Hater Season: Cal Newport on AI Reporting

26:24And then the clod 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 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 90 % of the really important valuable code on which everything runs. I don't think they would touch cloud code with a 10-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.

26:58And, you know, it's like not very good code. And then it's – I guess you could replace some of that with it. But it's functional enough. it's functional enough but I can't get my arms around it yet but I do get I have a lot of sources so I hear from I do hear from people that are talking about how cool CloudCode is I do think it's cool I hear from a lot of professional programmers that are like we can't use this we're trying to write serious programs we have to sell this software this is not solving a problem we have so I don't know but the problem is that's what the reporting should be hey here we are at this company looking over the shoulder of people what are they doing?

27:33Let'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 And we get super excited about stuff and we go crazy. You can't just launder our hype into this is what's happening.

28:07And 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 back in the Pentagon. It's not a way to report on what's actually going on. and i think the other thing as well is there is a if you don't do this hype laundering i think i'm i don't know how these if you're a reporter listening to this and you have a thought about this send it to me anonymously easy trump.76 on signal um 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 30 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 it 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 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 hacker 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 code code will replace them and ultimately i think it's just going to be a bloodbath for people's 401ks that could have been avoided except it would require reporters to do something uncomfortable and i don't think they want to do that ever i think there's two things going on is what i've decided with with reporting first of all it's asymmetric risk so a lot of reporters are like look there's not a major risk if i'm excited about this and it doesn't pan out because we We could be like, yeah, surprisingly, this didn't pan out and it was some factors we couldn't see.

30:42But they do feel like there's a huge amount of risk of saying this is not a big deal and then it is. And so it's definitely an asymmetric risk. 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 if you're 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 and change unrecognizably.

31:14It 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 like bad stuff that's happening. Who cares? None of this is relevant because this much bigger thing is coming. And so I think that gets wrapped into it as well. I think the economic reporters are more on this because their whole job is to try to... I mean, they're not on it. I think after your work, they're on it more. I don't know if I agree.

31:49I am reading big series. There's an article in Bloomberg I'm trying to shove through archive.is because they were so mean and unfair to my friend Steve Burke at Gamers Nexus that I won't pay him. But it's more shit about the software narrative, the fact that software is being disrupted by clawed code you get the same pallid reporting from bloomberg and even the financial times about anthropic and the ft is generally pretty good we're like yeah they're just gonna make 30 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 well there was a bubble reporting the last fall there was a period where everyone did the bubble reporting after gp25 you had a two-month period where every major publication did bubble reportings um but you're yeah i guess that did kind of die off it died off because they they hear one nice thing from jensen huang and they're like well i'm sold like that jacket that jacket's looking pretty sharp i mean someone who wears that jacket how could they be wrong would a man that has a shiny jacket like uh the lead singer of corn wears at concerts uh would he lie and it's just what really bothers me as far as the economic reporting though is the oracle because it's like oracle needs to be paid 300 billion dollars over five years by open ai a company that if we're to believe reporting which i do not uh they made 13 billion dollars last year in revenue and lost sorry they made 13 billion and lost they claim nine i think it's probably higher how are they meant to pay 30 to 60 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 going to cost 189 billion dollars they've raised 100 billion okay how they meant to do that and everyone's just like ah they'll work it out if i wish i could do this with the fucking bank i needed i need a 500 million dollar house i'll pay for it in 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 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 thing that were literally propping up sam bankman freed two weeks before ftx collapsed yeah who then went on immediately to cover ai yeah fucking and now they're peddling bullshit for anthropic and sorry i'm kind of hating well i guess it's hate 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 love the astonishment reporting of like oh well open claw does proven open claw has proven that ai is uh agi is here or they built their own social network so we should be scared software is dead insane is dead yeah yeah like singularity is here and it's like i i assume you saw the for the listeners by the way open claw had this this fucking clawed bot whatever it is they had their own social network where the quote-unquote llms would post but it turns out that most of those posts were just made by their owners.

35:15Yeah. I mean, this is a good case study, right? This one bothered me because writer friends I know who are not technology related were texting this to me. They were worried, right? They were texting me these articles like, this seems bad. This seems like something really bad. They were really getting the digital ick really strongly off of these articles. And I would start reading these articles and I say, well, there's no discussion of what is the technical breakthrough here and what are the concrete implications? because it turns out there was zero technical breakthroughs. There is no new AI technology connected to OpenClaw, which used to be called Mold, whatever, Open Mold, OpenClawed or whatever.

