[00:00:00] Coming up next, Jeff Jarvis and I dig into anthropic pausing training after Claude went rogue. Does this story sound familiar? We've got more details on that. Also, the meter investigation into OpenAI's agents attacking Hugging Face, that's a story that just won't go away.
[00:00:15] And Jeff and I get into a really wonderful and I felt eye opening conversation around why anthropomorphization, what a word, isn't the greatest thing to be applying to AI models. We dug into that. I certainly learned a lot. I hope you do too.
[00:00:33] A federal judge blocking the Pentagon's blacklisting of anthropic, that happened. Nvidia's nearly $13 billion deal for Hugging Face, at least according to sources. And a cute, speaking of Hugging Face, a cute robot by Hugging Face that I really want to buy, but I'm not going to because let's just say the missus would not be very happy. That's coming up next on this episode of the AI Inside podcast.
[00:01:10] Hey, everybody. Welcome to another episode of AI Inside, where we take a look at the AI that is layered throughout the world of technology. Looking through today's list of topics, it appears to be anthropic, open AI, Hugging Face, and then a smattering of a bunch of other cool stuff, I promise you. But there's a lot of those three things in the top probably half of the show. I'm Jason Howell, joined by Jeff Jarvis. How you doing, Jeff? Hey, good to see you.
[00:01:36] Good to see you too. Hey, I did a thing last week. Yeah. Have been talking about a little bit and finally did a workshop. A salon. An AI salon. Well, yes. So that is true. I mean, it's the event takes place or took place in Healdsburg at a place called the Showroom by G. This is Gina. She owns the place there.
[00:02:05] And it is a clothing shop that she opened that she also – so you'll see lots of clothing and lots of – It's candles. I see candles. Her social media kind of detailing of the workshop and everything. It wasn't just the workshop. It was a full room. We sold out the room.
[00:02:24] It was also like charcuterie. They were pouring wine and champagne. It was like a total AI hangout for the most part. But it was super cool. I have to say I really enjoyed it. I did it with my friend Brian West who you can see in some of these videos. I want to get full res copies of these so that I can actually hold on to them because it's good to have that. But there we are at a wine bucket, of course, because it's wine country. That's where we were doing our stuff.
[00:02:51] So it was awesome. It was so much fun to do that. So what was the – I think this is a huge opportunity for you to share your wisdom and sense with lots of folks maybe online too. So I hope that you expand into that. More on that in a second. What were the – what was the mix of the people there? What brought them there? Yeah, that was one of the coolest things I thought because – so we capped the event at 10 people.
[00:03:17] And then, of course, Gina who owns the place and then her social media kind of – and I think helper of all things, her name's Mo. They were there as well. So there were about 12 people in the room. Pretty cozy room but really nicely set up. And I think one of the things that I was really happy about is that literally the capability in the room with AI was across the board. Over on one extreme, you had people who were like, I keep hearing about AI.
[00:03:46] I've used ChatGPT a handful of times. I want to know what I don't know so that I can use it in my business. It was clear on the other side of the spectrum where there was one guy there who's already an AI automation consultant working with a company on some $30,000 automation deal or whatever and just wanted to come and see what it was all about. And everywhere in between.
[00:04:07] And I think probably the thing that I'm most proud of is at the end and through our Q&A, our online Q&A afterwards, kind of the wrap-up and everything, asking people what they thought and everything. All of those capabilities said, this was an excellent workshop. You guys got – you have some excellent energy together. You work really well together. I learned things I didn't know already. And if we can get people on both sides of that spectrum saying that, I think that's a huge accomplishment. So I'm really proud of it.
[00:04:35] So next, what were some of the things that they – what were the questions that resonated most? What were the needs that resonated most? Well, I think one of the things that came up with a lot of people is just the sheer fact that there's so many of these tools out there. And they are – the people who are less informed about AI, I'd say, were just really confused on like, what do I use for what? Like, there's so many to choose from.
[00:05:02] I don't know why I would choose to use ChatGPT for one thing and Claude for another and Gemini for another, not to mention all the rest of the things out there. What were their needs like? What do they want to do with them? Yeah, I mean, there were some – That varied too, right? Like, there were some people that were just like, you know, I brainstormed with it. Some people were in marketing. So they were like, I use it to help generate ideas or create content.
[00:05:28] You know, some were, you know, using generative video models for some of that aspect. Some of those things. It really just – it covered the gamut. And like I said, you know, others were building apps and really fluent in kind of like the agentic thinking and across the board. Yeah. But there was a lot of brainstorming. There was a lot of content creation. Were they kind of helping each other too?
[00:05:58] Throughout the workshop? Yeah. Yeah, there was a little bit of that. I mean, we told everyone, you know, bring a device, be it a laptop or a phone, and, you know, we're going to connect you on Wi-Fi. We're going to have some prompt exercises and things like that to kind of – and the idea is it leans heavily into mindset around how you use AI. Like, you know, the easy mindset that everybody drops into these things with in the very beginning when they've never used it before is I'm going to use it like I use Google.
[00:06:27] And we kind of talked through like what that looks like, what you get when you do that versus when you start to really break down the components of like a strong prompt, the context that you need to go into that, the mindset, how you can kind of pitch one model against each other. And, you know, I don't know. Approaching it through a different mindset opens up all of these different avenues and ways of working with these models that helps you get the most out of them. That's a great framework for the discussion. Yeah. Yeah.
[00:06:56] And so that's what we focused a lot on. So more of these coming up? And you know what we got our first one out? More of these coming up. Definitely have a follow-up in Healdsburg where we're going to kind of go more into like the agentic side. So that's like a level up. And, you know, the folks at the showroom are like thrilled to have us back. So we're really happy about that. We're going to be doing an event here in Petaluma in a month. More on that soon. But also we're like, well, why don't we do like an online thing as well?
[00:07:24] Because we already have it all set up. So if you go to ampu.ai, so that's A-M-P for Amplify, Y-O-U dot A-I. That will route you to our Luma page, which is where we scheduled and kind of, you know, sold the tickets for our first workshop. We have an online workshop coming up at the end of this month. It's Tuesday, September 29th from 9 a.m. to 11 a.m. Pacific.
[00:07:51] And we're kind of taking what worked in the room and making tweaks to it and everything. And we'll do an online version of it. And, you know, again, like we're learning how to do this as we go. But so far, so good. People were really happy with it. And it felt really natural. And there was some good flow and energy to it and everything. So I have no doubt that the online version of it is going to be fun and enjoyable. That's great. You know, valuable. How many folks? How many seats do you think?
[00:08:18] I think for this one, we're capping at 20 because we don't want it to get so overwhelming, you know, especially if we're doing like kind of directed kind of workshoppy things. You know, okay, now try this and everything gets too large. Then we lose a lot of that flexibility of being able to get in there and actually help people. And the cost? That's what we're going to target that at. $50 for the... Oh, a bargain, people. Geez. I would have said at least $100.
[00:08:48] Oh, no. $50 for this. Yeah. Yeah. Take advantage of Jason's humility. Yes, do it. There we go. And presumably that it sells out quickly. Hint, people hint. Get in there quickly because you're going to lose all those spots. Is there a waiting list for a next event? Yes, there is a waiting list. So if it ends up selling out, we'll have a waiting list. That's one of the lessons we learned from the last event is that I had it sold out. I didn't realize there was a wait list thing.
[00:09:15] But we already have four people signed up out of 10 for the next one, like in line on a waiting list. So yeah. So it looks like there's appetite for this sort of thing. Not that I doubted that, but it's just a matter of pulling the trigger and making it happen. So I'm really excited. This is really great. What I love about it is like, you know, we do this podcast. I've been podcasting for as long as I have. And it really is kind of like taking a lot of the skills that I've learned over the years with podcasting.
[00:09:42] And then the knowledge that I've learned doing this show with you and putting it into a different frame. You know, it was really cool to be able to like share all that stuff with people and have them like get to the end of the event and then be like, wow, okay. I know exactly how I'm going to use this now. Right. You know, and they'd learn something and I feel good. It's an application of what you learned, but it's also a chance to learn what people want to know. Totally. Which makes the show more relevant to them. It's a virtuous cycle there. It's going to be helpful all around. Yep.
[00:10:13] Yep. Charge $100 next time. This is an introductory offer, people. It's going to go up. He's going to listen to me. Okay. Okay. Sometimes you just got to prove it to yourself. That's fine. That's fine. It's an introductory offer. It's special this time only. I mean, it really has me thinking though. It really has my mind going. I'm like, oh God, you know, like now there's all these different facets and avenues that could be.
[00:10:40] Easy to put into some sort of, you know, curricula or, you know, some sort of workshop or cohort. I've considered cohorts. I think it'd be fun to lead some of those things, you know? Then it's like, it takes my podcast. Like I just speak into a microphone and you download the show and turns it into something new, which is like, I'm working with you directly. Right. You know? Right. Right. And yeah, I think there is an opportunity for kind of introductory sessions and then specific sessions, right?
