Hater Season: Cal Newport on AI Reporting

the tape. Tell me a little bit about what the bits that you found, because I watched it going yeah, yeah, yeah, like an angry person.

But I also write about technology. My main journalistic home is to New Yorker, where I write a lot about you know, I do AI journalism there, so I'm up on that as well. And so I'm often noticing things in journalism. When I see AI coverage that is faulty, it really catches my attention because I have a foot in both of these worlds. So there's a lot of good AI report out there. There's also a lot of trash, and I really wanted to help people figure out how do you sort it? Like how do you figure out

if you're reading something? Should I pull the ripcord on this article? Like this is not helping me, Like how do you know if it's good or not? So I was like, Okay, here's what I'll do is I'll come up with the three most common traps I see in AI reporting that makes me want to, you know, throw my iPad at the wall. And I came up with three and I'll just give you the three names. I made up. All these names don't. I don't think they're great. My producer thinks he loves them. But here we go

vibe reporting that's number one. So that is where you will omit certain facts and put loosely related quotes next to each other in a way that creates a general vibe that you want to be true but it's not quite true. So you don't actually make a concrete claim that's not true, but you imply that claim by what you omit or what you put in your story, or

what quotes you put next to it. So you'll put a quote, for example, about layoffs at the gaming division at Microsoft, next to an unrelated quote about concerns about AI and it's impact on jobs. Now you have the vibe, oh, man, all these people just got laid off because of AI. AI is taking jobs, where in reality the layouts had nothing to do with AI. But you put these things next to each other, you give that sense. Then I

had a mining digital ick. So to me, that's any AI story where you take an example from the edges of AI, like something that wireheads and San Francisco are up to, and you just tell a story that's unsettling without talking about any of the technical details like well, what's different here? Was there a technical innovation we need to know about, and not discussing any concrete implications. Oh, this means this is going to change in the future, or it's going to have an impact on this sector.

You're just telling a story to unsettle, and I think a lot of the coverage of molt Book and open Claw fell into that. And then finally this astonishment, which is more of a YouTube phenomenon than a print journalism phenomenon. But that's where every single thing that happens in AI

is insane, amazing, terrifying. Everything is going to change, and so you constantly create this atmosphere of something seismic just happened, so that the consumer of the infra ends up in a bit of a panic, like I can put my finger on exactly what's terrible, but like everything terrible is happening. Those three traps, to me, should be automatic rip chords from what you're reading or watching.

but I'm being nice today for some reason. All Right, you're a dip shit, because you are a dip shit. If you look at claude cowork, which is a thing for fucking around on your desktop, and you say this is going to compete with Salesforce, you just don't know

what you're talking about. You are wrong. But then one abstraction higher is this idea that claud code is going to destroy SaaS software as a service, and the idea being that software as a service is this thing that people are just going to stop building their own CRMs. They're going to stop building their own per seat software things, but they're going to build it internally. This just I don't know if you've seen this, cow It just reflects

a complete lack of understanding of how software works. Yeah, because you don't pay Microsoft for Microsoft three sixty five teams because you can't build your own word or what have you. I don't think you could. But nevertheless, it's also because they maintain it, because they make sure it doesn't break, or if it breaks, they fix the bit. So they make sure that it stays up all the time. They make sure it's accessible that has secure log in.

other command line interface coding tools became like that. For programmers, it was like really fun the watch it doing multi step execution of the construction of demos or this or that, and the reason why it could do those cool demos it was sort of well suited for that world because that's a world that exists only in text. So cloud code works on a command line interface. It's all text based, and it works with a filesystem. You can write files,

edit files, sind files to compilers. It's all exists in a small number of commands on a command line interface. It's all in a world of text, because that's where computer programs are built. That's a perfect case for lms, which love dealing with text, and they love dealing with

structure text like computer code. There was an extrapolation that then happened, and really this caught on January of twenty five, which is where you first began to get this sentiment of oh, it's doing such cool things over in the

world of command line interfaces in code. Certainly these agents can now soon do similar cool things and like all different things we do on a computer, and that is what laid this foundation of people were so impressed, programmers were so impressed by the coolness of what was happening with cloud code and the other command line interface tools that they extrapolated that vibe over to other computer usage.