35:54OpenClaw is, I think, where they ended up, right? Which is an open source library or framework for building AI agents powered by LLMs. There's no new AI technology involved in this at all. The agents you build are just accessing off-the-shelf LLMs that we're all using for chatbots anyways. You can aim it at whatever commercial chatbot 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 have a program. It's like a bit of Python code that asks an LLM, hey, here's what I want to do. Here's the tools you have available.

36:26Come 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 the LLM will give you some steps and then the Python code runs those steps and then you just go back and forth. That's like this basic React loop. Which is exactly how Manus worked as well. Everything, there's nothing new about this. Manus was this AI agent that Facebook might be, Meta might be acquiring. They literally, when you use it, it just writes Python code for every step. Well, yeah.

37:01So this, I mean, there's nothing new here technologically. The only thing that was new about OpenClaw is it was open source. so it made it easy for anyone to write bad agents. And the other thing that made it interesting to wire heads 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 OpenCloud, like you can just, you can give it access to anything you want on your computer.

37:32And so you could build really cool demos that are also like incredibly insecure and unsafe. safe 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 2023 right so chat gpt has 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.

38:05In early 2023, 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 LLM queries. You could have like a book an airline flight plugin and say, hey, ChatGPT, book me a flight. And 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 can interact with the real world that's following commands from an unreliable who's named LLM.

38:41It's 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 of 2025 and were failing to get this type of agent to work for anything outside of basically like producing computer code. So OpenCloud is nothing new. It was just a way for other people to build these things. The only interesting stories were security hole stories, but it was reported, man, I was listening to the All In podcast. And they said, this is the future of AI. Like, this is it. Everyone is going to have an open claw agent as like a personal assistant.

39:16And 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. Well, it's costing about$1 ,000 a day right now to run it. But like, I don't know why we're going to need employees in the future. Hell yeah, brother. Yeah. So anyways, like it's a non, if you said, what's the technical implications? None. If you say, what are the concrete implications? The best story they can have is like, maybe when people, I don't know what it is.

39:47All the companies have already been trying to build these things for years. So I don't know, but it was reported like, just something icky is happening. And that MultBook was an application that was built for these agents to communicate on 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 our own church or killing humanity. it's like guys this yeah python code asking llms 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 are they don't these people have things to do don't they have jobs like what is this yeah yeah it really it really my model train i added a tunnel like so that one really got to me i actually will push back on that model train listeners i think make up 20 of the listenership of this podcast of course also model trains are cool just signaling to my fellow autists out there uh but no model trains are a physical thing which you build it you build a little a little city i think they're delightful respect to those with this it's like if they and you know what if they were just saying hey i've been fucking around with some software and i did something cool i'd respect the shit out of them i'd be like yeah enjoy yourself it's going away this is all subsidized rates but have fun with that don't know why you need a max studio but good luck no they are like this is the future yeah but But I'm going to be honest, Cal, my real question is, what does OpenClaw actually do?

41:20Because when I went and read what it actually, I read so many posts, you said it books calendars and does this. I could find no proof that anyone successfully did that. And it doesn't, it's just a series of library calls that you can use to build your own program in theory that would do that. So OpenClaw itself is just like a series of like interfaces and hooks that makes it easy to write a program that talks to other services basically. and makes calls to LLM. So it's just like a wrapper in which you can write your own code. And some people are trying. Yeah, like you can write, you can have it talk to your calendar.

41:56You can have it talk to your email in theory. But does it work? But does it work? Well, it's just asking an LLM. You can simulate. I mean, I did this for my New Yorker piece. I was like, look, I can just simulate being an agent. Just ask any LLM. Here's what I'm trying to do. Give me like the steps for doing this or whatever. And like anything else you ask of an LLM, it will give you answers that sound very reasonable in general and then have like a lot of issues in the details, which is why agents based on LLMs have struggled. Because if you – it's fine if you as a human are talking to an LLM because you don't realize how much filtering and tweaking you're doing.

42:28Well, that's kind of – I'll 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 sense. or like this sounds generally reasonable, but with like issues in the details, does it work well when you execute things? I mean, in a practical sense, they said on the all-in podcast, three dumb bitches saying, exactly, that's a meme reference. Don't use that word usually. Those people sitting around going, blah, blah, blah.

