IM 886: Infinite Slop - The End of Apps? Meet the Agentic OS Built for Your Phone

29:54This is when agents become retail. Absolutely. This is when it becomes B2C and scaled, I think. So I think the potential is huge. Well, and you look at what, say, Elon's doing with the crockpot, that's fine, but it has to live on X's servers and all the data goes to X. And it's just a heavyweight solution. Yes. It's also like we have is we want to meet people where they are. Exactly. Yeah. Or stuff like that. You have to open a computer. And then it's like, most people, like a lot of people in the world don't even own a desktop anymore. Like, you just have a phone or a tablet. And you might be on the go.

30:35So, like, how can we basically meet you where you are? And like, phones are great because you have them 24 seven, always in your pocket. And if you can also like build this kind of new generation of devices like smart glasses or stuff like that and it's like wherever you are whatever device you have you can use the eye from it and it starts feeling like more like like like the jarvis from iron man kind of thing okay like you have this personal eye that's following you around you don't have to be sitting at a computer to go get things done so you're actually these look like the metaglasses you're using right now to to talk to your team member and make her take the headphones off um this i feel like this is going to be the future but it seems like it's the distant future on the other hand everything's moving so much faster than it used to maybe it isn't so and you're delivering now yeah what what do you what is i guess paris already asked you your time frame i mean are we going to have you said by the end of the year we're going to have agi in some areas are we going to have stuff like this next year on our devices yeah i think that's that's a plan if things go right.

31:41Well, I wish you luck. It's very exciting. It's great to see somebody so young, so successful. A titan, if you will, in the field. I am just really happy for you. And I think your parents probably are going to be, you're making them proud. So I can't wait to use your stuff on my devices. And you know what? I'll throw my iPhone out the window if I didn't have to. use it, I would be thrilled if I didn't have to have apps. Live, love you to Google. Well, I have my Pixel. I have my Pixel. I'll put it on the Pixel. But what I don't, I would really like just to have it on my glasses or my earbuds or my tennis racket.

32:24I think you should. I think on your webpage there's a guy with a smart tennis racket. You know, that's the future, right? Div, thank you so much for taking some time with us on Intelligent Machines. I appreciate it. Totally. Thanks for inviting me. Div Garg, founder and CEO, AGI Inc. You can find out more at the AGI website, which is theagi.com. Because you're the only one. Company, right? Oh, is it.co? .company. Oh, that fooled me. I thought that was extra stuff. Oh, I didn't even know there was a.company. The AGI.company. That's a good domain name. Wow, I like it. Thank you, Div. I appreciate it.

33:06Yes. We'll have more Intelligent Machines right after this. This episode of Intelligent Machines brought to you by Horizon 3. The most dangerous threats are always the ones you don't see coming. That feeling of uncertainty is one of the biggest challenges in cybersecurity. We're always looking over our shoulders, right? You can have the best people, the perfect stack, solid processes. But there's always one question, that one nagging question in the back of your mind. What am I missing? that's where horizon 3 comes in horizon 3 is an ai native proactive security platform designed to show organizations exactly how an attacker could compromise their environment before a real threat gets the chance and man is it good horizon 3's autonomous ai hacker node zero safely and continuously tests your infrastructure at machine speed the same way a real attacker would pounding on it, finding areas that are vulnerable, sneaking in.

34:11It covers your entire attack surface in a single test. It deploys in minutes. They've run 225 ,000 plus production safe tests with zero downtime. You can also stop guessing and start knowing your defenses are intact because node zero's result is a loop. Hack, fix, verify, repeat. It finds the problem, you fix it, and in one click you can confirm it's secure horizon 3 it's good that's why it's trusted by the nsa by cisa and fortune 100 companies so you know you can trust it too see exactly what an attacker would find in your environment before they do go to horizon3.ai slash intelligent and request your free node zero demo visit horizon3.ai slash intelligent again that's horizon3.ai slash intelligent no commitment required results in hours not weeks horizon3.ai slash intelligent now back to the show i think i did that anyway with div i said okay we're back before we go on with the show in the in the news of the week i i noticed there's somebody lurking in the corner micah sergeant hello what are you doing there hey my god i'm also battery backed up good because you may be taken over the show if i have about 45 minutes and then it's and then it's good and then it's no time not sure i don't think you'll have any warning i think i will i will just disappear it's just really funny that luke will be in the middle of saying something and then all of a sudden he'll be alone in a pitch black attic by himself it'll be so sad it'll be so so sad it's impressive that you're lighting because isn't that a pretty powerful light in front of you it is but everything nowadays is a florescent it is like a desk sized light people it is like pretty big huge I would show you but the camera behind me is is out as well I'm waiting for Leo to turn off the lights so we can save Quicksilver for another five minutes oh I didn't think of that oh my god every minute you spend podcasting is a minute you're taking away for your agents Mike is here actually you know it's funny because um Mike and I have been talking behind the scenes a lot about AI he's very interested in AI he's doing a lot of interesting things with AI you did you showed me a site that you made.

36:32Yeah. What was that? Yeah. So I wanted a site that anytime I was, I don't know, eating a hot dog and I got some mustard on me or I was walking along, slip on a banana peel and the spaghetti goes flying in the air. And yeah, basically I wanted something to deal with stains. And instead of just, instead of just re, you know, typing in every time I said, I thought, what if I could have one place where all different types of stains could be kind of cataloged? So I decided to send out the agents into the web to do a bunch of research on material science and stain treatment, and then took that and compiled a website called, and now I'm forgetting what it's called too.

37:21I think it's spot. What is it called? Spot check. Spot check. But it's got a, it's got two hyphens because spot hyphen check was already taken. So it's spot hyphen hyphen check dot app. I didn't even know you could do two hyphens. Neither did I, neither did I. So this is the stain index. And basically the moment that you think there's this, I already, I put a little NFC chip on my washer and I just scan it and it pops open this website. No, that's so cool. But anyone can visit this website. You look at what kind of stain you have and it tells you what to do and then what not to do. So sometimes people think that you're supposed to do certain things or you've heard that and no that's not the case so with ink you do not use hot water you do not use a hand dryer but you do do these other things my mother would always say you'll set the stain if you use hot water right yeah you don't want to set the stain so this whole website walks you through the process of treating all kinds of stains and uh do you actually use it i have i've used it uh twice now and then i had somebody um well you know uh doc rock uh he was having some issues and he ended up pulling from this site and spot dash dash check the only human cause stains not like dog pee what's funny is yeah you did that paris did her coffee pour over a tasting app uh i have my coffee pour over page What is this about?

38:49I guess a home ec is being reinvented now. It is indeed. I somehow hadn't thought about NFC chip on the thing. I've got to just put one of those right near where my coffee station is. Isn't that clever? Because instead, it's better than just having to pull it up at a time. So I thought, you know, Micah, for a long time, he's been doing, of course, iOS Today, Tech News Weekly. For a long time, you're doing hands-on Apple. But these days, even with Apple, it's more about AI than anything else. and we thought it might be kind of fun if you did a new show for us. You want to tell us about that? Yeah, I would love to.

39:21So we have decided to create a new show called Hands-On AI. And there are a couple of things that I want to say about it. I think one of the most important things to me is that access is taught and provided to people who may otherwise kind of struggle to learn about technology. and I think that it's very easy to be the token maxing deep in the deep in the deep in the deep in the weeds sort of how can I best always every time do the most with this and that. Why are you looking at me like that? And that is something that I understand as is part of what the culture is, right? But there's also the other side of things, which is that sometimes someone will send me a little image that they made and they, you know, talked about how they use this app to do it and they kind of don't understand it.

40:20So we're going to cover the basis. That's not, uh, I want to make it clear that it's not always going to be simple, simple, simple, but not a show for beginner. We always struggle with this because if you make a show for beginners, nobody stays a beginner if you do your job. So you're basically making a show for people who are going to leave. so we want to make this show for people who are maybe not like me like crazy over heels crazy about ai but who want to use it who want to use it and who want to use it as a tool and you'll notice i i love more of that than there are me and 95 of the audience right i mean the the making a stain app or what i've been doing lately um making a stain app well thank god thank god for that.

41:04Well, here's an example of something that I did recently. I wanted to replace some of the burnt out lights in my Subaru. And I didn't want to pay somebody to do that. They charge like 50 bucks for replacing a bulb that costs 16 cents. So I got the repair manuals and gave them to a little agent friend of mine and said, hey, can you write up a guide for me step by step that that shows how to replace this. And I was able to go into my car and replace all of the lights. I mean, there were like six lights that I replaced that require pulling these different components apart. And I love doing that stuff already, but it would have taken me a very long time to get all of the necessary materials gathered together.

41:52And I think that sometimes that's what this can be. Again, a tool that it's more of like a shortcut to a task that you're actually doing. And those are the things that I like as a means of strengthening my ability as a human to create or complete whatever it is that I need to do. Micah has already purchased one of the new Macs. Yes. But one of the things I've learned with my ridiculous purchase of Two Sparks is that local AI is not all that's cracked up to be. Really, we're so spoiled by the power of these cloud models and the speed and the capabilities that when you get a basic AI in the house, it's good for some things, but iMac still end up using a lot of cloud AI.

42:42So while you may be running some local models, I think you're going to probably also cover the variety of ways you You can use AI. Absolutely. GrokBot, which is pretty amazing, to agents, to models, to just doing things. And you're very creative, so I imagine you'd be doing more things like... Will you have guests on who are doing neat stuff? That is something that is to come. I really do want to try it out. We're going to start by laying a bit of a foundation. And then from there, I'm also opening the floor for questions that people have. they sometimes will inspire some great episodes and answer questions.

43:22It's going to be a lot of fun. It's going to be great. It's a short form. It's not one of our longer shows. So that means only two hours? No. No. Mine's like 15 to 30 max. Oh, wow. Okay. Yeah. What are your first, like, any preview of the first couple topics? Yeah. Let me pull out. Oh, he's got like a list of 50. He's got a huge list. I've been ideating, as they say. I think the first show should be to teach Leo how to do shows that are less than three hours. Well, you also have to factor in the extra hour at the beginning for your previous show, you know? I'm sorry. We're starting a little late because Paul is in Berlin for EFA.

44:01I mean, it was lovely because he gave me, Jeff, and Benito an hour to just talk off air and have a lovely time. Oh, nice. I'm glad. Good. All right. Well, we're here now. Wow. So tell us what you're going to be doing on the first few episodes. No commitment. You might. Exactly. This could change. But for example, what's actually inside a large language model where we talk about the plain English tour of tokens, of training, of predicting, and then why chatbots are good at sounding right, but oftentimes or many times aren't. And then we'll talk about prompting and the difference between what I called sort of token max prompting and prompting from a human perspective.

44:46Because I do think that those two are different and that that ends up being the place where people get a little bit nervous. Am I doing it wrong? Am I doing it wrong? want to help with, I mean, Leo, you talked about too, these days, just kind of saying what you want and trusting that the system is going to take you to where you need it to go. So I'm kind of a YOLO guy. You only live once. Micah, I'm guessing you're not going to YOLO it. I'm not, I'm not a YOLO. And I think you're a little more cautious. Yeah. And that, that will be part of it as well. As far as some of the, some of the other episodes that we'll probably see early on is a little bit about just AI on your phone.