Is the point of that article I reported was, Oh, it turns out like everything else we do on computer is much harder. It's not six text commands on a command line interface creating structure code that you can compile and test to see if it compiled or not. It's much messier. The interfaces are visual. I don't realize what complexity goes into the things we click and select doing something as simple as even trying to just book like a hotel in a new city or something like this.

And if you use a language model as to underline logic and decision engine of making actual actions in the real world, well, the language models make things up and get things wrong or a little bit wrong twenty percent of the time.

we're still vibes. This is Vibe reporting. Yeah, but cloud code is like really exciting, and so why can't we do that with everything else in the computer. It's actually a much harder problem than they were letting on.

just like a it's a project management tool. It's like I made some software that worked. This is everything now. But that's kind of what you've been talking about with this Vibe reporting, where it's like I did a small thing. Now all things will be done in this manner, whether it's possible, God no, God no. But she she, like many reporters, are able to find a lot of people who are invested in AI, who will absolutely go on TV and say, Yep, it's completely true, that's gonna happen

one hundred percent. It's just so strange because it makes me feel paranoid and kind of conspiratle when you look at so the majority of news about AI, and it is this viper POI. It's these vast extrapolations from sixteen thousand job losses Amazon. They mentioned AI. This plus this equals the AI is replacing people. It's just so it makes me feel like uncomfortable with the world.

of AI. They're being replaced by AI. They put the subhead of the article was the CEO of Amazon talking about how AI is going to increasingly disrupt the job force, and then the article itself, no alternative explanations are given for these layoffs. They kind of just give the details of like here's how many people are laid off and here's where they figured it out, and then they put a couple of quotes in there about AI being very

disruptive and being able to automate jobs. It turns out those layoffs had nothing to do with AI, and you can find other reporting that focused because it was reporting that was for the financial market, so it was trying to focus on what the hell is actually happening, and the deeper reporting was like, yeah, they laid off a bunch of managers because like a lot of tech companies, they over hired during the pandemic because cloud computing became much in demand during the pandemics, So a lot of

tech companies over hired during the pandemic and they're all shedding those jobs again. And Amazon's pretty ruthless about this, right, They're always looking for excuses to fire people, and they said, we have too much bureaucratic bloat, there's too many managers. We're going to fire a bunch of these managers we hired so that we can be more lean. Again, that

has nothing to do with AI. In fact, this is the second or third round of these firings that they've planned to do, the first round occurring before chat GPT was even released. Like, this has nothing to do with jobs being replaced by AI. But then you can vibe report it because like, well, technically speaking, Amazon also is investing money in AI products, So technically speaking, money saved by firing these managers could be respent on AI. So you can say with a like a semi straight face,

they fired people because of AI. But clearly, you know, the impression that you're giving to the reader is that they were replaced by AI and it had nothing to do with it. And I heard, by the way, so I wrote about that. I put in that video. I've heard on background from multiple Amazon executives since they were like, we were completely baffled by this coverage that was implying that people were being fired here to do with AI.

This is just what we do at Amazon. We're ruthless, Like if you're not earning your keep, we fire you. They were completely baffled by that coverage and like things for pointing it out, like what.

were pre pandemic. Consistently, across the board, these cuts are vibe reported is due to AI consistently. You see exactly that story. We see it, and I'm a computer scientist. We see this in coverage of computer science majors as well. Same idea. Computer science majors historically directly tied to the tech industry. If they're cutting in a downside, majors go down.

If they're hiring. I mean, it's just economic. It's not surprising when the technistry is booming, we get a lot more majors because they're good jobs, right, And so computer science majors went down in the last year or two as the tech company started cutting. That was reported in the Atlantic as kids are not majoring in AI because AI makes the the.

about it, and it's just over the whole days. Alex lieberm And had an idea what if he could create Spotify rapped for his text messages without writing a single line of code. Liberman, co founder of the media outlet

morning Brew, created I message wrapped. I just want to start here and say that guy doesn't do any fucking work, and I know people who work there like I just want to start by saying, if that's the best you've got, and that probably involves throwing a giraffe and the entire zoo into a vat of acid to make the GPUs move.