43:03Oh, we've made it and replaced our podcast producer with an LLM that can make appointments and send emails. In practice, is that true? I severely doubt that Also, you're spending $1 ,000 a day, so you're paying$365 ,000 A year For this, right? Are you really? Or are you just Is it complete? It probably isn't And that's the thing I get why the all-in guys do it Because they probably have investments And they're boosters, TBPN Same fucking deal It really rules that the two largest Tech things in the valley are just state media but for silicon valley what gets me is when you get like the atlantic cnbc business insider and places like that doing stuff like this and it bothers me because they don't even need to 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 i it's not an interesting story that's not the real story is not always that interesting it's like look hobbyists are building these tools that kind of do cool things but make mistakes and it's a little expensive.

44:12The most interesting story out of OpenClaw is the only thing I think is going to be impactful out of it is because it was so expensive, like that$1 ,000 a day, what it's forcing people, these hobbyists to do is to turn to 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 OpenAI or Cloud. But more and more people are running their own local models because it turns out for most specific uses, you might use an LLM. You don't need like a super fine-tuned trillion-parameter beast of a model running in some data center somewhere.

44:49It's like, you know what? I'm parsing my email to try to extract like suggested times. I'm fine with like a 20 billion-parameter model that can easily fit in a single GPU on a thing I have we share in our office or something like that. So to me, that's the most interesting story out of OpenClaw is the bad news it could be for the big companies as people get more comfortable with, we don't need these Formula One car versions of language models for the stuff we're actually doing. We're fine with the Ford Focus, right? And I think that's a transition that's going to – that's a transition that to me is more – that's what we should be writing about.

45:26Like 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 98 % 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 relevant to me.

46:23I'll see you next time. That's iHeartAdvertising.com.

46:56What gives us hope? We get honest about the big stuff. Identity when you don't recognize yourself anymore. Loss that changes you. Purpose when success isn't enough. Peace when your mind won't slow down. Faith when it's complicated. Some guests have answers. Most are still figuring it out. If you've ever felt like there has to be more to the story, this show is for you. Listen to If You Can Hear Me on the iHeartRadio app, Apple Podcasts, or wherever you get your podcasts.

47:30Segregation in the day, integration at night. When segregation was the law, one mysterious Black club owner had his own rules. We didn't worry about what went on outside. It was like stepping on another world. Inside Charlie's place, Black and white people danced together. But not everyone was happy about it. You saw the KKK? Yeah, they would have dressed up in their uniform. The KKK set out to raid Charlie, take him away from here. Charlie was an example of power. They had to crush him. From Atlas Obscura, Rococo Punch, and Visit Myrtle Beach comes Charlie's Place, a story that was nearly lost to time.

48:17Until now. Listen to Charlie's Place on the iHeartRadio app, Apple Podcasts, or wherever you get your podcasts.

48:48It has so much in it. It has Nazis. It has the German economic miracle. And it features one of the most famous ads of all time, an ad that really redefined what advertising was in the United States. The calculation was that there was some number of Americans who were ready for something different, who were ready for something that was counter to the culture, if you will. Perfect timing. In this new decade of the 1960s. Listen to Business History on the iHeartRadio app, Apple Podcasts, or wherever you get your podcasts, and watch episodes on YouTube.

49:47yeah i hear this data center out in north dakota losing a million dollars a day that's the thing this is even with these low these lower cost models on device could be interesting that's what's going to happen i think i think on device is what's going to happen i just remembered something this is a classic bullshit story that i see every so often so have you read any of the stories that are like yeah claude can now work for hours uninterrupted have you read about these you heard this you seen this work for hours oh yeah yeah we're talking about the yeah the multi-step agentic execution well no it keeps coming back yeah yeah ai on cnbc ai on the verge of eight hour job shift without burnout or break is it is 24 hour ai workday next gonna just censor myself what i was gonna say there because it's just like yeah it can work for for hours is the output good does it problem yeah it's just a loop does it just keep recall it recall it recall it like i can i They can write a Python program that calls an hour for 24 hours.

50:48If you need me to burn something for hours on end, just give me some gasoline. I got you, baby. I can sort this right out. It'd be much cheaper. But it's like, yeah, by September 2025, Claude Sonnet 4.5 was reported to run autonomously for up to 30 hours. Reported with, I mean, but that's 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 30 hours, that is equivalent to 30 human working hours versus they sat there and pissed their pants for like 29.5 of those at least. Well, and also that's, those are like specialized tasks typically that have clear milestones in testing.