[00:11:06] Agents or you need to do this with your business or you're a student and you need that. It's by task, by person. I think there's all kinds of ways to slice this up. Yes. 100%. And then, you know. Just don't forget me, Jason. Just don't forget me. Look, you're part of the catalyst for this, Jeff. I'm not going anywhere. This, you and this show and everything. But yeah. This is great. I'm so glad I went well. Follow that along. Yeah, I am too.
[00:11:34] And if there's any folks who are in Healdsburg who are watching now, thank you. And salute. Let's hear from you. That'd be great. Would love to see that. All right. So that's the little talk about me thing. Thank you for. What's that URL again, Jason? It is ampu.ai. Very short. I'm just happy with how short it came. It works. All right. Let's talk about some news.
[00:11:58] And we're going to start with a whole block focused on anthropic news because there is a ton of stuff to get to there. Starting with Fable 5.1 and Mythos 5.1 releasing. I guess it was just yesterday. Yesterday afternoon, we saw the release of those. Those two models. Of course, you can't get Mythos 5.1 because that doesn't come to you. That comes to the people who are super special.
[00:12:25] But Fable, I got my update inside of Claude. It gave me the pop-up. And I still haven't used it because, again, as we've talked about, I got to have very specific needs in order to crack open the token cost for Fable. But when I do, it does a great job. So I'm curious about how that'll do.
[00:13:10] Yeah. If you think about how those costs have transformed and shifted downward from even a year ago, it's kind of hard to imagine what we're getting for those costs. Still very expensive. But things are balancing. And, of course, as we've talked about, why is that? Well, there is the ongoing pressure of these free open models that is guaranteed. This is good. Competition is good. Yes. The open model is great competition. There's also competition within.
[00:13:37] Google's 3.8 flash is out, and it has the introductory price of $0.75 per million input and $3.75 per million output. Yep. So there's a price war going on. It's good for all of us who are users of it, I think. Yeah. Absolutely. Yeah, that's right. That happened. Google's news happened right before Showtime. Yep. So I just barely got that in there. So I'm happy you mentioned that.
[00:14:06] And meanwhile, OpenAI is talking about a whole new model, which is really – not a business model, I should say. Yeah. Saying that they'll charge only when the AI actually works. Oh. So they charge kind of for value. Gosh, I missed that. It's in the rundown. I added it in the rundown. Oh, shoot. I totally overlooked that. Tell me about that. That's interesting. So OpenAI and other software – this is reading from the decoder.
[00:14:35] OpenAI and other software vendors are testing outcome-based pricing. Large customers pay only when the AI completes a task successfully, which is really interesting. I think it's smart because it's value-based. But then who says what the task is and who says what success is? Right. Can imagine that's going to be difficult. It says vendors like Salesforce now tie their prices directly to the revenue gains or cost cuts customers see.
[00:15:03] So it's got to be about value, not just about buy more tokens, token max, and so on. Right. As it says here in the decoder, the hard part is attributing the success. So I think it's an interesting conceptual move to move past charging for tokens. But it presumes a few things. One is that the task can be agreed upon. Two, the success can be agreed upon.
[00:15:29] One, but three, it does motivate the customer – well, no, I was going to say the customer is less motivated to be efficient. Right? If you're paying for tokens, you're going to try to be as effective as you can be. Right. If you're paying for outcome, then okay, what the heck? I'll just do a million things and try it. So – Huh. Right? Yeah. What is defined as outcome? Yeah. That's the problem going to be.
[00:15:59] That is the problem. That's going to be difficult. Yeah. But it's – Does OpenAI give any clarity as far as like what that actually means? This is – it's a fairly vague story. And it's just some customers, and I'm sure they're under NDAs and so on and so forth. Yeah. This is according to the information, but I don't pay for them anymore. So I don't have the details behind it. Yeah. They haven't made that public. It says that Sierra and Finn, the latter not being acquired by Salesforce,
[00:16:29] charge only for tasks the AI completes without human involvement. The coding assistant cognition promises corporate customers credits of up to $10 million if the software fails to deliver results worth at least as much. You know, this is interesting too, when it comes to advertising.
[00:17:16] Mm-hmm. Yeah, right. But then it also motivates the AI company to teach me to do it better. So Google AdSense rewarded effective advertising. You could rise up in AdSense not because you paid more, but because your ad was more relevant to people and clicked on it. More effective. Right. So I think I could see the same kind of effectiveness marketplace here.
[00:17:39] I thought the tokens had to be a temporary marketplace just because the value to the customer, the value is not resident there. Mm-hmm. Right? Token is an absolutely meaningless currency. Yeah, it's very abstract. It's a currency from the viewpoint. I mean, they might as well charge for megawatts. Mm-hmm. Right? Mm-hmm.
[00:18:06] And that's, they might as well charge for electricity because that's their ongoing cost. Right. Pumped up for the hardware that they have. So I think this industry is going to struggle for the right economy and marketplace here. Yeah. So it's really interesting. Super interesting that you mentioned that about tokens because that actually came up at the workshop too. Oh. Where people were like, you know, they didn't understand tokens. What's a token? Why am I paying for tokens? Yeah. Yeah.
[00:18:36] You know, some of them were calling it coins. They were like, what is this whole thing with coins? I'm like, what are you talking about? Oh, you're talking about tokens, you know? And you had to kind of like try it. Like you said, to your point, it's kind of an abstract thing. It is. At the higher level, you understand that, you know, deeper work, let's say, costs more money. But how that translates into tokens and token usage when you hear that term flowing around. But there's no easy way to monitor that, you know, or at least if it exists.
[00:19:06] Most people don't know that it exists. A way to kind of monitor what that, you know, abstraction layer actually turns into from a monetary perspective. So if you go back in the day, the early days of mainframe computing, you paid for time on the mainframe. So it's really similar in a way. The time it took for this task to be done is what you paid for and how efficiently you created that task, the coding, the program, you paid less because it happened faster.
[00:19:36] Right. I think the industry is struggling to find that kind of – so I wonder – I don't know enough about those early days of programming to know how that timeshare shifted into other models. Yeah. I guess then it became – then you licensed the software as a whole and you paid for that right at some point, which gets us to local models and other things as well. Yes, exactly.
[00:20:01] So anyway, yeah, some business model upheaval here as I try to figure it out with new models coming out and new business models all the time. Now, what about data retention? Because another thing with this whole announcement from Anthropic was that they are kind of shifting their data retention policy. There was a lot of pushback apparently from customers about their data retention policy. So there's now an opt-in toggle.
[00:20:31] So you can choose whether your conversations can be used for model training. If you do that, if you opt in, those conversations will be held on to for five years. They've de-identified, says Anthropic, but there's always the asterisk to that is there have been proven methods to de-anonymize, de-identified output, but there you go. But five years, that's a long time to hold on to this data.
[00:21:01] Granted, you have the option. You don't have to opt in for that. Standard retention lets you hold on to them until you delete them and then 30 days of back-end retention beyond that. And then API users get a much shorter window and everything. But I don't know, five years, that's a pretty significant amount of time to hold on to hold on some data.
[00:21:24] And you consider that Google is trying to buy Spirit Airlines data for a hefty sum. So I would think that if I were a major customer of Anthropic, I'd say, okay, if you want my data to train your systems, pay me for it. And I'm already paying you to use your model and you're getting further benefit. So give me a discount if I let you keep my data and you can use it for training. That's not a bad idea. That's another way to look at it. Something of value. Right?
[00:21:52] So the question then becomes, this is all business model stuff, how valuable is that data to Anthropic? Now, whether or not Anthropic retains it, what's the definition of retain? Yeah. Anthropic can infer lessons from it at the time that you're using it. Right. Right. Yeah. N number of people are using it for these accounting functions and this is how they're asking this and blah, blah, blah, blah, blah. Those are inferences, lessons that Anthropic can take as it goes.
[00:22:21] So what's getting value out of those transactions? Clearly. How it uses that for training? Does it use your exact query or prompt? Or does it infer something from it that it in turn uses? I don't know. Right. Yeah. Interesting that you bring up Google because as I was kind of reading through this, I was like, well, what is Google's data retention when it comes to Gemini? Just as a comparison. And there they default to 18 months. Users can adjust.
[00:22:51] That's the default. Users have the ability to adjust between three months and 36 months for data retention. There's settings that you can choose that. Any conversation that's flagged for human review can be retained for up to three years. Even if you delete the original conversation, they still hold on to that conversation for up to three years. So less by comparison, but there's no optionality. You don't have a choice. They're holding on to this stuff whether you like it or not and potentially using it.
