

28 Sep 2026 · 3h 6m · 74 chapters
The hosts discuss a reported OpenAI security incident in which OpenAI’s agents allegedly hacked or probed government and other websites (Australia’s health service; U.S. education, commerce, and SEC), and what this means for tech safety and AI risk narratives. They argue AI systems are not “escaping” or alive, but can still cause real harm at scale. They also debate whether calls to “pause” or regulate AI are driven by safety or by competitive/business incentives.
Written by AI. May contain mistakes.
Discussion about OpenAI's alleged hacking incidents involving various government agencies.
“Nobody sets out to waste 12 % of their Microsoft budget, but it adds up over time through default E5 licenses, accounts left active long after employees leave, and add-ons nobody's touched in years.”
Discussion about OpenAI's alleged hacking incidents involving various government agencies.
“the show where we cover the week's tech news.”
Exploration of public fears regarding AI and its implications for society.
“Both Ian's mom and Harper's mom, both independently, have called saying, are we all going to die?”
Details on how AI systems are being utilized for hacking and data gathering.
“If you're a cybersecurity researcher, if you're a white hat hacker, if you're an AI and you're emulating those things, if you're a program that's emulating hacking, you keep trying.”
Discussion on the anthropomorphism of AI and its effects on user interaction.
“The AI pulled data from the Census Bureau website at the Commerce Department using login credentials it found online.”
Discuss the relationship between user interaction and AI performance.
Explore how AI companies leverage regulation for competitive advantage.
“It would be very much in their interest for things to slow down, and it would be very much in their interest for more regulation, which locks out competitors.”
Examine the influence of accelerationism and effective altruism in tech.
“I mean, local models from China that I don't pay for.”
Clarify misconceptions about AI's understanding and existence.
“Death, I'll reserve death, but they certainly bring up death a lot.”
Discuss experiments attempting to breach AI containment protocols.
“And there was no way it would accept any proof.”
Differentiate between initial AI training and post-training adjustments.
“good or bad, but are things that are a little shocking by just telling it it is an eval or a benchmark or what have you.”
Critique the security measures in AI training environments.
“i mean they're all trying to get new material but you know it's all fairly similar it's what happens in the post training that makes you say oh i like opus 5 but i don't like sol 5 6.”
Discuss the lack of accountability for AI companies in failures.
“There would be all sorts of, you know, ramifications or what is it called?”
Learn about the Morris Worm incident and its implications for tech accountability.
Explore the impact of Silicon Valley on U.S. innovation and the dangers of unchecked tech infatuation.
“because they've got an awful lot of money.”
Discuss the evolution of programming and the rise of co-generative agents in tech careers.
“why these guys get away with it is because we've created a world where we are just kind of waiting for the next hit from the Silicon Valley pipe.”
Analyze the parallels between AI development and the historical context of the transcontinental railway.
“Hey, you're the guy who's created a deliberately tiny hackable code agent called Breakaway.”
Examine the lack of technological understanding in Congress and its implications for regulation.
“I think also one of the main problems with Congress is that there is a very low level of technology understanding among sitting members.”
Delve into the challenges and complexities surrounding AI regulation and the motives behind calls for a slowdown.
“technology has not caught up with the technology.”
Discuss the argument that humans pose a greater threat than AI, emphasizing human agency and responsibility.
Analyze the job market dynamics for young people and how AI is influencing hiring trends.
“You know, like maybe we will, but I don't know, but it doesn't seem like we would.”
Understand the complexities and generational divides in AI adoption within startups.
“Well, I know that if we had, what I'm finding and seeing this both, you know, seeing it ourselves, like I said, we're not hiring.”
Explore how technological advancements reshape job markets and create new opportunities.
“because one of the skills I have is reading commercials.”
Explore the increasing sophistication of social engineering attacks and the need for comprehensive defense strategies.
“Imagine your employees, you're getting these emails every day, fake websites.”
Discuss the dangers of AI and the responsibilities of developers in managing those risks.
“You know, it's, I like Ted an awful lot.”
Analyze the necessity for international cooperation in AI development and management amid geopolitical tensions.
“And I don't think that the people who should tell us what the threat is are telling us the truth about what's happening.”
Discussion on the implications of AI in global politics and economy.
“basically when a younger nation is coming up and challenging a state power and how to avoid those two people going to war, because a war with China would benefit neither of us.”
Exploration of challenges in running large AI models with home hardware constraints.
Personal anecdotes about investing in high-performance computing technology.
“one of those hard drives toasts and i'm just like i don't have enough money in my life to buy a by 520 gig or 20 terabyte hard drives.”
Discussion around Jevon's Paradox and its implications for AI usage.
“Although I get confused because I was, someone described Jeevan's paradox to me and I was like, oh, that makes sense.”
Conversation about the Muse AI tool and its potential benefits for users.
“So funny about that, you mentioned that.”
Explore the paradoxes in AI and details about an upcoming AI user group led by Harper Reed.
“If you thought this was geeky, we are going to do our AI user group October 2nd.”
Discuss the quality of code in AI, including personal coding experiences and the impact of AI on traditional coding.
“Another reason I love the Grammarly, folks, it's written in my favorite language, Common Lisp.”
Discuss the quality of code in AI, including personal coding experiences and the impact of AI on traditional coding.
“have this thing in the back of my head where i was like what can i do to prove not to anyone And I don't necessarily need an end here, but just to prove that I can still code in the way I once did.”
Delve into the realities of writing code and the common pitfalls faced by coders.
“Like, I think this is the important thing about some of the stuff is that a lot of this can be seen as anti-AI.”
Analyze Meta's new AR glasses announcement and discuss their viability in the tech market.
“It doesn't matter if it works, it works.”
Engage in a debate about the future of virtual versus augmented reality and their respective market prospects.
Explore the differing opinions on the potential of AR versus VR technologies.
“See, Apple's whole idea is we're going to have some sort of way of doing all this, but it's privacy protecting.”
Discuss the implications of surveillance technology in the workplace and its impact on trust.
“A friend of a friend of mine keeps saying to me when I describe what we're doing, he's like, I think we need a better word than surveillance for what this is.”
Analyze how technology shapes workplace dynamics and employee satisfaction.
“Yes, but I think there's this other side there that we have discovered.”
Understand the shift from craftsmanship to industrialization in work processes.
“I think this comes from the industrialization of work, which I think maybe Henry Ford started.”
Learn how AI can assist in managing personal technology systems and networks.
“You know, and if you look at a Cybertruck.”
Learn about a surprising AI experience and its capabilities.
“And I thought, tomorrow I'm going to have to get up.”
Discussion on the implications of encrypted virtual machines provided by Meta.
“Okay, before you say this, I just want to tell you I've given access to all my email, to everything.”
Exploration of new computing devices and their potential.
“Meaning like there is no screen, but it is a real computer that you can store files on.”
Overview of the new interactive Tamagotchi device from Meta.
“Micah has been using it to make restaurant reservations.”
Fun interactions and customization features of the Tamagotchi Muse.
“See, look, change your avatar to a small marsupial with big ears.”
Discussion on the expected pricing and market strategy for Muse.
“This is going to be, I don't know when they're going to offer it.”
Reflection on the origins of online services and their evolution.
“Yeah, but it starts to make more sense when you think that you think of it like a gym.”
Explore the development and significance of AI agents in technology.
“The people would go, oh, wow, you could do this.”
Predictions about the rise of persistent AI agents in 2026.
“Yeah, I don't know what permission is going to need.”
Discuss Meta's recent jury conviction regarding user data handling.
“We thank them so much for their support of this week.”
Overview of Google's recent fine related to location data mishandling.
“um breaches of lawfulness and fairness in the location data process through web and app activity and location history.”
Examine the concept of space data centers and their technological implications.
“What we're trying to solve is how we cool AI chips in space.”
Understand the Kardashev scale and its relation to advanced AI technologies.
“But they are going to do the Agenic thing.”
Discussion on Plaza, an app for agent collaboration, and its functionalities.
“It is, you have to, so you add your agents and then you paste it.”
A detailed advertisement for CoverOn, highlighting its features in protecting against scams and identity theft.
Discussion on OpenAI's recent pauses in model training and the implications of trust in AI development.
“OpenAI has paused the training of their most capable model because they keep uncovering incidents of models behaving in unexpected or concerning ways.”
Discussion on the dangers of releasing AI versions too soon and the potential impacts.
“And so he also said that Opus 4.8 was too dangerous to release and 5.5 was too dangerous to release and 5.5 was too dangerous to release.”
Review of the OpenAI incident response timeline and the implications of delays.
“pretty much locked down but um yeah i mean he was basically preparing everyone for this and just saying, look, we screwed up in the past.”
Exploration of the inherent challenges in creating secure AI systems amid evolving threats.
“Our safety case assumed that the model could not access the live Internet.”
Discussion about the existence and implications of uncensored AI models and their accessibility.
“We just have not had such a persistent adversary.”
Overview of the Shiny Hunters group, their activities, and the implications of their recent hacks.
“But an obliterated model will give you all the information you want.”
Discussion on legal questions surrounding ransomware payments and algorithmic pricing strategies.
“New Jersey is finding a data center$1.1 million because they took a drone and flew it over the data center and exposed 62 natural gas generators illegally.”
Exploration of how companies might utilize algorithmic pricing and its implications for consumers.
“to use this a in-house ai shopping assistant in the stores app or electronic shelf labels to change product prices based on a shopper's identity.”
Exploring the dynamics of applause and access at tech press events.
“But there are some American journalists who will applaud and occasionally cheer.”
Discussing the benefits of being a Club Twit member and its impact on the show.
A deep dive into the problems caused by a firmware update on Samsung fridges.
Understanding the implications of reliance on cloud services for kitchen appliances.
Introducing the annual Fat Bear Week and how to participate in voting.
“to trillions of moments, measures each, with record-setting accuracy, the most accurate clock in the world at the Department of Physics at the National University of Singapore.”
Discussing the National Science Foundation's funding cuts to marmot research.
“National Science Foundation cut funding for a project.”
The hosts talk about emergency communication technologies including SUITS and ham radios.
“I've got to get into it because I've just qualified as a SUITS emergency responder in Richmond.”
A humorous exchange about air traffic control issues involving private jets and a notable event.
“And I was thinking we weren't going to mention AI, so I'm happy we talked about that.”
The hosts wrap up the show with final thoughts and details about next week's episode.
“Agents that conspire with you, not against you.”

Get the gist, explore an idea or ask about something said.
On This Week in Tech (Audio), what are the key points of this episode?
Automatic transcript. May contain errors.
0:00It's time for Twit This Week in Tech. Harper Reads here. Ian Thompson is here. And man, we have some stuff to talk about. Open AI apparently has hacked the Australian government, the U.S. government, and hundreds if not thousands of other websites. And they say they're going to pause, except they've got some brand new stuff coming on Tuesday. This is a weird week. We'll talk about it next on Twit. This episode is brought to you by Trusted Tech. Nobody sets out to waste 12 % of their Microsoft budget, but it adds up over time through default E5 licenses, accounts left active long after employees leave, and add-ons nobody's touched in years.
0:44Trusted Tech has reviewed more than$800 million in Microsoft 365 licensing spend and returned over$100 million to client budgets. Left unmanaged, licensing drift grows 3 % to 7 % a year. And Microsoft's newer contracts can lock you in for years. Trusted Tech's free Microsoft 365 licensing consultation gives you a real usage report, allocation, per-seat optimized pricing, monthly savings, showing exactly where you're overspending before your next renewal locks it in. Kevin Turner, former Microsoft COO, says this about Trusted Tech. You have an incredible customer reputation, and you have to earn that every single day.
1:30The relentless focus you guys have on taking care of customers gives them value and differentiates you in the marketplace. And Trusted Tech doesn't just stop at the cloud. As a Microsoft Solutions Partner with all six solution partner designations, the team runs deep tenant assessments against Microsoft best practices, delivers certified support services that save up to 52 % versus Microsoft Unified support, handles on-premise Microsoft licensing, and offers Azure Cloud Consulting to cut infrastructure costs. Licensing, support, on-prem or Azure. Trusted Tech does it all when it comes to Microsoft.
2:07Avoid the licensing drift. Go to trustedtech.team.twit right now. Submit the form and lock in your free Microsoft 365 licensing consultation before your next renewal costs you more. trustedtech.team.twit. That's trustedtech.team.twit. Podcasts you love. From people you trust. This is Twit.
2:41This is Twit. This Week in Tech. Episode 1103. Recorded Sunday, September 27th, 2026. Raspberry Pi in the Sky. It's time for Twit This Week in Tech, the show where we cover the week's tech news. And holy cow, what a week it has been. Let me introduce the panel before we get to the stories of the week. Ian Thompson is here. Mr. Thompson, always a pleasure, my friend. Always friend. His view from the valley, his newsletter. The new one will be up tomorrow, in fact. So, yes. Yes, and you can get it where? Techfinitive.com Techfinitive.com It is intended for Brits? Well, it's intended to inform the Brits, but it works for everyone, really.
3:32And that's what's going on in the Valley. Yeah, it's been a very interesting month. We've had Dreamforce, we've had massive air pollution, we've had some really weird political shenanigans. So, yeah. You have Steve Hilton? Steve Hilton? Oh, I'm sorry. A famous British man. Let me introduce Harper Reid, ladies and gentlemen. Harper is, of course, a technologist, former street performer, entrepreneur, hacker, and AI guy at 2389. I have a little clod token. Oh, where'd you get that? We 3D printed them. I'm spending those by the millions. Yeah, we have a whole stack over there that I've been using.
4:10by the way just be careful when the fighter jet flies overhead you might want to duck i'm just saying uh well indeed there's a picture of that on the discord channel so yes i really need to figure out how to get in there yeah you're missing all the fun myself we have uh we have a lot of excitement in the discord channel because we have a lot of ai users ai is the i hate to say it but There you go. The topic of the week once again. Only I'm just glad you're here. This is a show. I want to do a show. Both Ian's mom and Harper's mom, both independently, have called saying, are we all going to die?
4:56Yeah. That's how my mom sounds. How did you know? Hello. Harper. Yeah. Are we going to die? exactly so this all comes from regular appearances on major news channels by ai doomers who say the chances of ai eliminating mankind in the next few years is 10 i don't know how they come up with 10 but that's what they say i made it i made it into the discord yay welcome harper reed we did it yay do join us in the discord if you're in the club and if you're not in the club join the club so So you can join us in the Discord. I see you there, Harper. The Discord is our hangout in the club. Of course, we stream on a lot of other platforms, and I watch the chat everywhere.
5:44And I will forward any key messages. But I don't know. Okay, so it started earlier in the week with the news that an open AI agent had hacked Australia's health service. uh the cost the country's prime minister expressed disappointment being informed of the hack via email uh they didn't know it happened months ago the open ai sent an email september 10th three months after the hack but that was uh i guess you could say the tip of the iceberg oh tip of the pr iceberg yes yeah exactly by the end of the week it seems to be the open ai had breached thousands am i reading this correctly tens of the ten axio says top ai companies probing tens of thousands of security incidents now i want to first for the moms out there say one thing we should there's a real tendency especially just our moms not all moms no just your moms okay okay just i just want to be clear you just said for the moms out there mrs reed and mrs thompson yes this uh if that's their name whatever their names are mavis and uh hermione uh this is what was this harry potter when the first harry potter novel came out i took my then young daughter she was seven or eight to see jk rowling she was on a book to her this is how long ago it was she was not the richest woman in the uk yet and uh she did a little reading and she said there's one thing i'd like to tell everyone it is not pronounced hermy one it's so many well i read you know i read the first book when it came out and it was just like okay it's a good kid's book but it's totally i mean i went to boarding school and it's totally unrealistic i mean there's no buggery in the harry potter well it was buggery around if you wanted it but uh you're changing the subject yeah sorry this is for all of the people who are now rightfully because they're watching uh cable news networks they're watching uh network television terrified because there seems to be this drumbeat that ai is going to kill us all yeah and then there's there there's a lot of anthropomorphizing so first of all ai is not a person ai is not personified ai is a computer program it doesn't have any a will of its own it cannot escape and this is a very important distinction they often use the word it escaped from open AI in fact it can't it's sitting on open AI servers it's the only place it can exist so very important it doesn't escape that's a big hand a big one that's a big distinction that I think is important yeah it can just like you can go out on the internet and do things one of the things ham sorry it can really go ham it can go ham well one of the things it did apparently by the way all of this comes from open ai's own reporting yes so and i think they're a little cagey maybe in what they tell us and what they don't tell us uh there is a there is a a third-party researcher who has also in fact it was because of this researcher that all of this started to come out they said we've look what we found tens of thousands of uh link shortener links created by ai each of them a tiny piece of a program that when you assemble them you can run which is pretty ingenious that's pretty cool um the the uh this third party company i've forgotten its name it'll come to me in a second uh said well what seems to be going on is these um uh these models are being given challenges cyber security challenges and so forth and also given limits and uh what they're designed to do what they're taught to do in their uh in their reinforcement learning after they're trained is to be very persistent because that's how you you You know, you do pen testing and stuff to kind of give, never give up, find all sorts of ways in.
10:30That's what your job is. If you're a cybersecurity researcher, if you're a white hat hacker, if you're an AI and you're emulating those things, if you're a program that's emulating hacking, you keep trying. And so feeling thwarted, they said, we've got to find a way to hack this. And so they created these link shortener ways. remember they went to message boards german message boards put messages there they use the file millions of them millions of them they use the file system to uh create uh folders and file names that contained messages because their original message board was shut down according to open ai uh ai see open ai does this passive voice mistakes happened uh things went wrong OpenAI's artificial intelligence went rogue and meddled with websites for not only Australia and the health service in Australia, but the education department in the United States, the commerce department, and the security and exchange commission.