51:32So it's like it can do something, the loop can try and try until it passes the test and then you know that's done. And then you can move on to the next step and then you can keep calling to LLM and executing until it passes the next test and you can move on. And in theory, like it's – yeah, they're avoiding error cascade because it's testable and they can keep retrying or something 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 this sort of like super arbitrary decision of like, oh, here's a task that takes a human five hours.

52:03Now AI can do that. Here's a task where it's really more about like how many things in a row can you do without the errors cascading out of control or something like this, right? It's very different. you're being way nicer than you should be in my opinion just because they don't even explain they don't even say like yeah and we managed to make it do it and this they just go that's right hours cnbc just fucking just tearing their shirts off and screaming it could but it goes back to the two things you need we need in this reporting you need technical uh the details of the relevant technical innovation and a discussion of concrete implications near future implications of that breakthrough.

52:41That's what I'm always looking for. So if you want to cover a story like that, be like, well, what does that mean? What is the technical breakthrough? What does this mean technically you can do 30 hours of work or eight hours? 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 direct? This will now allow us to do. Tell us what jobs are going away. Tell us what tool we're going to see. But we avoid that, right? Because then you're putting your chips down and then those things don't come along.

53:12And 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. You're mining emotions or just helping boost stock. That's what really bothers me because I hate that they're scaring people. I really, really hate that they're doing marketing. like they're just doing like i run a pr firm i've dealt with early stage startups for like 15 years there isn't a single early stage startup that would get a percentage point of this bullshit like you you email a reporter about like a series a stop they're like all right all right motherfucker how much revenue are they profitable yet why not why not explain to me right now anthropics like ah we're gonna burn 100 billion on training i guess what do you think and they're like yeah i love it i actually think that's future 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 like that's 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 amaday who's 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 cal as we wrap up i did actually have like a technical thing that was an idea of being percolating so i think that this term training 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.

55:01But correct me if I'm wrong, training is everything from building a new model to updating a model's current parameters, correct? Yeah, there's pre-training and post-training is the right way to think about it. So the pre-training is unsupervised. That's where you take all the text that's ever been written. and you will take a real piece of text written by a human and you'll cut it off at an arbitrary point and you'll tell the language model, you're guess the word that came next. There's a real word that came next. This is 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.

55:32That's pre-training. And when you adjust the weights, what are you doing? You're adjusting. So you're running a training algorithm called backpropagation. This is a Jeff Hinton innovation where 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. And that's pre-training. That's pre-trained. It's unsupervised. So you take Hamlet or you take Dickens. It's like the best of times. It was the worst of. And then you give that to the thing.

55:58What were to come next? And it says bacon. We're going to adjust these weights now in a way that your answer gets a little bit closer towards times. 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.

56:31This is the right way to respond to this question, right? So you have pairs 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 it's called reinforcement learning as a general technology but you're using techniques from reinforcement learning to 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 this is post this is post training and so that's where uh you've moved past the the word guessing game which is where all of the sort of general smarts comes from these models.

57:08And now you're doing the sort of zapping and treat training around very specific things, right? So this is where - But they do that. Keep going, sorry. 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're 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.

57:46So 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-4. So GPT-4 was like the extent of pre-training making it smarter. After GPT-4, trying to make those models even bigger and pre-train them longer didn't lead to much performance increases. So everything we got between 4 in the lead up to 5 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-training.

58:24Now 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 prompts and correct answers so only certain things we can do this with but that's the game we've been playing since like 2024 but that's do they do that with models that exist now so this is basically updates right yeah they do it on a semi-regular basis yeah but usually there'll be a name change like GPT-5.2 is different than 5.1 different than 5 but don't they update the current models they don't re-pretrain them that's too expensive but they no no they i'm not saying that i'm saying do they post train the current models to make them better at stuff yeah to tweak things yeah so that's this is a very long way to get to a point i'm kind of making which is one of the problems with vibe reporting on this is training is framed as this thing like inference is framed as opex that is permanent that you cannot avoid inference being creating the outputs training is framed as this r &d mysticism which is just out there and you're not like training they never say this but you hear training you think oh you train and then you perform and so training would end but from what i understand training is as common and necessary and expense as inference at this point yeah it's the only way you improve or update things right like so if you had a microsoft 365 software you're constantly sending updates and patches and whatever as you like add new features or whatever in the ai model world it's it's yeah Yeah, it's more rounds of post-training.