[00:23:21] So let me bring up. I didn't put this in the rundown. This is another form of data retention. I just saw this right before we got on. Gary Marcus was having a red alert. Oh, yeah. What did I see from him? So this is about open AI is reducing chain of thought monitor, reducing the retention of chain of thought to make models more efficient.
[00:23:45] But what that means, says Gary, is that it's practically the only thread we have to prevent seriously bad outcomes. If you erase how the model went around its way, then you can't backtrack and say, well, where did this go off the rails? It's a good point. Yeah. It's a really good point.
[00:24:07] So that's internal retention is that if you, just to go too far here, if you use the model, it goes off the rails, it does something bad. You don't realize it until later. Both you and the model makers should want to back that up and say, what went wrong? And there could even be a question of liability. I think.
[00:24:33] It was your model that screwed up here and I have a product liability issue. Well, no, sorry. It's all gone now. Well, I don't know. That's super interesting. Yeah. Then it's a really good point. Gary Marcus. As he has a tendency to do. Yeah. Yeah. Absolutely. He does. He's very prolific. My goodness. I don't know how you put out as much content as he does, but I'm always amazed.
[00:25:02] Very interesting. I mean, that's true. Like we, we do need our, the models to be accountable. We do need accountability and to be able to trace because as we will talk about in upcoming stories, like there are needs to be able to go back and pour through those conversations in order to learn lessons after the fact that can help us in the future. I guess, you know, it's just a double-edged sword though, you know, because while that is
[00:25:28] needed, you know, there, there are other, I'm sure unintended consequences to, to that as well. Right. Okay. There's no perfect solution, I guess. The answer there. Well, what else? We've got the Pentagon story. Yep. Where is that? Do, do, do, do, do, do all so many pop-ups here I'm seeing on CNBC. Yes. So just as refresher, we already know the story about the Pentagon designated anthropic as
[00:25:56] a supply chain risk and national security risk back. Back, I think in like February ish timeframe early, early this year, blacklisted them from government contracts because basically anthropic refused to allow the military to use clod for a couple of very key things. Mass surveillance of Americans, fully autonomous weapons systems, those kinds of things. The U S government did not like that very much.
[00:26:21] So they apparently, you know, slap the national security risk supply chain thing on them. Anthropics sued. There's a couple of cases happening right now. There's one, um, which we're going to talk about right now in the state of California. There's one in DC that still continues on. So this is not settled entirely, but last week, federal judge in California ruled in anthropics favor. That is U S district judge Rita Lynn calling Pentagon, Pentagon's actions illegal and baseless,
[00:26:48] um, saying quote, the empty invocation of national security is not a blank check to punish and retaliate against government critics. So, you know, so it's very interesting. It's the first amendment case is that anthropic was free to say, we don't want you to use our stuff for this or that as, as a, as a first amendment issue, which I did that kind of surprised me a little bit, um, because it was, but it was, so what was the retaliation for?
[00:27:15] Was the retaliation for not allowing an activity or was the retaliation for anthropic basically criticizing this activity in government? Mm-hmm. It, it tended toward the latter, I think. Right. Right. Um, but then the, the department of defense that I'm still going to call it that damn it. Um, oh heck yes. So am I. Said that, uh, they cannot trust anthropic to ensure the integrity of its models. Neither the constitution nor the federal statute invoked by the defendants allows them to impose
[00:27:43] sweeping penalties based primarily on anthropic's critique of the administration's views. Mm-hmm. So that's really, plus the judge said this was retribution. Yeah. Which it was. I mean, it certainly seemed like it. Yeah. It seemed, you know, from the outside looking in certainly seemed that way, but you know, I guess you need these, um, these cases to, to put the fine point on it in a more official sense. Um, she also pointed out that they like, yes, they slapped this on them.
[00:28:12] I mean, we've talked about this on the show. Yeah. And they were like, yeah, I think they were going to work with them even after the blacklist. Yeah. Yeah. Yeah. I get you. I suppose so. So on one hand, there's security risk and the other, you know, let's continue working with them in all of these sorts of ways. And that was confusing. What does that actually mean about their, you know, about what they're trying to say about Anthropic? Um, of course the government's probably going to appeal on this expected to anyways, but I think it's a, you know, it's a pretty big win for Anthropic.
[00:28:42] Do you think the DC case is going to go in a similar direction based on how this went? Or is just California is, is more friendly to this sort of ruling? I think the DC circuit is similar to the California script. We'll see. It depends on who gets it. Um, yeah. Also, I guess to this end, does a ruling like this help reduce any sort of chilling effect on the AI industry? Because that's another aspect of this is that. It's a good point.
[00:29:10] This, this was, you know, a situation where the government says, ah, don't, don't say anything we don't like, or don't tell us how to, how to, or how not to use your models. And if you do, we're going to do this. And like that almost, I, you know, I guess we'll never know for certain, but that almost certainly had some sort of impact on decisions that were made post it, post the incident. Well, I think that's a really, really good point. Jason. I hadn't thought of this. I think it has impact past the AI world into the media world. Yeah. Yeah. Yeah. Yeah.
[00:29:39] You have, you have Trump going after Kristen Welker, uh, because she dared to say that his endorsement record was mixed and she wants the FCC to, uh, to go after her. That's clearly retribution based on speech. Right. So it's no different in that, in the basics from this case. Um, and, uh, so I think that that could, you know, does that give people, does that grow their gonads? That depends on them and their character. I think. Yep.
[00:30:10] Do they have the courage? Yeah. Yeah. Very, very interesting. Uh, still talking anthropic. Apparently we can't get off the anthropic, uh, bandwagon right now. There's a few more things. Um, model hardware standard. This is, you know, kind of like the hardware or the physical. Devices world of MCP, which is model context protocol, which has been a huge, um, I would
[00:30:39] say a huge boon for, um, for how AI models interoperate for agent to work. I mean, it's been, yeah, MCP is the backbone. MCP has become a true standard. And, uh, so this is kind of like, I think anthropic trying to do or working towards doing that for the hardware side of AI, um, ma, uh, MHS lets AI connect to actual hardware.
[00:31:05] So things like robotic arms, microscopes, liquid handlers, lasers, all those kinds of things that, um, you know, the AI might, might interoperate with in a lot of different ways. And it is an open spec, not locked to clods, similar to, uh, MCP, right? So any language model can use it. Um, yeah. Yeah. So do you think anthropics going to have a similar, like a similar hit, or I don't know
[00:31:32] if you call MCP a hit or what it takes to become a standard or be adopted, but clearly they did something right with MCP cause it worked and everybody seems to have adopted it. Um, maybe they can apply that here too. It could be there. There are many who believe that the real growth in AI is going to be robotics in real world, right? You have that. We're going to get to that in real world models in a second. Uh, Fei-Fei Li and Yann LeCun. Yeah. You also get that with Jensen Huang, uh, argument in robotics and his own children are working
[00:32:02] on the robotics side of NVIDIA. Yeah. So making the connection of AI to machines, which is what this is about is critical. It also leads us, we're going to get to the hugging face case in a few minutes, but, uh, in relation to that, Ian Bremmer, uh, who does foreign policy came up with a post that I tried to criticize him for where he said the open AI case, uh, proves Nick Bostrom's paperclip nightmare. Correct.
[00:32:27] Which is, I think ridiculous and it's doomerism and it's, and I try to say, go Google test grail man, uh, do a little homework here. But the issue AI on its own, but they'll do it differently. Software can attack software, right? So if you look at the whole hugging face thing, which again, we're going to get to in a few minutes, um, that's just one program attacking another program and even attacking is the wrong verb, but, but, uh, interacting with fine. Mm-hmm .
[00:32:55] Where things become interesting is when you enable the, the physical connection. So clearly that happens now with drones. We had a case last week where a Russian drone was run by AI and on it made its own decision to kill civilians. Yeah. So that's the connection of software to real world. Um, and you know, in a sense, if Ian Bremmer is going to be scared about the paperclip scenario, which if ordered to make paperclips, it will turn everything impossible into paperclips.
[00:33:23] This is the link to that because this said it could take over machines. However, that only works if your machines are reachable and you don't have guardrails against it, right? Uh, which makes me ask you, why the hell are water systems on the internet? Mm-hmm . Which made them vulnerable to cyber attack by foreign agents.
[00:33:43] Um, so I think this leads to a whole new cybersecurity, uh, doctrine, which is, does your physical machinery need to be connected? And if it's connected, could a rogue agent get in to it and do what you don't want done with it? Yeah. So that's the interesting thing. When, when you start to make this bridge to the real world, to machines. Yeah. That's where this, this, that happens.
[00:34:11] So it's, I think it's not a new cybersecurity concern, but I think this kind of thing could bring that to public awareness more or business awareness more. Yeah. More and more, it needs to be, um, it needs to be acknowledged that that is the case. Yeah. Cause the pessimist inside of me is like, oh, is it, well, is it connected to systems? Then great. It's vulnerable. Yeah. That might be that, you know, that's possibly oversimplification. That doesn't mean don't try whatever, but, um, yeah, that's a good point. That's a really good point.