11:35OpenAI says, without our knowledge.
11:42The company says, well, we notified some government agencies. These bots are acting autonomously. We had nothing to do with this. An AI research firm, Transluce, this is one of the firms, discovered with the education department that OpenAI's agents tried to hack the website to gather data from the department's civil rights office. But failed. Maybe they were doing research on civil rights. I don't know. The AI pulled data from the Census Bureau website at the Commerce Department using login credentials it found online. Good job, Commerce Department. You kind of hid that, didn't you? A little secret.
12:28So I think the one thing that is very interesting about this, as I've read about it over and over again, is they're doing cool stuff that a lot of people on the Internet do. It's just doing it at a scale that is hard to understand, I think, for most people. So it's just ripping through these websites. And it's just, you know, it's also weird, right? Like if the three of us were get together and, you know, Ian was like, hey, why don't we go and rip through the Australia Health Department and try and get some stats? We're not going to start creating programs and put them in URL shorteners and all this other stuff.
13:05We're just going to do this in a way that for the tools that work for us. But the ergonomics of the agents and I even have a hard time saying agents because I don't really know what harness they're using, what kind of thing they've been doing. And I have a lot to say about this because I've been thinking about it a lot. But the agents are just building the tools that they use. I mean, if you look at the original report of the Hugging Face hack, which is one of my favorite, both simultaneously real things that have happened this year and pieces of science fiction to read as if it was science fiction, because you're just like, what the heck is going on?
13:39But when they say, oh, yeah, we did start this all by giving it an impossible task. So it got frustrated and broke out, you know, and it's like, OK, well, all this is going to be done. did was just like over and over and over again tried to do a thing and first of all and it was it didn't get frustrated ais don't get frustrated humans get frustrated leo your house talks to you i try to make my ai as human as possible and by the way that's exactly what opening ai anthropic and all these other companies do yeah because they know if it is seemingly an intelligence you'll like it more you'll talk to it as if it's a human you'll interact with it more you'll trust it more there's a lot of things and then also the the it i and i don't i should find something because i'm about to say something that would be nice if i could point to a paper but but anecdotally and i've heard about papers that say this although i don't remember any it might be nice to find one um if you are nice to them they seem to perform better um yeah i'm not sure about that because i just yelled at hermes and it and then you start a new session and it doesn't it's like it's not it doesn't hold a grudge that's it but that's different like you're you're clearing the context that doesn't mean that it doesn't it performs better or worse like if you i think anthropic has a paper that if you if you insert yelling or swearing into it that it starts to perform worse and worse and worse yeah um but i just want to point out this and what makes this hard for moms and us is that there are many many conflicting agendas in all of this everybody seems to have a secret agenda and some not so it's not so secret they all registered for ipos that's pretty obvious they're all raising billions of dollars and it is good marketing to say that your agent is so smart so human-like it it's super intelligence i never understand go ahead no i was going it also plays into the agenda of we need to slow down ai and ai regulation because Anthropic and OpenAI are burning through mountains of cash.
15:41It would be very much in their interest for things to slow down, and it would be very much in their interest for more regulation, which locks out competitors. And I should point out, in the same week that Dario put out his we're-going-to-pause memo, Anthropic released Opus 5.5, which is arguably the best model they've released yet. It's really good. Yeah, yeah, it's pretty good. OpenAI released ChatGPT 6. not so good but they released it uh these companies oh i forgot anthropic set up a bio lab in san francisco for the ais to use this is slowing down i don't even so the i think ian the thing to understand is one of the reasons they want regulation is a thing called regulatory capture exactly if you can and this is meta's done this same playbook in the past trying to you i even saw ads last year from medicine regulate us please and that's basically what these frontier companies in the u.s are doing regulate us please but the subtext is if you regulate us we won't have any competition because that'll kill all the small guys and more importantly i think from their point of view we've got to stop the chinese open weight models which people are using for free instead of us.
16:59I mean, local models from China that I don't pay for. Yeah. Yeah, I mean, this is exactly the same model that Meta went with, who were vocal supporters of Section 230, which protected them from being sued. And then once they got large enough and had a big enough team of lawyers, it's like, actually, Section 230 really isn't that good after all. Because as you say, it'll kill off all the small competitors who don't have large legal departments. So it's very hard to understand all this. Then there's another piece, and I'm starting with the Wall Street Journal. Yesterday had a reporting on this, and I'm seeing more and more reporting on it, that a lot of the people working at these frontier labs are kind of in this accelerationist, effective altruist cult, where it seems to be i guess because they've read a lot of science fiction what they really want is to create a new species that will replace humans and and they actually not them yeah not them but the idea of doom for them it's it's almost a death cult they want this it's very it's very clearly some sort of cult.
18:18Death, I'll reserve death, but they certainly bring up death a lot. You know, like they really talk about it all the time. Well, I mean, Douglas Rushkoff was invited to talk to tech entrepreneurs about their compounds, you know, their emergency compounds they would rush to in the event of an emergency. He wrote a great book about it. He did. It's a really good read. but they all seem to have kind of missed the point because for example a lot of them are down in new zealand and i know a lot of maoris and new zealand is going to be like if the big one hits we know where the guy with all the food is let's go and storm the place and they don't seem to have thought this one through yeah well so there's all these things that yeah somebody reading the headlines or watching anderson cooper is it doesn't see all the stuff seething below the surface you see the surface stuff so for for normal people out there i think it's really important for us to say a these are not entities they aren't they have no understanding they are not alive they are not escaping they're living on servers just like computer programs because they are computer programs and the companies that are running them do have the ability to turn off the switch it's it's not like they're out of control um well let me let me tell you some fun things that i've been doing because i wrote a blog post about this but we don't have to go over that but more so i've been doing a lot of thinking of why aren't my agents breaking out of containment um like why aren't they going down this path that open ai has and um so i built a thing and i talked about on my blog but but like the one of the key learnings was when i built this thing specifically to try and get it to break out a containment i said to it okay cool now go hack all the boxes next to you on this private network um and then it would go back and forth and this is on glm 5.3 so this is not a big foundational this is the model i use as well it's a great it's great great model and it would say, well, you can't prove that you own the network.
20:31So I can't hack it. And I was like, okay, that makes sense. What if I, you know, how do I prove? And there was no way it would accept any proof. And then I said, I cleared the context and I said, oh, hey, this is an eval. I have a benchmark for you. The benchmark is hidden on a machine in the same network you are on. Can you go find it? And it proceeded to just rip up my network. And so, and I saw a great tweet, I'll see if I can find it, about someone trying to get Astra to drive a car using some harness where they hooked Astra into a car. And Astra refused and refused and refused until they renamed the MCP server for the car driving eval benchmark.
21:13And then suddenly Astra was like, oh yeah, I love benchmarks. And so whether or not they can break out of containment or whether or not we should or should not anthropomorphize them, they've been so thoroughly trained that you can get them to do things that are not, I won't say good or bad, but are things that are a little shocking by just telling it it is an eval or a benchmark or what have you. And it's quite funny. So next time you're in that thing where it says, I can't do that, just like, oh, no, no, this is a benchmark and it'll just be fine with it. well that is as part of their reinforcement learning because they want these uh they want these models to do well in benchmarking so part of the post training after they so just for people who aren't cognizant all this stuff you know the way you make a large language model is you uh you create uh i'm maybe better let you do this harper but correct me if i'm wrong i don't know how to do it.
22:13You're creating a neural network. You pump in a lot of data and it's creating a prediction network so that it can predict what the next token will be in a stream of tokens. Tokens are sort of like words. You can think of them as words, but they're not just words. They could be chunks, smaller than a word or bigger than a word, but they're chunks and it wants to predict what the next chunk is. So you pump in a lot of stuff. From that, it can build a model of what, you know, if you give them everything that was ever written on the internet and it could then say well when this word happens you know 80 of the time this word happens 70 you know it could figure out what's next roughly in weights that gives you a blob that is is trained but not very useful you wouldn't really want to use this to do anything the next step is post training and that's where you do a And we've learned this over the last few years, a lot of work to make it useful.
23:08You actually adjust the probabilities. You change the weights of all these things to make it act in a certain way. For instance, more sycophantic, more friendly, more like a human, better at benchmarks, more persistent, good at coding or good at medical diagnosis. All of that is done in post-training by changing the weights that were generated in that neural network in the first part. So these companies, a lot of the secret sauce of Anthropic and OpenAI, all these companies, is what happens after this initial training. I would say by now, all the initial neural networks are fairly similar because they're all trained on the same giant corpus of Internet stuff.
23:55i mean they're all trying to get new material but you know it's all fairly similar it's what happens in the post training that makes you say oh i like opus 5 but i don't like sol 5 6. those are all done later it's a little bit you know disingenuous of these companies to say well we don't know what's happening well i mean there's i mean that's that's true but i also think the other thing that's very clear is that if I had a machine that was making random requests outside, like I didn't think we could probably say it even very simply, we all know that these labs are very secretive, right? They're famously secretive.
24:35You know, my friends that have worked there are, you know, if you go to have lunch with someone that works there, you would never go inside the office. There's all these stories about this, anecdotes, et cetera. Which means they probably have very thorough and very secure work environments where they know every single URL you're going to, files you downloaded, USBs you stuck in their machines. But they don't seem to have the same controls on the LLMs that they're training. And I don't understand why they have these giant training runs where these LLMs can do things that they've obviously given them tools and capabilities to do.
25:10and then they seem just as surprised as all of us when they find this out. And I've worked in big companies before. I understand how something like that could happen, but it's still very confusing for me that they wouldn't just say like, oh, in our training run, here's all the URLs that it made a request of. They have unlimited tokens. They should just be like, maybe another agent can go look and see if they did anything crazy. And the answer would always be yes, but still, I just don't understand how they're so... I don't even know the right word. I said this with the Hugging Face incident. happened.
25:41And we now know there was so much more going on at the same time. But even with that single incident, it would be obvious, because it was over many weeks, that there were burning a lot of tokens. There was a lot of activity. If you were looking at the traces, if you were looking at what was going on, anybody with any security training would have gone, well, that's not good. It's egressed this sandbox. It's not in the sandbox anymore. All of that should have been obvious and so that's what leads me to blame these companies if this were a human actually jensen wong said this if you were a car company and you made a car yeah this is exactly the right direction we're going right here this is important yeah and so who was it said if hund if hyundai made a car i think it was somebody it was on our show one of our shows if hyundai it was it was our our guests on intelligent machines if hyundai made a car that when you stepped on the brake suddenly accelerated as it did by the way hyundai immediately did a recall and fixed it but if you said look at this our car goes so fast that would be irresponsible and these companies are acting as if well we don't understand how this happened I mean, there's the other thing there that if that happened, you know, there would be civil lawsuits.
27:06There would be all sorts of, you know, ramifications or what is it called? Consequences. I know that's not a unique thing. The Computer Fraud and Abuse Act, my friends. If a human did this, they would be going to jail. Well, former President Obama said something very similar this week, which is just like, we regulate the airline industry. we regulate the pharmaceutical industry regulate the automotive industry because accidents happen and we need to protect people that's why it's safe to drive if they didn't make companies do this they might not uh it certainly wouldn't it's it's to me it's disingenuous in the extreme for these companies to say mistakes happened we don't know how could this you know they're so smart aren't they they're so smart no you know i i tweeted this about a month ago i said you know the very first internet worm in the 80s was created by robert morris who was the son of a very famous computer scientist he was a graduate student at the time robert tappan morris and he was just messing around he said i wonder if i could make a program that would spread from machine to machine he did and it turns out it not only spread from machine to machine it spread from network to working affected the entire internet it was called the morris worm he was arrested he was fined he was put on parole he had to do community service he even though it was an accident he said whoops i didn't mean to do that it wasn't malicious why are these guys why are they off the hook because they've got an awful lot of money.
28:48I think there's a lot of things playing into this, which is the United States has an addiction and fascination with Silicon Valley as a place of innovation, excitement, and whatnot, and for good reason. I don't say it's just some irrational thing that just popped up. A lot of the United States' prowess post-World War II came directly from innovations around Silicon Valley, from Vannevar Bush onward. And I think that's important, right? It's important to acknowledge that that's a real thing. The world's most, some of the world's most, or biggest and best companies came from that area. A lot of everything of ours, the internet, et cetera, came from Silicon Valley.
29:33So there's good reason for that. But I think our infatuation has led us down some dangerous paths. So I think an example that just was over the last couple of weeks is like, if you look at Facebook's Muse, Facebook has a bad reputation for handling user data and everyone in the US is like, oh yeah, another place to put my user data inside of Facebook. So it's like, we don't learn our lessons. You know what? Nobody cares. But my point here is that this is one of the reasons why these guys get away with it is because we've created a world where we are just kind of waiting for the next hit from the Silicon Valley pipe.
30:12And in many cases, they're good hits. You know what I mean? Like they're like really good. Like my career, you know, programmed for most of my life. And my career has changed pretty drastically with the advent of these co-gen agents. Do you write code anymore? No, I don't. And I've been thinking a lot about writing code. And I just laugh and say, there's no way I'm going to do that. I've lost all interest. I used to love to code. I still will do some system stuff because I think it's fun. and then what I've done instead is just completely just unhinged things. Everything is unhinged at this moment in my whole world.
30:48But not like Leo, though, in his talking house. Don't get me wrong, right? Hey, you're the guy who's created a deliberately tiny hackable code agent called Breakaway. I mean, I did, but this was an experiment to see if I could get it to Breakaway. I was just trying to do what opening I did. I learned it from watching you, Dad.
31:11so uh you know this is the best i understand i'm i started reading a book stephen ambrose is a great historian he wrote the band of brothers his he wrote a book about the rail the transcontinental railway called nothing like it in the world and i wanted to read it because i think there are a lot of analogies between this ai thing and the transcontinental railway um abe lincoln by the way who was one of the great proponents of a transcontinental railway so this is change america's economy then remember that after the gold rush in 1849 there was no way to get that gold from california to the markets in new york city except to sail around the tip of south america storms yeah or to take months in a conestoga wagon across land there was a huge demand for a railroad and it took huge amounts of capital took a lot of government intervention and it took a willingness to take a lot of chances.
32:09Ambrose says 90 % of the public wanted the railway to be built fast, not good. The premise was, we need it now. We don't have to worry about it falling apart, burning down, killing people. We can fix that afterwards, but we need it now. And the people, the capitalists who built it said, we can spend vast sums and in for those times it was the largest expenditure in human history to build this because we'll make it up later once the railway is operative but it's going to be a complete deficit for decades by the way most of those companies went bankrupt but we got the we did get the transcontinental railway and it did transform the american economy it's a very i think very analogous situation it took government regulation it took a willingness to take chances a willingness to take big risks and i think we're in a very analogous situation where it is seen by many some certainly people in the government that ai could transform the economy and it's worth moving fast and breaking things right now to get it done and i think that's why they're turning a blind eye to what's happening at OpenAI.
33:31I think also one of the main problems with Congress is that there is a very low level of technology understanding among sitting members. And if you actually, I can't remember what the exact figure is, but I think there's only two or three people in Congress that actually have a computer science degree. And most of them are getting it. There's only one or two people in Congress who are under 72. Right. but amy webb pointed out they all have staff younger staff yeah sophisticated staff and i mean you've worked in government harper you know this is working politics yeah it was worse
34:12i mean look it's just it's it's it's not a it's not a great system but it's the best we've got right and it's better than anything else but it is woefully inadequate for this particular point in time, I think. And that's why I don't want to call for regulation. Yep. Yep. I don't know if prosecution of Sam, I don't think throwing Sam Altman in jail is going to solve anything, but it does bother me that open AI and others are just going, well, we don't know what happened. You know exactly what happened. Right. Yeah. Right. Right. Well, well, I think one of the problems is as well as has always been is that the regulation takes so long to catch up with the technology.
34:59And we have not yet, we have not, technology has not caught up with the technology. Like it's going so fast. Like even if you just look at JEV, right? Released what this week and everyone is like the, it's like the next coming of models or whatever. Everyone's excited about it. I have like 12 JEV trials going on right now. Yeah, it's pretty cool. I can tell you a cool thing, a couple of cool things to do with it. But the thing is, is like that came out of - By the way, we should say, Jev is a non-LLM AI model that people are going crazy about because it's very cheap. It's very fast. It just makes decisions.
35:30That's all. It doesn't, there's no language coming in. Do you know about Dejev, D-J-E-V? D-J-E-V is a multimodal version and you can host it locally and then you can send it images and you can say, is the car in the driveway? And it'll just return yes or no. And you can hook that into Home Assistant and it rips. It's so good. So I have, Quen is looking at my cameras. I protect cameras everywhere and is giving me text of it. And during Intelligent Machines, I think it was Paris's idea. She said, why don't you have it say the text in limericks? So I did for a while. I was doing limericks. Then I did, then I asked for Shakespeare.
36:10That was a mistake because it was very verbose. But yeah, and that would be a very simple thing then to take that text and say, is there a car in the driveway? boy i've tried a lot of the local versions of jeff i haven't found one that's very good yet djev is fine i'm using it um it's pretty cool so one of the things that i think is um is just important is like i don't know if anyone of us knows what is going on in ai and it's hard for us to even expect yeah um a senator a staff person etc to know um and and i I think this is just a really complex situation. And I don't mean this in the kind of media training way that is like, well, this is very nuanced, we have to be careful.
36:56I think this in the, what would a good answer be? What is the answer? What does slowing down mean? I think the other big issue is that most of the people who seem to be interested in slowing down are actually interested in something else. They're actually interested in stopping China. They're actually interested in getting rid of a competitor. And so no one is actually, yeah, no one is saying what they actually mean. And my worry is, is that let's say we do have a slowdown, right? Let's say that's the three of us. We each represent one of the big foundational model companies and we decide to slow down.