You're meant to read this, and the deliberate effect you're meant to have is you're meant to be scared and excited like it is a kind of a mishmash between the vibe reporting and the I guess it's the folksting. In fact, this might be a triple score because this is meant to feel you, make you feel with discomfort, but also make you freak up but also be excited, a rare triple score. It's just stories like these pissed me off because I'm fine with people going this could

do this. I don't mind it happens, But when it's just like, hmm, feed me the slop right into my mouth, asshole, make me feel, make me feel all the marketing things all at once. First of all, I'm not impressed by the idea of doing I message s grapped. That's not That's not a thing a human being with that with like friends and hobbies does. I've been I've been watching the first season The True Detective. I've got many more

shows I could watch. Never once would I think what if I could get it wrapped of all my messages? What a fucking psychotic thing to do. But when you read this, it's like yours. You spent your holidays with your family, wrote one tech Polticy expert. That's nice. I spent my holidays with Claude code.

It's a fun hobby I can put. Look at this like I made a thing that can Like I was talking to a friend of mine yesterday, computer scientist, and you're like, oh, I built a thing where I can, you know whatever, email an appointment to a thing that gets parsed by an LM and then goes to my calendar. That's just fun for him, in the same way that someone else might be like, hey, I built a cabinet and like it's really nice, Like look, I got the

wood to go together, and like I'm proud. I could just on a cabinet or whatever.

so they know who to charge. You could just steal everyone's API key who was using it because the guy just vibe coded it and so no one was actually there looking at the code. But what but what I think is going on? Okay, So here's what I would like to see. I would like to see more reporting. That would be how are people using X Like to me, that's very interesting, yeah, right, how are.

Most people aren't interested in that, but for some people they are. And you know, I think it's fun. Computer programming is fun, right, so like the ability to make a program work is and some of those tools are useful. But this is that's not a major industry on its own. You know, I'm designing a tool for my small company that makes it easier for us to do whatever. That's cool,

but that's not like a trillion dollar industry. I can't get my arms around yet exactly how professional computer programmers are using it. They're all talking about it, really yeah, but I can't get I can't get my arms around it yet. Yeah, what do you what's what's your sense of.

no one actually says. There's always these these kind of vibe store, the vibe tales, the mythology where it's like I had a problem vaguely that would have taken me X number of hours, but when I used claud code, it solved it immediately and caught two bugs that I didn't know about. It's very Marine Todd's style. Then everybody clapped yeah, and it's you were a computer science teacher and you you don't know either, which makes me think it's not as big a deal as people are saying.

My other bit of evidence is that at the end of last year, Anthropics claud Code revenue was one point one billion annualized, so about ninety to one hundred million a month for a revolution. That feels low somehow.

going to organize information and efficiently access it? When performance comes into play, there's a lot of really subtle decisions to make about, you know, how am I going to store use things in such a way that we don't get bogged down when we're trying to execute things. It just becomes I'm I don't code as much anymore. Exo'm a theoretician, but I did my whole life since I was you know, seven, And it's a there's a it's

an art form. Right When you're using cloud code, you're not really supposed to look at the code, and so I think that takes a lot of uses probably off the table. It's like the use cases. You're supposed to have these different instances of cloud code running, and this one's going to write code, and then this one's going to write test for that code, and then the cloud code is going to run the test and then try to fix the code if it doesn't matches the test.

But your eyes aren't on the code at all. And there's I mean, obviously for a lot of programming, that's an issue. But then the other thing I've heard about the computer programming industry is it's very stratified. There's a smaller number of like really good, serious programmers that produce like ninety percent of the really important, valuable code on which everything runs. I don't think they would touch cloud code with a ten foot pole, like they're good at

what they do. And then you have these like huge strata of people writing like JavaScript and sort of hacking together Python, and you know it's like not very good code. And then it's I guess you could replace some of the functional it's functional enough, but I can Yeah, I can't get my arms around it yet. But I do get you know, I have a lot of sources, so I hear from you know, I do hear from people that are talking about how cool cloud code is. Is cool.

I hear from a lot of professional programmers that are like, yeah, we don't use we can't use this. We're trying to write serious programs, like we have to sell this software, like this is this is not solving a problem we have or you know we so you know, I don't know, But the problem is that that's what the reporting should be. Hey, here we are at this company looking over the shoulder

of people. What are they doing. Let's talk to the engineering teams on background of this tech company exactly what role is going on here? And I think what a lot of reporting has become on AI is your hype laundering.