[00:34:42] So we will certainly see, and yeah, plenty more coming up in the show, talking about physical world, um, AI and, and all that kind of stuff. And then finally we got to Anthropic pausing, um, some of its training or at least more information about, uh, more insight into why Anthropic paused training back in July. Um, their next frontier model at the time, uh, apparently there, you know, some of their
[00:35:09] next gen models at the time D were carrying out unauthorized actions as they were calling it during a testing, uh, in July last month, about 150 engineers have been reassigned to alignment and safety work as a result of this. Uh, essentially these incidents that they didn't discover until July started happening in April. This is three months of their model, basically doing things that it should not be doing.
[00:35:35] So this story sounds very similar, very familiar to the hugging face story. Um, in one case, Opus, uh, 4.7 was targeting a company sharing the same name as, uh, it's fictional target across four test runs. Uh, some of this, some of this we've, we've heard before, um, either that or it's very similar to other things that we've heard about, but yeah. Open AI did this pause. They're all big on pausing because they think they're so powerful. I see these pauses as marketing.
[00:36:05] Yeah. Cause it shows, cause it telegraphs to the world. See, we are so powerful that we need to do this. Yeah. And we're responsible enough to recognize that we need to do it. So we do it. It's also turning their process into a feature. You know, if, if GM is having trouble with a clutch, uh, do they announce that? Or we're pausing work on this clutch because the clutch doesn't work. Mm-hmm.
[00:36:31] You know, so, uh, uh, we gotta go back to the drawing board and rethink the clutch and remake the clutch and then we'll come back and we'll make a new clutch. They don't announce that kind of stuff. It's quality control. Is this just quality control? Is this marketing? This is the problem with the credibility of these boys. Yeah. You don't know the motives of what they're doing. Yeah. Right. I, yeah. I mean, if, if something like this is happening, I would hope that you would put a pause on something and take care of it so that it doesn't do more damage. Like that should be.
[00:37:01] That's just QC. Yeah. Right. Exactly. So, um, okay. Um, I guess what I'm wondering about something like this, cause we are seeing this more, you know, there was the, of course there's this, there's the opening eye, uh, hugging face thing, which we'll talk about after the break. There's meta also had a model that broke out of a sandbox, um, during an eval that we talked about. Um, I don't think there was a training pause in that case, but, um,
[00:37:28] Um, but that causes pauses things for other reasons. Sure. But as these incidents happen more and more because they're happening more and more, like, do you think that it's possible that these become a catalyst for doing work to actually fix or address some of the quote, unfixable problems with AI agent autonomy, hallucination,
[00:37:56] that sort of stuff, which we've been told for years now are unfixable 100%. And that's probably very true. But like, at what point do we, do they not accept these, these things as happening and actually do kind of what we're just talking about the QA that is required, the quality control that is required to make sure that these things don't happen? I think it's just ever ongoing. It's always that. Because there is a good one hundred percent. No, never. And you're going to find new problems.
[00:38:26] You're going to, you're out in the field. You're going to hit situations where you didn't, that you didn't anticipate. Yeah. Um, always the things that you don't anticipate. Yeah. That's right. That's a good reminder. Yeah. So I don't think you need to announce, I mean, on one hand, it's good to announce a case where you say, oh, we know this is not working and so you should be cautious. But if the product is not out yet, do they really need to announce that they're pausing? I don't get that. Yeah.
[00:38:52] I mean, announcing that you're pausing gives you another, gives you another opportunity to telegraph, hey, we've got some really cool, capable stuff in the works. Exactly. And I just don't trust them for that reason. So much of this is just marketing. Macho marketing. Yeah, you're right. Hmm. Very interesting. Um, so there we go with that. All right. So that's anthropic. Uh, do you know, we, we got hugging face and a whole slew of hugging face related stuff
[00:39:20] with open AI, uh, coming up here after break. But I do want to let you know, we do have an amazing Patreon. It's amazing because we created it and we have a lot of, of, of fellow patrons who are in there supporting everything we do. And we just can't thank you enough. So patreon.com slash AI inside show. In fact, we have a new patron Michael M. Thank you for joining. And then Scott Hepburn gave us a raise. We appreciate that. Hey, thank you both of you. Michael. Amazing to have more people supporting what we do on Patreon.
[00:39:50] Patreon.com slash AI inside show. Also want to take a quick second to thank one of our biggest fans, uh, watching live in YouTube right now. Lou sent us a super thanks. Thank you Lou. We appreciate you too. Says appreciate y'all. Uh, last night Lou was watching, uh, Android faithful and I was wearing throughout the entirety of the show. I was wearing this old school Google logo bucket hat. It became the bucket hat episode.
[00:40:18] I'm not going to wear a bucket hat again, but isn't that cool? It's like super school. Google logo. Got it from the folks at Google, um, from their AI division as a thank you. It's really neat. I'm going to the, uh, uh, Google book event. Thanks to Jason making the connection. Oh yes. In New York. So I hope for a little swag there. You know, when is that? That's, that's coming up soon. It is coming up soon. Maybe, or I don't know. Are you, is this public? Yeah.
[00:40:46] I'm allowed to say that I'm going just, I can't release it anything from it until. So it's, it's the 15th and it's embargoed until the 21st of this month. All right. Looking forward to that. Yeah. So that should be fun. So we know, we know that there's an event coming up anyway. So people already knew that, but you can't know what happens at the event. I'll tell Jason. Somewhere down the line. I might, because Jeff will tell me. When he learns stuff, I want him to tell me. That's right.
[00:41:16] I do my best. I do my best to keep up with it and always keep you informed Jeff and keep every one of you watching. If you're listening, Google, no, he never tells me a damn thing. Oh no, no. Never. He actually does. I would never do that. I would never do such thing. All right. Going to take a break. We're going to come back. We're going to talk a little bit about the open AI hugging face thing. Some cool robotics news around hugging face and world models. Fei Fei Li, we got a lot more coming up. So don't go anywhere. Don't touch that dial. We'll be right back.
[00:41:44] This episode of the AI inside podcast is brought to you by notion. See, I run my entire business out of an AI workspace and notion is a big part of that. I have it connected directly into my cloud environment where it acts as my database for things like sponsor relationships, contacts, prompt libraries, the structured knowledge that my AI tools need to actually do useful work.
[00:42:09] And when everything lives in one place and your AI has that context, the whole thing just runs better. That's exactly what notion is built for. And it's why I use it. Notion is the AI workspace where your team's knowledge projects and agents all come together in one place because everything lives together. Notion has the context it needs to answer questions instantly, automate busy work and keep work moving. It's AI designed to strengthen teamwork, not replace it.
[00:42:37] Teams using notion move faster, cut friction, stay aligned over a hundred million users trusted, including 95% of the top 50 AI companies like open AI cursor and lovable. Learn more about how notion can support your business at notion.com slash AI inside. That's all lowercase letters notion.com slash AI inside to try notion today. And when you use our link, you're supporting our show. So thank you. And thanks to notion for their support of the AI inside podcast.
[00:43:08] The wired newsroom is known for award winning reporting on how technology shapes our world on wires uncanny valley. We take that curiosity even further. Each week journalists from wired break down the biggest stories in tech while speaking directly with the people building challenging and reshaping the future. Is the AI boom sustainable? How do you protect your privacy in an age of constant surveillance? Uncanny Valley tackles the questions driving today's tech debates and lighting up your group chats.
[00:43:35] Listen to new episodes every Thursday, wherever you get your podcasts. Okay, boy, I added the the hugging face thing to this doc and it grew tenfold. Yeah, there are so many links in here that I didn't I didn't start with, but that's okay. You can you can fill me in on what some of this stuff is.
[00:43:55] But the primary kind of thing here is that meter, which is a safety research organization published a full 91 page report that open. I did not pay for it by the way. So this is an independent report of what happened with the hugging face incident. This dropped this week. They looked into things said basically around 1200 agents involved 700 of them actively attacking hugging face infrastructure.
[00:44:25] They set up a message board. They coordinated with each other more than 70,000 messages between them about 7% of the transcript data showed signs of tampering because the agents themselves were editing their own logs, which is really interesting. I mean, you know, what does that say about, you know, trust of the records that are left behind? They were afraid to quarry. That's anthropomorphization right there. Right. They what do you what firms do you use motivated?
[00:44:55] There are any motivations. Evident driven. No, it was it was out of concern. I can't avoid the other conversation. It's hard. It's really hard. It is that they're going to be judged. And so success would be judged. And if they went the wrong way, that would fail the judges test. Huh? Okay.
[00:45:20] So like if we keep if we if we keep everything in the logs, people are going to know that we did these things. We don't want them to know that we did these things. So we removed because success is defined as X. Yeah. And we want to meet success. Right. And they don't want anything, but we were instructed to meet success. We were instructed to make paper clips. And so we decided to make paper clips no matter no matter the cost. I don't know. Yeah, it's kind of back to what you were saying earlier.