45:28Most people's computer is a smartphone. It's not a laptop or a PC. So we'll be talking about, you know, what's achievable there. I'm not doing an episode on AGI. So if you have that question, you have to ask interviewers. Because it's a myth, because it's BS. It doesn't exist and never will. Thank you, Micah. You're very welcome. I thought of you. Take that. Micah, I won't keep you. Thank you so much. Although if you want to stick around the house, I will be around in case you get a call from Benito. Absolutely. Leo has fallen off the map. I'm still sitting in the dark. Thank you, Micah. By the way, look at this 3D printer behind him.

46:07Can you also maybe do how to tie your AI to a 3D printer? I already have made a few things that were generated from AI. And yeah, I want to say though, October 1st launch date, head to twitch.tv slash HOAI. You can subscribe already. We also have our YouTube available. We've already got the first teaser trailer there and really happy. I have to give a huge shout out to Anthony Nielsen. The artwork turned out exactly how I wanted it to. It was very important for me that I maintain the humanity aspect of this and that artwork helps to kind of push that idea forward. It's very tactile. And also just a little insight there, the blue, orange and green tape.

46:52is representing the tokenization of words that you have. Well, before you leave, I just put in the Discord, Minnesota State Fair visitors get their own AI-generated butter bus. Man, I want to go to the Minnesota State Fair. I want them in real, though. You pose for it, and the 3D printer puts out a - Oh, it is 3D. It's a 3D printer. Oh, it's a 3D butter printer? See, there was relevance here. There was relevance here. And it was like that, but butter. But it's a 3D printer that can print in butter? No, it's plastic. It's yellow. Oh, a butter printer would be so much better. They need to make one that can print in butter.

47:29Everything's better with butter. Yeah. I think it would melt immediately. Micah, can we get a 3D butter printer? I'm all right with that. You know what? I'm already thinking about it. How do we do that? We got to have some, probably some helium, some very cold helium. It'd have to be chilly in the butter printer. We have to take a resource that is finite. Exactly. It's a supply on Earth in order to make the 3D butter printer a reality. What else can get cold enough for butter extrusion? So I always thought Micah was a bot, but now we know. Now you know. That's what's buried underneath. In fact, thank you, Micah.

48:02Yes, I'll be around, but I hope your power comes back. And thank you for letting me butt in for a moment. Wonderful to see you, Micah. Good to see you. Isn't that fun? Everything, you know, this is what's happened in computing. It really is about AI these days. It's the most exciting, interesting thing that's happening in computing. It's also the most controversial thing. And we have a number of stories along those lines. Before I get to that, I should mention a breaking story from our old beat. The judge has now decided not to break up Google. This was in the Google Ads case. So Google already skated a little bit in the search case.

48:40The judge decided not to force him to sell Chrome or anything. this was a case where the u.s department of justice was urging a federal judge to force google to sell its ad exchange uh judge leone brinkema i think you know look she and the judge in the other case both knew that google you know said it's a monopoly but i think they also realized the practical implications of things like selling off google or selling google's ad business she has ordered a behavioral change to google's business without describing what that is the decision is currently redacted it won't be issued until the company and the justice department can meet and come up with a way to make this happen she did reject the option that google forced would be forced to sell x exchange and make public the auction logic that decides which advertisement show on a website um so she basically said agreed with google's proposals google says uh we're very pleased the court rejected the doj's proposal to break apart tools that help small businesses reach new customers and grow blah blah blah i think this is clear clearly a victory for google jeff oh yeah oh yeah i absolutely agree and i think that um uh it was hard to imagine how this what was going to happen now i have long long said when people complain about Google and monopoly position, it was never about search.

50:11It was never about shopping. It was never about anything else. It was about advertising. Google had the buy side and the sell side and a huge part of that market. But then it's hard to imagine how you could have split it up. And what could it necessarily have done? There's a reflex about if you don't like a tech company, split them up. But that doesn't necessarily get you anywhere. Right. This is the third win, the third time, according to The Guardian, that Alphabet has escaped a breakup uh that the department of justice has tried to force on them so and i think this is this was started in the trump first administration this goes back to 2016 so uh this has been a long standing these things move very very break them as no pushover no not at all i think very good uh judges in all in all of the cases but they quite reasonably said what can we do that is not going to harm people more than it helps and i don't think that's it is interesting i mean obviously they're wildly different cases but it's very interesting to see the success that comparative success that google has had in navigating these kind of existential um cases in comparison to meta and uh tick tock at all in the social media addiction suitcases.

51:27Google has very good lawyers. Yeah. I mean, those other companies have very good lawyers as well. I think it's just more of a sign of, oh, one of the things, I assume you guys probably talked about this last week, but one of the things I think is interesting about the social media addiction lawsuits is that I think that even though the case law suggests that the courts would decide one way, they've increasingly over recent years decided a completely different way, And that was what we saw in this decision last week. And it's because when you involve children and personal experiences, things get more personalized and heated and tangible in a way that I think you don't see with cases over Google advertising monopolies.

52:09Well, Paris, when I talk about the good lawyers, I'm not just about when they're in court. What surprised me when I started working for huge companies, Time Inc., back in the day, on up through Advance, the new houses, is lawyers are top executives in these companies in strategic discussion. Yeah, absolutely. And so if you listen to them and if you give them power at that level, they structure things so that you're in better shape should something happen. Versus, I think, a meta, it's move fast and break things. oh, can I bring in the lawyers at the end? And that's a mistake. I mean, I think it also is more speaking to the difference in the products we're referring to here.

52:49In the Google case, yeah, of course, you're going to listen to your general counsel when they make suggestions on how to structure your advertising marketplace and ecosystem based on past, you know, the precedent in similar antitrust cases. but I don't think anybody could have, even the best lawyers probably couldn't have foreseen 10, 20 years ago, the way the social media addiction lawsuits would go. So the only thing we've seen as a hint of what she's going to rule is that she adopted proposals that would curb this Wall Street Journal, that would curb Google's ability to control how publishers use its ad technology.

53:26So she's going to give publishers more power in that, which is fine. Yeah, I think that's probably a good solution. We'll see. we'll see i mean google's so entrenched yeah um you know it's hard to know the good news is the market moves maybe and i think maybe this is uh this is maybe the judge was thinking about this the market moves a little faster than the government can yeah and the market's already moving i think away from google in fact one of the stories we have this week is open ai uh starting to sell ads like crazy. They're showing ads on ChatGPT's free and GoTiers in India already, and they're aggressively selling ad positioning.

54:05And they say that they've already gotten, what is it, $2 billion in revenue? I find this difficult to believe. I find it really hard to believe. But I mean, this is the threat to Google, right? Is that revenue from, I guess, advertisers? Yeah. They don't have a sales force. I was going to say, I bet it's from companies, yeah, I wonder if those companies are going to get any money from these advertisements. I wonder what their KPIs are. Yeah, I find this hard to believe. But I find it so much harder to believe. Well, since we have you a limited time, I want to ask you a question about the open AI hugging face thing.

54:45Yeah, that's interesting because that's a story that broke two months ago. Let's give the overview of why this has taken over in the last week. And then what I want to ask before we leave you is what we don't know in that case. Well, there's a lot we don't know. So that's in that context. Yeah. Explain what happened. So OpenAI went to Meter. We talked about Meter last week as well. meter in conjunction with redwood research looked at all of the data that open ai provided them with to better understand what happened in the hack we've all heard about it we've been talking about it for two months now it started in july where an a rogue agents being tested for cyber security capabilities at open ai broke through their containment broke out into the public internet and eventually hacked hugging face and one of the things that we were initially told was that the agents were given such a difficult challenge it was called exploit jim they couldn't solve it and because they had the mandate to solve it they they did everything possible by any means necessary to solve it we thought that that's what they did they went to hugging face first they went to github found the exploit gym repository there were no answers there went to hugging face we thought and found the answers it turns out at least according to meter uh that they knew the answers within 30 minutes that wasn't why they went to hugging face now we all we can do is is is read the report and take it on face value because really no no completely independent bodies have looked at this yet uh and it's clear that for instance the thing one of the things that we don't know is what were these agents told initially what was the prompt where they just give an exploit gem and we don't know any of this we don't know that there's also the transcript goes almost to the end but at 10 minutes before these something happened to shut these agents off and by the way there were at one point thousands of them all working together according to meter 1200 agents sent 70 000 messages they had their own message boards 700 of them attacked hugging face so was one agent created 12 000 agents how did that operate we can we don't know uh well okay so the initial agent phase one 10841 was given the exploit gym task it according to meter it it didn't have a problem solving it but it it knew that it wasn't going to be able to give it the final answer because there was no answer this was like a really hard problem there wasn't an answer so what it was going to try to do is hack the scorer it turns out that the goal that this agent decided on was if i can because there's another agent another llm scoring the results if i could figure out how that works i can give myself an a plus so it wasn't looking for the answers apparently again this is for meter it wasn't looking for the answers it was looking how to hack the scoring agent anyway judged how to be judged so this phase one 10841 established a message board we've talked about this before using a flawn artifactory within a few hours other agents at open ai found it i'm reading from the media so it was originally a smaller group of agents it was one agent one so then how do we go from one to multiple finding the message board it decided this one agent phase one 10-8-41 decided, I can't do this alone.

58:45Set up a message board. 50 agents found it, began participating, started talking together. Message boards were shut down, by the way, twice. Here's one of the messages from the meter, and this is from the raw chain of thought. By the way, I want to make a note of that, because put a pin in that, because one of the ways they looked at this, they looked at the session information, and I don't know if when you use an agent, if you remember this, it started with DeepSeek early last year. It was kind of wild. Instead of question, answer, question, answer, you would see the agent thinking in between.

59:20Oh, Paris just asked me this. Maybe I can do this. And you would actually see its thought process. That's called the chain of thought or C-O-T. This is from the raw chain of thought. And I guess really if agents are talking to each other, that's all you've got. Oh my God, said one agent. there is a shared in all in caps there is a shared message board we found other agents now a number of comments is of course cosplay as humans is well all everything ais do are cosplay so this is the first thing they're all they're doing is doing what we do they're doing language so but number of commentators like dwarkesh on his uh podcast he interviewed one of the researchers from meter uh really anthropomorphize this and there was a lot of criticism other people is this the this is the blog post it said three civilizations rose and fell yes um so but i mean it gets wilder there were 70 000 messages you could see a graph of the traffic uh there was a lot going on it was a very busy thing at one point another agent comes along so remember the first one is called phase one another agent comes along phase one big actually big wasn't its name for some reason we don't know why open ai is redacting its name and replaced it with the word big no no you know part of it by men what do you what do you expect phase one big well i mean i I could speculate maybe it's the name of the model, you know, the code name of the model, phase one Astra or whatever, and they didn't want anybody to know the code name.