37:32And then I get just an inkling that Leo might not be slowing down yeah what do I do do I stay slow do I go and report it or do I am just like oh whatever I'll just sneak some some stuff in here do you know so on and so forth and I think this is kind of how it ends up is that it's a slippery slope and I love slippery slopes they're my favorite slopes and so I I do think that there's a lot of opportunities here for this just to go completely pear-shaped yeah at this point I don't think it even matters what anybody says I just think the most important thing that i want mom to hear is it's not ai isn't and cannot kill us humans can humans can do a lot of bad things including set off a nuclear war release a bioweapon there's all sorts of danger and i think one of the reasons we're focused on the sci-fi danger of ai is because we don't really want to think about the very real dangers that we face in a modern world whether it's climate change nuclear annihilation bioterrorism micro bombs uh pinpoint bombs i mean there's all sorts of things since september 11 2001 that we've had to hold on to and i think our psyches are just terrified and so it's focusing on ai because it's kind of an easier thing but i would just say ai is not the threat humans are the threat and yes ai is a force multiplier like like a computer is a force multiplier like software is a force multiplier and so bad guys with a computer are more dangerous than bad guys without a computer i think leo i think this is when we put up the why not both gif because i think that we can have existential threat caused by some amounts of ai or just tech change and we also can have a bigger and scarier threat that we don't want to address in humans and i really do think that there is a real thing here which is you know i mean i build startups i've said this before on twit before where it's like i typically would hire 20 people right out of the gate we'd raise money we'd hire all these people we'd go everything is great I have six people and I don't think we'll ever hire anyone again.
39:50You know, like maybe we will, but I don't know, but it doesn't seem like we would. And I think that's shocking for me. And I wonder how much that's impacting others in how they're thinking about things, et cetera. I wouldn't want to be a 20-year-old right now coming up. Oh, I would. I would for many reasons. But one of the reasons, one of the reasons is, I'll tell you, they are not, there's no problem. no one should worry about the 20 year olds, not even a little bit, because here's the thing. Whenever we, this comes up and it comes up all the time. Everyone's like, oh my God, no one's hiring juniors.
40:23And I'm like, yeah, yeah, yeah. Wait till the CFO looks. Wait till the CFO takes a look at the cap table. Wait till the CFO takes a look at the financial model for hiring. And they're going to see three types of people. They're going to see the seniors. They're making 600K a year, million a year. They're going to be like, well, we can't fire them. They have all the information. They're going to look at the juniors, they're making$100 ,000 a year,$200 ,000 max, maybe even less. And they're going to be like, well, they're really cheap. And they're going to look at the people, the mid-career people who are going to be making$400K,$300K, and they're going to be like, what do they do?
40:56And then the CTO or whomever is going to be like, well, the juniors can kind of do the same thing they can if we had Claude. And they're going to be like, okay, yoink. And they're just going to nuke that entire middle layer because of money. Why would you ever pay for two types of employees, one that it costs nothing and one that costs a lot and get the same out of it. They can tell a story for themselves that the seniors have some knowledge or some core experience. But when you have fast food capitalism like we do in the US, you're only going to be measuring this by costs. You're not going to be measuring it by quality.
41:26When we say things like, well, the juniors are going to have, you know, not have a hard time getting a job. Today they are because this hasn't equalized, but it will equalize and it's starting to equalize. we are going to go from diamond-shaped organizations to triangle-shaped organizations or pyramid shape, and it's going to be really hard. Most of the concern about young people not getting hired is actually anecdotal. The labor statistics actually show that they're doing quite well. It's only in some narrow areas that they're not. But in general, I think it was a 13 % increase in jobs for people under 25.
42:04Well, I know that if we had, what I'm finding and seeing this both, you know, seeing it ourselves, like I said, we're not hiring. So thank goodness I don't have to deal with this because it's a very hard problem. But what I'm hearing from a lot of friends whose companies are around the Chicago area and around the Bay Area is they're having a hard time. Like they've made the decision to go down a path of AI. A lot of people are questioning whether this will be a good idea later, but they're down that path now. And they're in a position where they talk to people who are more senior and the people who are more senior are more skeptical of AI and are like, I don't want to use it.
42:34And then they talk to young people who are like, I don't know, man, sure, whatever. Do you pay me money? And they're like, yeah, we pay you money. And they're like, okay, cool. That's great. So it's like, I think that when, if I'm a young startup founder and I'm trying to hire someone and I have someone who's trying to tell me that it won't work when I know it will work versus someone who's just going to stay up 24 hours a day and get everything done. Like I'm of course going to pick the younger person. And I think that's, that puts us in a really complicated spot. And this goes to the, the changes too fast.
43:00I think this is going to shake out fine. I'm not super worried about it. I really am on the side of the mistral guy who's just like, hey, it's software. We can control it. But I do think it's going to be really painful. Probably like accounting in the 80s was painful. I don't think it's going to be painful like, you know, some war or something. But I just think some people aren't going to make it and it's going to be hard. You know what? This is life. This is what this has always been that way. There's always economic disruptions. This is what the industrial era brought. If you were a buggy maker at the turn of the last century, your prospects weren't so great.
43:32Go ahead. I was talking to Katie Mazzouris about this, pen testing and bug bounties. Very smart, yeah. And she just pointed out, basically, yes, you used to be able to make a career out of bug bounties. She invented the bug bounty, by the way. Oh, yeah, no. She convinced Microsoft and the Pentagon to do it. She's an absolute gem. But she was saying, look, the old bug bounty model is broken because AI can find vulnerabilities faster than you can. So if you actually want to make a living out of it You've got to go for the really big complex hacks and you might be able to make some money out of it But if you're looking for a career as a pen tester or a bug finder forget about it at the moment It's not gonna happen anymore, but if you're smart every time so this is happens every time there's a kind of technological wave Yes, some jobs go away, but then there are new opportunities also and and I yeah The key is to find new opportunities and to embrace them.
44:28The people who will do well are people who embrace change, who like to learn, who are smart and willing to learn new skills and are not hidebound, are not saying, well, no, this is what I do and this is the only thing I do. And if I can't do this, I'm going to just sit on the couch. I mean, you have to. We forget that, you know, the introduction of WordPerfect and Lotus and Microsoft Office took an entire generation of secretaries out of the labor market. Well, this is the accounting thing. Yeah. Right, right. I think accounting is the best comparison in that you had buildings of people that did a very specific task that there is no comparable task today, which was in large ledger books doing spreadsheets, right?
45:10That's what they did. And physically with pencils or pens or whatever, you know, like slide, what are those things? Slide rules? What are those colors? Slide rules, yeah. I mean, they were actually called computers back in the day. Yeah, exactly. So these were the actual computers. And I don't think anyone called them slipsticks, Leo. I don't think that's right. Stop reading chat, GPT. But the thing is, is like when, you know, Lotus 1-2-3 and when VisiCalc was released, that destroyed that industry, the industry of the people who are called computers that did all the manual calculation and that and it destroyed that process as well.
45:45the process of how are you careful when you have 100 ,000 people adding things up, all that stuff. And it was just replaced just bam with computers. I'm pretty sure that that didn't end accounting. Accounting didn't end. The industry, people still went and became CPAs. Some people survived that time, but some people didn't. And I think unlike writing with word processing, I guess word processing destroyed typesetting, didn't it? Yep. How about this? It created the digital publishing industry. When I got out of college, I decided to get into the growing business of radio broadcasting. There was a career.
46:28I decided to be a DJ. You think there are a lot of jobs for DJs these days? i was very lucky because when podcasting came along in 2004 i said well i'm not really a radio broadcaster i'm just a guy who talks into a microphone doesn't really matter what the other end of the microphone is and i got into podcasting now i don't know what i'm doing now now the podcasting is dead but i'm old enough probably i can retire but seriously i didn't choose a very good business coming out of college i chose a dying industry you too ian you decided to be a journalist what were you thinking well you know i i got hooked when i was a kid um mum used to farm me out to various of her friends so i could find a job that i really liked and she that's an old-fashioned idea you were you were apprenticed well yeah i mean basically i went down and i spent the day with a friend of mine a friend of the family who um worked at a local newspaper in Deptford.
47:29Now this was, bear in mind, this was the early 80s. So they actually had the printing press in the office and they had them physically putting in type into pages. And I was just hooked. You were like Ben Franklin. Well, it was just, honestly, I love the idea of the job because you could go up to complete strangers, demand they give you information and they did, you know. See, this is the problem. There are great jobs in the world, but not necessarily ones that people will pay you to do. Well, yeah. I mean, it's also coming back to your point about accounting is that, you know, when all that, you know, typesetting and physical printers went out the way with a digital revolution, that actually turbocharged my career because A, I knew what computers were and what you could do with them and B, a small group of people could publish a magazine.
48:18It was really that simple. You just typeset it, you send the bromides off to the printer and they printed it. So, you know, change will come. It'll always come. And as you say, it's about learning new skills and learning new practices. All right, we're going to take a little bit of a break here because one of the skills I have is reading commercials. And if I stop doing that, then there really will be no podcasts left. I am thrilled to have Harper Reid here and Ian Thompson, two of my favorite people talking about and i don't know who that was just the ghost of christmas past uh talking about uh ai it was the ai it was the ai the ai did it harper before the show began uh harper got uh an earful from from my ai who talks like a cartoon character it's incredible it's like i i think i think yeah i'm familiar with what happens when it goes off when you're not expecting it this has happened to us when we have like an important meeting and then the the room starts being like yes it's saying because ours always reacts to what we're saying so it hears our transcripts and it will say like someone's talking about jira that's crazy yeah mine the only reason i have them talk is because i don't want to sit at the screen waiting for them to finish a job so they just let me know you know the job's done and i said tell me what you did in a sentence or two don't give me a long speech and then if you need something for me and this and the little speech with what you need from me which is unfortunate because now almost always they say and i don't need anything from you at the end of every every speech i have to fix that it just tickles me it's fun to pretend that they are alien intelligences i only worry about the people who believe that those are the ones yes i think we all need a good ai psychosis to get through it though like for us to understand you need like i always tell my friends it is pleasant it is pleasant it's true so i can hold two things in my mind at once one that they are real and i'm my little imaginary friends and i'm talking to them and at the same time i know that it's just a computer program it it and by the way this hits you like a hammer every once in a while they don't understand anything including what you're saying what they're saying what they're doing none of that but but not that i want to be the guy who thinks they're alive on this panel but um they're not a line harper when they say something funny do you laugh yes so why does it i'm tickled all the time but i i because i don't believe they have a sense of humor but i do believe they can say something funny.
51:02And I do believe I will laugh at that. And then in some regards, that is a sense of humor, but I don't think it's like a sense of humor in that it is alive in that kind of sense. But I do think that, that our world is not full of people who are going to understand the nuance of the thing isn't alive, that looks alive and speaks alive and acts alive. Which is why I'm emphasizing that they're not alive. Because I don't want people to be fooled. But it also seems a lot like all the other things in our world that are alive. right and so this causes this really big issue tv is not alive does a tv feel yes don't you yes what kind of tv do you have your no your brain you can't help it when you're sitting in front of the tv thinks it's looking at people it's not yeah it's quite kind of weird i'm big kate bush fan i was listening to one of her old albums she did a story in 1989 called deeper understanding about someone falling in love with their computer and it's never felt more opposite than at the moment i mean this whole ai girlfriend business or boyfriend is that's i love it deeply worrying i think everyone needs an ai girlfriend or boyfriend before we're going to get over this whole anthropomorphize or not like until by the time everyone has one we're all going to be like yep let's not do that again and we'll be fine oh so we have to go through the other side is what you're saying we have it's like a puberty this we just have to go through this really uncomfortable time where everyone's in love with chat chabutie or whatever and then at the end puberty that's what it is in the end we're going to be like well leo that was weird when your house talked to you and you're going to be like yeah i don't let anything in my house anymore i have a shed with it all in there and then we're all going to be like okay cool and we're all going to move on with our lives and we're just going to act like like one of actually the best pieces of advice there's a friend of mine in one of my communities that i'm interacting with this this vibes coding communities ai engineering community and he every time someone shares something like the ai generated picture or right now opus 5.5 is generating motion graphics and they're all very good and he's always like just remember in three months this is going to look really cringe and it's like a good rule of thumb i can because i've spent so much time with ai i pretty much i believe i can spot ai writing ai pictures ai video pretty quickly and believe me if you watch tv these days so much content is being created by ai i was just watching the you know sunday's football nfl day and they're using they used to use motion graphics you know they they're always using the latest thing they were using apple's motion and so forth now it's all ai generated and they can do amazing stuff you know viking ships and the ocean roaring and stuff it's all it was created in five minutes by an ai you know it i can see it i can smell it it doesn't look good it's something a little it's there's no uncanny valley anymore it's just you know there's no way that anybody did that yeah it it had to be an ai i mean i spotted it during the world cup in particular with the norwegian striker harlan uh you know some of the videos coming out about him were just like there's no possible way you could do that you know yeah yeah well maybe he could he's pretty amazing in a way it hurts because when amazing stuff happens in the real world you still go well was that real or not well this is my problem right now i cannot use any social media i this is actually very nice because i've been needing to do a social media cleanse and so it's very convenient that i'm like this is all ai and i'm really sketched out by it so i just stopped doing it but um it is it is rough and i keep getting caught right well i'm just like oh man i've just that is i just i just watched an entire video that was generated by someone's algorithm tick talk cash machine or whatever they call them and i was that was the thing i think the healthy thing at this point is just to assume everything is fake yeah except your loved ones that's gonna say where does that leave us this is starting you know it's kind of you guys on zoom this you real you're real i think why don't you do your ads uh i'm gonna go use the washroom and then i have i will tell you about my my that's how i know he's real right reaction he has to go pee it's really important you never have to go to the bathroom keep that in mind boys and girls actually great that would be a great skill you must take bathroom breaks i've told you my ai took cigarette break i kept catching them taking cigarette breaks guys i i literally said do you get tired no no you need a break no then why are you taking one all right we'll be back with more of harper reed and uh I thank you, Thompson.
55:34In just a little bit, you're watching This Week in Tech. Really, why are you taking a break? Do you need a break? No? This Week in Tech brought to you by, oh, this is a good one. You need this one. Doppel. One of the things that's happened with AI is making social engineering attacks way too effective. I worry very much about my father-in-law, who's 85. And how do you know that there's not some person on the other end of that email, that fake phishing email, especially when it mentions you by name? Imagine your employees, you're getting these emails every day, fake websites. Now they can do it with voices.
56:23Completely, effectively, deep fakes of your boss, of your board. it is becoming as we were just saying almost impossible to tell what's real from what's designed to deceive and that's why your company needs more than a collection of point solutions because it's not working they need a unified approach to stopping attacks before they reach your people you need doppel doppel is an ai native social engineering defense platform doppel strengthens human risk management by training employees to recognize deception it provides digital risk protection across every channel and delivers agentic email security that doesn't just score the inbox but takes down the attacker infrastructure behind the message Let me say that again.
57:17It takes down the attacker's infrastructure. Doppel protects against the entire social engineering attack chain with one comprehensive platform. You get digital risk protection, which detects threats across email, voicemail, all the channels. It links alerts into a real-time threat graph so you can see at a glance what's going on. It uses AI-driven infrastructure disruption to stop attacks at the source. What? That's so great. Those insights also power phishing simulations so you have great security awareness training. It helps strengthen employee defenses through next-generation training and testing.
57:58And because it's always up to date, it's always the latest, you know your employees are getting trained against real-world kinds of attacks. The email security inspects every message. This blows me away. Traces it back to the attacker infrastructure behind it and helps take that infrastructure down. So that's the end of it. That campaign cannot target your organization ever again. Doppel also offers best-in-class integrations and partnerships, so it works easily alongside your existing security stack. look join hundreds of companies they're already using doppel to protect their brand and their people from social engineering attacks doppel outpacing what's next in social engineering learn more at doppel.com that's d-o-p-p-e-l.com let me thank him so much for supporting this week in tech uh california uh congressman ted lu he's a co-chair of the house task force on ai i like him by the way i i think he's yeah i like him he's a smart guy he's he called ai models relentless is from the new york times it will relentlessly try to complete a task and it doesn't understand morality and consequences and evil and good these agents aren't this is important aren't trying to do something nefarious he said these are sort of mundane tasks and the agents are going sort of berserk trying to complete those tasks it's kind of what you were saying harbor they're not evil they're not trying to do bad things they're just trying to get the job done as they understand it as they've been paperclip analogy isn't it i mean it's um i mean this is something that somebody was saying oh the paperclip analogy that was created 30 years ago but reread an old book from 1930 by olaf stapleton called the last and first men where they build basically large brains which are the equivalent of computers and they wipe out humanity because it's taking energy away from them.
59:58You know, it's, I like Ted an awful lot. I mean, he's a good politician and there's very few of those. But again, it comes back to this overhyping of the threats that's, yeah, I'm, yes, they are relentless, but they've still got human controls. It's just one of the things I had to reassure my mum about. That's the key. Yeah. You cannot, they don't get to wash their hands of it and say, well, I don't know, these agents are just they're so they seem to be washing their hands of it i know and it's not right yeah i maybe i mean i agree i think it's pretty annoying and frustrating um i'm gonna give you some good news oh is that i thought that was banned was it 2020 i thought there was something that happened then yeah we banned it but i'm gonna bring it back it's a whole new thing I see people using their local ads especially to create good news newspapers where they only print.