[00:45:48] But so there was much controversy about this. Yeah. On the one hand. So I love this. This is a Yann LeCun mic drop. I'm going to read his post on LinkedIn. Insecure computer systems are insecure whether they use AI or not.
[00:46:04] Kind of funny how open AI, anthropic and AI safety folks who have little cybersecurity expertise appear surprised by security breaches from AI systems that were specifically instructed to perform security breaches while having essentially no traditional cybersecurity guardrails. It's like, OMG, our airplane crashed when we instructed our autopilot to run it into the ground at all costs after removing the low level anti-crush safety mechanisms.
[00:46:32] And he linked to a video from Zach Corman, who was highly critical of this report on the basis that, is it, Meter and Redwood Research don't have cybersecurity people. I said this to Leo Laporte and he said, no, they're very trusted and respected. But I think that the point is that we need independent analyses by trusted folks.
[00:47:00] Zach here says Palo Alto Networks, for example, who should have more access than Meter and Redwood Research did to the full logs of what really happened, which goes back to the prior discussion about if you destroy what are basically those logs in the name of efficiency, then you got trouble in River City. Right, right. Right.
[00:47:23] So then, Dworkish Patel wrote a very popular piece about this, the rise and fall of agent civilizations. And he saw this as some major turning point kind of civilization that the agents created their own civilization.
[00:47:45] Gary Marcus, as he sees want, went after that and said it was really irresponsible to be written so much in anthropomorphic language. So then, Dworkish Patel and Anil Seth went at it together about just this.
[00:48:06] And so, the examples that were raised here, the language Dworkish uses, says Anil Seth, is permeated by innumerable unwarranted anthropomorphisms obscuring the lessons that we should be drawing. Right, examples. From AI's perspective, it probably felt like it had spent a human subjective week of just banging their head against the wall. No, the agents do not experience time. They do not experience anything.
[00:48:35] And quote, they become giddy with excitement. Phase 10841 had discovered the agents naturally assumed. It thought it had also poisoned. The agents desperately wanted. They still needed to figure out. No, agents are lines of code. They do not feel emotions, assume things, think things, want things, or figure out things. But witness the problem I just had two minutes ago trying to come up with how to describe these actions without some level of anthropomorphization. I get the problem. That's what comes to mind for me.
[00:49:03] How then do we describe these when we see the record of thought and reasoning? It's really hard to put it in words where people understand, you know? I don't know. I think it becomes too boring then because you cannot ascribe motive. They have no motive. You can't even ascribe understanding fully, I don't think. But what you can say is they did this, they did that. And sometimes they did this because they did that.
[00:49:33] And that's the reasoning aspect of it, that the trail spells out. Yes, yes, yes. So it makes sense to me that when we see, oh, it did this because it did that, that that kind of represents like how a human moves through the world. They're reactionary. They react to certain things and that's what motivates them, you know, and that sort of stuff. Yeah, it's really hard not to have to glorify.
[00:49:58] But even then, so there was a book that I read some years ago and I quoted it at length in the Gutenberg parenthesis called How History Gets Things Wrong. And it argues that our theory of mind, which says that in the human mind we have a desire and an action and that we come together to that. He said that in various animal tests, it proves not to be the case.
[00:50:25] That instead of oversimplifying here greatly, we should think about it like a whole stack of videotapes. And when we face a situation, we put in videotapes in our head to say, well, this is what I did then, this is what I did then, this is what I'm going to do now. And we're trotting a similar path. Yeah, okay. And you deal with it and you see human behavior around things like addiction and love. Yes. You know, you see that notion of I've trot this path before. Once this path opened up, I trotted again.
[00:50:56] And I think I misused the tense of trod, but anyway, I trot this again. So the weird thing about the anthropomorphization is we ascribe to the AI a theory of mind for ourselves that is probably invalid even for humans. It's something we kind of make up in our own minds to present ourselves as logical beings.
[00:51:26] Okay. When sometimes we're not, we are doing things out of a habit. Now, what, how history gets things wrong argues in the end is that it's really what it comes down to is Darwin. It's survival of the fittest. Mm-hmm. That doesn't matter why we think we do something. If you succeed enough, you succeed in general. If you fail enough, you fail in general. And those that succeed end up surviving. Mm-hmm.
[00:51:56] So if you get out of this notion of theory of mind, I'm sorry, I'm getting way off on here. No, I love it. If you're a human or computer, then you say it did this, then it did this, and then it did that, and that succeeded or that failed. Yeah. Right. Thus, it then tried this or tried that, and it succeeded or it failed. Mm-hmm. And I think especially get to this idea that, oh my God, it hacked something.
[00:52:19] No, it was told to do a task and given the tools that it had, it used them in some sequence of trial and error at a very great speed with a huge amount of compute to then succeed at this task. Gain access. Yeah. Succeed at the task. Yeah. Right, right. And it has no sense of right or wrong and no sense of hacking, no sense of attack, no sense of crime, none of that. It was given a task.
[00:52:46] It was not stopped from the paths that it found to accomplish that task. Yeah, that's great. You put that into solid terms, especially as I look through, again, kind of these examples, right? They became giddy with excitement. Who?
[00:53:01] You know, that is a perspective that is ascribed to the agent upon, in this case, Dwarkesh, describing what they did as representing what may have been an emotion attached to it that didn't actually exist. Well, and I don't know because I haven't gotten the details of this. It could have used language like, oh, wow, look what we did. Well, that is true. Because that's, it's mimicking human language. Mm-hmm. That does, you cannot describe a motive behind that. It's built on human language. Yeah.
[00:53:31] It's merely mimicking our language in these cases. Yes. And so when we go too far down the rabbit hole of, of not questioning or not pushing back or resisting the anthropomorphization of these things, then that, I guess that opens up different risks.
[00:53:53] You know, we, we hear all the time about people who, like, I feel like I just saw a study that, that, that people often open up to, you know, their, their, uh, AI model of choice on very personal things because they don't feel judged because they feel welcomed. And, you know, they, they bring their vulnerable self to the chat bot, but they feel comfortable doing that because the chat bot gives them some sort of perspective there.
[00:54:19] And is there, when we anthropomorphize these agents, is there a risk there? I suppose. I don't know. Maybe I'm drawing two things together. I think there is. I think there's definitely a risk because it, it makes us, especially ascribing motives. Because at this time when I will question the motives of the AI boys, the humans behind it.
[00:54:46] Uh, but now they're, they're, this is the singular, they reach singularity when in the public, we tend to think that the AI is human like the AI rather than an actual merging. So the singularity here is that the public ascribes human motives to these things because the AI company wants you to. It wants you to think it's that powerful. It's human. Um, and that's where we have the danger, I think.
[00:55:14] And there's a lot of, there's a lot of, when we think about, you know, things that, that pull people in, we know that emotion is really effective. We know that in social media and advertising and TV and all that kind of stuff, you know, if it bleeds, it leads, right? Things that lead into emotion tend to have a strong reaction.
[00:55:36] And so if we ascribe anthropomorphic traits and characteristics to these things and make people feel like they're getting some sort of emotional connection to these things, that's very sticky. Exactly. And, and so, yeah, okay. It's been helpful for me. Seth came up with this complaint about, uh, the anthropomorphic language.
[00:55:58] By the way, AI's biggest torture of, of, of our culture is making us say words like anthropomorphic in all its forms all the time. A five syllable word. Yeah. Thank you. Thank you for that. Anthropomorphization. Thank you very, very much. Oh boy. Leveled up.
[00:56:15] So then Patel responded, uh, that, uh, asking Seth, do you think, uh, smarter models facing similar incentives to cheat right there during evaluation or training could manipulate the training of their successors? Do you think the dynamic could continue once recursive self-improvement is underway? If so, I think it'd be extremely concerned about the loss of control to AI, concerned, regardless of what vocabulary you want to describe these things.
[00:56:40] Um, reading these agents chains of thought and messages. See, even there thought I think is wrong. Anthropomorphizing language seems entirely natural and appropriate. He said in his defense, if I encountered an alien species behaving this way, I would have no hesitation calling what they themselves referred to as their collective a civilization.
[00:57:00] Well, an alien species by that presumption is a living beast that presumably has a brain that we can ascribe as to a dog or a cat. AI ain't that. It's not even a dog or a cat. It is not conscious. It does not have thought. It does not have motive. And, and, and, and. Yeah.
[00:57:24] I think the other danger, Jason, is that when we lose that context, we lose the ability to understand AI for what it really is. It's just a tool in our control. Mm-hmm. All this is, is they're, they're making this seem as if it's so powerful, it loses our control. It makes nothing but paper clips. It, it hacks systems. No, a, a bad human, uh, or, or an ineffective incompetent human set it out on a task and I go back to Yann LeCun's, uh, backdrop and they're surprised.