So you look at the discussion about the technology happening from more engineering minded people, you convince yourself as a reporter, I can't understand the engineering, but I will trust the people who do, and then you launder what you're sensing from that hype into your articles, not realizing that like nerds like me, we get hyped up about stuff, and we get super excited about stuff and we go create like, yeah, you can't just launder our hype into this is what's happening.

And so it's like reporting on a war where you have no one embedded, there's no one actually on the ground where the battles are happening. You're just responding to the press conferences that the generals are holding, you know, back in the Pentagon. Like it's not a way to report on what's actually going on.

But my thought is as well, is there's probably a problem with rationalization as well, because if you look at this and you say, Okay, well it's kind of cool, it's fun in whatever indeterminate way, it doesn't seem like serious software, like actual real deal software is being made with it. But then the CEO of Google says thirty percent of code is written in AI, which is bullshit.

And I've heard from so many software engineers, well, it couldn't be that everyone's just wrong, right, It couldn't be a case of that everyone is making the most egregious capital expenditure fuck up of all time. This will be historic, I believe, worse than railways, digital beanie babies, but done at the scale of laser tag arenas. Now, that can't possibly be that, because everyone else is saying this is

exciting and good. And at that point they choose instead of being like worried, instead of being a bit anxious about this, they say, well, Amazon Web Services spent a lot of money, so this spends a lot of money money too, so this is actually good. It's actually good. And indeed these people seem emphatic and excited, and I, as a non coder, can build a fudged CRM that probably would not withstand even the laziest haacker. I can do this, and thus it will extrapolate further from there.

And what sucks is what the people that I believe actually will be hurt by this are retail investors. I think, regular people buying stocks in these companies or selling software stocks because they believe that color code will replace them. And altomately, I think it's just gonna be a blood bath. For people's furrow. One case that could have been avoided, except it would require reports to do something uncomfortable, and I don't think they want to do that. Ema.

We saw a lot of this during like COVID as well. Right is like you wanted it was less There would be less harm if you were too alarmist about something, but there could be a lot of harm. They felt like reputationally, I feel like this is not a big deal, and it was, and so there's definitely an asymmetric risk profile. There's also like a meaning defining profile. It's just really exciting to think everything is going to be disrupted in

chain unrecognizably. It sort of gives a focus and meaning to like an otherwise somewhat chaotic and disrupted world that we're in right now. And so there's that aspect too, is that people want to believe there is something massive about to happen because in some sense it wipes away all the bad stuff that's happening. Who cares, none of this is relevant because this much bigger thing is coming. So I think that gets wrapped into it as well.

I think the economic reporters are more on this because like their whole job is to try to I mean, they're not on it. I think after your work they're on it more.

they're just gonna make thirty billion dollars next year. It's just fanciful. It's there's the skepticism, the cynicism doesn't exist. And I get I agree on that there is a bubble reporting.

probably higher. How are they meant to pay thirty to sixty billion dollars a year in a year, And everyone's like, well, they'll work it out. How's Oracle meant to build those data centers? Those data centers are gonna cost one hundred and eighty nine billion dollars. They've raised one hundred billion, Okay, how they meant to do that? And everyone's just like, oh,

they'll work it out. If I wish I could do this with the fucking bank, I need a five hundred million dollar house, I'll pay for it at some point, all right, it's fine, and the news would run articles about my genius housing purchases. It's just it's one of those things where and you said, well the asymmetric risk doesn't exist, it does for some of these reporters because

I've been saving their their bylines for years. Because I actually think that there needs to be some sort of reckoning with this, because if you look back, there are major financial outlets that did the same, that were literally propping up Sam Bankman freed two weeks before FTX collapsed, who then went on immediately to cover AI fucking and now they're peddling bullshit for anthropic and sorry, I'm kind of hating. Well, I guess it's hated season, so I can.

It just frustrates me because regular people are being scared. They're being scared by the kind of astonishment which I actually loved the astonishment reporting of like, oh well, open clawd has proven open claw has proven that ai agi is here, or they've built their own social networks that we should be scared, scared insane did Yeah, And it's like, I assume you saw that for the listeners. By the way, open Claw had this this fucking claude bot whatever it is.

They had their own social network where the quote unquote lms would post, but it turns out that those posts would just make by their owners.