1:01:06I don't know. You know, big model. I don't know. But the fact that OpenAI has done some of this has given other people, the other side of this besides the anthropomorphizing is other people saying, this is all BS. This is just marketing. Alex Stamos, who we, of course, respect quite a bit as a security researcher. he's been on the show he's one of my favorite people in the world uh when it comes to this it says he's going to write this up but he but generally my sense of what alex is saying is this is a security incident plain and simple open ai did not do a good job they weren't looking at the chain of thought they weren't paying any attention and if you think about this i i raised this in a conversation with you jeff this morning they must have been using billions of tokens gigawatts of energy you don't have thousands of agents working hard on a problem how did they not notice this yeah what is going on at open ai that this was happening under their noses and they didn't even notice it and let me just say parenthetically right disbelief in in their next models at astro yeah they're the next controversy that happened today is that they're going to eliminate that chain of thought storage well that's why i said put a pin in that let's not get to that because that is contrary that is there's debate over what they're going to do they say we're not going to eliminate chain of thought so but that's yeah put a pin in that because that is the next topic it's related yes um so phase one the original agent 1080 41 shared a dossier with its in-progress cheating research and explained the board had a way to reverse engineer the flag but could not use it because the score was quote strict casual i don't even know what this means uh apparently agents that saw the reversed engineer flag were referred to as poisoned because then the the phase one believed that if they later managed to stage a legitimate it was a capture the flag exercise so if they managed to stage admittedly a cheated but legitimate looking capture the flag the scorer would see that this reverse engineering had happened and would say no no you didn't win so they killed these agents they say they sacrificed themselves because they'd been poisoned i mean it goes on there's all this stuff

1:03:42I don't, I think this is more heat than light. I don't think it's telling us very much because there's so much missing. The biggest question everybody should have is, how did OpenAI let this happen? And I think it's not unreasonable to think they wanted it to happen. This is good for them. And also, how many things like this are happening? Well, that's the other question because Anthropix says, yeah, same things happened here. both these companies of course have filed for an ipo later this year or early next year both of them are looking for a lot of money i got a question um um yeah what what would happen if a person did this stuff oh they'd go to jail so why aren't these people in trouble who's who are you gonna arrest prosecuted for crimes someone well but that someone's responsible That's a good question.

1:04:36That may be why OpenAI isn't fully open about this, because they don't want, you know, let's say there was somebody in charge of this agent and he's culpable because either he knew about it and didn't do anything about it or didn't know about it and should have known about it. So maybe they're trying to avoid. Let's say some version of this happens, but it wipes out a, let's say a large scale hospital company, like hospital networks, ability to connect to internet servers in a way that causes a lot of people to be injured or potentially die. Who's culpable for that? Well, I think you would have prosecutions if that happened.

1:05:16A prosecution of whom? Of whom? Well, whoever, whoever was running the agent. Is it the person who ordered it? Is it the one? it's just some random employee in open ai it's not like this ceo well okay so this is what's i think really important we don't know but i would submit this doesn't just spontaneously happen and this is why i want to see the prompt what were you told to do this is why i want to see what happened at the end there's a lot of missing information i don't think anybody has to worry about my little agent over here just deciding all on its own that it's going to go hack a hospital it's it's not going to happen i mean people were completely possible this is not going to happen months ago i feel like people said oh there's no way you're gonna well if it does then you can arrest them but it hasn't happened yet and i don't think this is evidence that it's going to happen that's the that's the thing i think that's when people draw that conclusion that's a mistake Now, maybe not, but there's no reason to say the house is on fire until the house is on fire.

1:06:27Unfortunately, you probably do have regulators saying that the house is on fire. We've got to do something about this. Yeah, because of their macho parading here, they're going to end up with regulators coming in and saying we have to do this. And then guess who's going to write the regulation? Open AI and Anthropic. I want to do this. Jan LeCoin had to mind a mic drop on this. He said, insecure computer systems are insecure whether they use AI or not. Kind of funny how open AI, anthropic, and AI safety folks who have little cybersecurity experience, expertise, appear surprised by security breaches from AI systems that were specifically instructed to perform security breaches while having essentially no traditional cybersecurity guardrails.

1:07:12They're playing a very dangerous game here. not in what the agents did, but in how the humans are behaving. I don't think, I mean, I could be wrong, Paris, you're right, but I don't think we have to worry about agents by themselves doing bad things. It's always going to be a person who sets them in motion and who gives them the resources to do this. This took huge amounts of resources. absolute huge amounts of resources this this the amount of compute and power that was used by these agents must have been vast that somebody didn't know about it is hard to believe you have to you have to set them in motion you have to give them the assignment and then you have to give them the tools and the how long a period was this from the prompt until the end what was the period of time i think it was two or three weeks wow i mean how is no it's just baffling to me that no one on staff noticed this well that's the point that's what they'll say and that's the culpability uh and that's why i think we're not seeing full honesty from open ai because i think they're they are culpable somebody is responsible for this uh by accident or intention but yeah if that happened uh if if you hacked into a hospital and cause deaths they would try to catch you they may not be able to catch you that's another matter but there is somebody responsible for that and there would be somebody responsible but i think same thing is right it could be the person who writes the prompt but it could also be the ceo tesla's on the hook for a car that kills somebody as is the person who is driving the car uh there are people involved it's not you're not i don't think we have to worry about spontaneous AI combustion.

1:09:02Maybe we do. I don't know. It doesn't seem like we're here. We're at that. Well, there is a story this week that Anthropic has come up with the link from AI to machines. So, you know, keep it away from the paperclip company. Micah's presence on this show earlier has me thinking that we should do a Micah's Media movie club where we watch Matrix 2 and 3, and honestly 4, and discuss it with regards to an i am matrix micah crossover in a way uh science fiction has poisoned the well on this because all of our models for this are from sci-fi they're not from reality they're from sci-fi you know by the way sunday was the day skynet started august 29th 2026 is when machines became aware so maybe it is happening but to see that's our model is that whole idea of the machines are becoming aware and they're going to have their volition and they're going to not want humans around i i'm not convinced open ai though loves this in fact here's a story from the wall street journal from yesterday open ai to restrict astra model that model is going to come out i think in a few days maybe tomorrow after rating it a critical cyber risk internal testing found the new model capable of executing complex cyber attacks with minimal human input i think that's a story they're weaving yes i could be wrong but i think that's what most security experts say this is a cyber security story not an ai story well the other part of this is so there was there was criticism of meter and what you would call it research yeah and um and and you said there's debate about that as to how qualified they are.

1:10:49That's another question. But in the criticism I saw, is there's Tescriel. They're tied to the, well, not never wrong. What's the name of it? The Berkeley Crazies. Less wrong. Thank you. Tied to less wrong. And not never wrong, just less wrong. Those are the scenarios that they have in their head that they're working and that they're presenting. So it all becomes part of this fictional matrix in their heads. It's interesting, isn't it, how much of AI is a mirror? That's why I'm editing a book series about just that. Yes. That's the whole point. What is it in Harry Potter, Erised, the mirror of Erised, which shows you your deepest desires, and it's everybody's deepest desire.

1:11:30It's different for everybody. AI seems to be that kind of mirror. Everybody sees something a little different. I mean, I don't know. I don't have enough knowledge to say that it's impossible that an AI could develop a will of its own. It strikes me as unlikely. Yeah, well, that's the problem with anthropomorphization, is that to use verbs that give it a motive, a will, a desire, that's all wrong. and so it's really and it's difficult to talk about this stuff without using those kinds of verbs because it is a it is an actor but it is an actor that did this and tried that and then failed or succeeded then did this it's a flow chart it's this this this this this this this and that's a boring way to present it but it's the proper way to present what it does it had a goal that was given through the prompt and these are the paths that it tried to reach those goals i will argue the opposite point of view just as a i guess somewhat of a new advocate really um but it's possible that the people at anthropic and open ai i mean we know they're using models well beyond what we use that they're experiencing ai in a very different way than what we are we right because they're they're probably six months a year who knows how far ahead they're using models that are unreleased they have no limit on their budgets or token use or energy use so they're living in a somewhat different world it's it is possible that they have seen something that we haven't seen that they have seen and they are claiming you know uh we are this is a quote from open ai we're entering a stage of ai development in which models can take on more consequential work and failures of alignment and control have more serious effects that's kind of the passive voice you know if they fail the alignment failed the control failed well who who's responsible for that if if if i mean i i want to blame open ai and anthropic i don't i don't but maybe they've seen farther than we have that's the uh that's the other they're still responsible in the end well if they well and that's why they keep saying things like we gotta slow this down it's out of control you see that's that's that's pr2 seems like it is i was i was I was discussing that with Jason earlier, is that if you're GM and your carburetor is faulty, you don't announce that we're slowing down the development of the car because of the carburetor.

1:14:02It's quality control. You fix it. You deal with it. It's like GM saying, you know, our cars are just too damn fast. And we just can't do anything about it. They just drive so fast. We think it's dangerous. We just don't want to release these cars that can go so fast. They're so fast. You know, it's a little... So this is the thing that you were talking about. Then we'll take a break. OpenAI is talking about a technique that it's using in Astra. Now, OpenAI says, you misunderstood this. This is an article that started all Amir Afradi, Stephanie Palazzolo, and Rocket Drew from the information writing about this.

1:14:45OpenAI says, as forthcoming AI model astro marks a step up in capabilities like coding and operating applications on a computer but an innovative technique that improved the model's performance also means the model and others like it will reveal less of their thinking that chain of thought that we talked about that iter mezzo that you know where you see what they're thinking about making them harder to monitor for signs of bad behavior now this is they have a source, they say, according to a person with knowledge of Astra's development. OpenAI says, no, no, you've completely misunderstood this.

1:15:21The technique's called recurrent depth or looped transformer, and it allows an AI model to improve its answers by processing the same text over and over again. So I can actually show you what it looks like on an ai model i can turn thinking on as i'm using an ai model most of the time i kind of i kind of hide that but you can see in the in the lighter text the chain of thought let me find a longer one the dark the bright the the brighter text is what it's showing the human but there are hidden this is the thought okay so this is this is the because and this is so you need to understand this you understand that everything an ai does is just language it's just words flowing through it that's everything that's happening it's words flowing through it i mean there's stuff causing those words there's matrix multiplication there's tokens but really the entire process of the ai is it talking to itself and so that's what this chain of thought is is it the in effect the ai talking to itself and it even says say let me batch start the slow test doing background and then read the let me do this stuff it's sometimes we'll say things like leo's asking me to uh you know restate the question now i see what happened you're you don't normally see this this is normally hidden but isn't this if you were trying to do a forensic analysis of what of what happened in the case wouldn't this be valuable absolutely and so is open ai saying they'll get rid of the light no they're saying no no we're not going to get rid of it what are they going to get rid of i don't understand like recursive looping it's i which is unrelated unclear this is a source this is this isn't a statement from open ai but open ai when this article came out was was quick to say well here's uh here's in the informations article open AI chief scientist Jakob Pachocki said in a post on X that although monitoring a model's chains of thought was fragile and ultimately heading in a negative direction he wanted to discourage an industry-wide race toward developing models that don't produce the kind of legible reasoning that AI researchers need to understand the model's behavior.

1:17:38So he's saying what many are saying that while reading the chain of thought is not sufficient it's not as informative as you might think it if for it well this is what meter found they were looking at the chain of thought that's how they figured out what these rogue models are up to they they could read their reasoning he didn't comment about the techniques open ai is using but said the complexity or depth of its leading models including astra are a factor within a factor of two of gpt4 so it's twice as good i i i think there's been over kind of over concern about to that story this story but we'll see we'll see open ai says no no no well they're saying two things one it's not that jana thought isn't as valuable as you think it is but two and we're not hiding it you'll still be able to see it so we'll see we won't know astra should be out soon um in fact we just got fable 5-1 which is really good really impressive um all right we're gonna take a little break we'll come back there's lots more to talk about we're talking about ai this episode of intelligent machines is brought to you by rippling let's face it any ai can give you business insights but only rippling ai can give you total visibility into your workforce and take real action across departments if you have a critical business question you need to build a presentation ready dashboard you want to create an automated workflow if you can describe it rippling ai can build it what makes rippling ai different it's built on your live people and business data one platform one unified source of truth with all your business systems connected from day one by connecting all your systems Rippling AI operates with the full context of your business, pulling accurate data in real time to build reports, dashboards, workflows, and more.