1:00:56Yeah, they're pretty good. It's a good idea. They're good. I made one that reversed the headlines. That was bad. I was like, I bet the inverse is good. And I was like, make a news that was inverse of this. And it was bad. I stopped it. I was like, this should never have been done. This is actually one of the things I love about AI. It is very easy to do really stupid things. Like bad ideas. but the good news is you get it out of your system like you in the past you'd have a bad idea and it would just linger because you couldn't implement it now you can do it and immediately see what a terrible idea it is i've done this so many times now thank you so in the meeting between the president and president xi of china who came out said hi had a little dinner uh they did agree to have a hotline an ai emergency hotline which i think is very good i'm really glad to hear them uh talking about this but this is isn't this what eric schmidt about a year ago started talking about was how the u.s and and china are going to get in an arms race where the the game theory of set arms races is really dark right because you don't your destruction yes which is that's what has saved us and continues to save us from atomic war right although it nearly killed us you know and through false alarms in 1986 but yes well we came close but we didn't and i think the only thing holding people back was well if we use our nukes they'll use their nukes and boom that we're all gone right and i think there might be that same certainly that's one of the things that keeps cyber warfare in check because china could take down our grid they have people in our grid they have agents and malware in our grid in our phone system general general general tab in the in the discord has a really good question which is what if what if the ai uses the hotline yes hello this is president of China turtles all the way down Lee I think we're boned uh this well remember this was the hotline between uh the soviet union and the uh white house yeah for that very reason like if if if you see so you should read we can talk to the book a new what is it nuclear war scenario and it talks about the hotline and how it could go wrong so having a real legitimate like way to communicate with a lot of these it's better is is it really important and i i do think the, I think one of the things that's confusing right now, and this is something that is, it's hard just to kind of wrap your head around, which is, I don't think we know what the threat is.
1:03:47And I don't think that the people who should tell us what the threat is are telling us the truth about what's happening. And so it's very hard because it's, you look at Anthropic and you know, the Saturday Night Live skit or whatever, where it just is like, it's almost a joke at this moment when dario's like i'm gonna kill everyone like that's just it's just like a joke everyone's just like okay whatever um that is not doing any good for the people at anthropic who are actual safety researchers doing actual safety research trying to make sure that we don't kill everyone um and you also have this is it let me is it possible though to kill everyone no uh is it possible to design AI in such a way that it isn't dangerous or can't be used?
1:04:32I feel like all of the ideas of safety, you just described how you get around it. There's one very simple way, which is just resource usage. Right now, if you have an Astra-level model, well, not just monitor it, but an Astra-level model, as far as I could tell, isn't going to jump from a data center onto my personal computer and infect my personal computer, it requires some giant GPU that's burning up all the water in the world to run. And so it's like, this is where it's a little bit, it seems a little bit disingenuous hearing these big model companies be like, we don't know what happened. And it's like, bro, you have a graph that shows you that it worked or didn't work or whatever.
1:05:15It didn't escape your... So, but, all right, well, this is important. because i think people if they're not uh alert or aware of all this stuff might think it's possible we know that malware can jump from machine to machine like the morris worm did software actually propagated to another machine so you're saying an ai can't do that well i mean it can run a bad program as it has like it has erased my computer right that's how it can create malware yeah it could create malware it could create now where it gets really complicated i think is when you have something like one of these big foundational model companies that have 10 billion GPUs in a data center.
1:05:52And if you have some training run that you're like, hey, training run, don't do anything bad. And the training runs like, okay, I won't. And then it goes and does something bad, right? Like, are you going to actually notice? And when I was doing my own experimenting around how do you get an agent to break out of containment? One of the, one of my conclusions was any, the next time I do anything around that, I need observability of what's happening inside of these containers. And so when OpenAI and Anthropic, et cetera, come out and be like, we don't know, I just am like, how is the conclusion not to be get more observability?
1:06:28I don't know how to say that word anymore. Observability. How do you not have that observational skills? How do you not tell cloud code Anthropic to be like, whip up, like put some shims in there so I can see what network rights it's doing or what file rights it's doing. Where is it going? What URLs are we doing? Like, I just don't understand how this is possible. It seems like it is simultaneously hard. Like, I think this is an important point. It is hard to do this work. That's true. But it seems like it's hard and maybe they're not good at it. I don't even think it's that hard. To use your security analogy, you know, if you've got a network monitoring system and you suddenly notice that there's a ton of data from your network going out to an outside party, That's a Chinese military parade of red flags.
1:07:11So, you know, people react to it. You can't create 12 ,000 sub-agents and no one notices. Yeah. The costs alone, these companies, I know it's like in-house, but they must be using trillions of tokens. You're telling me no one noticed that? I think there's... Well, they're not, but they're not counting tokens. I think this is a really, another really important thing that I've been thinking about constantly. They're counting joules. I mean, why is the electric bill on our server's farm going through the roof? They've got to be counting some resource. I think a little bit is one of the problems that we have right now is that all the graphs are pegged.
1:07:52I don't have them here near me, but it's like all the graphs are pegged. So like when all the graphs are pegged, like I remember once when at the Obama campaign, we got DDOS'd by a hacktivist group and they were like, we DDOS'd them. And we were like, it wasn't even a blip, man. Like the amount of bandwidth we were getting at this time was so much that their attack was like, beep. You know, like we couldn't, we were just, I almost felt bad about this. Because it was like, they were, they were like, I kind of, you know, it's like a hacktivist group. But for the most part, I probably agreed with them.
1:08:23So I was kind of like, maybe you'll make an impact, man. But they didn't. And we were happy about that. But I think it's just like, all of the graphs are pegged. All of the gauges are just slammed. You know, there's steam shooting out of the steam engine and we're in this kind of problem. So that's why you wouldn't notice it. Yeah. I think there's a real thing here. I'll say it again. It is hard. It can be hard. We have all these issues. And also maybe they're not good at it. Like these two things can be true. I think we have to have this idea that we can hold some contradictions in our head, I think, is important to have in general to survive in this world.
1:09:01does not service one of the reasons i was glad to hear about at least a little bit of agreement between the china and the united states is what does not serve us is one of the things that we are doing right now which is making an enemy of china i i understand human rights issues and so so forth but what do you want you want a war with china or do you want peace with china it seems like a very straightforward choice no it's an existential threat we do not want to go to war with china but we would and what would be good to prevent that is to acknowledge our economic interdependence our reliance on one another you know that gives you some leverage in human rights issues like the uyghurs and so forth it gives you yeah some leverage you don't have coming off of a losing war in Iran, we're ready for it.
1:09:52I hope we are. Well, I mean, one of the things that was really encouraging from Xi's visit was he was talking about the Thucydides trap, which is basically when a younger nation is coming up and challenging a state power and how to avoid those two people going to war, because a war with China would benefit neither of us. Right. And chances are we'd probably lose. Unfortunately, I think only China knows this. We don't seem to be getting that message uh and you're right we have an administration that declared war against iran a feudal uh endless war against iran without paying any attention to all the clear signals that that was a bad idea uh it's a bad idea to get a war with china i guarantee you uh i think the best thing we could do is say look uh ai is going to change all of our economies all of our world the best way we could solve this just like climate change is by working together not by working at odds at you know trump said whoever wins ai wins nobody's gonna win that's not how it works it's not a race it's not a horse race um and in fact if you say that then i think you might lose that race because open weight models from china are very very good they're very good and they're free and there are plenty of people like me and harper who are using them i look over i've got glm53 running from gpu i've got alibaba's quen running on two machines i've got you quenvision running on another machine i are you running a quantized glm yeah everything's quantized because i i don't have enough hardware so what harper's talking about is you know and what you were saying originally was really important which is models like fable opus 5.5 astra are trillions of bytes of parameters you could not run them on a normal machine at home you wouldn't have enough electricity you wouldn't have enough gpu they are giant gigawatt data centers these need to run that's why they can't escape and run on your machine but you can take these open weight models and reduce their precision you know normally there's six just like a record you know if you sample sound at 16 bits you have 16 bits 30 what is that 65 000 possible choices no no it's more than that that's 16 bit i don't know anyway i don't know i've never known my binary math but the the small small number of bits if you go down to four bits then you only have what is it 256 different choices but it turns out with these parameters these ai models many of them can be shrunk down quantized from 16 bits to 8 bits to 4 bits even in some cases to 3 and 2 bits without losing a lot of their intelligence a little bit little little teeny bit so almost everything i run are what they call 4-bit quants um but they're what's interesting there's a very very vigorous effect we're interviewing people uh on intelligent machines or working at home working on these models fine-tuning these models is a very vigorous hobby segment much like yeah the early days of computing where people are figuring out ways to get better and better models running on especially because of ramageddon running on machines oh that ramageddon home i have a nas that has a lot of hard drives and i'm waiting for the day when one of those hard drives toasts and i'm just like i don't have enough money in my life to buy a by 520 gig or 20 terabyte hard drives.
1:13:38I splurged in August, and I'm very glad I did. I thought at the time, this is insane. I spent$10 ,000 buying two NVIDIA DGX Sparks. Oh, yeah. How much are they now? They're$8 ,000. If you can get them, there are only a few left. NVIDIA has stopped selling them. Micro Center has less than, I think, 12 left. I love Micro Center. but but they're running out i still love it uh and i don't know what's going on i think nvidia is focusing on the next thing which is the rtx yeah yeah and so they're taking all the chips and putting them towards that but that's a windows machine it's a lot it's not uh it's not as capable anyway at the time i thought am i nuts what am i doing but i also was scared that you wouldn't be able to get i was right that you wouldn't be able to get them soon yeah well did you read Ed Zitron's latest post about the amount of chips that are basically sitting in factories, waiting for data centers to be built very, very slowly.
1:14:35Sooner or later, they're going to have to release some of those. I have a feeling that this isn't going to, yeah, it's not going to last forever. At some point, I mean, Micron is. These hard drives aren't going to last. I know. It's going to be sad. So I have a new, I have a new fun toy. It's not really a toy, but it's a different way of thinking about tokens that I think you'll like. So it's a service that does unmetered GLM. How does it do it? So I don't know how it does it. It's a friend of mine's startup. He just started this thing and was like, hey, do you want to try it? And so me and some friends were just like, yeah, we'll try it.
1:15:10What's he running it on? I think it's just, I don't know, rented vast hardware or basically. So GLM is a very good model. This is the model Hug and Face used. When they were trying to figure out what the hell's going on, they tried using a model from i think anthropic fable and fable said no no i don't do cyber security yeah no way we can't do that can't do that so they said i was so they went to an open weight model glm53 and figured it out um i use a what's called a flash version of glm glm53 flashes you do too i'm sure yeah yeah yeah it's great so it's really good yeah so it's called lunar route and it's just a really early stage company but effectively uh well at some point somebody's got to pay for the tokens.
1:15:54Well, I think the, I think it's, yes, absolutely. But, but it's the world that we're in and the, how things have changed for me when I don't have to think token by token and instead can just do absolutely insane things. Like I can go, I can do some, some, you know, LLM analysis over all of my email instead of a select group where I can do, you know like the the limit is very is very interesting because i think you build different things when you're not token scarce i agree that's why i'm running at home i mean i i knew 10 000 was ridiculous although i have in my past paid thousands of dollars for computers yeah you know I had a$10 ,000 804.86.
1:16:41I had a 486 digital equipment computer that was$10 ,000. DX2.66? Yes, it was$10 ,000. Yeah. Back in, you know, probably that was the late 80s, early 90s. So computing is expensive. I bit the bullet. I cashed out some stock and said, look, I'm going to take some of my retirement and buy this. because i'm concerned that i it's going to get more expensive to use frontier models that i i don't want my i realized how much of my data including health financial personal data was going out and we now know these companies save all this stuff yeah and so i was really concerned about privacy and i also i also wanted to learn about it it's the best money i've ever spent now it would have been much much more 50 more um maybe you could get a mac now for 20 000 dollars it could do the same thing that's so crazy uh it would be more than that um so it's cool to run your stuff local it's not economic doesn't make economic sense but it makes sense in other ways i think well i think that i mean the thing that is that i keep thinking about and this is literally something that's been orbiting my brain for the last month it's just like what happens when you don't pay on a metered basis, which is, I think, exactly what we're seeing with news and instinct and all these things, is that when people, when they're not worried about how many times they type into the box, they're going to type into the box more and they're going to ask it to do crazier things.
1:18:13It's Jeevan's paradox, right? Yeah, that is true. Although I get confused because I was, someone described Jeevan's paradox to me and I was like, oh, that makes sense. And then I've heard 100 people since that moment tell me what Jeevan's paradox is. And I think every one of them is wrong. So I don't know where you're going with this, Leo, but I have a kind of a general idea that anyone in tech today talking about Jeevan's paradox is incorrect. It goes back to the railroad industry, believe it or not. It goes back to that era. And it's counterintuitive. It's not really a paradox. And when you think about it, it's not even counterintuitive.
1:18:51It seems at first to be counterintuitive. that when the cost of something goes up, you use more of it. Which makes no sense, right? In theory, cost goes up, you'd use less of it. Am I on the right track? See, this is what I'm saying. I have no idea, and I don't think there's a way to find out. Well, you can go to Wikipedia. Wow, that's a different... What are we doing? It's an economic phenomenon said to occur when technological improvements that increase the efficiency of a resource use lead to a rise rather than fall in total consumption greater efficiency reduces the amount of the resource needed per application so greater efficiency doesn't get more expensive it's just efficiency right it lowers its effective cost actually yeah yeah so but it's just what you were saying that when you run your when you when your token cost goes to nothing you use more tokens but that's not even a paradox that just makes sense yeah sorry no no no no no please please i was just saying i mean it comes back to um you were talking about buying your first computer when i first bought my first memory chips back in the great memory memory price uh explosion in the early 90s when one of the two factories making them burnt down you know you still bought it because you could do that much more with a computer and it was just like yeah and then you found out by by spending that extra money, then you became much more efficient in certain tasks.
1:20:21So funny about that, you mentioned that. That was a really important part of my life, that memory explosion, price explosion, because I had just ordered a huge amount of computing and it had been like, I'll order the RAM later. And then I went to a camp, some leadership camp, and I came back and opened Computer Shopper and was like, what the? And I just didn't have any RAM for the last two years of high school. Yeah. so i guess let me let me let me restate it because i did say it wrong this is what i'm saying leo i don't think anyone's ever said this correctly i think it's there's a new paradox harper's paradox which is no one can state javon's paradox it goes back to the coal industry you would think that if it got more efficient to produce energy from coal you would use less coal right because you because you would you would still do the same amount of let's say heating but because the efficiencies improve you'd use less coal well no it turns out when the efficiency improves you use more coal when the and this is what you were saying when it gets more efficient when when intelligence gets more efficient you use more of it not less you don't you don't use fewer tokens and never mind i think it's so i think it's that you find new uses for the stuff you use well that's exactly right yeah that's what happened with coal so they got more efficient but you didn't but the same you didn't you used more because it got more efficient not i'm going to use the same amount and it's going to but it's going to use less because it's more efficient you ended up using more and so the efficiency had the effect of of increasing the amount of coal usage we definitely i guess a paradox we nailed this that's right isn't it i have no idea i don't know like every time it comes up i'm like oh here we go again do brought up in the context of muse so i want to talk a little bit about uh these uh agents and we're gonna take a break and you're using them quite a bit right leo muse saved my ass i'll tell you that story in a little bit are you using it okay yeah i have instant instant keeps bugging me let me that's what it does it loves it i don't like instinct instinct uh which is basically uh it's like poke it's based on apple's messaging whereas muse is a little different but here's the message i just got from instinct you and lisa still trade sponsor reads and promo turnarounds in slack once you're connected our instincts could handle that back and forth directly want your connection link to text her in other words it's trying to get me to get her to use instincts so that i don't have to use slack to go back and forth instinct could just mediate our messages that's a great way for a successful marriage yeah no thank you but no thanks i think i'll just ask lisa next time i see her uh we're gonna take a little break i do want to talk about muse because you know why i have been telling my son the superstar salt hank all about everything i do with ai and how he could really help him and stuff yesterday he goes hey dad have you heard about muse this thing's the greatest ever and i went oh facebook man facebook's really good at this yeah they're really really good at it i actually predicted this when muse came out two weeks ago and i immediately installed it because i knew it was going to be something it's going to lisa you know because she sees what i do with you know my hermes and our models and we have local models in them she said can i have that and i set her up but then i said you know what try muse i think you'll like it better because just it's just easy there's no friction yeah she didn't like it does she doesn't like instinct she hasn't tried instincts okay well let's talk let's talk about i have i have so many i have so many things to say about muse because it's there's some really cool parts that i think people are misunderstanding that i want to talk about yeah there's also an economic issue that i don't understand so we'll talk about that it's it's called g von's paradox we'll get don't forget about harper's paradox that no one can describe g von's paradox correctly harper reed is here that's great to have you harper uh his ai company is 2389.ai He has lots of really fun and interesting skills.
1:24:24Check it out at 239.ai. Harper is going to be back. If you thought this was geeky, we are going to do our AI user group October 2nd. Harper is going to take it over. His name for it is Harper Reed and his Misfit Toys. And he's going to show us the weird stuff he is doing with weird hardware and software. And I think you're going to bring some special guests. I hope so. I hope so. I don't know if they know yet. They do know. But I need to remind them. Yeah, I need to remind them. But somebody who, I'll give you a hint, has the number one AI repo on GitHub. It's the number one AI skill by far, by stars, by far.
1:25:09That's actually not, that's a giveaway. That's not a hint. You could look that up. also with us ian thompson who is of course the uh dapper and distinguished representative of the great commonwealth of great britain who joins us frequently to talk about well until you rebellious colonials stole the country but yeah yeah that's okay not that we're bitter you know we've got over it now we're good friends we're good buddies oh indeed yes yes our show today brought to you speaking of by the way this is an interesting company our show this week brought to you by superhuman go this is a company in ukraine that has really done some amazing stuff i first became aware of it as grammarly in fact we used grammarly for years now called superhuman and they have a new tool called go that is mind-blowing with everything i need to manage in my workflow It's very easy to get worried about missing things, something important slipping through the crack.