Because it turns out there was zero technical breakthroughs. There is no new AI technology connected to open claw, which used to be called molt whatever mold or open clawed or whatever open claw is, I think where they ended up right, which is an open source library or framework for building AI agents powered by elims. There's no new AI technology involved in this at all. The agents you build are just accessing off the shelf elms that we're all using for chatbots anyways. You can aim it at

whatever commercial chatpot you want. There's no new framework for how the agents work. It's the same sort of react loop that we've been trying for the last two or three years, where you just you have a program. It's like a bit of Python code that asks an LM, hey, here's what I want to do, here's the tools you have available. Come up with a plan, and then it sends it back a plan, and then the Python code takes the first step out of that text and says, okay,

here's the tools available. What should I do to execute this step? And then like the LM will give you some steps and then the Python code runs those steps and then you just go back and forth right. That's like this basic react, which.

easy for anyone to write bad agents. And the other thing that made it interesting to wireheads in San Francisco is that the commercial products where people are trying to build these companies, they have common sense security, like, well, probably if it's just a Python program blindly doing what an LLM tells it to do, like it probably shouldn't have access to like your credit cards or to your

hard drive or whatever. At open cloud like you can just you can give it access to anything you want on your computer, and so you could build really cool demos that are also like incredibly insecure and unsafe. And so it was fun for hobbyists, but there was no new technology there. Nothing was new, So it was just a way for other people like hobbyists to build their

own agents. That were like less safe than the more carefully built Like what people don't here's a story people don't know is like in the immediate aftermath of chat GPT. So we're in early twenty twenty three, right, so chatchepts come out. Open AI goes on a road tour of major publications, right, and they're to try to Hey, here's what's going on. You need to write about this. In early twenty twenty three, they're like, the next thing we're

going to offer is something like these agents. They call them plugins, and you can install plugins that basically can do actions on behalf of your LLLM queries. You could have like a book an airline flight plug in and say, hey, chatchipt book me a flight, and a language model can't do anything but produce text, but then the plugin could take the text and go and book you a flight.

And they're like, this is the future, obviously, and that project disappeared because oh, it's incredibly unsafe and unreliable to have code that could interact with the real world that's following commands from an unreliable who's stayed an LLM like that went away not because the technology was hard, but

because it's not safe. So there's nothing new. My whole article on agents, they tried this all at twenty twenty five, and we're failing to get this type of agent to work for anything outside of basically like producing computer code. So open clause is nothing new. It was just a way for other people to build these things. The only interesting stories were security whole stories. But it was reported. Man, I was listening to the All In podcast and they were tasing why this is the future of AI, Like

this is it. Everyone is going to have of an open claw agent as like a personal assistant, and this is like the biggest thing in open AI. And I don't know who it was. One of those guys was like, yeah, we replaced our podcast producer, you know, with this agent who can like email guests on our behalf and book it on our calendar. And well it's costing about a thousand dollars a day right now to run it. But like, I don't know why we're going to need employees in the future.

like a Reddit style social network. They vibe coded that framework and it was full of security holes as we mentioned before, and there was like a small number of users that created a huge amount of agents, and they were just kind of like prompting and prodding their agents to produce, like let's talk about creating around church or killing humanity. It's like, guys, this you have Python code asking lllms to like write text in the style of like the matrix, and you're posting it on a fake

social network. The real story here should be where they don't these people have things to do? Don't they have jobs?

I'd respect the shit out. Then I'd be like, yeah, enjoy yourself. It's going away. This is a little subsidized rites, but have fun with that. I don't know why you need a max studio, but good luck. No, they are like, this is the future. But I'm gonna be honest, cal My real question is what does openclaw actually do? Because when I went and read what it actually I read so many posts you said it books callen does and does this. I could find no proof that anyone successfully did that.

to other services basically and makes calls to LLLM. So it's just like a wrapper in which you can write your own code, and some people are trying, Yeah, like you can you can write, you can have it talk to your calendar, you can have it talk to your email and theory and.