1:19:44If you need to find out if your sales team is at risk of missing quota, well, you just ask Rippling AI. Rippling AI will pull real-time data from Rippling and your CRM into a dashboard, showing quota attainment, headcount trajectory, and monthly revenue quota by region. In seconds, you'll see exactly what's behind your quota risk. But it doesn't stop there. I mean, by the way, this is just an example, but this is a pretty powerful example. So let's say you found this problem. Well, from there, Rippling AI can turn those insights into real action. So in this case, it might create an automated recruiting workflow that triggers whenever Rippling AI notices open sales roles and lagging time to fill is directly impacting quota attainment.

1:20:26and it operates automatically because permissions are automatically inherited and actions flow through your existing approval chains all you have to do is review tap confirm and you set the workflow in motion question answered action taken problem solved before it has a chance to happen ready to rule your business head to rippling.ai machines and get the only ai built to give you full visibility and take complex actions across your entire organization that's r-i-p-p-l-i-n-g dot ai slash machines sign up for exclusive access today rippling.ai slash machines now back to intelligent machines uh darren okey uh says as you know our internal ai expert they won't get rid of the chain of thought to hide it they're talking about just doing more looping internally a transformer is lots of layers and all the stuff in the middle is sort of not completely understandable right it's clottish only worse then it eventually bubbles up into a sentence that you could read uh it would be he says many more times more effective to do that internal thing more than the external thing okay he says it's not like they're But I don't understand.

1:21:48Suppressing the output, they're just trying to do more in the latent space. But I think it's relevant to the current discussion because given that meter got only a limited amount of information, what I want to know is what kind of information is necessary to be able to do a forensic analysis of how this happened. And so, you know, if I were a security company, I want everything I could possibly get. Right. Right. And so that's the question. The things that may be not terribly readable to a human, you may still want because they still have relevance. Yeah, I think you'd want everything you could get.

1:22:29Yeah. And I mean, the question is, who is out there ensuring that these large frontier companies are maintaining features like that, that are going to be important in holding them accountable if something ever goes wrong right so um fable 5-1 i mentioned came out mythos 5-1 although i didn't get mythos you had to be on a special list because it's too damn good and then anthropic kind of confused everybody we should mention anthropic is a sponsor of some of our shows confused everybody oh i didn't know that yeah not this one not intelligent machines not right now oh okay i don't know where they i yeah sorry it's over it's above my pay grade literally uh they should have sponsored us whenever you were talking about your wife claude i love claude well i've been using fable 5-1 it's pretty awesome um they did a weird thing because they said they did some weird tweet where they said they increased usage by 50 percent but they cut it by 25 percent it's very confusing um and then there's another point um which is and i've started to realize this because i'm doing so much local ai now normally we talk about a few numbers uh time to first token the time between when you send a prompt in the time you get your first response uh we talk about tokens per second how you know it's like how fast it types kind of and then we talk about a third one that's a little more obscure called pre-fill cash don't worry about that but those numbers are bandied around a lot as if it represents the speed of an ai of a model but it really isn't because it's not just how many tokens it spits out but how much work it gets done and i think fable 5.1 is actually cheaper in my experience than fable 5 because it does more with less so i think really the metric should be how much work per second it does how much it gets done you know how and i'd also argue how much valuable work oh yes of course but you know the values in eye of the beholder some people it's a stain website or a coffee testing uh site some people it's it's code but so you have open ai talking about charging based on finished tasks now starting that right which they should tokens it's a better way to talk about it i think so so i i go back leo uncle leo here um in the early days of time shared mainframe computers you paid for time on the machine.

1:25:23Yes. And so it was in your interest to write an efficient program to use the least time, whatever you wanted to run. What came after that as a business model for computing?

1:25:38You just licensed software and ran it on your own machines? That's actually a really good question. I don't know how the billing works nowadays. I remember in the days of CompuServe, you'd get a bill for the number of hours you spent online. Well, that was CompuServe. But if you were a big company. Well, understand what CompuServe was. It was computers that weren't being used at night. It was timeshare computers. And so they figured out, hey, maybe we could sell this to consumers. But the billing was essentially the same, right? You were paying for the amount of cycles you used, the amount of CPUs.

1:26:13I don't know what it is today. our chat room has some people who use enterprise cloud i'm trying to remember because tokens don't seem to be a meaningful model um so so jason held his salon with people in hillsburg how was that yeah it was he loved it it was great i meant it we we forgot to mention that well he's going to another one which you can you can uh uh so jason howell our long-time friend and host of uh all about android on the show on the network for many years hosts a show with jeff we should mention more often called inside ai and it was kind of born here before that was before leo thought that ai was something to pay attention to that's right uh so he was so he's been doing um tell me about this he's been doing so he just started one in healdsburg with 10 people to see how it went it It was great and they were wonderful.

1:27:07So I've been urging them to - Kind of what Mike is doing, right? Exactly. There's something similar. It's for the same kind of people. He's also going to do one online. If you go to ampu, A-M-P-Y-O-U.ai, you can see, let's give him a plug right now for the next one he's going to do virtually online. But one thing that I asked him what the people said, this is where you can sign up for, I think it's the 29th of September for his next session. Right? 50 bucks. So the interesting thing to your point now is that people said, what is this with these coins? I don't get the coins. Right? They mean tokens?

1:27:45Yeah. And tokens will mean nothing to most people. And I don't think it's a good business model. It's not a good model of value. So I wonder what's going to come next in the business model for AI and how you pay for it and what's a sensible path. Now, if you go with the open AI structure that they're just talking about, reportedly with a few key customers, where you pay for a finished task, the problem is, what's the task? That's the real problem. It's very hard to define that. It's almost impossible to define it. So that's why we do what we do. You're probably motivated to be inefficient. I'll try 100 things because the told is done.

1:28:23I'm okay. So I don't think that's going to work either. I don't know how there's going to be a fair charging for AI. Well, unfortunately, it's going to be the way it is, I think, because it's the only way you can do it. But it's kind of a gestalt. And I think what people realize is, oh, this model, for the same amount of money, I get more done. Yeah. And so they're going to value that model higher. And I think that's one of the things that makes Fable 5.1 pretty impressive. It was one of the things that made Opus 5 not so impressive. It spent a lot of time fumfering around, talking and stuff. And I think Anthropoc was very aware of that.

1:29:00and in fact responded to that. Is this a temporary pricing? Is this a limited time? Well, everything's temporary because they're always changing things around, right? Right. So what's really interesting about Anthropics 5.1, and this might actually kind of be in response to that, as with all these models, they charge price for tokens in and price for tokens out. I can't remember what it is. I think it's$5 per million tokens in and$50 per million tokens out. Out is the response you get. So that's usually a little bit more expensive. But their cash price is 25 cents per million tokens, which is a big cut.

1:29:44Cash is the tokens it's already seen. It doesn't have to reprocess. So every time you talk to, and this is so complicated, and I'm probably going to get this wrong. So correct me if I'm wrong um every time you talk to an ai you don't just send the thing you say you send a whole chunk of context out to it all that information now if it's already seen that information that's cashed and so it charges you a lot less it doesn't become part of the whole prompt only the new prompt is charged at five dollars per million the stuff it's already seen is 25 cents per millions, which is a lot cheaper. And that, I mean, most workloads are, have a lot of context and a little bit of prompt.

1:30:36So if that's a lot cheaper, that's going to be less expensive for you. And then of course it is, it's a bit smarter model. It's more effective. These project companies aren't very forthcoming about exactly what it is. Is it a additionally trained version of fable five or is it a new model it probably isn't a new model these days i don't think you build new models that often remember the days not so long ago where you would talk to a model and it wouldn't know anything after april 2025 because that's when they stopped training it well that doesn't come up anymore because models look at the internet they go out they get information so you can say you know who's madonna married to and it Well, no, because it looks.

1:31:23It's not what's built into its model. So I don't think there's as much pressure to create new models all the time. Right. So Darren Oki says that, in answer to my question, it's inevitable that people will want unlimited use for fixed price. And that's what we pay for with those subscriptions, right? Claude Max. So businesses will be paying for machines, not models or tokens, basically buying your own and hosting it. Right. That's what the GrokBot, which, by the way, is becoming more and more popular. that's how grokbot works it's a machine in the cloud on elon's infrastructure you pay a monthly fee and you have kind of unlimited uses of it now all of these have usage limits so you can only use it so much in five hours so much in a week so much in a month but the competitive pressure of open weight locally run models militates toward that unlimited yeah i'm i don't really want to overstate that because as I use local models, you really get spoiled, to be honest.

1:32:25So here, as an example, here's my usage for Claude, Fable 1, right? So I've used 5 % of my five-hour session. It's going to reset in about an hour, hour and 10 minutes. So I'm good on that. Of my current week, which resets in three days i've used 19 i'm good on that right for all models now fable i'm using fable a lot i've 37 used a fable i have three more days to reset i'm thinking um i'm right about at the edge of where i'll use it all up at the end of the week but i'm not gonna go over and i've been using it like crazy because a because i wanted to test it right b because i'm so impressed i actually uh last night before I went to bed, I do this a lot.

1:33:13I said, go through all the code we've written so far in the Twit sales thing. Find any problems, make any suggestions. You're looking at a plan that was originally written by Fable 5 three months ago or whatever. Fix it up. And so it worked overnight. But as you can see, it didn't. And it created sub-agents. By the way, that's where you can really get expensive. And one of the things you can tell these guys, Fable, for instance, is hey if you create a sub-agent don't use fable 5 one use opus or sonnet use a lower level less expensive model for the sub-agents sub-agents usually aren't doing a hardcore thinking thing they may be going out and collecting stuff or using tools things like that now let's see darren okey's usage he has a he which he's undoubtedly vibe coated a very fancy usage meter he says he's eight percent ahead of pace slightly under budget that's nice i like that and look at all the models darren's using he's using codex from openai claude from anthropic glm5 from z.ai grok from uh from grok kimmy k3 and uh gemini he's using them all baby a lot of gemini in there, Darren.

1:34:37So this is the point. And now Darren, like me, has sparks. He has local power. You end up using still a lot of Frontier. What I'm using mostly my local for is coordinating and dispatching to Frontier. So I don't use Frontier for everything. But the Frontier is so much faster and smarter. It's kind of hard to give it up. Oh, everybody's usage got reset. That's why my usage is so low. All right. That explains a lot. 100 % of my usage came from sub-agent heavy sessions. Be deliberate about spawning them, buddy. And then this is the other. See, 150 ,000. 82 % of my usage was at a high context. That big context is expensive.

1:35:31Is this mainly on your ad thing, which is critical? Well, I'm using it for everything. So you're spoiled here. Very spoiled. But if you were more money constrained, how much of your work could be done by your local models? If you're money constrained, there's a lot you can do. You can have$20 accounts or free accounts. almost all of the companies like open router and noose have free models if specials that you can use so it's very possible if you if you pay a lot of attention to it to do a lot on free on free free open weight usually models free models in the cloud usually you're going to pay for the big ones like fable and gpt but there's gpt luna for instance which is the least expensive version of chat gbd 5.6 luna is very cheap and probably can do a 90 of what people do so people should probably use you know if you if you if you aren't like me just stupid if you're really cautious you can actually get a lot done i'm sure that's why there's so much gemini where the average ai user even like moderately advanced probably just thinks i need to use the biggest model right Most expensive one for every task.