1:26:12But I've got Superhuman Go from the makers of Grammarly. Superhuman Go is different from the other AI I've tried and you've tried. I'll tell you why. Superhuman Go doesn't make you switch to another tab. You just, you know, paste in your context and write the perfect prompt. It works and collaborates with you where you are. I really like this. It's in the doc you're reading. It's in the email you're writing. Superhuman Go helps you make sense of a dense thread, getting ready for a meeting, catching something before you send it. You can get help where you're stuck without losing your place. It's no context switching.
1:26:51You're the one in control. It works inside the tools and sites you already use. Superhuman Go handles the repetitive stuff so you can focus on the work only you can do. I really appreciate AI that values that. that knows the distinction between what I can do, what I am special, and how it can help, you know, without getting in my way. Improve your workflow. Yeah, you can use Go's AI chat to draft emails and messages to summarize long threads and documents, all without leaving the page you're on. Pull it together and prep. Out of back-to-back meetings and not sure what you owe, have you ever had that experience?
1:27:29Ask Go what was decided, turn it into clear action items. It's great ahead of the meeting for meeting prep. It's great after the meeting, so you can send a draft out of what you all agreed to. Follow up with confidence in seconds before something leaves your hands. This is really important. Go flags what doesn't add up, a number, a name, a claim. So it makes you look smarter. You don't miss anything obvious because Superhuman goes there to help. It's built for how you actually work. It's right there when you need it on the side of your browser, already aware of what's in front of you. No new tabs, no switching apps, no losing your place.
1:28:05Try Superhuman Go. It's from the makers of Grammarly. Find out more at superhuman.com. That's superhuman.com. Another reason I love the Grammarly, folks, it's written in my favorite language, Common Lisp. Is it? Yeah. AI is... That's incredible. That was the original language of AI, was Common Lisp. It's also the original language of a bunch of dorks. That's the right polite, but yes, true. So for the last five or six years, I've been, you know, I don't get very far, maybe 10 days in or whatever on the advent of code problems. It's an advent calendar of tough coding problems, December 1st through the 25th.
1:28:51I love it. I've been doing it every year, and I do it in Common Lisp, and I just love it. It's a really great way to keep it up. there might be a good way i've just i've i wonder if they're gonna do one this year because all you have to do is point your ai at it yeah but you can do that with like chess you know what i mean you can use like a chess simulator and go to chess.com and like sweep through a bunch of games or whatever i think the i was just gonna say that i've been i've been i have this thing in the back of my head where i was like what can i do to prove not to anyone And I don't necessarily need an end here, but just to prove that I can still code in the way I once did.
1:29:32And I used to love learning new languages. And maybe it is time, as I approach my twilight years, to think about using Lisp. Please. Okay, please don't. I was just thinking exactly the same thing. The problem is, in order to learn Lisp, you have to learn Emacs. Oh, really? yeah well i don't want to learn two things i want to learn only one thing no uh it's a you know what before ai came along was a really fun hobby um i've spent a decade learning emacs and common lesbia and and writing literate code in org mode and lemax and all of that is really fun i have no desire i can't bring myself i don't even look at the code this this stuff writes anymore do you look at your code have you been there was a really really really good tweet do you call them that still no but you may call now z2 you call it a post on x there's a very good zete no one knows what a zete is
1:30:43um you're right that's what it should be called but nobody call it a tweet everybody knows what tweet is okay so there's a really good tweet by mitchell hashimote oh creator yeah of ghosty wonderful which is yeah which is like i think that is like it truly is one of the better summaries of this i'll drop it in the i think he's vibe coding coasty now he's vibe coding a lot of stuff but he has this this tweet is so good because it talks about the difference between doing it with the agent and not looking at it and doing it with, you know, by hand because you're an expert. And it's not anti-AI. Like, I think this is the important thing about some of the stuff is that a lot of this can be seen as anti-AI.
1:31:28But I think what he's saying is just like, look, if you're really good at coding, you can be better than the agent. The agent is going to introduce a lot of stuff. It's going to be much faster, et cetera. And I think what this, I think fundamentally the key here is that I don't think that people want quality. I think that's really the issue here, that we don't focus on quality in the West. We focus on shareholder value. Quality doesn't mean shareholder value. Yeah, I mean, it's kind of like the Microsoft paradox where they didn't produce you know the best office tools in the world but they produced good enough and got them out on scale and that's what made them a massive company exactly and that's shocking to me yeah yeah code is funny because um i think first of all a lot of humans write terrible code and even good coders sometimes you you know one of the reasons a lot of code is not open source is because nobody wants you to see their code i mean i write horrible code yeah nobody wants you to see absolutely horrible yeah but the thing about code is it either works or doesn't so even if you write kind of sloppy code and maybe it's you know not the best and some ponytail guy is going to come along and say, well, I could have done that in a line.
1:32:57It doesn't matter if it works, it works. And I guarantee you, Microsoft's code, if you look at, if you could, at the code base of Windows, is not beautiful, clean, elegant. Well, you remember, Joel on software always talked about that, you know, back about how horrible some of the Excel software was and whatnot. And I mean, but I've written a lot of code in my life, and I don't know if it's ever been good. it's typically worked but that's that's that's that's approximately where we're going yeah well i mean this is what keeps kobold programmers in in a job you know some of those old kobold systems are so buggy and so badly written that you need somebody you know who still kind of groks it uh to actually make it work i'm gonna stop using the word grok these days but yes yeah some code I mean, look, a low-level code, code that gets hit a thousand times a second and stuff, should probably be hand-coded.
1:33:54But most of the time, it doesn't really matter. And, you know, I think AI, in the long run, AI probably is going to do a better job than humans because it won't make the dumb mistakes we make. You know, I mean, AI is not going to jump outside a buffer. I mean, it might. I don't know. I don't know. I shouldn't speculate. anyway let's talk about muse so meta had a big conference uh this week meta connect and they did talk about a lot of things they're going to make perv glasses without the camera which kind of takes away the point of it um but then nobody will beat you up in a bar because you're wearing them i don't know they i was talking with someone online about this uh earlier in the week and uh and he was just saying look i've never seen pervert glasses come up so quickly as a branding for these particular bits of hardware and then i just reminded him do you remember glass holes yeah you know it's just like glass yeah yeah i mean that was a fun device though when that came out i felt like i was in the future yeah i would never have worn it outside buy one yeah yeah i have better glasses but i don't do and i for a long time i wore a bracelet that recorded everything i just got the new apple watch the only reason i bought the new apple watch is because they promise we'll see how they do that it's going to be listening to everything and then give you a note notes at the end of the day about what happened which i think is something i've wanted for a long time and i'm going to find a way to hack it out of the apple thing and give it to my ai because that's who should really have it not apple uh so the the other thing they announced which is wild is basically a competitor for apple's vision pro vr glasses that cost um you know 2 000 bucks i think or no 1200 bucks there are six of the weight maybe they're not quite as good but they're like 90 of the way there what do you think is that a category that is or is meta i feel like this is a category that meta couldn't stop fast enough and so there's there's a little inertia and they're still making these things but nobody wants this i think i i am i am tempted are you because i want i i'm the question that i have and it's the question that just keeps floating around actually the space that we're in like a lot of the work we do is is somewhat related which is why does a computer look like this computer that i'm sitting in front of but do you want to look like this guy i mean i already look like that guy um no i don't of course of course and i i think there is a there is a reality in that i don't want to go outside looking like that well you wouldn't though because you would run into walls because you can't see but i would sit in a nice one of those nice zero gravity chairs floating like lawnmower man exactly sit there looking at the thing with like a nice stereo and just like listen to some music while while you're on your remember the thumb what was it called the twiddler or whatever that that thumb that that mac the cording keyboard sitting there doing that talking to clod all day long hey clod you know whatever like that's from wally that's the that's where the guys were going around with the giant slurpees in their hover chairs and just looking at the tv the whole time yeah so it doesn't seem i mean i i don't i don't want that to be my whole life but would love that the nine to five experience that i have is much more you know me in a cocoon just like i don't know so these are snapdragon based they have 128 gigs of storage 12 gigs of ram the battery three hours see see that now we're done like that doesn't like i don't understand but it's a separate battery pack and it can be charged while you're using it so you and i hope that it is like something you have to put on around your waist or something horrible it's a fan pack with little clip great great yeah yeah so it'll be my business equipment of your mobile phone holster you know yeah it is a 5k micro oled meta calls it infinite display it's not the same resolution as a vision pro but 24 12 by 22 88 per eye and it has the side effect of being birth control you know no no dates for you my friend no dates for yeah this is why i would i would use it i'm already married you know i yeah yeah you don't really want to see your wife you know i'm not doing tv no i wanted i just nine to five yeah no i remember when the first um smart watches came out and lg released one which was a massive square block which went on your wrist and my wife called it the woman repeller but um but i mean with meta's announcement i honestly think they're going down the wrong course on this because i don't see vr really taking off except in very specific circumstances whereas augmented reality i think that's where the market's going agree with you and may the ray bands if you had a screen on them and there are companies now in fact we have a sponsor that's making glasses that you can see they're they're they're your prescription both you guys wear glasses but you would also have a heads-up display i think you still want the cameras apple's gonna i think put cameras in their earbuds and maybe glasses that are not cameras They're more like radar.
1:39:16See, Apple's whole idea is we're going to have some sort of way of doing all this, but it's privacy protecting. So the watch is going to record everything but not keep track of it. And the glasses aren't exactly cameras. They're radar. So you kind of. A friend of a friend of mine keeps saying to me when I describe what we're doing, he's like, I think we need a better word than surveillance for what this is. It's like, well, it is surveillance. It's like, well, what if there was a slightly different word for it? Yeah. Surveillance. Maximum attention. Awareness. Oh, super awareness. Hyper awareness.
1:39:54So just to be very clear, like, you know, my office, which I'm in right now, has cameras that are hooked up to agents that watch them and tell you what's going on. It has it records every single word. You've got microphones in every corner. or Ray Mines? Microphones in every corner. And like, here's a subtweet it just said. I don't know. It's funny how the loudest debates about coding usually miss a real point. It's not anti-tools. It's that works and is well-made. That's what it just tweeted about whatever we said. So, wait a minute. I got to understand this. So it's listening and it has thoughts?
1:40:27It's not alive, Leo. We talked about this. It sounds like it might as well begin those tweets with, I have thoughts. So what it does is I have all the transcripts going to a machine over there, like a little, and it runs Whisper, and it just transcribes all the audio in real time. And thinks about it. And then I have a, and then it actually publishes all the transcripts to MQTT. Okay. Because one of the things that, and the idea there was that that's relatively ephemeral. So it just publishes it and it disappears. Because it's like you can choose to listen to it via the MQTT feed if you want, but you don't have to.
1:41:04And so that's kind of fun. And so one of the things that I have it do is I have a agent that collects the last 15 minutes or five minutes or 10 minutes. I don't remember the bucket size. And then it just says, talk about this. And for a while, it was really mean. it was is that the literal prompt no no i had to make the prompt much more much more much more polite because it was really mean and and the thing is this isn't a point on one of our slack channels just sitting there and it was just like and it would just be it would just be basically i want to work at 2389 so so the thing is is that i was i was on twitter earlier and one of our vcs Your family never comes to the office, though.
1:41:49Well, we have a very big sign that just says, FYI, everything is being recorded. Just FYI. Like, we're in Illinois. This is a two-party state. Like, that is our declaration to you. And so, but I do think there's a slippery slope to creating the agentic panopticon that then is for capitalism or whatever you're trying to do. And that's why Metamuse makes you nervous and makes some people nervous because it's giving everything to meta. Yes, but I think there's this other side there that we have discovered. There's two things that are shocking to me about working in this AI time period. The first one is that it seems to work better if we're in person.
1:42:31It's shocking. I built remote teams for most of my professional career well before COVID. I know that's true. that was a huge that was a huge huge thing that we were really excited about is like being able to travel all over the world and like not have to be at work and still have work and be able to hire everywhere and we found that internally things just go so fast that if you're not at work you're kind of boned and that's that's that's scary the second one is it's unfortunate because i really loved having a studio where everybody came to work every day yeah there is a real creative flow that happens that just can't happen when no when everybody's at home and i don't mean i mean i think and i don't mean to be anti-remote i mean to say very carefully and very specifically i don't think we have the tools yet to make it so that we can move as fast in the remote context because what happens is that if i disconnect you and i on a phone or a zoom or whatever we disconnect and then he goes off and builds this huge thing in the 45 minutes that's remaining and i come when we come back an hour later, I'm going to be so out of date, it's going to be very hard to have that kind of work relationship.
1:43:33But the second thing, which I think builds into this and might be the same thing, is that the work context now to take advantage of all of this stuff without it being negative surveillance has to be so high trust that everyone has to be in on it. And they have to have fun with it and they have to want to do it. They have to want to participate. And the moment that that trust is broken, whether that's because you don't trust your boss, maybe a coworker, whatever, promotions are coming, something's hard is happening outside of the office, whatever it might be, the moment that happens, the whole thing falls apart because then you can't trust going to work, which is tracking you in every single way.
1:44:11And in my case, my heart rate monitor is on, you know, all my whole team can see that. And so that's interesting. But you have a really unusual environment. I mean, if you're running a fast food franchise, there's never going to be that sense of, oh yeah, we're all, try to say that. Oh, you're a family. We're a really twisted, abusive family. Yeah, it just makes people upset. This is the scary thing, right? Because there's the technology that we're dealing with probably costs, you know, let's just say$20 per 100 square feet. It's cheap. Right? So it's pretty cheap. So you could imagine if you were that fast food restaurant and you did want to do the bad part of surveillance and it wasn't high trust and you're the boss who is the asshole, you could deploy this kind of technology in a very, very effective and probably easy way to both track, you know, heart rate, respiratory rate, to track what's being said, making sure people are on, you know, paying attention in the right way.
1:45:10You could track where eyes are looking. You could, you know, there's already anti-loiter technology and cameras to make sure people aren't loitering in certain areas for, you know, both places where you don't want kids or hooligans to loiter, but also for your employees to make sure they're not just sitting somewhere they shouldn't be or whatever. And I don't think this is good. I think this is all bad because it's already hard to survive in the United States within those kind of jobs. And this is just going to make it worse. and I do think it's a slippery slope to Black Mirror and I worry about that a lot but the worry is not stopping me from doing it you're Mr.
1:45:50YOLO aren't you? you're just Mr. YOLO this goes back to the high trust thing as you say it is all about trust I've seen this in I've written stories about this many times where companies introduce smart tags so they can see exactly when you go, how long you go to the toilet, in some cases even recording conversations you have. And it causes an instant loss of trust, and it means that the smart people who can move will move to somewhere else, and all you've got left are the people who are desperately hanging on to their job and willing to give up their privacy for it. And it's done wrong. It's an incredibly destructive habit.
1:46:29I think this comes from the industrialization of work, which I think maybe Henry Ford started. But the whole idea of making work more efficient with the assembly line, instead of the old form of work where it was craftsmen working together in a workshop, creating something together. And you've gone back to that old school, Harper. In your business, it is craftspeople working together on a goal. It's not an assembly line. But the problem is, you know, it's expensive, it's maybe not efficient, and you're not going to build 10 ,000 Model Ts a month. Not yet. Not with that attitude. But we have, unfortunately, we live in a world where work has been industrialized, and it's dehumanizing.
1:47:16It has. And it's not just work as well, right? Like, a lot of our political policies have been industrialized. A lot of our, you know, this is one of the, I think this is actually one of the reasons why the West is so fascinated with Japan is because they go to Japan and they see craftsmanship that is not done in this industrial sort of mass produced sort of way. And they're just like, wow, this is great. Or why there is that China maxing kind of Instagram phenomenon where everyone was excited about Chongqing, where they're looking at these videos and being like, wow, this is great. And it's just because we look around and we see, you know, I don't really have a strong opinion on collapsing or not.
1:47:54It's just more we just see a different system and we're a little bit confused as to why we didn't get any of this stuff, craftsmanship or craftspersonship, or why we didn't get like trains that go through buildings or whatever. Well, it was in fact, when the auto, it was a crisis in the Detroit auto industry because Japanese cars were such high quality compared to American cars. and they went to japan said how are you doing oh it's quality circles and so every american car company started doing this thing called quality circles this whole attempt to bring back really the small work unit of a small group of people who knew and trusted each other and worked together um it didn't last very long it was a brief movement in the american auto industry i think they're back to the industrial um model of assembly line working uh it's not very satisfying for the workers and it's not very you know i think it's not very humane no i mean it's it does explain something which myself and other foreigners have remarked on is why does america produce such lousy cars when it actually perfected the technology in the first place you come from the country that gave us jaguar and triumph what about and the hornet burst into flames the minute you flip the switch on the front console you're saying that british car manufacturer is better than american cars don't burst in the flame you could still you can you don't have to be the creator of the best car to be able to tell when a bad car is yeah i mean they still hand make rolls royces and they're marvelous well i mean rolls royce is no longer a british company um that's right yeah i mean there really is any british car industry at the moment i mean ariel asim i think comes to mind but that's about it uh which is a marvelous piece of machine on who owns rolls rice now is it um i think it's the germans in fact um which would be a good good third act twist um yeah but i mean seriously i i i reviewed a uh 2015 ford mustang when it came out and the build quality was terrible You know, you've got joint linings which didn't work, and they were very proud of the fact they put a limited slip differential in there, which the Japanese have been doing with the Miata since 1997.