Which is why agents based on llms have struggled because if you it's fine if you as a human are talking to an M, because you don't realize how much filtering and tweaking. You're like, well, that's kind of ignore that and this is good, or let me ask you to redo it. It's really an issue if you write

a program that just says I will do whatever. I'll ask the LLM for a plan and then just do whatever it says because it doesn't know that doesn't really make sen or like this sounds generally generally reasonable, but with like issues in the details. Does it work out when you execute it? I mean, in a practical sense, can it? They said only all in podcasts, three dumb bitches saying exactly, that's the main reference. Don't use that word usually those people sitting around going blah blah blah.

it is it completely it probably isn't. And that's the thing I don't I get why they all in guys do it because they probably have investments and they're boosters TVPN, same fucking deal. It really rules that the two largest like tech things in the valley are just state media book for Silicon Valley. What gets me is when you get like the Atlantic CNBC Business Insider in places like that doing stuff like this, and it bothers me because

they don't even need the hate on it. They could just say, yeah, I could do this, it's pretty cool, right yeah, But I guess that that doesn't get the clear.

dollars a day. What it's forcing people, these hobbyists to do is to turn to like cheaper alternatives for models, and that I do think is significant, this idea that there's open source models out there, they're significantly cheaper than trying to use like open ai or claud. More and more people are running their own local models because it turns out, for you know, most specific uses you might use in LLM, you don't need like a super fine tune trillion parameter beast of a model running in some

data center somewhere. It's like, you know what, I'm parsing my email that try to extract like suggested times. I'm fine with like a twenty billion parameter model that can easily fit in a single GPU on a thing I have, you know, we share in our office or something like that. So to me, that's the most interesting story out of open claw is the bad news it could be for the big companies as people get more comfortable with we don't needed these formula one car versions of language models

for the stuff we're actually doing. We're fine with the forward focus, right, And I think that's a transition that's going to That's a transition that to me is more interest.

That's what we should be writing about. Like, that's an interesting idea to me is that you have these huge, high valuation companies, but you also have all these open source models, like the weights are just out there in the public domain that can do ninety eight percent of what people care about, and you're beginning to get low cost competitive services where people can just spin those up in cheaper data centers. That's interesting to me. That's an

economic story. AI creating its own church is not as not as relevant to me.

That's the thing. This is even with these low these lower cost models on device could be interesting.

more vibe reporting. It's you're meant to read that and go, wow, this thing is performing tasks that are useful, that execute code, that do something, and in that thirty hours that is equivalent to thirty human working hours versus they sat there and pissed their pants for like twenty nine point five of those at least well.

like that. I mean, this was like the meter had this problem with their graph of how long of tasks AI can now do, and it was like the sort of like super arbitrary decision of like, oh, here's a test that takes a human five hours, Now AI can do that. Here's a task, whereas really more about like how many things in a row can you do without the errors cascading out of control or something like that. It's very different.

What is the work? What was changed? What did they figure? What was happening before? What technical breakthrough made this possible now? And then what are the concrete implications? What specific things now? Can we directly this will now allow us to do? What? Tell us what jobs are going away? Tell us what tool we're going to see, like you have to, but we avoid that, right because then you're putting your chips down and then those things don't come along, And so

I'm always looking for that. What's the technical innovation and what are the concrete implications? If you don't have that, you're mining emotions.

how much revenue are they profitable? Yet? Why not? Why not explain to me right now? Anthropics like we're gonna gonna burn a hundred billion on training. I guess what do you think? And they're like, yeah, I love it. I actually think that's huge, that's great. I'm not And to be clear, I think that this scrutiny should be from the beginning to the end. I think that everyone should face this scrutiny. Yeah, I'm not saying that I

should get an easier I'm saying, actually everyone should. Anthropics should face the most brutal scrutiny, and I guess that they don't because they want access. It's just very dull and annoying, and it's just helping already rich guys like Wario Ama Day, who should not have more money. Listen to him speak, he needs he needs to face some stress. I think some stress would help him grow. But caw, as we wrap up, I did actually have like a

technical thing that wants an idea of being perclaying. So I think that this term training with models is being misused and used in a way that is kind of vibe reporting style, which is they use training as a word that suggests that it will stop, that they will stop training these models. But from correct me if I'm wrong. Training is everything from building a new model to updating and models current parameters. Correct.

came next. This is a real writing Guess the word that came next, and then it guesses, and then you adjust the weight so it gets closer to the right answer. That's prehen you adjust the waits. What are you doing, You're just so you're a training algorithm called back propagation.

This is a Jeff Hinton, you know innovation where you're you're going through and you're adjusting the weights all the way through the layers in such a way that the answer it gives for this particular test gets a little bit closer to the right answer, because you know the right answer that's pre training, that's pre trains unsupervised. So you take Hamlet or you take Dickens like the best of times it was the worst of and then you give that to the thing what were to come next?