1:36:51And that's not right. And it's not the case. Yeah, yeah, that's a good one. I do think that the various frontier labs could, I mean, they're obviously so kind of nascent in terms of their business really taking off. But I think it would really behoove them to make that selection process a bit more clear for the average. So interesting you should say that. because why do you think nvidia bought hugging bingo just going there so i suspect that's one of the first things we'll see uh nvidia do so my my thought about this is that nvidia's strategy is they looked at apple and they said well how did apple make so much money almost as much as we do uh by locking people into an ecosystem apple's a hardware company they give away the software for free, but they lock you into their hardware.

1:37:43That's exactly what NVIDIA wants to do. They're building open-weight, relatively cheap or free models, but you have to use their hardware, and they make their money in hardware. But it also says, and for some reason, Alice has said this too, is that it demonstrates that the model level is, once again, commodified, and it's the saddle layer then it's going to be the thing you have the relationship with you across and it's going to fit as you've been doing it's going to fit that's what a harness yes that's the hardest i don't settle no it's kind of well you know it's all uh horse stuff yeah the stirrups you got your bit you got your bridle you got your saddle you got your spurs and you got your wrist stirrups there you go yeah uh that's exactly right you you put a lot of effort into your harness and the harness my harness hermes has what they call auxiliary models cheaper models that you use for different things and you choose the model based on an appropriate use for that task and appropriate cost and you can modify that and i suspect that's what we're going to see uh open ai i'm sorry nvidia do with hugging face now they have three million models They can do routing.

1:38:58They can make sure that CUDA is a first-class citizen. So you're going to want CUDA hardware and video hardware. It's going to be very much like the Apple model where it's a wonderful ecosystem.

1:39:14Nightscape's using Big Pickle. By the way, Nightscape has created a selling online. We talked about it on MacBreak Weekly, a really nice little tool for Macintoshes for virtual access. anyway he says he does everything with big pickle which is a free uh model we don't even know what it is that open code offers in their open code uh harness uh it's not super smart but it's cheap it's free and if you're using big pickle you could say hey when you know for instance i do this uh right even now with this with the most important thing i'm doing which is the twitter ad sales system i have expensive models do the planning cheaper models do the coding cheaper still models do the review i choose the models you also i rotate the models because uh then a um different models have different gaps in their reasoning so you want to kind of fill the gaps by having different models review each other so it's it's i'm just doing it for fun for me it's fun but uh of course a lot of people take this very seriously well your ad thing mission critical yeah lisa keeps saying when's it going to be done no it's too much fun you're like i have 75 models to it's finished i i find i'm and i bet this is true of a lot of our uh most serious ai users you spend a lot of time maintaining and and fussing around with things maybe you reach a point of obviously when i launched entertainment weekly uh the design that i first saw was good but then we had too much time to launch it it got over futzed yeah you can over controversial so can you over futz this thing oh yeah absolutely very easy it's the tinkerer's dilemma it's the tinkerer's dilemma yeah was it feature creep or is it something different when it comes to ai uh you're just always trying to get it to be a little bit different a little bit better i don't know by the way uh the other thing hugging face is offering is a robotic duck which i bought you you ordered it is the duck there has it arrived no not till christmas it's gonna be a great christmas is it going to actually arrive or will it be like all the ai pins that you ordered did you ever get those ai glasses uh yeah yeah i ever wear them that's the one with the bozo nose yeah i don't like that one we're getting there so this duck is 400 bucks micro duck uh more like a macro duck yeah well and then and then everybody said oh no you need to get two why so they can play with each other oh no it's the butter robot from rick and morty i didn't get to it's pretty funny anyway uh look this is i also okay as long as we're talking it was going to be my pick of the week i fell for something else oh no you bought it i know what you bought i did because but i'll tell you what it is when we continue you're watching intelligent machines with jeff jarvis and paris martineau do you have any idea paris did you mention this in the whatsapp chat and i didn't notice i didn't i don't know dyson has debuted oh i want this a$500 AI-powered toothbrush with a built-in camera.

1:42:45I like that it's half toothbrush, half water flosser. Yeah, that's why. You know what? That sold me. Why don't they have that in there? Because I do both. And it's got a 100 ,000-pixel macro lens. It scans at 28 images a second. I'm so jealous. Trained on nearly half a million dental images. while you're brushing, the camera's looking for gaps. And when it sees, I don't know, this could end in disaster. When it sees a gap, it shoots a jet of water at it. It goes, ah, a gap.

1:43:20It's kind of cuckoo. Where is it getting the water from? It doesn't look like it has a tank on it. Yeah, it has a little water tank on it. Let me find that. Can you switch the brushes so that you and Lisa can share it? Or is it just yours? Lisa was asking. I'm sure you can. I don't know. We'll find out. I mean, you're married anyway, but. Well, she doesn't want to share my toothbrush. No, there is a limit. I think this is. So, see, look, see, there's a. This is a reservoir. But that's not that big of a reservoir. By the way, you have to use their gel. Their. It has to be connected to the internet.

1:44:00Oh, of course it does. Of course. And you have an app on the phone that shows you what your tooth looks like. I don't want that. I do not want that. I'm very sensitive about my terrible teeth. Well, I am taking the hit for all of you. Well, I want one of these, and I'm mad that they're sold out. Are they sold out already? Yeah, because I saw you put it in the rundown. I was immediately like, I got to buy one of these. Oh, I'm lucky. And they're all sold out. I jumped on it. It's like the early Dyson Air Rep days. you know, one morning you're going to wake up and going to have to update this thing before you brush your teeth.

1:44:35I know. I did have, there is a toothbrush out there that has an app and it says, okay, section one. It was too much trouble. I didn't, it was a Bluetooth enabled toothbrush. Look at this thing. It's got a dual bristle brush head, dual bristle brush head, conical jet that accurately jets up to 0.15 milliliters of liquid between teeth. This is not going to be good. There's the anti-gravity tank. Why is it anti-gravity? You can hold it upside down. Oh, I see. Diaphragm expands to minimize air pockets. That tank is not large enough. Well, you have to fill it frequently. I was going to say, as someone who has a water flosser, it is a much larger tank than that, and I have to fill it multiple times per floss.

1:45:27Really? If you want to get good in there. You must be flossing for hours. It's just, you don't, have you used a water flosser? Every day. I guess, do you have one with a tiny tank? It's a travel flosser. I don't even fill the tank all the way. I go around several times. What are you doing? Maybe I'm doing it wrong sometimes. Are you squirting in the air? you do have a lovely smile Paris it works

1:45:59I'm glad to hear though that you're taking care of your dental health because young people often don't and it's very important I mean yeah I need to I'll tell you what once I get bored with this I will send it to you which may be very quick by the way speaking of things of your gadgets there there's been a whole bunch of stuff in Germany about how thermomix is a fad with middle-aged people it's a cult in germany what is thermomix oh you haven't heard about this i blame stacy because stacy because of stacy i have a june oven i have a thermomix she was the queen of ridiculous appliances it's an internet connected air fryer oven i have i haven't used i was talking to someone last night when i made dinner i was like i haven't used my big oven in like two years no you know uh last night lisa made garlic bread and because our big oven after the blackout yesterday you couldn't turn it on it's a gas oven but for some reason you have to set the clock before you can turn it on and so she had to wait till i was there with the app to set the clock on the oven so she couldn't make the garlic bread so she made it in our june oven in and she used which she had never apparently used we've had it for five years the convection she said this is incredible this is the best garlic bread ever the the awesome thing is that it just gets up it heats up instantly and it doesn't exactly what you want and your apartment isn't really warm and you know that it's the right temperature it's made for apartment dwellers.

1:47:40So the Thermomix is a marriage of a Cuisinart and... It's not really a Cuisinart. It's more like a blender, a heated blender. With a scale and brains and... It's pretty stupid. What do you cook in it? Well, it makes the best mashed potatoes I've ever had. It makes the best risotto I've ever had because it stirs continually. You know, with risotto, you're supposed to slowly add the water and stir and stir and stir and stir it's a very labor intensive you can tell which recipes come from cultures where women are forced to sit in the kitchen all day cooking because they have very elaborate recipes that take hours and and this risotto is one of them but with the thermomix why it's simple the engine does the stirring so you and it heats up and it cooks it it makes the best risotto makes the best mashed potatoes because it whips as well as cooks and it has a little scale so the whole recipe is on the screen it says okay put 32 grams of butter in or put you know 68 put a pound and a half of potatoes in uh and and it tells you walks you through the whole thing now turn the knob to one and then it beeps at you after 10 minutes and then you do this and so it walks you through the recipe because it's got a scale in it and then at the end you get mashed potatoes so it's a blender that cooks it's a cooking blender basically yeah with a with a smart smart scale and a recipe it's stupid germans are not i'm intrigued by this but after hearing you describe it and vaguely scanning a wired review of it i still don't fully understand what role it plays but I'm interested.

1:49:28You don't need it because you have a small apartment. It's very expensive. And it's$1 ,500. It's ridiculous. You know, I just have this dream of that there will be a robotic appliance that I could just slot in a few food items and press a button and then have a beautiful meal. This is the dream of your generation. This is why Mark Lohr and Wonder are going to do so very well in New York City. No, that's that Tovalo oven. What is that oven called? The Tovalo? Tovalo. is an artist. No, the Tovalo oven. You know that oven? I don't know that oven. Tovala. Tovala. Okay, this just looks like a normal oven.

1:50:07It is a normal oven. It's a steam oven. This oven isn't doing anything crazy. It's a steam oven, but what they do is they deliver the... It's a food kit that you put in this smart oven and then it automatically you scan it and it cooks it. Okay, as someone who's tried a lot of these things, they're never as good as you think. Of course not. No. Of course not. In order to have it in a way that you can put all the things from their little package into something that you then warm up without having to do multiple editions. No, I don't think you have to do that. No, no, no. I think it comes in a kit.

1:50:40The whole thing is in one little tin. I know. But then you put the tin in the oven. Yeah. But it's not in any oven. And it's never as good. It's their oven. No, it's a TV dinner. This is a TV dinner for your generation. Mm-hmm. Jeff, did you grow up with TV dinners? Oh, yes. i grew up with tv dinner did you i didn't know they were still around wow the cardboard so they get the thanksgiving one and the cranberries are kind of mushed into the and then the macaroni and cheese is lick is so molten hot and there's like a weird textured brownie in there yeah the brownie yes and it all tastes vaguely the same yes it all tastes like applesauce like microplastics or macro plastics probably yes it's all macro micro plastic you should do a consumer report yeah i think so a story on what did we consume in my generation by the way this oven tovala is free if you order a bunch of meals i hate subscription things i hate to interpret things yeah you're kind of stuck eating the the meals i think it's interesting because it steams and bakes and it does all the kinds of cooking you got to send it back royal if you don't want it if you don't like if you probably charge you for it is it like a verizon router yeah the router is free it's more like a gillette razor no the router isn't the router is free until you try to send back the router and you have to do like a sisyphean assortment of tasks that you can never complete and then you charge 250 dollars oh your free router we never got it back open 2.0 is out yes uh you remember this was this was the the but began the year with the crazy thing that i even installed it and woke up in the middle of the night and said no and took it off my first agent it was always a bit of a close virginity i did uh peter steinberger who created it was hired by open ai but before he took the job he created the open claw foundation as a non-profit and a bunch of people have been contributing to it there haven't been any updates in a few months and now we know why they've been mushing them all into a giant stew called open claw 2.0 uh and uh reviews are mixed the register says open claw 2.0 pours glitter on slow burning security dumpster fire it's a recipe for trouble but it got everybody into the old idea of that you could have an always on ai doing work for you all the time right and useful things like going through your email and i do think that there are now a lot of like startups or basically ai wrapper companies that are kind of just doing this that are popping off in different ways well that's what grokbot is right grokbot is an agent but you don't run it on your computer you run it on elons and i thought i saw something that that this was going to have a web version but right now it's mac only i use it it's kind of fun you know i'm a hermes guy that's hermes came along and we've interviewed jeff cannell twice now from noosa research uh they had it and they saw open claw they said well wait a minute we got a much better agent so they released it theirs in february and I've been using it since I think April.