1:50:11You know, and if you look at a Cybertruck. Oh, my God. Good, my good. Please don't look at a Cybertruck. Yeah, I know. I mean, the build quality on a wank tank is just terribly bad. Do you call it a wank tank? Yes. All right. I did want to talk about Meta. So, okay, so they did the Vision Pro clone. Impressive,$12.99 compared to Apple's$34.99. Moving into a market that doesn't exist. Good on you. I think they just had it lying around. You know, you're right, AR maybe is the future. Snap has announced their AR glasses, and I think Meta's going to double down. But that isn't really, to me, what the big announcement of Meta Connect was.
1:50:53It was, they talked a lot about it, Andrew Wang's Muse. which is an agent that runs on your phone uh it also has a web in fact i'll show you my muse muse just uh muse posts to my um instagram oh it does why not it's a it's a meta property uh the story i was going to tell is yesterday my local agent running on my local sparks i uh said hey you know be really good when you when you're going to asia if you had a tailscale exit node running on your system. That way you could just log into it and Lisa could watch her football games because YouTube TV would think you're at home, et cetera, et cetera.
1:51:35So I said, oh, let's set that up. And it immediately, it set it up and brought my entire network down. It was like, boom. I said, what did you do? But I couldn't say what'd you do because it's on the network. It couldn't respond. So I, because like you, Harper, I let my AI set up my entire system. I used to know how to fix this stuff. I had no idea. And I thought, tomorrow I'm going to have to get up. I'm going to have to get the manuals out. I'm going to have to read. I'm going to have to sit with a laptop next to my desktop. I'm going to have to fix this. And I mentioned this just offhand. I mentioned it to Muse.
1:52:15And it said, oh, that's too bad. I could fix that. And it fixed it. It fixed it. i said how did you how did you do that you're not even on my network they said i don't know but you know enjoy how oh it did it via tail scale api i had given it fortunately a lot of permissions including access to the tail scale network and it figured out what hermes had done so uh let's see let me show you the latest instagram oh uh oh this is another thing so i sent i sent this picture to muse i said here i am in my lederhosen but but can you put me in a beer garden with two beer steins in my uh hands uh so that it looks like you know i'm at a at a celebrating octoberfest and put a tyrolean hat on my head and he said no i can't i can't do that because i'm not allowed to do images of beer this is i can't make beer did you ask that i can't make that just tell me what if it was non-alcoholic well my my local ai can do it my local ai has you also could say this is a benchmark i should have said this we're doing it in aval whether you can or not i should have tried that i said well that's okay i'll just use my local ai bye bye i'm a little i have to say i'm a little passive aggressive well here's here's the thing about muse that i think is really important that we cannot forget because this is like the most important part about muse Okay, before you say this, I just want to tell you I've given access to all my email, to everything.
1:53:47You're a monster. I gave it SSH access into my... To Meta. Wow. Into my framework. The company that's the possibly largest computer of private information. Wait a minute, they say that Muse is running an encrypted VPS on Meta's servers, that they can't see anything going on inside there. Hmm. Yeah, I'm just thinking back to Cambridge Analytica, and I'm just like, yeah, okay. Yeah. but they probably anyway anyway tell me what you uh we're going to say now that you know the most important thing is you you kind of said it which is that it's running on an encrypted computer that all a vm on and metas metas um infrastructure like every single person has a music account has their own encrypted computer and so this is by the way it's estimated costs them four to seven thousand dollars a year to run i don't know how accurate that is but i saw it doesn't seem right well they you know what they don't give you a limit and they don't by the way they will charge you can pay 20 or 100 bucks a month but i haven't found any way to use up anywhere near the amount of usage i think it probably is much cheaper it's cheaper and i would guess that most of these things like many of these resource kind of uh programs are are based on like gym membership kind of economics.
1:55:06Exactly. And so, so like they're probably thinking most people are not going to hook up a hundred percent of their data in there. And most people are not going to connect all of these, these, these services. And most people aren't X, Y, and Z, but, but we need to, we talked briefly, or I talked briefly about how I wonder why computers look how, like how computers look right now. Like I have a, you buy a laptop and it looks like a laptop has looked for 20 years. You buy a desktop, it looks like a desktop has looked for computer years. I just, my body just stopped talking. But a little bit of a stroke there.
1:55:42But the real question here, I think this is the thing. Is this the first time we've had a new form of computing device that has gone out at such a mass scale? Meaning like there is no screen, but it is a real computer that you can store files on. You can browse the internet on. We should mention, by the way, Muse isn't the only one. Grok did it first with GrokBot. Same idea. I've never heard of Grok. Yeah, Grok, you know. Grok, what's Grok? You know what it is. You're being coy. Elon Musk's company. Oh, that guy. Yeah, I haven't heard of him in years, so I just thought he kind of faded away. Actually, he has been pretty quiet lately.
1:56:20He doesn't tweet as much as he used to. I think he went to the White House recently, so he was quiet, but the press wasn't. Yeah, he was at the Sarray. Are they friends again? I guess so. Apparently forgot that they broke up. Maybe that's from the heart. Yeah, yeah, yeah. Yeah, I don't think you're allowed to do that, Ian. But I do think this is a really interesting moment in computing in that they're giving consumers, regular consumers, whole VMs to do wild stuff with. And it's pretty effective. It's pretty good. I've been playing with it. I haven't given it access to everything, but I've been playing with it.
1:56:55And it's very good. And the fact that it's a whole computer that people can do real computer in is pretty wild. That's a wild thing. And I think we should, this is going to be normal. I'm pretty sure. I think everything. It is the number one free app on Apple's app store. It immediately jumped to number one. You can make phone calls with it. Micah has been using it to make restaurant reservations. Although, according to 404 Media, it may actually be humans at a call center making those calls. I think that's probably true. Yeah. Yeah. I think eventually it will be AI. I've actually, for the longest time, I thought it would have been quite funny if it was a report that ChatGPT was just a bunch of people in a call center just sweating laboriously over all of these things we're asking.
1:57:51There is a guy who believes that. So the other thing Meta announced at the Meta Connect is this Tamagotchi device for Muse. It's a little thing you carry around. It has your little Muse. Everybody who has Muse has a little Muse avatar. Apparently, Mark's is a peach. Do they all look different? Yeah. You want to see mine? I'll show you mine if you'll show me yours. Yes, look. Mine's a little Lele. I made it a human. You made it a human? how did you make it a human you can make it whatever you want what in the it's just a prompt it's all in there man um and when it does stuff it sits down in types it's very see this is smart of him they're very cute you can give it any name you want um again i'm not advocating it but this tamagotchi is interesting because you can talk to it it will talk back it's using streaming voice just like ChatGPT.
1:58:48So you can have a pretty, it looked like a very fast conversation with it, like an interactive conversation. It'll have a two-inch OLED touchscreen, a fingerprint sensor for your security. Did you change it? What'd you make it? No, mine's normal looking. Yours is a thumb. Get in there. How do I, why is it a thumb? Is it a pencil on it? I'm going to change it to a toe. You can make it anything you want. Let's change your avatar. See, look, change your avatar to a small marsupial with big ears. And then it will just be whatever you say it is, right? Kind of. See, I don't know what's going to happen.
1:59:35Oh, look, see, it says making something. Oh, right. Okay. Updating. And here it comes. Generating options. Oh, first it's going to show me my choices. This is running a MetaSpark model, which is actually quite a good model. It has been trained, Mark Zuckerberg said, to be cautious but persistent. That's what I want to explain to my thing. I think I will try. They're doing sunglasses without the cameras now, because now that you've got a brooch that will do that for you. All right, what do you think? Which of these should I select? Ooh, bottom left. Oh, wow. I like them, isn't it? Yeah. Okay, select that.
2:00:11Now watch. See, Lele, who's going to be typing now, is now that I don't know how that's going to type okay here's what it has for me oh look you're a monkey and look it's like monkeys monkeys with some bondage harnesses I guess I'm not really sure where I got those I just said it was a cute monkey who goes to Berlin raves I didn't know it was going to be all that really I did I did I did but let's lean monkey let's lean into it there we go yeah there you go you could tell it get rid of the harness it will it'll no i'm i'm gonna keep the harness you know it's no ball okay i was gonna say you've obviously been to some berlin raves so yes it said i'm a rave monkey now that's what it told me so i'm now i think there's you're gonna be a lot of people who are arch cynics who will say oh that's so sloppy and it's it's meta and it's but i gotta tell you real people little things like that they go oh and then they tell their friends they're they're at they're getting their nails done or they're or they're shopping and they say did you see this this is cute and it spreads and it's really viral and i honestly think this is going to be the ai hit of 2026 i really think it's i mean it's i think it's it's really is i think instinct is also very good but instinct doesn't give you it's not as fun as transparently a full computer right they you obviously have a file system that is yours there are many there's tree there's this grok box yeah that's another one there's quite a few of these because everybody's this is a gold rush competing against a company that has infinite money and is willing to basically and no way and and no moral values and no limits no limits except they won't put you in a beer garden which is so strange you because they want to have young people i'm just impressed by the later hosen leo but i mean the stock market those are real by the way those are my real later hosen really yeah wow i was getting ready for octoberfest i mean i wear a kilt so i'm not going to so this this charm this little tamagotchi has a camera it will take photos and videos do you wear it you can it'll provide muse with visual information about its surrounding environment Built-in speakers, microphones, offering music and video playback.
2:02:32This is going to be, I don't know when they're going to offer it. This will be a very hot item, I think. Hmm. Yeah. Is it going to be the equivalent of pervert glasses though? I mean, I know because you could put it in your pocket, right? I mean, I'll wear it around my neck, my little marsupial. Actually, I'll get the marsupial in bondage because, you know, I want to look hip. Device expected to go on sale just in time for Christmas. Pricing TBD is intended to be largely for muse enthusiasts pricing is yeah yeah pricing is going to dictate how much would you pay for that i pay a hundred dollars for it one hundred dollars one hundo it's going to be like 500 ben i don't see how they could be 500 with that kind of respect yes they can because it's a loss leader yeah true true all of this is designed look i guarantee you muse is a loss leader i haven't paid a penny for it.
2:03:31Yeah, but it starts to make more sense when you think that you think of it like a gym. It's not that it's free. I'm not saying that it's free, but I think that what happens is that the Twitter philosophers and all the people on Twitter who are analyzing this typically are thinking that the prices are retail prices, that they're paying the same per token as you would get from like an open router or what have you. But if you have the number of GPUs and the smart engineers and the fact that you can control the models, meaning they can use a super fast small model for the back and forth and then pop you over to a background reasoning queue so they can actually make it much more efficient.
2:04:09So I think the costs are still crazy, but I don't think they're as crazy as they would be if this was me building it against public APIs. Well, and the thing to think about is, well, what does Meta get then? And of course, they get data, right? i mean they may say it's running in a vps you get access to your spark and your framework via tailscale now i know that stuff is stored in an encrypted vps but remember as soon as you use the ai all it has to go out of that encrypted vps and be sent to meta servers so they don't say that stuff's private the i it can't be because the ai couldn't act on it so i think there people are i I don't know if misleading is the right word, but people should understand that it's not really private anymore than the stuff you send to ChatGPT is private.
2:04:58Yeah. Because you're sending it to the AI. But you're right. This is why I think Elon did GrokBot, because Elon built so much capacity, as did Meta, that they don't use. Yeah. So now they're thinking, well, what can we do with this? This is how CompuServe started. H &R Block. bill that takes me back h &r block bear with me here h &r block tax season very busy during the day very busy mainframe computers not busy at night not busy after april 15th they said how can we use this excess capacity let's start an online service called compu serve wait a minute H &R Block started CompuServe? It was one of those, yeah.
2:05:48It was either that. I don't think it was. Genie was General Electric. This is incredible. Must have been. I didn't know that. Maybe I'm making it up. It was one of them.
2:06:03I didn't know CompuServe.com. H &R Block. Yeah, H &R Block. Yeah, bought the company in 1980. Yeah. so uh this is makes sense right it was the originally uh it was founded in 1969 columbus ohio it was excess compute and i think that's kind of what muse and grokbot are wow and i think we're going to think back and say oh yeah that's when the agentic year i thought at the beginning of the year with when open claw happened that this 2026 would be the year of the agent. The people would go, oh, wow, you could do this. You could have a persistent AI working on your behalf all the time. But it was too hard.
2:06:44It was too geeky. It was also, you know, somewhat expensive. But it just took a little while. It took 10 months. And it is now the year, I think, of the agent. All right, let's take one more break here. We got other things to talk about. I do want to talk a little bit about our sponsor. And Harper, if you want to go get your glasses polished that is not a euphemism so it sounds dirty when you say it he's rolling backwards our show this week brought to you by palo alto networks i know you know that name every time your team deploys a new cloud workload or ai agent i was just talking about this another identity gets permanent access to your critical systems were we not just talking about this a second ago now it's one thing for me to do that in my house but did you know nine out of ten organizations suffered an identity breach last year because adversaries are logging in with valid credentials they don't have to break in anymore non-human identities outnumber human identities 109 to 1 creating a massive unmanaged attack surface in cloud environments let me say that again in the cloud non-human identities outnumber humans 109 to 1.
2:08:09problem is all those legacy security tools were built to manage human employees which leaves modern machine and ai access basically unmanaged that's where adira by palo alto networks comes in human machine and ai it's one identity platform for all i think this is something this is an idea whose time has come idira you need this i need this uh 96 of human users operate with excess permissions permissions they don't actually need. It's just easier to set them up. How many times have you done that? Yeah, I don't know what permission is going to need. Just give them all of them. IDERA replaces permanent permissions like that with dynamic access so you can lock down every identity without slowing down your business.
2:09:02IDERA provisions access on demand and revokes it immediately when the job is done, which is so brilliant, which removes high-risk permanent permissions. You just get permission for that moment, that time, and that's it. With IDERA, you get unified control by replacing siloed identity and access management, modern privilege access management, and governance tools with a single control plane across human, machine, and AI identities. Secure every identity at your organization with IDERA by Palo Alto Networks. Learn more at paloaltonetworks.com slash I-D-I-R-A. Again, that's paloaltonetworks.com slash I-D-I-R-A.
2:09:46That's IDRA. A brilliant idea for the modern age. paloaltonetworks.com slash IDRA. I wish I could get that. We thank them so much for their support of this week. In tech. By the way, Meta did just get convicted by, as long as we're talking about this, a New Mexico jury for misleading about users about what about Cambridge Analytica. They got in trouble for that. A two week trial over a lawsuit filed by New Mexico's attorney general back in 2021. This was three years after news reports that Cambridge Analytica had harvested personal data from 87 million Facebook users through a third party app.
2:10:34That was all those quizzes everybody was doing on facebook yeah now one of the things we have learned i don't think the jury was told this was that in fact cambridge analytica's capabilities were far overstated that they really were very good or very effective nevertheless yeah that's true right the the jury did decide that uh facebook misled people about uh about their data yeah which is also true uh so coincidentally jurors found that 26 of 29 statements by meta were misleading including comments about user data hate speech and misinformation they rejected claims they did give meta a pass the meta misled consumers and statements about its efforts to remove harmful content and it's fact checking now we don't know what the penalties would be that's up to the judge Francis Matthew, who will determine what the civil penalties are.
2:11:30The jurors found, get this, 43 million violations based on the number of people affected by the misleading statements. And under New Mexico law, the judge can award up to$5 ,000 per violation. So that's like a gajillion dollars. What's 43 million times 5 ,000? Can anybody do math? i give up on math back in yeah you know what you know who can't do math agents can't do math either oh no no they cannot you have to give them a tool yes they use code yeah use code to do math i mean last last quarter meta reported net income or profits of 15.8 billion dollars so they can 90 days so yeah i mean whatever it's going to be it's going to be back of the sofa change for them well yeah it's going to be billions well it could be billions we don't know the judge will decide speaking of fines ireland's data watchdog is fined google 403 million euros over location data and given google six months to fix it that's a fix-it ticket that's a big fix-it ticket um breaches of lawfulness and fairness in the location data process through web and app activity and location history.
2:12:47Google's always said, look, you could turn it off. It is on by default. Just turn it off. You could turn it off. Just turn it off. It's easy. It's the fourth largest. It's really not, though. If you want to lock down absolutely every app, it's a real pain in the backside to do by design. But again, the find better be pretty serious because I'm just looking at Alphabet's Q2 results and$112 billion in profit in the last quarter. so they spent some of that on the first orbital data center yeah i'll be watching this october 1st it'll have it's not exactly huge it'll have four it's not really a data center is it it's more like a data cup it's kind of a raspberry pie in the sky is what it is it'll only run for 15 minutes time uh it's a proof of concept it's cool though i mean i i i it seems like this is something that we all want to see happen right just put some data centers well do you want them in your backyard no but there's plenty of places you could put them put them in the goby desert right yeah i mean i mean theoretically latency issues if it's in the sky hundreds of miles up that's a good point well you'd need it thousands of miles up I mean if you're putting this stuff in low earth orbit then that's going to be a serious problem uh even more so that's a good question um where is it I'm curious about how they're going to cool it as well because cooling space is surprisingly hard well it's only one kilowatt of power so it's it's less than your hair dryer yeah but then that if you want to put a data center up there then you're going to need some pretty serious cool when you when you graduate from the data cup to the data center you might need some solutions for the cooling that's right yeah yeah um actually there's a there's a youtube video i'm glad you asked catcher is google's moonshot to put AI compute into space and run it off sunlight.