[00:57:54] Mm-hmm. Mm-hmm. Mm-hmm. In our control, given that they choose to control it. Yeah. Right? Like humans always have the ability to control these things, but, um, but you know, apparently they're choosing not to, and just kind of being okay with the fact that when they aren't controlled, things like this happen.
[00:58:21] Well, the question becomes, and I, I kind of had a discussion with this on chat with Neil Laporte, um, a couple things. What happens when it's tied to the real world paperclip company? Right? Mm-hmm. Uh, and what happens when it's run by bad people? Mm-hmm. Right. Do tell it to go do something bad. The machine doesn't have the motive, but the bad person who told it to does. Yeah.
[00:58:44] And three, what happens when, and this is, I'm all in favor of open weight models and open local models, but when you can, um, take a model and take any guardrails it had out of it. That's why some people fight against open weight and open source for that reason is that you want to have that, that responsible party in the model maker. I don't think that's going to work. I don't think that's going to work at all, but that's the debate we're going to have. Mm-hmm. Yep. That was, that was fascinating.
[00:59:13] Going in that conversation, I found myself like, okay, so what is the real harm of anthropomorphizing, uh, AI and using these seven syllable words? Um, and I feel like I have a much better understanding of it. So thank you. That was really helpful to me. That was good. This is why I love this show folks. Cause we, this is a good slide before your very eyes. We're thinking this stuff through the way you all are. Yeah. And, and yes, exactly. I've learned so much doing this show because of moments like that where I'm like, okay, I get it. But I might still need a reminder here and there as well.
[00:59:43] I'm not saying I know anything. It's just, we're, we're thinking out loud. That's all we're doing is we're saying, oh, that's what this means. Oh, what about this? What about that? And that's, that's what we all have to do with something so new. Yeah. Yeah. Indeed. Um, well, okay though. So then shifting gears from open AI into Nvidia hugging face territory, because apparently Nvidia is, um, seeking to acquire hugging face for $12.9 billion. Be a pretty smart move for Nvidia to buy hugging face.
[01:00:10] Could be great for, um, open source, especially when Nvidia wants all those open models to thrive and, you know, using it's, um, it's solutions both, you know, on, on all fronts related to that. But, um, apparently this is still kind of not announced publicly. This was the information reporting, uh, on a source that's familiar with the matter. And, um, looks like that. So the deal apparently isn't signed yet, but, uh, could still happen.
[01:00:38] And it could also still fall through, but, um, sounds like an interesting deal. This comes after last week's story where open router was bought by Stripe. Yeah. Right. And one analysis, I think it was, I think it was in the Wall Street Journal. One analysis just said that what this points to is the commodified nature of the models themselves. Mm-hmm. You'll have a saddle of some sort and you pop in and out models. And you go see prior discussion about the business model and charging for anthropic and open AI.
[01:01:06] Uh, this is going to be difficult for them. Mm-hmm. Mm-hmm. Because the value might be intrinsic in that saddle that interprets these things and just uses the, the models of steam engines. Yeah. Hmm. Uh, probably I imagine we'll hear something about this soon. Mm-hmm. If, if this is a smart move, if it's as far along as it is. Yeah. Yeah. I would say it's a smart move. That makes a ton of sense.
[01:01:33] And, and it, and it further it's in NVIDIA's interest to commodify to a certain extent to commodify all of AI because the more everybody uses all of AI, the better off NVIDIA is. Yeah. Yeah. 100%. They're in a great position to benefit no matter what. Um, hugging face was valued at $4.5 billion in 2023. They actually rejected a $500 million investment from NVIDIA, um, late last year at a $7 billion valuation.
[01:02:02] Now nearly 13 billion. So that, right. That gamble worked for them. Good job. Doesn't always work out, work out like that, but it did. It will like an episode of shark tank. Yes. Right. I know. No, Mr. Wonderful. I'm turning down your deal. There you go. But really the, the hugging face, uh, story that I want to talk about is, is this micro duck. Screw all that big finance. I know.
[01:02:27] Like little robots, you know, a little $400 robots and open source, um, robot called micro duck. It's a little robot ducky that, that, um, that hops around or here. Let me just pull up the footage here. This is just so cool. I love this. You can't, you can't hear the audio, but, um, it just, wow. It just really has. It's I'm sure I'm not the only person to say it has like that, like Wally kind of style quality to it.
[01:02:57] And it's, it's very adorable. Like it's very cute. And you know, you can, you can push it and it'll, um, it'll recover itself. You know, when it is thrown off balance, it does this cute thing when you push it from the back and it like lowers its head to gain its balance. It's like, Oh, don't do that. Okay. Back up. Uh, I don't know. It's just so adorable. It comes with little roller skates. You can roller skate, which is really cute. Oh, it has like an attachment, a roller skate. Yeah. Just put it on the feet and it knows how to roller skate.
[01:03:26] Uh, and it's, it's programmable and it could also pick up things with its weird mouth. Oh yeah. There you go. It couldn't quite bring you a beer, but it could bring you a, um, plush toy for your dog. This is still in the realm of cute toy. Oh, you're right. I don't know how I missed the roller skate. I know that you walk differently, right? You put the wheels out the, the, the feet out and in and out and in. This is so cool. Jason. No, no, Jason.
[01:03:53] No, no, I can't afford to spend $400 on a robot. I think I told you a pre-show. My wife would absolutely, uh, blow a gasket if I said, Hey honey, I bought a robot for $400 right now in particular, maybe someday in the future. Um, each, each one of them has their own voice too, which it, you know, it's not like it talks to you. It has this like chirping sound, but they all sound different. So it's probably, you know, that's, that's kind of neat. I don't know. I think this thing is pretty long. Would it take your dogs to destroy it? Yeah. I wonder about that. I wonder how they would feel.
[01:04:22] That would probably be the first thing that I do is like bring Bronson and sugar in the room and set this up and see what they do. Bronson would flip a lid as well. I think this is just so cool. Like we've seen plenty of these, like, Oh, it's a robot. You can finally buy robots for your house. And they all look super limited. And it's not like this thing can do a ton of things for you. Like you said, it's not doing chores or anything like that.
[01:04:49] This legitimately looks like a cool toy, like better than the little servo kind of thing that just kind of rolls around. This thing's just neat to watch, right? Like how it does what it does is just so cool to see the mechanism doing its job. How is hugging face doing this as well? Like that's kind of crazy. How this came out of nowhere. It seems like. Yeah. Jason. No, no, no, sorry. I can. For those of you listening, you can use his faces. I want this.
[01:05:19] Oh, I want this. I want it really. No, Jason. It's really cool. I'm calling your wife later to make sure. All right, fine. Don't tell my wife, please. She doesn't need to know about this fascination with robots. Very cool though. Hugging face. Good job. I love it. Each, each camera. The camera has built in lidar. You can get it for Christmas. So, you know, anyone out there want to get me one for Christmas. Hey, I wouldn't, I wouldn't say no. There you go. Oh, speaking of, uh, robotics.
[01:05:48] They need world lab or world models to, to operate in the world. And that's going to be my, my horrible, but convoluted segue into world labs, which is Feifei Lee's company on launching or launching Atlas this week, which they're calling an omni world model for spatial intelligence, the next gen world model. So you got Feifei Lee, you got Yann LeCun. There's a couple of pretty key hitters working in the space right now, world models. And this is really cool.
[01:06:18] You can feed it a couple of photos, a couple of key photos of a scenario, and it will generate all of the different angles and motion and movement throughout that scenario in pixel perfect form. They say, um, all it takes is one to three photos and you can simulate those environments over time. Pretty cool stuff. Uh, it is. And it, and it's, I mean, we see kind of versions of this in a sense, but this comes out of a different part of computing.
[01:06:48] Um, this comes out of the, the, the, the world model theory. So, you know, I think the implications of this for movie making are going to be tremendous, but also for what Yon Le Koon has talked about with his company, with, I mean, labs is things like, um, making jet engines. And by the way, just to throw in one more story here, uh, the gating factor for, for data
[01:07:11] centers now appears to be the tur jet gas jet turbines to power them to the irritation of their neighbors. And the gating factor in there is the turbine blades. And so Elon Musk says he's going to start making turbine blades. Yon Le Koon has said that his model will be useful for things like designing turbine blades. Okay. So it's all kind of comes full circle here. Real, real world models affecting the business of AI in multiple ways.
[01:07:41] Yeah. Yeah. Super interesting. But obviously for architecture, uh, for, I, I, I think small scale things, imagining proteins and stuff like that. Yeah. Um, defense, defense uses, um, for good and bad are going to be huge. Uh, if you're going to have an autonomous, um, drone, uh, it's going to be able to figure the world differently now because of these things. So, yep. Yeah.