1:00:00It's the only way to make any sort of improvement or fixing bugs. Like you're like, oh, here's something it's saying that we're really upset about. Okay, let's go in and do some zapping. Let's get out the zapper, give it a bunch of examples of saying that bad thing, and let's zap and say don't do that. And now it's like very unlikely to do that. So, yeah, they're constantly doing that. Otherwise, you're at 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 as you talk which is absolutely not it's just absolutely not true that as you talk to it it's learning and it's getting it's getting better and they're like but wait a second it remembered 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.

1:00:57The model doesn't have memory. 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 doing inference now. But that's the thing. The reason I bring it up as the kind of closing vibe story is because training is very clearly getting more expensive. They say profitable on inference, which, again, doesn't really make sense to me, But putting that aside, they're not. But even if they were, if training never stops, then who cares about profitable inference?

1:01:33Like, it just means that you will get more expensive forever. Inference is just, yeah. Inference is P, training is poo. I'm not putting that one in an article. Yeah, okay. It's an interesting question, right? Because it's also getting like, was it Altman who was making those comments about, well, if we just didn't have to pay for the training, this would be profitable. If I didn't have all these expenses, I'd be so profitable. The one distinction that's maybe relevant there, not to be an apologist, but the one distinction that's relevant is pre-training versus post. So pre-training is insanely expensive, right?

1:02:07Because you're training something on all the words in the world, and it takes months. So it's just like you're running a data center that's going to have nowadays up to six-digit GPUs 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 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 it on. It's like, all right, we generated like 10 ,000 examples of people responding to questions in a racist way.

1:02:47And those 10 ,000 examples we'll 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. Unless you're doing it all the time. Yeah, that's true. That's the thing. If you're doing it all the time and it takes months of pre-training, but you're constantly doing something like post-training for months, is that not functionally the same thing? It's the same thing. I think this is a fair point.

1:03:21When you have a particular example that you're post-training on, like one prompt with an answer, it's kind of like you're doing inference in reverse, right? So you're back propagating from one side of the network to the beginning as opposed to going from the beginning to the end. Now, it's more expensive than that because when you're just doing inference, the fundamental operation is basic multiplication. You're just multiplying numbers in a big table. Backpropagation, which you use to training, 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.

1:03:54I mean like it doesn't really matter, but you kind of need to – you want the derivative because you want the gradient descent 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. It becomes a little bit more complicated. So, yeah, it's like inference and reverse but also like a little bit more expensive. So if you have 10 ,000 sample question responses you're going to use to post train, it's kind of like 10 ,000 users sent prompts.

1:04:26And they're 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. It's also an ongoing cost. Like it's, everyone is leaning on this idea that this stuff will magically become profitable. I don't know if you've seen the cashflow diagrams of Anthropic and OpenAI, but there is a mysterious math going on where year 2028, 2029, they just become profitable. Wait, I was going to ask you about this last month. this announcement that open ai had some massive increase in revenue yeah well this is actually a great vibe reporting thing and that's annualized revenue they said they hit 20 million billion in annualized revenue which would mean 1.67 billion in a month now important details we don't know how they're defining a month we don't know if they mean 28 days 29 days 30 days we don't know if they mean a calendar month so the month of november or december or do they just mean any 30 days we don't know if they're doing insane math which happens very rarely but this company feels like one that might do it just my gut instinct is they may be doing here's a 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've seen this no this formula yeah okay yeah it's an insane formula that does not map to any economics like it's just it's the kind of thing that if we had a functioning business and tech press that would just be scrutinized to the bone that would just be ripped apart and say what the fuck does that mean because if this were true if you simply add more gigawatt add more revenue yeah then you would simply print more money like it would be a money printer like my movie theater did well therefore yeah if i build a hundred thousand movie theaters we're going to make a hundred thousand more you know times the amount of money and it's like well wait that theater was in manhattan and it was like really well run and they're yeah um well okay i i assume i assume this much um trying to understand that story okay yeah the annualized revenue stuff i think i even used that term talking to someone i credit you as like ed zetron would say look out for annualized revenue it is the funny thing is with that as well is it's more viable reporting because ARR standard it's very standard in SaaS companies that sell on a per seat basis so a sales force would do well even they are doing annualized now with the AI revenue but it would be a software company has 100 seats they sell to a company and they charge 15 bucks a bit a seat per month and they charge that annually and the actual cost of each user is fairly measurable 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 multiple clip 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 yeah because they blow past whatever whatever their monthly revenue is whatever their cost is, they blow past it.

