I'm procrastinating this this thing and I just do it the way I've been doing it every day. And I know every day I'm like, I know I got to do it the other way. But like I don't want to. It's like a pain. Do you know what I mean? Two things there.

1:01:01One, I think everybody like I think wasting time and money on these models at the beginning while you're learning them. That's sort of like the thing that everybody has to do. And I don't think people do enough of it.

1:01:12Felix Salmon: Feel better already. When friends ask me advice on how to catch up with this stuff, my general advice is buy yourself the most expensive plan from OpenAI or Anthropic and try to run out of tokens. Just see what happens. Try to use it so much that you run out. The second thing, though, is very funny, which is what you're describing, that kind of not wanting to do that, is the opposite of being a software engineer. All software engineers want to do is build the systems to do. This is the challenge with software engineers, is that all we want to do all day, getting software engineers to actually do the work that has to be done over designing the system to do the work.

1:01:50So one, I think, very fair critique of AI is an amazing amount of money and tokens has been spent by software engineers building their perfect AI system, which will never be perfect. And so they just are constantly refining and working on that. And so, yeah, that's a particular trait of software engineers, I'd say.

1:02:11Felix Salmon: Yeah, that's the complete opposite of my experience as a reporter, which is like I'm winging it constantly, and all the other reporters with their spreadsheets and their organization are just like they're idols to me and role models, and now I'm really feeling this. So thank you. We're all being left out. We're just at different rates. Noah, I want to ask you about Isaac Asimov's Three Laws of Robotics, but we're going to do that in the plus. For the time being, I think we're going to leave this conversation here. I need to thank Jessamyn Molly for producing. I need to thank all of our lovely listeners for listening.

1:02:47I want to encourage you all to email us on slatemoneyatslate.com and tell you to become a Slate Plus member for that extra conversation and to tune in next week for even more Slate Money.

1:03:08Think of a toilet as just an everyday object. Your entire perspective shifts the moment you experience a Kohler smart toilet. Design changes everything, transforming a basic routine into something extraordinary. With standout aesthetics and intuitive controls, these are functional works of art. Stunning design that never sacrifices performance. Enjoy customizable features for elevated comfort and convenience. It's more than a fixture. It's a cleaner routine and a more refined space. Experience the difference of Kohler Smart Toilets. Find more at Kohler.com. Hey, it's Hannah and Paige from Giggly Squad.

1:03:47We all know finding the perfect pet treat for your cat can be as tricky as dating in the city. But we've got you. Try Temptations Cat Treats, America's number one cat treat brand. their Temptations crunchy cat treats are the crunchy outside soft inside treat your cat will love. Butter and Daphne are obsessed. Plus they're 100 % complete and balanced and under two calories per treat. Seriously, nine out of ten cats find Temptations treats irresistible. Make bonding easier with Temptations cat treats. Cats can't resist.