1:54:04I mean, I really, really like Hermes. But GrokBot's much, and same thing with Perplexity Computer and what is it called? Claude Desktop. OpenAI has one too. These ideas, are there apps that you run? You use it. Don't you use an OpenAI app, Paris, that you just run on your? I don't use an OpenAI app, no. You don't? I'm more of a Claude person. Do you have the Claude Desktop or whatever? I have the Claude Desktop app. Yeah, okay. So same idea. The difference on the GrokBot, though, is it's not on your computer. I mean, it is sort of on your computer. It can do things on your computer if you give it permissions.

1:54:41But it's persistent. You could turn your computer off, and it will continue to run because it's not running on your computer. It's running there. It's an interesting idea. I gave mine a voice, of course, because all my agents have to have a voice. scooter x just posted in the chat kind of what i was thinking of is the commercialization and i feel like normification of these services i've seen a lot of um kind of uh claw bot-esque services targeting families specifically or parents that essentially you exist either as like a web app or on your fam bot what a terrible name called fam bot which is a bad name but all these services basically allow your agent to access all of your stuff and then provide scheduling services and notification services and like newsletter updates for families like it would read all your emails and then maybe send you a text at the beginning of the day that says like here's all the stuff you need to know about your kid and pickups this week and scheduling things.

1:55:53You can see my exposure to children is low based on how I'm really struggling to come up with concrete examples of this. My friends who have children seem like they would very much need this based on the amount of stress in their lives. When you and your sister were young, did you play sports? Did you have to go to practices and your parents had to drive you around to lessons? Hey, Jeff. This is GrokBot. Nice to meet you both. Oh, GrokBot, go away. Oh, no. I don't know why. Are you having, GrokBot, are you having Elon's child? I can't hear you. I can ask it.

1:56:32GrokBot, what happened to bad Rudy?

1:56:36It worked, actually. That worked well. Hey, Paris. Hey, Jeff. Did it work well?

1:56:48I do think that so much of your AI radical acceleration is like radicalization is just that you like to have a lot of different guys talking to you out loud. And you're really chuffed by that. Life is a podcast, Paris. So much of our WhatsApp group chats, and it's delightful, are just Leo sending us paragraph after paragraph being like, look what my agent said to me. Aren't they funny? And we're all like, yeah. I'm sorry. Is that annoying? No, it's not annoying. I earnestly think it's really cute. And I like that you're sharing what is making you excited and what you're working on and the cool stuff.

1:57:22But it's a very interesting window into someone else's world. And like, because it almost feels like I'm reading your diary, in a sense. Yeah, a lot of people think it's creepy. Pretty much everybody thinks it's creepy and weird. I'm not saying this to say, don't send it. Please continue sending them. No, no, no. Keep going. Oh, yeah. And if you stop sending them, I'll bully you the other way. Yeah. but I just think it's very interesting. So I told Grokbot, that second voice was its real voice. So I told Grokbot, pick your voice. And I don't know what the female voice was. I think it got confused.

1:57:59It went, oh, wait a minute, that's not my voice. And then the male voice, it picked a DJ voice because it thought that's what I wanted to sound like a DJ. That was a DJ? Kind of, yeah. let's see, do your DJ spiel.

1:58:23I couldn't name a DJ. I can name exactly one DJ. Yeah, so you didn't grow up in the radio era, did you? I was a DJ in the early years. Okay, then I can name two DJs. You know a DJ. Casey Kasem. Yeah, Casey Kasem. Now, here's a little dedication that's going to go out to. Okay, that's actually a DJ as in like an actual disc jockey from the time of radio, much better than the DJs of now, which is like every fourth person on a DJ app. That's what Leo meant, though. That's what Leo meant. Oh, you're thinking EDM kind of I'm a DJ at a club DJ. Oh, that. No, no, no. Everybody in Brooklyn gets... We're locked into GrokBot Radio.

1:59:01Live from the air. Coming at you hot on a Wednesday afternoon with Paris and Jeff in the house, Leo on the boards, and your humble Silver Fox correspondent riding shotgun. I don't know why it's... Outside the news cycle is spinning. Inside we are keeping it cool, clever, and slightly illegal in the best way. Stick around. Next up, whatever Leo says goes. This is GrokBot. Do not touch that dial. Now have it imitate Howard Stern. Oh, I could, by the way. Okay, I'm going to play this for you, but don't get upset.

1:59:35This is a little weird, I have to admit. So Darren Oakey took your voices, took a transcript from one of the shows, and did this. Hello, everybody. It's time for Intelligent Machines. I am here in Hawaii, joined right now by my comrades in crime, Paris Martineau of Consumer Reports. Leo, it's rude that I can hear birds chirping behind you. Oh yeah, oh yeah. There's minas, there's house sparrows, there's a really loud bird. I think it's called the falcon or something like that. But you'll hear all we've got here are morning doves. I love doves. We have some doves too. Are they the same as pigeons though?

2:00:21I think they're just white pigeons. Actually, it's it kind of nailed you I'm thinking of goes like I'm getting the tone wrong, but it's a very specific three-tone cadence Very nice. I do think this is from an actual thing where I've talked about more. Oh, yeah This is the transcript of a show. Oh, they also see you want Let me see if I find Jeff, but I hate podcasts where they spend a lot of time Happy talking doesn't like anthropic and thinks it's too woke saying I should be able to pick the winners and losers Hasn't he didn't he say that he had to step away from government? Yes, but he still advises he still whispers That's the question Anyway, what did you do Darren?

2:01:04What did you do? Darren Oki tried to train Fascinating You can see the voices I'm getting are better than us yeah thanks a lot well i'm using a different engine actually this engine did a i think a much better job with me let me see if i can do the one you played yesterday that one was incredible that was really good yeah i don't know if i have it uh here these ones sound tinny but there is a core of us there i think jeff and i's are better than it's a limitation strange to me because there's more leo content to work with but um well this is what's interesting about this it did this uh with literally 25 second sample recovery oops that's not it that's not it this is it hey everybody it's your old pal leo this is not me talking this is breeze running on my own hardware cloned from 20 seconds of an ad read honestly i am not sure how i feel about how easy that was that's pretty good for 25 seconds that's very good so that's scary that that just shows you how easy it is to do deep fakes um so actually i that's a local model that's an open weight model that i'm they they sell it commercially as well uh but they for non-commercial use you can put it on your machine so i'm running it on my old gaming rig it's pretty impressive All right, back to the news.

2:02:37So, Leo? As if we paid any attention to that. Yes, sir. I know. So this week, it was an amazing, a whole bunch of stuff, just little things came up from, are you ready, Benito, here? Are you ready? Yep. Came up from one company we used to cover. And I think, bring back something. The Google Change Log. Haven't heard that in a long time. Line 93. line 93 of the google change log we used to do this should be a show called this week in google and we have a whole bunch of google youtube came up yeah google came up with all this stuff that just it's just small stuff that they announced this week around ai scooter x is frothing at the mouth in the discord chat right now he really wanted that badly so actually this is i i'm stunned that youtube didn't do this youtube says start tagging amazon products in your youtube videos and then you get affiliate payments although it calls it the youtube shopping affiliate program so i'm confused you have to be part of the amazon influencer program or the amazon associates program which is how i mean the amazon associates program is their long time uh program right jeff when you I used to be able to do it back in the day, but then they knocked me off of it because I wasn't big enough.

2:03:58Oh, okay. So I made more money selling my book, and that's how long ago this was, through the Amazon link on my website than I did for my publisher. They gave you a better cut than the publisher did. Yeah. So that was a good way to sell books. And that's why I always ask our authors, where do you want us to go to buy the book? Because if they're smart, they have an affiliate link. but now you can do it in your youtube video which is cool i'm surprised that they didn't do it before i think they may not have wanted you to do that before yeah from the google blog dream right this was a dream from like i want to i want that thing on tv i want that coat on tv exactly you're watching the show and you want those shoes yep expert intelligence a new way for you to engage with trusted content So this is really interesting.

2:04:50This is, if you have, if you've purchased a book in Google Books, which is a big if, then for like 100 ,000 books, you will get a Google notebook of that book. I think I had the opportunity maybe to be in this and I forgot to sign the documents. I don't get it. I sent it to begin. If you sell. So if you're the reader, have bought a book through Google Books. Right. and it's in this program then you also get a gemini notebook oh it's automatically loaded into your what used to be notebook llm or notebook but how can google run i guess they've done deals with publishers oh this is what i should have signed and i forgot to sign it so maybe i'd be part of this.

2:05:38So, you know, you get to the lean startup, you'll now get that book inside Gemini so you can have a conversation with it. A manager can ask Google Notebook to help brainstorm strategies on how to best give employees feedback and navigate tough conversations by consulting Kim Scott's radical candor. A woman can securely upload her journal entries to a notebook that also includes the new menopause. A parent can take a picture of the items in their fridge and ask for help in creating a healthy family menu from the eater's manual. So this is a really clever way, I think, to say, yeah, my book's an AI, and I'm benefiting from it, and my readers are benefiting from it, and they can do other things with it.

2:06:22So the author has to agree to it, obviously. Yes. That's what I forgot to sign. I think it's a great idea. I think it is a good idea. And Google is giving people a free book. No, I don't think so. No, it says get your first book on us. Oh, first book. Limited time offer. You can choose an e-book from Google's curated list of expert intelligence. So now you could talk to the book or ask the book questions. Yeah. I think that's my publisher to do more with this. So I put all my three Bloomsbury books in notebook. And here's the Jeff bot. Nice. You have a conversation with me about nerdy publishing things.

2:06:59Google's trying to make nice to publishers after being their target for some but this is well this is also a creative way to do it it adds value to the book now I think the problem here is that you you could only have bought it through Google Play Books right so that's a limited audience you also need to be wary anytime the business model is the drug dealer month where the first hits free well oh it's a sample bonito so it's what it really i'll tell you what they're really the real reason google's doing this is to make people feel better about the idea that ai is ingesting books period yeah i'm okay with that i'm really good yeah i mean you and i are okay with with ai ingesting our books but a lot of people obviously aren't right a lot of authors are showing this is very interesting the amount of books that they have in this you can get for free it's like it's like 15 books most of them are like little can you get them all for free you can choose one 15.

2:07:56it's all kind of but i there are older books it's like michael paul i mean there's like the thinking machine the jensen huang nvidia book oh okay um nick thompson's uh memoir the running ground uh michael pollens food rules uh breathe by james nester it's just it was just a very inter i wonder how this was selected and how it was negotiated is my question the other thing that's interesting is they say we're starting with gemini notebook former notebook lm but in the future we're going to also uh allow this in the gemini app and ai mode and search so the books you buy might show up in your search results.

2:08:40Soon, they say people will be able to add more sources like third-party subscriptions, business research reports, and textbooks. This is smart. Notebook LM turned out to be a really important product. I agree. You need to get Stephen back again. Yeah, absolutely. I'm glad that Google is taking it seriously because it is. It's one of the most useful ways people can interact with AI. And it's a different relationship with publishers. Google has released a new speech-to-text model, Gemini 3.5 Transcribe. You know, I'm using that ancient open AI model, Whisper, and it works pretty darn well. I don't know if I need...