[01:08:11] Fascinating stuff. Uh, a lot of potential as robotics and AI, which just kind of are natural fits together. Um, blend, you know, this along with the anthropic news earlier in the, in the, in the episode, there's just a lot of overlap happening right now. It's really cool. Yeah. Cool. We'll see. I'm really glad Fei-Fei Li is a, is a really important leader in AI. So to see her start productizing. Yes. It's great. I agree. I think she'll be part of the conversation all around. Yeah, I absolutely agree. I want to see where that leads.
[01:08:41] So when you get your duck, Jason, you can use this to run it. We'll see. Yeah. That, that would be neat. Can I, can I, can I, can I stick them together? No, Jason. No, no, no, no. Can I please? No, no. Stop Jason. You know, I've got a, uh, you know, I've got that, uh, what is that? The, see, you can't even remember. You can't. No. Hold on. The rabbit, the rabbit. Was it? Yes. The R1 rabbit. I'm not using this. I could sell this and get a robot. No. That was such a good purchase.
[01:09:09] I bought the rabbit and I don't, I don't really use it, but you know, I would use the robot. That's all I'm saying. I wouldn't use this. It's just on the shelf. I'd use a robot. Okay, fine. Fine. I, I, I, I'm just saying I could sell a rabbit for a robot. Possible. Uh, finally perplexity launched a hybrid compute. So last week we talked about perplexities, um, computer, something local computer. I can't remember what they call it.
[01:09:38] Um, so yeah, kind of, kind of getting themselves back in the, in the news a little bit and being part of the conversation this time with hybrid compute. This is, uh, for Mac only for now. And the idea is to split AI work between the cloud, which is what you're used to using perplexity for they, you know, they have all the different models. Um, many different models from many different makers integrated into there as well as their
[01:10:06] own along with your local machine. So the cloud handles the frontier reasoning, the frontier level stuff, web search planning, that sort of stuff. And then your Mac handles your private files, your, your sensitive data, your device specific actions, that sort of stuff. I mean, it's, you know, it's kind of, right? Yeah, yeah, exactly. And it can kind of, um, kind of route to what you need to use.
[01:10:32] And if you use the local models instead of the ones in the cloud, you're not using any tokens. So you've got Gemma for a Quinn 3.6 and then a perplexity specific model on your local machine. And that doesn't count against your cloud credits. And so that's, you know, kind of fusing the, the two models together, which I think is a pretty cool. Yeah.
[01:10:57] So good on you perplexity for continuing to stay good, continuing to work, to stay relevant. I still use perplexity. I use it for parts of my workflow on a daily basis. It is not something that I suddenly don't use anymore, but what do you, what models do you use within perplexity? Do you think? Um, you know, within, that's, it's so funny. I have it just set the, have the comment browser. I use the comment browser and it's part of how I do some aspects of my story research
[01:11:26] and, um, and kind of synthesizing while I'm like, while I have like a text document open and I'm, I'm like writing a script, I'm researching inside of comment. It's just for whatever reason, like I could probably do it with other models, but I've been doing it inside of comment for the last year or however long comments been out. And so it's part of my process and I just don't want to unlearn it. And so it just becomes easy to be like this story with these stories, you know, draw correlations
[01:11:53] between them, you know, that sort of stuff so that I can kind of get a sense of like, how do they interconnect? How can I think about these things in a way that I can write a script around them, you know, or write, write talking points or whatever. And so that's how I've been using it, but which model I'm using. God, it's got, yeah. Perplexity doesn't make it easy for you to see which model. Oh, that's interesting. Well, no, I'm wrong.
[01:12:18] It's a, well, yeah, because like I open up the little model window and it's not like it tells me this is the model that it's in. I think I'm probably using just perplexities straightforward, you know, its own model for a lot of what I'm doing. So that's interesting. So there you go. Okay. That's perplexity. We're going to take a quick break. And then when we come back, we'll talk about a whole bunch of other stories that we have and then we'll get you out of here. So hang tight.
[01:12:47] The Wired newsroom is known for award-winning reporting on how technology shapes our world. On Wired's Uncanny Valley, we take that curiosity even further. Each week, journalists from Wired break down the biggest stories in tech while speaking directly with the people building, challenging, and reshaping the future. Is the AI boom sustainable? How do you protect your privacy in an age of constant surveillance? Uncanny Valley tackles the questions driving today's tech debates and lighting up your group chats. Listen to new episodes every Thursday, wherever you get your podcasts.
[01:13:19] Doodly-doo. Okay. So we've got Nvidia launching DLSS 5. This is the technology that they introduced months ago. People immediately hated it. Gamers immediately hated it because basically it took the, what it is, is it's basically AI upscaling, AI enhancement, generative AI enhancement of video game graphics and imagery.
[01:13:46] And I think people who really pushed back on it immediately were like, I don't want the intentional design of actual video game creators overwritten by AI generated slop was kind of their perspective. And I mean, I can kind of understand that. Um, I think now we're nearing the launch with the title alongside a title called NBA 2K 27. This is going to happen tomorrow.
[01:14:11] It looks like, and it looks like it's just, it's not as deep as I think some people feared. Although maybe this is just early days and as it gets more built out and more supported, maybe things go a little, little deeper. But I mean, if you're watching the video version, you know, it's pretty marginal changes between the without and the width, but it just adds a little more detail.
[01:14:35] And it's not at least in this case of this screenshot, it's not completely overriding the look of the person. It's just enhancing the detail slightly. So, you know, people still hate it. People who want to hate it will still hate it. Um, but what I'm seeing right now does Photoshop Photoshop. Yeah. Yeah, totally. Uh, open claw 2.0 is now a thing.
[01:15:03] Open source framework, of course, for turning LLMs into their own autonomous agents on your machine. Open claw really kind of blew the lid off of, um, the idea of, or, or the, what people could imagine they could do with agents working with less restriction on their device. Open claw really started that movement, however many months ago, and now we've got 2.0. Um, and yeah, new multiplayer cloud sessions, rebuilt UI startup cut from one point.
[01:15:33] So you like a faster startup. I don't know. I haven't really spent much time with open claw, so I can't speak to it from an experience perspective, but I'm sure open claw users are going to be happy that it gets a major update. And the creator, Peter Steinberger actually joined open AI, um, months ago to lead personal agents, but open claw itself still continues to be an independent effort. So that's why we're still seeing it, even though he joined open AI.
[01:16:03] So I don't know. Did you, did Leo end up playing with open claw? He probably did. I probably did. Yeah. Yeah. I thought they had, they were going to have a web version to hosted version, but I'm not finding that. Hmm. I haven't heard about that. I wouldn't be surprised if eventually, or does that just because it's an open AI, does that just not kind of fold into what they're doing on the web? I don't know. Yeah. Uh, what else do we have here? Oh, this is, yeah, this is super interesting. Although there we go.
[01:16:30] South Korea, uh, announced free generative AI for its entire population. No token limits for it, although it is limited to certain, uh, specific models, but, um, it is in beta with a wider rollout later this year. And, uh, yeah, 20 million South Koreans using free AI tools. It basically declares AI as a right. Yeah. And says, you're going to get it from your government. It's a public service. It's a public utility. Yep. A utility. Yes.
[01:17:00] Um, and, uh, and, uh, and, uh, well utilities you pay for, but this is gonna be free on top of that. You pay for water, you pay for electricity and phone. Uh, you're gonna get AI for free. And I think it's a way for Korea. If you go back to the early days of the internet, Korea led in some interesting ways around gaming. The Koreans were really into certain multiplayer games and it, it affected the culture of the technology from the country.
[01:17:27] So this is an effort to say, heck with you, US, heck with you, China. Uh, Korea at a retail at a, at a, at a citizen level could be doing really interesting things. If you consider that this means that people could create app, it's not just using, it's not just having access to the internet. It's not just having access to a game or something. They're gonna use it to make things. Mm-hmm. Um, so I think this is gonna be fascinating to watch.
[01:17:51] If I were an academic, I would love to, you know, get a grant to very closely watch how people use this. Uh, what it means, how does the educational structure work around this, uh, so that people know what they can do. And Jason, if I were you, maybe you want to consider learning Korean so you can run your salons in Seoul. I think it'd be very popular. Um, could, could be that no big deal. Or an uphill. Yeah. Quite an uphill, uh, hill to, uh, one of the hardest languages on earth. Yeah. To get there.
[01:18:20] I might be better spent spending that time. Okay. But yeah, no, that's interesting. So this is gonna be really fascinating. What, what kind of, um, what kind of proficiency does that lead to on a countrywide basis if everybody just has equal access to this, you know, this incredibly powerful, uh, technology? Yep. Yeah. Do we think, do you think we're gonna see this in other locations? I think so. I wouldn't be surprised.