1:07:38So you can't even do a per seat revenue. It's just every, all of these things, when I say them out loud, I'm like, I feel like this should be more obvious. It's why I loved your video. Cause it was like, thank you, someone else. Well, here's the question. Here's my, here's my, here's like the dangerous question. I actually put out a video last week. It was like dangerous question. We're now on year three or four of new, like the, I'm counting new year, starting like new year, 2023 or whatever. of people saying, oh my God, this is so cool. These massive disruptions, they're going to change everything is imminent.

1:08:13And year after year, we've said that. I'm not yet seeing the massive disruptions. Like 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 do we have to go? So I did a video where 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 are the big things that have come out of this technology that are changing things?

1:08:56And I read through this whole thread, and it was interesting. People don't have much like, well, you know, these are like really small case studies. I used it to help gather cleanup 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? To me, the scale would go like this. Like blockchain software, then Oculus VR, then maybe internet, then electricity, right? So we're going to have a scale of disruption, right?

1:09:32Blockchain software is something where the premise made no sense, and it was never going to get off the ground, and there was going to be no impact on the world. And because my training in CS is in distributed system theory, I was there in 2020 saying, guys, let me just tell you, this is nonsense. No, Web3 is not about to take off. None of this makes sense. And that was true. That did nothing. Then you have Oculus VR. It really is cool. You put on these things. That is awesome. I love this technology, but it's having a hard time having any real major impact because people aren't sure what to do with it.

1:10:04Also, most people don't necessarily have a great experience initially because it's extremely dependent on where you are, who you are, the size of your skull, that kind of big head. For a limited group of people, it's really cool, but it fails to come out. Then the internet is really disruptive, changed a lot of things. It's not so much as whole industries disappeared or whatever, but it changed the way a lot of industries actually function. and then like electricity, you could say like it just completely changed what day-to-day existence of business was like. Yeah, how existence of the right. Exactly, yeah.

1:10:35And so the big question like everyone should be asking is where is Generv AI going to fall on this? And, you know, I would say right now, this is what gets me yelled at. I'm not saying this is the prediction necessarily going forward, but right now I don't think it's got past, much farther past the Oculus part of that scale. Yeah, I actually - Where it's really cool. There's very cool things. ChatGPT 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 to do with it. Comprehension. It's – no, it's a – Text comprehension.

1:11:07I mean we take it for granted. But for CS people, the ability that like I can – hey, give me text that like whatever in the style of a poem that does whatever and that includes a character from Star Wars. And then it can give you text that does that. That comprehension, like for a computer scientist, was like, oh, we didn't really know how to consistently do that. technical level yeah yeah that's 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 there's 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 there's not big markets in any of these yet and will how far will it go from oculus to the internet.

1:11:52To 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. We'll extrapolate it. We'll 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 a really interesting, significant story. If it ends up electricity, obviously, that really matters, but I don't know anyone who actually thinks it's going to be that disruptive, not the current technology.

1:12:25That's the story to me, not let's extrapolate, hey, these things are creating a church, or let's hype launder, extrapolate that, and react to our extrapolation. 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 where this is going to go? positive and negative i don't have enough talk on that i fully agree cal it's been such a pleasure having you thank you for joining me always happy to talk shop always happy to hate with you i guess oh yeah season i like it hater season's the best we will be back this week with either a monologue or an interview i have not decided because i've got a wonderful cory quinn interview i just did so i'm considering putting that in the monologue you'll find out on friday anyway this has been Better Offline.

1:13:15I'm at Zitron. Subscribe to the premium, download a t-shirt, whatever you desire.

1:13:44email me at ez at betteroffline.com or visit betteroffline.com to find more podcast links and of course my newsletter. I also really recommend you go to chat.wheresyoured.at to visit the discord and go to r slash better offline to check out our reddit. Thank you so much for listening. Better Offline is a production of Cool Zone Media. For more from Cool Zone Media, visit our website, coolzonemedia.com or check us out on the iHeartRadio app, Apple Podcasts, or wherever you get your podcasts.

1:14:21We'll be right back.

1:14:44on what we had to do to get to visit with him. I've seen some people do some crazy things, but nothing like that. Plus what Hudson almost had in common with Billy Bob Thornton. Crook and Chase Nashville Chats with Hudson Westbrook. Listen and subscribe on the iHeartRadio app, Apple Podcasts, or wherever you listen to podcasts. This is an iHeart Podcast. Guaranteed human.