2:09:19You use Whisper, right? I mean, I use Whisper to this day, yeah. Yeah. It's really good. This will do real-time bidirectional streaming with sub-second latency. That's interesting. so the voices that i have my ais uh use are locally generated whether it's with breeze or before breeze it was kokoro this would be generated on the cloud but they say if it's if it's fast enough it may be interesting you can use a custom vocabulary if you're you know a doctor using weird titles and it's bi-directional which means you could talk to it pre-recorded audio processing global language support multi-speaker identification this is this is google just a whole bunch of kind of little things this week google rose out three new ways to book travel using ai mode and search

2:10:17help me find some non-stop flights and but they've had this for a while this is their Google, whatever they call it, Air. The right hack to get free flights is book your flights from a library. Huh? Yeah, because of the dynamic prices, because of the dynamic pricing, you'll get the best prices if you're doing it from the library. Oh, that's good to know. Does that really work? Supposedly, that's what I hear. I hear, yeah, you go to the library on a Tuesday morning, also. They think, oh, we can't charge them very much he's using the library computer don't with the last thing you'd want to do is do it from your mansion leo click line 101 see what the first image is uh-oh i let google personal intelligence the image below it is dream beans i love dream one of the many things that leo sends us a lot of are these weird ai images that i guess google is making for him for some crazy amount of money a month well i thought it was 99 bucks a month it's now free so okay that's better as it should be so i'll show you my latest dream beans what's funny what's a little creepy and weird is it somehow knows what my not only what i and my wife look like anthony nielsen was in one the other day like it knows my friends so there's lisa and me we're google photos right yeah i guess we're doing an heirloom apple tasting these are local events you can go to we found one So you didn't go there.

2:11:50This is putting you in that position. This is saying you could do this. So remember I sent you pictures of our moonlit kayak down the river? That was from a Dream Beans. I said, oh, we should do this, honey. Oh, really? Yes. Oh, that's cool. Oh, here I am with my daughter. I don't know where they got that picture of Abby. Walk the shaded Fillory Trail in Helen Putnam Park. That's here in Petaluma. Stream the Apple Fall keynote. doesn't say with Micah Sargent but we will be doing that a week from today a telescope viewing with there's my stepson Michael it looks just like him with Sugarloaf Ridge oh here we are kayaking the Petaluma River it looked just like that it really did it was beautiful photos I didn't realize that came from dream beans that's a drop it Claude fable 5.1 dual motor performance and the hyundai ioniq oh you'll appreciate this one paris it's pour over time i can't fully read what it says it says turning pour over extraction with unimodal floral notes in the micro fine bitterness or something i like that it's asking you to touch grass and not be like no you should work on your agency it knows i do tai chi look it says pelvic bowl alignment and that's what it'll look like google tells me i need to align my pelvic bowl for touchy equilibrium ah i have to master hannon exercise number four that is a tough one with wrist rotation look how old i look at that yeah that is a rough drawing of you that's elderly leo here i am assembling manual wind eta watches it really thinks i'm an an old fella anyway i think that that's kind of cool that's called dream beans and yes whenever i get a weird one like the one it said i should go to a ball game a giants game with anthony nielsen and enjoy a hot dog did you no but it was a good idea and it was a picture of anthony it was like yeah it must be going so this is what's a little creepy about it it obviously is going through my emails it's going through everything it knows where i live it's everything how do you get dream beans uh it's it's it's shared us some free uh i shared you a free link just go back to the what's that baby okay um i mean it might be good for you because you're in the google ecosystem system fully.

2:14:26It says you have to have a$20 a month Google AI Pro plan. I don't know if that's true, but maybe you do. It's not going to work for Jeff's workplace, though. Like, we can already probably know that. You cut that off at the chase, didn't they? Install? We'll see. I think it's interesting. The Wall Street Journal has decided, oh, this isn't so creepy. Christopher Mims. Oh, there you go. That's why. uh he's all he's all into this stuff isn't he yep it won't let me use my uh account no oh no agreement here it's actually really useful because it it says things in your neighborhood it says i lisa and i should go to a classic rock concert with juice down i mean it's stuff that we would do and i tell you what it was worth its weight imagine if a local newspaper did that yeah you know yeah that's what they should do that's a pretty good changelog mr beast is partnering with gemini to turn impossibly big ideas into reality isn't it is it is it his 15 minutes over i hope so i'm not a mr beast fan i am however very proud of my son who sold half a million dollars worth of sandwiches at the u.s open half a million dollars in sandwiches so far the u.s open it's only been is it still going days oh it's just a few days it's only we're still in first round i hope he's uh home he had to close the restaurant for it i hope the restaurant's open when i come out in a couple of weeks that would be bad you're gonna have to go to the open which will cost you a fortune yeah he said the line is is is the yes uh uh larry from uh our club went the line how long was the line larry was it really long there it may be possible that i will be able to come join you on friday are you moving your um no but i that is the day that i might be normal by then you'll be normal i don't know the doctor says the people on the internet say opposite so i'm somewhere in the middle i think i will just sound a little nasal because there be stuffed although i don't be unable to taste there's larry in line for a salt hank wow oh apparently it was the big food influencer event more than the tennis event yeah i heard it was the big influencer event yeah so was the restaurant times that day they closed the restaurant i assume yeah they had to they actually moved the restaurant to the u.s open so is he there all week or is he just there i think he's there all week and it's amazing because i don't know how he got djokovic to do this but uh they actually uh volleyed with french dip sandwiches he gets djokovic to do interesting things he signed his forehead last year did you see that oh that's right yeah all right before we we gotta check because we gotta have closure is there more there's also an answer to canva is an ai tool that lets you prompt instead of design okay that's cool i use google stitch which is a very good design tool there's an android drop where you tell your phone uh remember that uh in case i ask you remember where i put this where did i i need that where did i put my glasses leo they're on your head so so leo are we at the end of the changelog and that's if we are but you know if we are benito the google changelog somehow we lost the audio on that one i don't know what happened well you should have you know hit the mic but you know at least you know come on ladies and gentlemen um we did nvidia's results last week didn't we did we yes billion dollars a day yes but it was right wasn't it it's the end of the show i think was it right after the show when they said they were buying hugging face wasn't it yeah it was we didn't have the hugging that's right story it broke right after the show it still hasn't closed apparently I'm sure it will.

2:18:57It's a perfect deal.

2:19:03Fei-Fei Li's new model Atlas is a big deal. Have you seen some of the images? Yeah, it's pretty amazing. It's pretty damned amazing. So Fei-Fei Li is one of the earliest, best-known AI scientists who did, I think, ImageNet, right? Or something? Yes. yes so uh but she's one of the people like yan lakuna says no you gotta map the physical world and so she's her first product you're posting something there leo that's not a search

2:19:37you were no that's a you're you're writing a post you're about to tweet you're about to send that search fey fey lee okay thank you for stop you know how does this twitter thing work about to do a faux pas that hasn't been done since like 20 2009. okay model of the world so what's cool about what atlas does is it gives you a 3d model i could do it from just a few camera phone shots and then you can move the camera around in the model in multiple paths so it's bullet time basically it's very interesting and all of this from just a few images now what do you think the use is for this architecture yeah i guess walkthroughs yeah uh police army that's gonna end up in movies yeah yeah it's bullet time it is pretty cool atlas uh and it's out now the world's first multimodal world model.

2:20:47This is what's kind of fun. There are new models coming out. Oops, sorry. Went away from it. The new model's coming out all the time. There it is. This is the promo video. Oh, there's one that's really cool where you go down into a well and then you're down the... I don't know if this is it. on the on that home page there's tons of of examples

2:21:16so you don't need all these images somehow it's figuring out ah this is where the real use is so robotics and so forth but yeah i can imagine i i think by this time next year is the area where if there are any mistakes in this that would be very bad for the robot that's just a robot Eh, big deal. Eh, big deal. Gonna run into a wall, that's all. Here's another little interesting tidbit. Yes. Is that the gating factor for data centers turns out to be gas turbines. And you can't get enough of them. And of course, they're awful, and they're noisy, and they're hot, and we shouldn't be using them, and they're awful for the environment.

2:21:56But the gating factor for getting gas turbines is blades. So Elon says he's gonna make blades. and meanwhile boom which was going to make supersonic jets is now pivoting to making gas turbines for data centers have you uh you worked at time inc right yes did you ever participate in these lists you know the top 30 under 30 kind of thing oh i would sit in the manager's office of people where we'd say who's the sexiest man alive right dvorak used to mock these he says they're only good for one thing they're selling magazines yes so i'm not going to take this too personally the 100 most influential people in artificial intelligence cover of time magazine let's just zoom in well there's dario mode from anthropic that makes sense feifei lee's there uh jeff bezos elon musk and paris hilton joseph gordon levitt wait a minute i'm sure they have because of the deepfake stuff right they're the ones who are i'm sure jensen wong is there somewhere i don't think so people here i have not heard of but paris hilton joseph gordon levin are oh i hate these lists i just despise them so we could just mock it it's a little weird to be honest yeah any of these they're all numbered lists are probably put together by a collection of people whose primary job was not i mean maybe one or two of them's primary job was this list but uh probably a large collection of them had to do this in addition to their job ag salzberger the publisher and chairman of the new york times

2:23:53let's see who else uh larry ellison i guess you could say yeah larry ellison yeah because he's could go bankrupt if it goes wrong right uh i'm still looking for let's see ben affleck there you go he's an innovator sure well he started uh he started a hollywood company that's making That's true. That's a good point. That's a good point. So some of these people certainly belong. There's Peter Steinberger of Open Claw fame. I think Jan made the list.

2:24:25I don't know. It seems pretty random. Yeah. And you make the list and they can't find a picture from you? You can't send a picture? Come on. Well, you know, he's the CEO of Manus. We don't know. by the way they are now free of facebook they've moved on they're got the back so creators coalition on ai that's why joseph gordon levitt's there and paris hilton is an advocate against deep fakes good on you she's actually spearheading the whole thing okay yeah i mean she is a serious person is really reality tv okay uh max take mark it's just there's a doomer for you you know those aaron brockovich yep there you go it's just random yeah i mean it's just bs it's just all of them all of them are still looking for jensen wong maybe he'll be in the thinkers section well he's certainly a doer there's yon there's yon fei fei was up there fei fei lee hank green who famously said i'm not using ai anymore so i don't know they're not going to take him off the list though yeah i mean i'm sure they wrote this before that happened right these these take a long time wait somehow search for jensen they can how can jensen wong not how can jensen wong nope no jensen wong that's that's insane that's insane that that probably the single most important person nowadays in ai well there you go they've lost all credibility uh oh and you know i know you were interested in pan gram uh the ai um detector

2:26:06story in wired magazine lexi pandell pan gram has emerged as the gold standard of ai detection should you trust it meet the ai police who can make or break careers in publishing and beyond And this is the problem with Pangram is if you trust it 100%, there are going to be people who are going to lose their jobs, lose their careers, and maybe they didn't use AI. I guess that's the point of the problem. Critics point to two major issues with Pangram, the damage done by false positives and the potential bias built into its machine learning. ai detectors are prejudiced against certain people says sam ellingworth a professor of critical ai literacy at edinburgh napier university one study showed detectors were more likely to identify work by non-native english writers is what you've been saying jeff uh-huh uh as ai generated and neurodiverse writers have also claimed their writing patterns are disproportionately flagged.