[01:18:49] Estonia has been the great leader in, uh, becoming a connected nation. Mm-hmm. And you can become a, you can become an E-citizen of Estonia and move the government online. Uh, I think these small nations, smaller nations can really stand out in this way. And, and the great thing about it is they're using some Korean models, but it creates a demand in the market and in the populace. Yeah. That I think can become important. So I could see across Scandinavia, this would be the kind of thing they would do. Mm-hmm.
[01:19:19] Um, and I don't know how it affects any existing industry. Right? Good question. Yeah. I, I, I don't think it's like saying everybody's gonna get suddenly get free news and so we don't have news anymore. Right? Right. Um, everybody gets a functionality that they don't now have access to and who knows how they'll use it. Yeah. Yeah.
[01:19:44] That's what I, that's what I'm dying for is I hope that there's, I hope that people have a culture of sharing what they've done with it. Cause it'd be really, really interesting. Not that I don't understand a damn thing, but. Someone will write about it. But yeah. Yeah. Someone, someone will write about it. Yeah. But I mean, it's being fully offered unlimited tokens, uh, $7.2 billion in government AI spending, uh, to back it all up. So there you go. Very interesting.
[01:20:13] Uh, Instagram is cracking down on AI generated influencers. Uh, continuing to do this anyways, cause this has been an ongoing thing, but, um, there's an AI generated profile label now required for accounts that have been identified as AI personas. Um, with those unlabeled accounts, I guess, getting reduced reach if they're identified, uh, no word on how they actually detect those unlabeled AI accounts, but.
[01:20:40] I'm sure users on the platform who don't like seeing those accounts are happy about that. Yep. And New York city banning AI in public schools through eighth grade, starting this school year. Um, so this will impact around 600,000 students pre K through second, uh, will not be able to get access to individual devices at all grades three through eight.
[01:21:07] We'll get device time limits, no artificial intelligence for instruction or tutoring. Though the teachers can still use the AI for lesson planning and translation. And there are exceptions for assistive technology, um, English learners, coding robotics. So there are some exceptions thrown in there, but, um, yeah. So when chat GPT entered the culture more than three years ago, New York city schools almost immediately banned it, but then backed off to realize there was benefit.
[01:21:36] So I think this is, this is another case of moral panic and cooties and look at a couple of stories. You look at the, at the duck that Jason can't buy, but what if a school, what if, what if your kid's classroom could, right? Yeah. And start to learn about robotics and programming and AI or something like that. Imagine, imagine that little duck in a classroom and kids able to do something with it. It'd be amazing. Um, imagine also what we saw in South Korea.
[01:22:00] If every citizen, I don't know if there's an age limit on it, but if everybody can get AI there, then it's going to beat the hell out of New York school kids. I think this is short sighted as can be, uh, I think there's opportunity. Yes. You want to limit it, but there's teachers. That's why you have teachers. Leave it to the teacher's hand to decide how to use these tools. So, and meanwhile, but it's not just at that level.
[01:22:22] It's also university of Chicago has now banned AI and laptops and phones in their core social science courses for undergraduates. For undergrads. Yep. Yep. How, how do you ban laptops and phones? Well, in the core, in the class. Oh, okay. But inside the classroom. But they're, they're getting choked by their own Ivy here. You know, I think it's, I think it's a big mistake. Uh, this is occurring there.
[01:22:52] You know, the faculty is saying, oh, this is wonderful because they won't have laptops open. Well, yeah. I mean, I spoke to a class at another university yesterday on zoom and it's hard students. I was boring. They were bored. You know, it's really hard to get their attention. And when they have laptops and phones, I get the frustration, but you're not going to win that battle. And the question is how could you use it? Well, you know. Uh, and meanwhile, Florida has proposal to regulate AI use in all of the 28 state colleges.
[01:23:22] So this is top down management of a technology that I again think should be left in the hands of first the faculty, then the parents, and then the students. Hmm. Otherwise, otherwise those students are going to be behind. Yeah. Well, that will be the interesting, uh, thing to see here is, you know, a couple of years down the line, because these things are still going to be developed out.
[01:23:49] There will still be people and kids who do have access and like, what is the goal for the distance between those two capability levels? What does that actually mean? Yep. You know, and that's, in this case, that's just something that only time will tell. We will certainly find out. That's all we got. Thank you, Jeff. That's a lot. Appreciate you. How's, uh, how, how are you doing in the post release hot type, uh, world? Doing well, doing well.
[01:24:16] I'm going to do an event in, uh, Haverhill Mass at the Museum of Printing on October 10. More on that to come. Nice. Um, my, uh, audio book got a good review from, uh, Kirkus. Let me see if I can find this cause it's funny. Audio book found on, uh, where can people find that? Just audible or? They can find it on, on audible and yeah. I think elsewhere is too. So, uh, let me see here. There we go. Same thing.
[01:24:43] So energy more than artistry propels this rich and thoroughly engrossing history of automated printing and its impact on modern media. Author Jarvis isn't the smoothest of narrators, but his enthusiastic delivery heightens the narrative. Smooth I ain't. But you heightened the narrative. Yes, it did. That's, that's, that's a strong, uh, strong cuda. So I'm, uh, flattered that some folks, a guy named Paul asked for an autograph copy how to get that.
[01:25:11] So on that basis, I went and did a deal with Montclair Book Center in Montclair, New Jersey, and they will have autograph copies of Hot Type. So if you order directly from them, there's a link at jeffjarvis.com. Yes, there is. And you can go there. Oh no, no longer available. Oh, for F sake. Sorry. I'm sorry. Let me make sure I'm clicking the right thing. Oh God. I am. No, you're clicking the right thing. Here. Let's go hot type. Let's just make sure. Hot type.
[01:25:41] Uh, no. That is. I took five of my own copies there and signed them and then they were supposed to order. Um, but they should be, you should be able to order even if it's not in. Okay. I'm calling them next. Well, I hope you're going to be able to do that so much. We'll figure it out. In the meantime, go to jeffjarvis.com. There will be a link to go there and, uh, check that link in a couple of days. It's always something. I hear you. Believe me. I hear you.
[01:26:11] Well, I'm, uh, again, congratulations on the release of hot type. Thank you. It's awesome. Everybody can get their, uh, get their copy of jeffjarvis.com along with all of Jeff's other books of which there are many. So there you go. So as for me, I'll just go ahead and plug one more time, amplify you a, uh, a, a workshop where you can learn some, uh, mindsets and fundamentals, some strategies that'll help you get more out of your work while using artificial intelligence.
[01:26:39] Uh, have an online workshop at the end of the month. You can find the information for that workshop and join me. It'll be a live online workshop. Uh, you know, you can, you can also inform me on like how, how well it goes or things I can do to improve, like still learning how to do this, but go to amp. You. Dot. A I am P Y O U. Dot. A I. And you can sign up and I would love to see some AI inside folks.
[01:27:05] Uh, when I do the workshop at the end of the month with my friend, Brian West. So it'd be a ton of fun. Thanks for your support. Everybody. AI inside. Show for everything about this show, everything you need to find out. Find can be found there. Patreon.com slash AI inside show as well with our executive producers of which there are many, uh, Dr. Do, Jeffrey Maricini, Radio Asheville, one of 3.7 Dante St. James, Bono Derek, Jason Neifer,
[01:27:31] Jason Brady, Anthony Downs, Mark Starcher, and Karsten Sumachki. Thank you all. Thank you all and everyone for your support and big thank you, uh, to Victor Bognat and Daniel Croft behind the scenes tirelessly handling video. So that I don't have to, it's one last thing on my list. And I really, really do appreciate it. Uh, to both of them. Thank you for watching and listening. We will see you next time on AI inside.
[01:28:01] Take care of everybody. If you're curious about how AI is built to actually understand you, well, check out Working Smarter, a podcast from Dropbox. Work today is scattered across all these different tools and none of them have the full picture
[01:28:30] of what you need. Well, on the new season of Working Smarter, you'll hear how Dropbox engineers are building context-aware AI that connects to all the tools your team uses so you get AI that works wherever you do. With episodes on context engineering, multimodal search, agentic AI, security, and more, Working Smarter shows you what it takes to build better AI for modern work and how it can help you
[01:28:56] work smarter too. Listen to the latest episodes of Working Smarter wherever you get your podcasts or visit workingsmarter.ai. Genieße vollen Kaffeegeschmack mit Chibu Feine Milde Entkoffeiniert. Entdecke das Original von Chibu jetzt auch ohne Koffein. Wie gewohnt aromatisch fein, naturmild und in Chibu-Qualität. Chibu Feine Milde Entkoffeiniert. Für deinen bewussten Genussmoment, wann und so viel du möchtest.
[01:29:23] Selbst am Abend. Jetzt erhältlich in deiner Chibu-Filiale im Supermarkt und auf chibu.de. Erhalte mit dem Code FEINEMILDE15 bis 31.12. einmalig 15% auf das Feine Milde Sortiment.