And you know it says Bacon like, we're going to adjust these weights now in a way that like your answer gets a little bit closer towards times. Right. Okay, that's pre training. Then you get post training, where you

already have trained, so you have this network. All the weights have been set through this massive multi month you know, billion dollar pre training, and now you want to go through and you want to tweak this to avoid certain types of behaviors or to influence towards certain types of behaviors. So for post training, it's almost always based off of you have inputs like prompts and correct answers. This is the right way to respond to this question, right, So

you've you've you have of questions and answers. You give the prompt to the LLM, it spits out some answer, and now what you're doing you're using is called reinforcement learning. It's a general technology, but you're using techniques from reinforcement learning the sort of like zap it like you would zap a dog when you're dog training it to if it's a bad answer, you zap it, so you get those weights away from that answer, and if it's a good answer, you give it a treat like Okay, this

is post this is post training. And so that's where you've moved past the word guessing game, which is where all of the sort of general smarts comes from these models, And now you're you're you're doing the sort of zapping and treat training around very specific things, right, so this

is where you keep going. Yeah, So like so you'll go through and like ask it questions where the answers might be like about building bombs or whatever, and every time it spits out an answer about a bomb, you give it like a really bad negative shock, and you like definitely turn off those circuits, like we don't want you to spit out answers by bombs, Like that's where

all the guardrails come from. Or if you want it to get better at doing like a particular math exam, you can give it like lots of questions from that math exam and then you have the right answer and you can kind of zap it to move it towards what the answers look like on this math exam. So so post training is more focused. You have particular types of behavior you're trying to sort of instill in this already pre trained massive network. That's mainly where the focus

turned after GPT four. So GPT four was like the extent of pre training making it smarter. After GPT four, trying to make those models even bigger and pre train them longer didn't lead to much performance increases. So everything we got between four in the lead up to five was post training, and that's when they began focusing on metrics because if I have a particular metric, I can post train a model to do well on that test.

And so everything became about metrics and post now we can do this thing better or look at this thing we do better, And so that's kind of the game that's played now is we do lots of post training that requires like much more specific data because you need like right answers, pairs of props, and correct answers. So only certain things we can do this with. But that's the game we've been playing since like twenty that's one.

so training would end. But from what I understand, training is as common and necessary an expense as influence at this point.

and do some zapping. Let's get out to zapper, give it a bunch of examples of saying that bad thing, and let zap and say don't do that. And now it's like very unlikely to do that. So yeah, they're constantly they're constantly doing that. Otherwise, your a stasis, which

is different than the way most people actually think of it. Differently, they think that the model is somehow like learning online that like, which as it's absolutely not true that as you talk to it, it's learning and it's getting uh, it's getting better. And they're like, but wait a second, something I talked about earlier, it must be getting smarter. The model doesn't care about you, just your local software. You don't realize this. It's including like huge bits of

stuff you've talked to it before. In the prompt that's going to the model. It's like, here's a bunch of stuff that Ed has submitted to you in the past. Okay, now here's his current question. So it's the model hasn't changed. The model doesn't have memory a canunt of justice. It doesn't adjust its memory in real time. It's all static. There is no dynamic memory involved in these models. They're fixed until you go out and post train it and then you bring the new weights back in the data center.

And that's the thing. Do an inference now.

differences pooh, training is pooh. That's just that's I'm not putting that one in.

words in the world, and it takes months. So it's it's just like you're running a data center that's going to have nowadays up to six digit GPUs uh running full time for months just to get that pre training done, and you have to pay for all of that, right, So that's all right time, you're not getting money, you're just paying for training. Post training is also expensive. It's not that expensive though, because as expensive as pre training, because each post training session it's a way, way, way

way smaller data set that you're post training on. It's like, all right, we generated like ten thousand examples of people, you know, responding to questions in a racist way, and those ten thousand examples will use to reinforcement learn and try to move it away from answering those type of questions in a racist way. That's like not that big of a data set compared to we're going to train this on every word written that we have access to. So the post training is not as expensive as pre training.

operation is basic multiplication. You're just multiplying numbers in the big table backpropagation which you use to training. You're it is multiplication. When you do a bunch of it's a bunch of derivatives because you're constantly you need to calculate like the derivative of these. I mean like it doesn't really matter, but you kind of need that. You want the derivative because you want the gradient to scent to be towards like better and away from worse and derivatives.