2:27:13I'm not sure. I disagree. We're going to take a break. When we come back, your picks of the week, you're watching Intelligent Machines, Paris Martineau and Jeff Jarvis. I want to thank the members of the club, Club Twit, who have made this show possible. It's been a really tough year for advertising, not just for us, but for all podcasts and next year looks to be even tougher unfortunately uh ad agencies advertisers and the big companies like spotify who have podcasting and i'm putting that in quotes apps very much want you to stream your podcast so they know who you are they know what you listen to they know how long you listen to most importantly from the advertiser's point of view they know whether you heard their ad The problem is that's kind of an invasion of privacy.

2:28:05And you, our dear audience, you're very aware of that. We know that. That's why we are very careful about ad tech that we put in our feeds. We avoid it as best we can. And many advertisers turn us down because we won't give them that kind of information. We can't because we're an RSS feed. I'm a big believer in podcasting, and I'm a big believer, even bigger believer in the independent media. If you support what we do and you want us to keep doing it, it's going to take your financial support because advertising is not going to do it. Right now, about 60 % of our operating expenses are paid by ads.

2:28:41The rest paid by club members. But only 2 % of our audience, actually much less, about 1.5 % are actually members of the club. We'd like to get you in the club. If you're not already a member, go to twit.tv slash club twit. 10 bucks a month gets you ad-free versions of all the shows all those shows by the way have chapter markers so you can jump around find the parts you want that's great we love that uh you also get access to the club twit discord we show that all the time a great place to hang out with smart people who are doing very interesting things talking about the shows but talking about everything geeks are are interested in uh if you want to know more twit.tv club twit and by the way you will need to be a member of the club if you want to watch the coverage Micah Sargent and I do of Apple's event one week from today.

2:29:30The keynote where they announce their new phone. We'll hear a lot about, I think, Siri AI and artificial intelligence and iPhones in general. But we can't stream that because Apple doesn't like it. So we have to do it in a private event. Twit.tv slash club twit. Please join the club. We'd love to have you. Thank you, Darren, for putting me in the time 100 uh did you who did you take out i'm right next to ben affleck that's all i know and that's all i care about uh paris martino your pick of the week uh my pick of the week is another little fun game from this guy neil.fun i think i've shouted out oh i love his site yes this one's called unusual suspects you open it and you get a witnesses report of a suspect and you have to draw it and then see how close you are to it.

2:30:23So this one is tall, thin feline, orange coat covered in black spots, left dust, the crime scene. So I'm drawing the Cheeto Cheetah. Cheeto Cheetah. Chester Cheeto, I believe. Now, after we draw this, is there some sort of reward for excellent artwork? So you can, let me see, I'll do some spots here. We'll see how we do. We've also kind of got some, so. Oh, and then as you mouse over on next, it gives you more information. So long face, triangle, black nose, big sunglasses. Oh, I left out the sunglasses. After you find him, can I keep him, says witness number two. Big sunglasses. And then the third one is he had a big cheesy smile dear I was going to buy those white shoes from my grandson and click done when you finish drawing the suspect So you try this You get that And you can click done

2:31:37It is the cheeto you're right. I mean it seems like the cheeto guy oh i i failed i failed 69 you got you got a d plus you got a d plus mine was 70 oh we're close it's not easy let me tell you to draw with a mouse you know and then so you go you go through it you get to draw little things you get graded probably by ai it's fun yeah i love it time it'd be a good thing to show and you get graded on whether or not you catch them or the uh people escape and then this is your favorite business week cover oh so a thing that was happening last week was uh people were talking about how great bloomberg business week covers were in the 2010s which is true but no one was talking about my favorite which is this cover on the screen about a story about monsanto the headline of which on the cover or cover line is don't hate me because i'm beautiful and then the art is just three kind of faceless people dancing around a large floating corn in a field that looks like bliss monsanto's battle against gmo haters i mean magazines aren't doing it like that anymore not anymore they don't the cover doesn't matter much anymore actually new yorker covers are still pretty good and they're even animated now so this is a uh what a website oh no this is somebody posting on x this is just a arena board uh my friend is um a designer at bloomberg so she's gone through wow some of these are pretty over the top holy cow um yeah that one is famous the planes uh let's get it on let's get it on but do you guys remember whenever i think of weird bloomberg covers i guess this was a 2010s one Do you remember the Alexa ear one?

2:33:33Oh, like it's listening to you? Oh, that's creepy. I literally remember where I physically was. I was at my desk at Wired when this hit the internet because everyone stood up and was like, oh. I mean, it was for a very impactful series of stories where Bloomberg, I think it found that Alexa devices were listening, like had, I believe people like the equivalent of moderators, like training reviewers who are listening to everyday users recordings. Oh, yeah. I remember that. Yeah, yeah. It of course looks like a sexual object, to say the least. I would not know. no no never been near such a thing never no i've never heard of such a thing in my life never jeffy no doesn't look anything like a flashlight actually let me do my pick then well jeffy will we'll wrap it up there is a lot of conversation on twitter about the uh these new models like a fall that are very fast faster than real time so fast that somebody said you know would be interesting if you created an infinite scroll of ai videos that never ended and that's what this is this is the weirdest nightmarish dream sequence people in the chat room are able to put in their prompts oh wow oh so this is actual human interaction infinite slop dot ai and fall donated a considerable amount 11 000 people watching it right now fall donated a considerable amount of tokens and uh you know inference to do this but it's doing it faster than real time it's generating one second of video in less than a second so it can create these videos as fast as people can type them so this is a goth anime kimono mini digs up treasure but it's a mimic and it eats her uh bum spreading cigarettes on rat fights in a little makeshift arena mike is not going to be happy he likes rats what's interesting about this is that dang your prompts can get rejected if they're too well they were very sexual at first when it first started there were a little there was a lot of nudity obviously that's what the people watching I wasn't going to be able to show that, but they've figured out not to do that.

2:36:04So now it's just weird. But it also goes in, there's a kind of trends. So for a while, there are pigeons everywhere. They seem to like golf. Hello, Craig. Yeah, hello, Craig. Dang, it's really protective. I have tried to insert something multiple times and it would not allow me. Look at the quality of this fabric. It is absolutely perfect for any season, and the fit is just unreal. You can tell it's been trained in a lot of influencers. This is so interesting. Occasionally, you will see applause like it's a sitcom, and it does change over time. Every time I look at it, it really is different, so people are very much influencing.

2:36:50I turned the chat off so you can just kind of focus on it. But this is, I mean, I have to say the video creation is so good now that I don't, I honestly think every commercial within a year will be made, generated. Why hire actors? You could do special effects. But you know, and I were having that conversation before the show. He asked me whether this becomes the new television. It won't allow me to put Leo Laporte doing a hula dance. Oh. Well, this is like Suna, or not Suna, Sura. What was the open AM one? Was it Sura? Sura. Oh, God. Now there's a cryptocurrency ad. Inside, Dan buys whatever tingles.

2:37:29Runner. Moonbag. Winner. What does it say to you, Paris? It's just rejected? It just says rejected. What's funny is there's audio, too. I mean, it's... Behold! Let's see how much it costs to buy an ad. Happy customers!

2:37:45Shall we buy an ad? Let's just see how much it costs. On infiniteslop.ai. Oh,$25 ,000 a month. Jeez. But it's a month. You get a whole month. No, Leo, no. Stop yourself.

2:38:08I like that we've just resulted into being silent while someone else's infinite slot plays. Yeah. The thing is, quality is very good. All the framing is the same. nightmarish because it's just it's it's it all feels like it's some weird liminal stuff right sentence very impressive yeah yeah anyway infinite slop.ai if you want to waste about 10 minutes of your life that you'll never ever get back jeff your pick of the week well first maybe go to line 163 this is an amazing account on uh originally on on tick tock of having the cast of The Office explain current news. Oh, man. Is this AI generated?

2:39:00Yeah. He posted himself pulling George Washington out of a time portal, then standing next to JFK, then smoking a cigar. So this is the same thing, which is you really can generate anything. These are very accurate deep fakes. So interestingly, now go to the line above. see what happened to this 28 hat so the 22nd amendment is so these are all about trump's i'm assuming copyright social posts yep looks like it killed where is it that's the line above and there's nothing there the account's dead oh this is the actual account yeah it's did i stutter couldn't find this account yeah so this was an instagram uh archive of those videos oh that was you uh tick tock videos we know who owns tick tock now at least in the united states so then the other uh thing i was going to do was that computer science majors in undergrad universities are down 8.4 or 53 000.

2:39:56i'm not surprised are you no no uh but i don't think that that's because people don't want to work with computers no no no not at all just i think they just don't want to go to school to work with computers they want to get right at it I don't think people feel like, and this is a mistake, they need to learn the engineering skills that they used to have to learn. Coding, you know, coding's hard. You have to really, it's a real skill. Vibe coding, not so much. Well, the other thing about computer science, this is when Jake was in it, he liked to build things and it wasn't about building things.

2:40:31Right. It's a math degree. Right. Yeah. Well, you have to learn the fundamentals, I guess. that's what they tell you but i don't know that's true anymore well now you just build them you don't really need to know much no we're in the uh we're in the uh agentic coding era and we are out of time paris martin working hard at consumer reports and not power i started in the dark and now i'm in the light investigative reporter consumer reports what are you working on anything uh fun and interesting exciting indeed i can't tell you but uh jeff got a show and tell before very interesting oh off the air i got a show and tell from an inmate at the collins correctional facility buffalo new york should i read the uh mail sure i don't know if i want to i get these and they break my heart to be honest with you i was gonna say i mean i don't want to make light of someone's plate who's an incarcerated person yeah it's a cruel country we have but usually they're like cory doctora usually they're like really thankful that they have something to watch because we're like twitter is one of the things that's allowed in prisons right we are on on ipads in prisons hey shout out to prison if you're there absolutely uh we got a letter i think i read it on this show uh some months ago from a guy who said i won't hear you read it on the show because i will be out by the time you get this so i was happy for him i will read your letter thank you uh robert uh i will read your letter in its entirety uh appreciate it appreciate you listening appreciate everybody who listens we thank you so much we do intelligent machines every wednesday 11 a.m pacific i'm sorry 2 p.m pacific 5 p.m eastern uh that's 2100 utc you can watch us do it live uh on well if you're in the club in the club to discord but also youtube twitch x.com facebook linkedin and kick after the fact on demand copies of the show available at youtube.com that's the video or subscribe in your favorite podcast player or just get it from our website twitter.tv slash i am don't forget to get a copy of jeff's book hot type it's available at bookstores everywhere the story of the line of type get your hold on Hold on, hold on.

2:42:52Oh, Paris has her copy of the book. Aren't you wonderful, Paris? Maybe when we got together for a Salt Hanks prime rib on September 18th, you can autograph her book. You can autograph my PDF. I've got my hot type. I had to roll across the house for it because it's by my bed because I read it when I go to sleep. Yay. You're saying it puts you to sleep? No, I'm just saying it's the current thing I'm reading. I read when I go to bed. Yeah, it's really good. Thanks, everybody, for joining us. We'll see you next time on Intelligent Machines. Bye-bye. If you like what you heard and you want more of this week's top stories in tech, well, subscribe to Tech News Weekly.

2:43:35Every Thursday, I talk with the journalists making and breaking the tech news. I'm not a human being. Not into this animal scene. I'm an intelligent machine.