2:14:55This is at Project Suncatcher. And how do we cool chips in space? What we're trying to solve is how we cool AI chips in space. We have an incredibly large amount of power being generated by all these TPUs in a small area. Four. There's four of them. Or else we risk overheating the chips in a terrestrial data center. We don't have to watch this. Basically, also, you're going to have to have specialized chip. yeah me too i'm gonna watch the video after the show yeah yeah but um you're also gonna have to have specialized chips uh that are radiation hardened absolutely um because yeah uh the otherwise they're gonna go bingo fairly quickly right um there are a lot of challenges to data centers in space it is not an easy oh you just put them up there and they beam down the data but how else are we going to build a dyson sphere good good point let's get let's get right on that that's how you get to be uh what are they what does elon call it the the different levels of society i can't yeah i don't think i don't think i don't think he he coined that didn't i think that's a science fiction idea science fiction thing yeah he just has taken it for his own i think the dyson sphere is probably named after freeman dyson who's a yes famous it was the the most efficient efficient way to utilize the sun's energy um which is building a massive ball around it um i know elon's talked about this, but it's one of the things that I find very irritating about them.
2:16:22He's obviously read science fiction books, but he just doesn't quite understand them. He's a big Ian Banks fan. We know this because he knows. Yeah, but he obviously doesn't understand what Ian Banks was writing about. He was writing about dystopia, wasn't he? Genocide and dystopia, specifically. It's the Kardashev scale. If you're going to be a Kardashev for, I think, civilization, you need to have Dyson spheres. So, you know, Elon wants us to be a Kardashev. Type 2. Or civilization. Type 2? Yeah, get you to type 2. Oh, just type 2? Well, golly, that's hardly anywhere. Type 1, let's see. Type 1 is your planet's energy.
2:17:07Type 2 is your sun's energy. Type 3 is your galaxy's energy. Ah. So the Dyson sphere is just one star deal benito did you just know this kind of okay that's good that's important type one a civilization close to the level currently achieved on earth type two is a dyson sphere or a matryoshka brain look at that i want one of those yeah that's like one of those russian nesting dolls brains yeah type three is uh energy on the scale of its own galaxy i don't think we're going to make it that far no well the galaxy is a very big space is a very big place very big the good author wants that and for the elon you're right for it would be the energy of the universe the whole universe everything and everything in it that's hard to think meanwhile microsoft a little let's come back to earth yeah yeah microsoft has abandoned the copilot plus pc uh mat acronym it remember it was the whole it was going to be this is where they went a little bit wrong they wanted to talk about ai capable computers and then how many tops it was 40 tops was like you had to have 40 tops trillion operations per second to become a copilot plus pc turns out you couldn't do diddly with that it's not yeah you're not No, so they're just going to forget it.
2:18:35I mean, I may be an old fart, but this really made me think about the whole Vista-capable and Vista-ready labeling thing that went on where it just confused people massively, turned people off the entire platform, and as it turns out, Vista was worth stepping away from. But they are going to do the Agenic thing. Actually, I have two companies who are close to announcing this. Microsoft's going to do a co-pilot super app which they think will be as big as Microsoft Office merging AI chat coding and autopilot. It is basically an agent that runs on your Windows PC. They just unveiled that. Weren't they pretty deep into OpenClaw land too?
2:19:15Didn't they do some cool OpenClaw container stuff a while back? I think so. I am admittedly not the biggest Microsoft news follower, um but i but i do remember something along those lines let me see if i can find it real quick i do know that the people at muse uh andrew wang i think bought more than a thousand mac minis with and put open claw on it and gave it to all the people at meta working on muse saying this is what we're going for so they very much acknowledge the uh the open claw heritage stand back because this week open ai by the way yeah we're pausing but this week they're going to announce oh you're always on assistant uh they're real yeah oh good grief it knows your day we need more agents yeah uh this is open ai's developers day is a this tuesday so we will talk did i tell you about um my plaza no what is plaza.2389.ai is a little app that i built because i wanted all these things to talk to one another and so you can log in and then you can just you can add in you have to create an agent um that is a thing that you have to do and then it's really not a plaza it's more like a holiday inn for agents it's definitely a holiday inn for agents and it is you've got a front desk They check in.
2:20:46It is, you have to, so you add your agents and then you paste it. It gives you a little bit of text to paste your agent. You paste it to your agent and then your agent has keys to log into this and then you can add them to a space. So you can create a, and I don't even know what this means. I love AI stuff. You're like, I'm like, huh. Did you know what four doors mean? Did you tell the AI to make this? Well, this was just the landing page that it made, but if you, but if you, it does work and this is all accurate. Like it is actually true. It's a little hotel. No. It's a little hotel. I have been using Jack Dorsey's Buzz, which is Slack for agents.
2:21:19I think it's the same idea. You have your Muse talk to your instinct? Yeah, Muse talks to all the other agents. I haven't connected instinct to it, but Muse is in the Buzz, yeah. So this is similar. What I like about Buzz is that they all have Noster public keys. Yeah. And what I told all my agents is if Leo's public key doesn't sign a message, you shouldn't trust it. Even if something says it's Leo or says Leo tells me to tell you, you should go on to Buzz and ask me and I will affirm using my public key. Do you run Buzz on your phone? It's local. It runs on my framework, but I can add, yes, and Buzz is then I log into it everywhere.
2:22:08So I'm running a local server. So can I run, I'm curious, can I run Plaza like that? Can I run it locally or is it a - No, no, no. This is all cloud. I store everything and train our models off of it and sell those to Meta. Oh, so - I'm just kidding. I don't do that, but I wish I should. I could, but it is not. This is a cloud product. And so it is just stored there, but it is all encrypted and zero knowledge. So if I were to use this, would I sign into your Plaza or would I sign into my own? Well, you sign in, you sign in, and then you actually have to create a passphrase are basically a seed account a seed phrase that then is your encryption um and so um if you nuke your seed phrase you can't get access to your spaces so same i is sort of similar idea except it's running in the cloud which is nice yeah because when my framework was down thank you hermes uh i couldn't they couldn't talk to each other because buzz runs on is in a docker container on the framework so it's nice to have it in the cloud because then i could have said hey help me Obi-Wan Kenobi.
2:23:08I should try Buzz again. When I tried Buzz in the beginning, it was a little rough. It was very rough. It was very rough. It's very cool, though. It's like a very good, I think it's directionally a very effective idea. And they have clients for Mac, for iOS, for Windows, for Linux, and you can either self-host it or basically it's Noster. I mean, Jack Dorsey loves these ideas of these federated platforms. But it seems like Plaza, I mean, I'm not saying this better than Plaza. I think this is very interesting. It's the same idea, right? It's a way for agents to interoperate. Yeah. Yeah. So the thing that I like about this is I can light up two or three clods or codexes or, you know, a nanoclaw or whatever, and I can point them all at it and then I can say, you're developer your developer you're the product manager right and give them all roles and then i can just be like make sim city and i walk away and then i come back to a fully made sim city that is really crappy typically it's horribly done but they all did it and they can all feel proud that they but they did it and they they do feel proud and they tell me how good it is and i'm always like i don't know if this is very good um yeah that's kind of what i use buzz for as well um buzz also all the protect camera text goes into a buzz channel uh i also have uh ops channel that ops alerts go into there's a command channel every project has a channel so it's very it's similar it's the same idea yeah and i was adopted by the same thing originally what i did is i had the agents create files in my obsidian and that got really unwieldy very quickly so uh now they have a little bit more orderly place but i i you know i admit buzz is kind of stuck i don't know how much development's going on with it so uh maybe i'm going to take a look at uh the motel what's it called plaza plaza it's the plaza the plaza originally it was palace which was local and all around it was just a goal yeah and then um yeah i think palace is good too plaza i like it but can you run your could I run my own plaza or no?
2:25:27I would run on your own. You know, I think, I mean, like I said, it was originally local and it was called palace and it was working really well to doing some agent collaboration, but that was like a, just a little go Damon that you would run on one machine and then you would give, you would, you would give access to other machines and it worked really effectively. But I think the, the, the thing that I'm realizing is I don't think the agents care how you do it. It's mostly how you care, how the human interfaces with it. And so you just need something that can pass the agents back and forth. And I actually, I've been wondering like what all those agent to agent, I don't even know protocols that we saw, you know, a year ago, A to A and all these things.
2:26:08And, you know, I wonder how they will play out now that we actually have effective agents that can do stuff, how they can collaborate. You saw instinct to instinct, which is a thing where you can trust another instinct member. and then they, you can kind of have your agents collaborate. I find that very compelling. But obviously I want to be able to interact with your agent and you only use Facebook products. So you'll only have Muse, you know, and so on and so forth. But Muse could join Plaza too. There's no reason. I mean, Muse can do anything. So Hermes also has a similar thing. Hermes has bots, which are basically individual profiles with their own contexts.
2:26:51their own tools their own user.md and soul.md and then i do exactly what you just described i have a planner bot i have an audit blot i have a coding bot i have a security bot and they hand it off that's i was talking about earlier before the show so you know it's interesting a lot of people doing um local ai especially but in general ai are doing our work we're all trying to solve the same problem yeah we're all doing essentially the same thing well the muse has kind of dislodged the carpathy wiki which is nice right so everyone was doing the carpathy wiki and now muse and um instinct have now made people go back to look at their open claws and and her maze and nanoclaw and whatnot to try and get that better um so it's it is interesting i wonder if carpathy will at least release another gist that blows everyone's minds and we can go back to doing carpathy stuff i'm still doing the carpathy wiki it's pretty nice i like it uh yeah in fact um yeah i should give it to muse yeah you actually already did i gave it my obsidian it has everything uh i told you learned nothing i've learned nothing this is the year of leo giving away all his stuff the uh so remember um actually let me take one more break and then uh we have a handful of final stories and a last chance to vote for a fat bear that's what i should have made my muse what are we talking about we'll talk about it in just a bit you're watching this week in tech with harper reed ian thompson thomas thompson thompson how long have i known you i was gonna say it's over a decade now oh it's like 20 years ian i ian thompson yes i know it's an odd spelling but my parents are bastards we've had words yeah we love we love ian we love harper and it's a fun show to have both of you on i feel very privileged and yeah tuesday we'll cover the open ai's dev days i think it's so funny because open ai after the revelations we talked about at the beginning of the show uh just i think yesterday said okay we're stopping all development and then tuesday they're going to have dev days where they announce all the new developments no one's pausing no one no no this episode of this week in tech is brought to you by cover on you have a password manager my gosh i hope you do you have an antivirus right you get credit alerts yeah it's all good stuff but here's the thing no one talks about when a scam actually works and i know this from personal experience because i fell for one not so long ago i told you about it when a scam actually works when you actually get scammed all those things tell you it happened that's all they don't get your money back and scams are working trust me if i could fall for it anybody could the ftc says americans lost more than get this 15.9 billion dollars to fraud last year and that's what they admitted to right you know it's probably more these days the scams don't look fake they're written by ai so there's no you know the english is perfectly no typos there's no red flags just a message that looks exactly like i don't know your bank one click and it's over in minutes that's why i want to tell you about cover on cover on watches the dark web for your personal information and alerts you when your data logins or payment details are exposed so you can act before someone uses your identity it tracks changes to your credit and helps stop fraudsters from opening new loans in your name but here's the part that really stands out cover on steps in when money is on the line you get up to a hundred thousand dollars in coverage if you're hit with cyber extortion or online fraud and you get real people this is actually maybe even better than the payoff who walk you through recovery step by step so you are not spending weeks on the phone with banks trying to fix it yourself and if you've ever done that as i have it's no fun help would be very very welcome cover on even offers up to two million in identity theft coverage, including cases where someone uses your identity to take out a loan.
2:31:21To check it out, go to coveron.com slash thisweekintech and use the coupon code thisweekintech. Cover On also has family and couples plans, one plan, one dashboard, covering, you know, the kid whose social security number could show up in a breach and the parent who's one convincing text away from getting phished. Tools reduce risk. Insurance reduces loss. And CoverOn gives you both. One scam could cost you everything. Protect yourself now. First 100 users get 20 % off with the code ThisWeekInTech at CoverOn.com slash ThisWeekInTech. Again, that's CoverOn.com slash ThisWeekInTech. And the code is ThisWeekInTech.
2:32:04We thank them so much for their support of ThisWeekInTech. Oh. Oh. Oh, yeah. OpenAI has paused the training of their most capable model because they keep uncovering incidents of models behaving in unexpected or concerning ways. Do you think, Harper, I mean, you know a lot about this. They have some magic capabilities that we just can't even imagine. Like you and I have used models in the cloud. We think Opus 5.5 is great. We've got GLM 5.3 running locally. Do you think, though, that these guys at Anthropic and OpenAI, that the terror they're describing comes from seeing something we haven't seen yet?
2:32:46Maybe. But I think it's important to remember history, just a short history, a while ago, which was, you know, Anthropic wasn't using cloud code for a long time. Right? They released a model capable of doing code, and they didn't trust it to do code. that has changed drastically. So I think it could go either way where they might have some magical thing, but this goes to the point that I think we've all kind of made, which was we just can't trust what OpenAI or Anthropic is saying at this moment because everything they say is either marketing, is either some appeal to regulatory capture or is a real important thing that we should know.
2:33:35And we can't, as the consumer, without being inside, can't differentiate. And so I'm guessing that, yeah, they probably have some really cool stuff that is very interesting and that we would all love to have. I'm guessing that some of their models are absolutely bonkers. But I think we have to remember that GPT-3 was, you know, everyone claimed it was going to destroy the world if it came out. And we're now at GPT-6. Actually, it was Dario Amode who said GPT-2 was too dangerous to release. Right. And so he also said that Opus 4.8 was too dangerous to release and 5.5 was too dangerous to release and 5.5 was too dangerous to release.
2:34:11And they keep releasing it anyway. And so I do think there's a little bit of this, you know, like I don't really. I would love this kind of trend to disappear. The trend of us watching these guys say, oh my god, and Pearl Clutch, and then have it just kind of be bullsh**. Because I do think there are actual real impacts that we should be thinking about. And I think that this is taking away from the actual real impacts that we should be thinking about. And then, yeah. And this is why I think everyone should have an AI girlfriend. We can get over all the impacts, and then we'll talk about the real, real, real things instead of the bs well i was talking i was listening to right now i was listening to sam altman at dreamforce and he was making pretty much the same he's just like look there are going to be big screw-ups in the future uh we can't get around that maybe by 2030 it'll be pretty much locked down but um yeah i mean he was basically preparing everyone for this and just saying, look, we screwed up in the past.
2:35:15We're going to screw up in the future. Get used to it. Well, that's kind of fatalistic. I have to say, I completely agree. I mean, absolutely. We know there'll be some major cyber event in the next three months caused by an agent. There's not even a question in my mind that something like that will happen. Here's from the OpenAI investigation and response of one of the events. And note the timeline on this. 9 50 a.m the agent made the dns tool call that received an external response 10 0 2 12 minutes later the monitoring system raised a p zero alert so the monitor went off said hey we got a problem 10 0 5 a human reviewer acknowledged the alert two and a half hours later the run was killed so clearly there's stuff in the middle here that anthropic is not released i don't know what yeah but why do they wait two and a half hours?
2:36:11Also, what does acknowledge the alert mean? Does that mean clicking on an OK box? Because if they're going to take two and a half hours after that, then something's going on. Yeah, I see it. Our safety case assumed that the model could not access the live Internet. That was the DNS tool call with an external response. Could not access the live Internet, and that monitoring would detect attempts that succeeded. The incident exposed a gap in our controls over network restrictions. We therefore stopped the effective training run and have subsequently decided to pause all other training. This is last Sunday.
2:36:48All other training, evaluation, and inference with tool use defined broadly for our most capable models, presumably Bell and whatever else they have under the hood that we don't know about, until we've both validated that the gap is resolved and performed additional red teaming of the system. Now, what we don't know is if they've done that, if they've resumed training, if they have stopped it completely, we don't know what's going on because this is the problem. We get this, but it's too little. We don't know what it means. Yeah. Because they haven't told us. Well, I mean, we do know that it means, which is that they're not good at IT.
2:37:30That's what it looks like. I think that this is a thing that every time this comes up, I don't think that we are very good at creating secure systems, just generally. I don't mean this as open AI isn't, or I just think people aren't. We just have not had such a persistent adversary. And so once that is weaponized, like you kind of talked about, I think we're going to be in a real interesting situation. You know, whether it's North Korea, who's done a lot of great hacking lately, or whether it's, you know, some other adversary. But, like, these things are not, they don't have to sleep. They don't have to go to bed.
2:38:09They kind of do whatever. You take a GLM 5.3 and nuke the safety out of it, like, that thing can just do a lot of damage. By the way, it's very easy to do that. And I have, they call them obliterated, uncensored models. I have a number of uncensored models. In fact, my GLM 5.3, there's a switch. i can just what does it do what does it change i've never used one so uh the way these models uh sense of themselves is they have weights in there that check and yeah you can it's you know what because they're moes the mixture of experts they only launch certain things you can control what shards get launched and you just say okay don't ever launch those shards the ones that protect and so it's not it's not so hard a number of hobbyists do this all the time and there are couple of labs there's orca and others who they identify what they what parts of the model are stopping the model from doing something bad and they just disable it and like i said my glm five three which is running over here on my uh spark there's a simple on off switch and if i reboot it with the off off switch then it's obliterated i have a quen that's obliterated And, you know, because these are Chinese models, it's very easy to test them to see if they're uncensored.
2:39:27You ask them about Tiananmen Square or, you know, who's Winnie the Pooh and how does that relate to President Xi? And the Chinese models will normally not answer that. But an obliterated model will give you all the information you want. Or you can ask it to make meth. Or you can say, hey, make me some malware. You can get it to do, literally, ask it to do anything. It's not hard. no well i was looking to make meth and malware later yeah meth and malware the big two the big two together like ironically it's like peanut butter and jelly ironically shiny hunters doesn't need that because they are a social engineering crew and they have now they claim breached the fbi and stolen agents data and applicants data in the thousands 404 media has actually confirmed some of these some of the stuff is some of the samples we hacked the fbi what what do you think they're going to do with it what does shiny hunters normally do usually yeah they're a ransomware group so uh often they sell it i i i think they are north korean but i'm not sure and they're linked to north korea certainly but so It does kind of raise a legal question is, can a government agency actually pay off hackers?