My understanding is this is like more expensive operations per weight that you're trying to change because you're not just multiplying a number, You're having to calculate derivatives and becomes a little bit more complicated. So yeah, it's like inference in reverse, but also like a little bit more expensive.

So if you if you have ten thousand sample question responses, you're going to use the post train, it's kind of like ten thousand users sent prompts and there are particularly expensive prompts, and you had to pay for that instead of like them paying for it. So yeah, it's good to think of it as like inference and reverse.

any thirty days. We don't know if they're doing insane math, which happens very rarely, that this company feels like one that might do it. Just my gut instinct is they may be doing his seven day period and we're going to turn it into a month like they We don't know how they're doing this. And they also coupled this by saying that as compute grows, revenue grows. I don't know if you seen this.

understand that story. Okay, yeah, the annualized revenue stuff, I think I even use that term talking to someone I credit you as like Trum would say, look out for annualized revenue.

because they're doing CPU based stuff. Like it gets expensive at scale because there's a lot of people, but it doesn't get multiply multi I can't say the word. It doesn't get much more expensive as you grow with AI. It's actually because of the way large language models work, your most excited customers are your most expensive.

oh my god, this is so cool. These massive disruptions they're going to change everything is imminent, and year after year we've said that I'm not yet seeing the massive disruptions, like not the not the stories of what might be disrupted or the stories of what's different, but like, how many years do we have to go without industries crumbling or major new economic players that didn't exist before, or complete restructuring of huge companies around this technology? How many

years we have to go? So I did a video, right I went, I found the Reddit thread where someone just asked this question. This was from like earlier in the month. They were like, outside of the vibe coding stuff, what what are the like, what are people? What are the big tools? Like, what are the big things that

have come out of this technology? They're changing things? And I read through this whole thread and it was interesting, there's not people don't have much like well, you know, like it could these are like really small case studies. I used it to help, you know, gather clean up data that I got from whatever. I was like, this is like such a nerdy specific use case, and so I went through that thread in a video and this

has kind of been my question. It's like, it is very cool technology, but how do we know where this is going to fall? Like to me, the scale is it would go like this, like blockchain software, then Oculus VR, then maybe internet to electricity. Right, so we're gonna have a scale of disruption.

That did nothing. Then you have like Oculus VR. It really is cool, right you put on these things like that is awesome, Like I love this technology, but it's having a hard time having any real major impact because people aren't sure what.

very cool things. Chat Gipt is very cool. It's very cool that it can like have that comprehension and no one thought it could do that, But we haven't yet figured out what.

like very cool. Yeah, but we haven't got past this is like the the surprising thing of this field is we're not really past the oculus stage yet, like where they're like, for certain this is vibe coding is really cool. The comprehension is cool, so like sora is weird, but like it's cool, you can do that, But the markets are not None of these have markets yet, right, Like that, there's not big markets in any of these yet. And will how far will it go from oculus to the Internet.

To me, that is like the number one question, the number two question, the number three question of all reporting on this, and almost no one's talking about that. It's just hype laundering. We'll take this hype, will extrapolate, it will react to that extrapolation. That's kind of what reporting is, right, now in AI, Whereas to me, this is the hugest question. If this ends up oculus, retail investors are going to get screwed. If it ends up Internet, all right, that's like,

that's like a really interesting significant story. If it ends up electricity, obviously that really matters. But like, I don't know anyone who actually thinks it's going to be that disruptive. Not the current technology. That's the story to me. Not let's like extrapolate us. You know, hey, what are these things are creating a church? Or let's let's let's hype launder extrapolate that and react to our extrapol That's not really reporting so much as speculative fiction writing. I guess

I don't know quite what to call it. But this is the real question, where exactly are we now and what are the possibilities of like where this is going to go positive and negative? I don't we have enough talk on that.

Offline theme song is Matosowski. You can check out more of his music and audio projects at Mattasowski dot com, m A T T O S O W s ki dot com. You can email me at easy at Better Offline dot com or visit Better Offline dot com to find more podcast links and of course, my newsletter. I also really recommend you go to chat dot Where's youreaed dot at to visit the discord, and go to our slash Better Offline to check out our reddit. Thank you so much for listening.