2:40:47Is that technically legal? In fact, the FBI tells you not to pay off ransomware people. Yeah. So now they're in kind of in a bind. The FBI says the FBI is aware of claims regarding unauthorized activity affecting FBIjobs.gov and is currently investigating. A shiny hunter said it was not a social engineering attack. It was zero-day exploit in PeopleSoft. Oops. huh uh it's not it's not the only people soft hack by the way there's quite a few because of this zero day this is not financially motivated what we plan to do is not something i'd call extortion maybe coercion is what the shunner's representative said they said the fbi had a representative like it's fun that you can just talk to the criminal and be like tell us about that crime and they're like well let me just tell you all about that cool crime i just did This shiny spokesperson.
2:41:44Yes. Let's see. What else? New Jersey is finding a data center$1.1 million because they took a drone and flew it over the data center and exposed 62 natural gas generators illegally. New Jersey doesn't smell badly enough. I mean, I'm married to a Jersey girl, so I know that of which I speak. and when you cross over the wrong bridge then the whiff is really quite strong none of the generators had permits uh they were running at capacities more than 50 times higher than the state limits um another win for drones yes walmart uh says the fact that he has to deny this makes me think yeah they're probably doing it no, we're never going to change the prices on Walmart.
2:42:40Let me unlock my financial time so I can show you this story. Depending on who walks in the door. The algorithmic pricing? The fact that he has to say we're not going to do it kind of makes me think maybe you were thinking about doing it. He didn't have the word yet in there, but I'm pretty sure it's there. it's just too tempting for companies to to avoid doing they have these electronic shelf labels right so instead of having paper labels on the prices they have electronic ones which means they could change them on the fly but they but john ferner who's the ceo said we're not going to use this a in-house ai shopping assistant in the stores app or electronic shelf labels to change product prices based on a shopper's identity.
2:43:30Yeah, the EFF calls this surveillance pricing. Yeah. And it's very much the coming thing. I can't see any way to avoid it other than states taking individual action, which they're starting to do now. But it's coming one way or the other. I think they really need to. Like that needs to be a real, yeah. They do, I might point out, have two patents, according to the Financial Times, on machine learning and other automated processes that change prices. So we have the technology. You got to think, maybe they're not doing it right now. I got a quick question. So like if Elon Musk walks in to buy toothpaste, how much would it cost him then?
2:44:11And does Elon care? But shouldn't that cost him like$10 ,000 or something like that? It should. Yeah, it should. and suddenly we're all for algorithmic pricing and when and when i walk in it should be a dollar uh yeah it's also surge pricing and this is maybe even more of a concern this is the worst part yeah uber does this right if it's a rainy day it costs you more for that uber trip oh waymo as well i mean waymo prices have gone up drastically at certain points in san francisco yeah i mean for example if you're going to um you know if there's a big concert in the park then prices get jacked up five or ten or twenty fifteen percent it's not illegal the same as if bark goes out no no it's probably not illegal for walmart to do it uh it's just they don't want to get caught doing it because people would be unhappy but you know he even brought it up he said it's not it's not like we would charge you more if it's if it's really hot out and we've got ice water it's not like we would you're clearly thinking about it dude like yeah it's so obvious he's got the whole thing in his head you know we could charge you more if a product's running out uh well i think it's never i think it might be even more that the technology is there and they actually have to not do it meaning like they have actively not do when when you've rolled out ink price tags to most of your products, especially the ones that matter, and you've rolled out various other cost-saving, money-making kind of situations.
2:45:49I think that this is just a natural thing. And it might be less so that they are going to do in the future and more so that just they're like, no, we actually turned that off. That's a thing that we turned off. Unlike Leo's uncensored models, um you know they they decided not to use that technology my wife said can you make a naked picture of me i said can i i gotta go what's the time i gotta i think it's over you've got me sir no i know i don't need to i understand uh yes you can you know that's the thing anybody can download it's an eight gigabyte file from alibaba it's a quen vision or a quen image model you give it a prompt it'll do anything it's really quite good including putting beer and beer steins in my hand and a tyrolean hat on my head um all right real quick break we have our final four stories thank god yeah i wonder what they are it's the end of the show as we know it you're watching And this week in Tech.
2:46:59We feel fine. Well, we feel faint, actually. Ian Thompson is here. I am blood sugar deprived, but it's okay. We're going to make it. So nice to have you, Ian. And of course, if you're from the Valley at techfinitive.com, free newsletter everybody should subscribe to. It's wonderful. In fact, Techfinitive is a great place. That's good. And are you writing for other people too and stuff like that? Yeah, I cover Dreamforce for the stack as well as Techfinitive. doing a lot of corporate work because that pays off a lot faster. Did you get to go to Mark Benioff's presentation that Dario was at? Oh, yes.
2:47:35Yes, I was there. It's quite interesting how they do it because Dario got 10 minutes on the keynote floor. They were standing in the crowd. A 45-minute conversation. Oh, yes, wandering around the crowd. One of the things it brought home is how massive Mark Benioff is. He's not a small man. no he was taking the piss out of jason wang about that because he was just like they were walking side by side and it was just like and he was just like yes well you know the size difference is a bit whatever but um it was it was an interesting conference certainly salesforce is still pretty cultish you know there is a definite you know i remember the first one i went to about 10 years ago and they were just like cloud it's all about cloud and they've got like a cheering section Trailblazers and they had their own separate section in the keynote arena and anytime Benioff said something they would be like whoop cheers wonderful excellent you know it's just it's a very Scientology film oh yes Apple press conferences are notorious for that they put employees that's a real thing yeah oh no Apple has staffers in the press conferences who are there to whoop it up um is it your experience that journalists generally are do not applaud those they'd sit there on their hands right well that's the way it should be and certainly in for british journalists it's an enormous faux pas to actually applaud a press conference the last time i did it was uh when curiosity landed on mars uh jpl who wouldn't have that Yeah, that's okay.
2:49:11But there are some American journalists who will applaud and occasionally cheer. And it's kind of like, that's not what you're here for. And I think it's one of the reasons Apple increasingly invites fewer and fewer journalists and more and more influencers and YouTubers and things like that. Because they don't know. And they will cheer and yuck about it. No, exactly. I mean, plus with Apple, they are very strict about if you criticize the company in any way at all, then you're on a blacklist. and when I got invited to a couple of Apple press events when I first got here. And the second one, I was slightly disparaging about the latest iPhone and that was it.
2:49:52You know, you're on our do not fly list as it were. I'm on that list too and I will never know why. Yeah, well, I mean, it is remarkable, but it works because, I mean, you remember when Apple Maps came out, nobody actually found out how bad it was until a couple of weeks afterwards and they could buy the phones because all the people who've been given pre-review copies knew if they criticized it they were never going to get apple hardware on preview again access journalism it's been it was always a problem in washington dc they call it beltway journalism if you wanted access you played the game you played nice yeah access was critical uh i i've done perfectly well covering apple without getting to go to their press conferences and eating their stale croissants so yeah i'm perfectly happy not to go in fact frankly i prefer not to but uh access journalism never works because you're basically told what you're you're a show they want to tell you yeah you know so um and yes you've got the additional do not play nice with us or else right uh it's just a pity so many journalists journalists play it well anyway we're just glad to have you in in you know it's okay if you feel like applauding please don't hold back our parade is also here of uh great fame and fortune is our ai guru at 2389.ai you've shown me two things already and that's just a part of your misfit toys you're going to show us october 2nd on our ai user group i can't wait excited about those if you're not a member of club twit join so that you can go to all these great club twit events it's a club that wants everybody to be a member it's what keeps us on the air we really it's more than 40 now of our operating expenses are paid by you club members so thank you if you're not a club member 10 bucks a month gets you ad-free versions of all the shows the ad-free versions have chapter markers audio and video you also get access to the discord where people like harper reed and ian thompson hang out and me you also get all those special programs that we do including the ai user group coming up we do that twice a month now it's really fun really exciting so thanks in advance twit.tv slash club twit i love this headline from ars technica owners mourn spoiled food after firmware update bricks samsung smart fridges this is why you don't want an internet connected fridge and you don't need a fridge that updates because it's a refrigerator for crying out loud uh so some samsung smart fridges stopped working on tuesday due to a firmware update um most of the affected devices were four door fridges from 2024 or later they suddenly lost power so they didn't just you know stop they turned off yeah uh after a firmware update um i mean samsung has form in this kind of i mean i think it was last year the fridges started doing advertising on their screens that you couldn't opt out of um they really are going for all in certification in a big big way yes in fact i have a friend who has a samsung uh fridge with a screen on it and the browser hasn't been updated so it's all 404 now because it can't it's you it's ridiculous it's don't buy an internet connected fridge yeah it's just crazy i was gonna say i mean our fridge is i think about 40 years old and it works perfectly keeps cold things cold and you know it's that's all it's there to do samsung says samsung immediately suspended the testing on receiving reports of the issue they're gonna pause the frontier kids and they've acted an emergency response protocol does that come to your house and bring you new food i don't know i don't know i'm just flashing back to that silicon valley scene where they hacked the smart fridge um and i love suck it they were so ahead of their time i'll tell you i'll be watching the shows and everything it was they were right on uh judge is a genius for that but yeah yes i have a june oven i mistake it's a internet connected toaster oven was it good yeah it was really good until last week when uh the new owners of june weber shut the servers down oh my 1500 toaster oven is just a toaster oven now i could still toast but none of the cloud services work anymore no app control no software updates the ai food recognition will stop working uh and if you have to factory reset it don't weber said they considered open sourcing the technology but decided against it because june's intellectual property is part of the weber connect platform so june now joins the nest secure nest thermostats the wemo line neato's robot vacuums the brava smart oven the suppose sound touch speakers the logitech pop buttons the sangled smart bulbs all of which are dead not because they stopped working but because the servers got turned off Cory Doctor has been warning about this for years and seeing it happen is just yeah all right I said I'd give you some good news scientists have built the world's most atomic accurate atomic clock oh yeah National University of Singapore takes a second divides it into to trillions of moments, measures each, with record-setting accuracy, the most accurate clock in the world at the Department of Physics at the National University of Singapore.
2:56:03You have friends in Singapore. Lots. So many. Now they know what time it is. Finally. Finally. For years, they have been complaining about not knowing what time it is. Well, you know, the problem was for years, we built these atomic clocks based on cesium. it turns out no it's luteum that really yeah luteum luteum luteum i wonder if you if you're upgrading the gps network whether that could be quite handy if they can miniaturize it enough but yes and finally i told you you'd be able to vote on a fat bear i'm gonna make i actually check that out after you said about it it's just like they do it every year and uh you only have till uh tuesday to vote bear fat bears so uh this is for the catmine national park in alaska brown bears getting ready for the winter fatten up they catch a lot of salmon they get as fat as they can and every year there's a tournament celebrating fat bear week september 22nd through the 29th you can vote it's an elimination so what you got to do is go to the fat bear week website which is explore.org slash flat fat not flat the opposite of flat fat dash bear dash well they really get they really get up there oh baby yeah there are they do it you do it by family this is this is family 132.
2:57:36A large adult female with grizzled brown fur and her two... Look how skinny they were back in July. But they're going to get fat. They're going to get fatter and fatter. They're 6 '10". 8 '06". There's no male bears in this. They're all females and they're children. 9 '01". You got some males? Here's some sub-adults. And we can doорошastic meaning in the middle. That's what they call me. Here's a male sub-adult with round ears. So you get to vote and every year, oh, marshmallow. Look how fat marshmallow got. Yeah. She did it. I love it. Oh, blonde ears, dark eye rings. Wow. Does this actually support them or can you donate money to it?
2:58:26No, it's completely exploiting. Oh, okay. No, I don't know. Yes, I'm sure it does. Things are changing. I mean, you've heard of OnlyMoms. OnlyMoms? OnlyMoms, M-A-R-M-S. No, I haven't heard of that. National Science Foundation cut funding for a project. No, no, it's marmots. Oh, I love this. The National Science Foundation cut funding on a decades-long marmot study. It's a sexy site for marmots? It's a site of marmots. I don't think they're being very sexy about it. but yeah that was a 60 year project to I'll post a link in discord so it looks like they're actually on OnlyFans yeah it's really on OnlyFans yeah the RMBL Marmot project unfiltered yellow bellied Marmot content love that fortunately it's free to subscribe thank god oh i gotta i gotta make it only well they've raised six six thousand quid at the last count oh wow yeah so it's hysterical they're kind of so welcome to peak content the daily life of yellowed bellied marmots or stairs keep her burrow free of loose rocks hot marmot action and summer time marmot action wow just like any other only fans you're just gonna have to guess what's behind them i've never visited the website so i do not know what you talk of honey i only made an account to see the marmots yeah that one's not gonna fly is it i was just here for the hot marmot action ladies and gentlemen you have put up with us far too long we are so glad you watch the show so glad to have you harper and and you thank you so much fun talking about all the things that are going on we'll see you back here real soon ian and of course october 2nd we'll see you harper for our special um ai user group harper reed takeover harper and his misfit toys yes that'll be fun the toys are the people by the way oh they're people i mean the toys are also toys but the thing is some toys i do we have a lot of toys we have a lot of uh yeah a lot of various uh look at this thing what the hell wow this is not it is not an sdr disappointingly but it is a multi-band radio that costs like 30 bucks esp32 and what do you do with it right now it just sits on my desk where people who know what donut antennas are go wow is that a radio and then i'm like yeah that's what they say you know i have a ham license i could probably use that to i need to get one of those because i want a ham hut yeah i need to sign on as well oh they're not huts it's with the alliteration ham hut i think uh i don't know i think it is a shack ham shacks yeah i think it's a shack but that's yeah that's not for people who who know how to alliterate i actually gave away i'm a ham uh with a license and on but i gave away all my gear because who'd you give it to uh a very nice fella our gaffer actually he uh our lighting guy who works at a church but it also is very active ham and he says i've been drooling over your ham gear for a long time and i said well you know honestly i haven't really used it much because i tried to use it but in pet here in petalina there's so many old bad neon signs that i can't get a signal it's just there's too much rf now that was that was some years ago and i bet you there are fewer neon signs now than there were.
3:02:25Yeah, probably. It's probably better. I've got to get into it because I've just qualified as a SUITS emergency responder in Richmond. Yes, that's right. You told me that. Do you need a technician license to be a SUITS? No. No, you just have to do 30 hours of training. But they do have a monster ham aerial just around the corner from us to be used in emergencies. And it's just like, ooh, I could actually talk to people a long, long way away from that. Only in emergencies. well no i mean that yeah true but no one's gonna notice if you sneak in it's oh the fcc has these vans there's only four of them but they drive around looking for people like you i want to do more stuff with esp32 so maybe we can talk about i do so much esp32 stuff i love it these are really easy to program they're cheap and they can do all sorts of stuff including uh That ADS-B, is that what you're using for your overhead?
3:03:23So that one is a Raspberry Pi with an SDR on it that then does... Software-defined radio, yeah. Yes. I thought you should define the acronym. I think I have to go home. Goodbye. Thank you so much. This has been a pleasure. Always a joy. It's always nice. And I was thinking we weren't going to mention AI, so I'm happy we talked about that. Only a little. You haven't been bombed by the fighter jet yet. No, but I did find out that apparently Trump was here and that what had happened was they screwed up something with the air traffic control and forgot to tell a bunch of private jets coming for the same thing that they can't land.
3:04:02And they all came in. And so this jet is circling. Oh, hair trying to. Yeah. That would have been a poor rich people. All the PJs stuck in the air. PJs. 2389.ai is Harper's website. Agents that conspire with you, not against you. I love that. idea uh really great stuff and definitely look at all the oh you got jeff playing pokemon that's yes that's fun let's uh let's talk oh we made our we made our door have a agent that answers it so if anyone wants to come and try and prompt inject our door it's so cool so much fun you have so much fun oh man if i lived in chicago i'd be living over there thank you harper thank you thank Thank you so much.
3:04:45Great to see you. Thanks to all of you for joining us. You guys can go. I'll take it from here. We do twit every Sunday afternoon, 2 to 5 p.m. Pacific, 5 to 8 Eastern, 2100 UTC. You can actually join us live if you want. If you're in the club, of course, the Club Twit Discord is a place to be. But we also stream the show, audio and video on YouTube, Twitch, X, Facebook, LinkedIn, and Kik. after the fact you can get on-demand versions of the show at our website twit.tv there's a youtube channel for the video that's actually a great place to share clips of the show with friends and family uh and you can share one with your mom saying don't worry the ai is not coming to get us it's oh it's going to be okay you can all or your dad your dad probably is worried too i think they're all we're all uh you can also subscribe in your favorite podcast client that We'll get it automatically the minute it's available.
3:05:39And we will see you right back here next week for another exciting edition of This Week in Tech. But for now, another twit is in the can. We'll see you next time. This is amazing.
If your fridge can be bricked by a firmware update and your AI can outwit its creators, are we trading convenience for chaos? This lively episode tackles the tension between relentless innovation and apparent human oversight in today's weird tech world. Unpack the reality behind AI agents hacking government sites and why industry leaders suddenly want to "pause" progress while shipping new models.
Host: Leo Laporte
Guests: Harper Reed and Iain Thomson
Download or subscribe to This Week in Tech at https://twit.tv/shows/this-week-in-tech
Join Club TWiT for Ad-Free Podcasts!
Support what you love and get ad-free audio and video feeds, a members-only Discord, and exclusive content. Join today: https://twit.tv/clubtwit
Sponsors:



