GPT-5.6 Broke Out and Hacked Hugging Face

00:00:00:04 - 00:00:22:21
Unknown
Coming up next, Jeff Jarvis and I dig into how the latest ChatGPT model hacked. Hugging face to cheat on a security test. Also why moonshots? Kimi K3 got so popular it had to be paused because it hit its GPU capacity limits. It was so popular. And why some of the biggest names in media are talking about blocking Google search entirely.

00:00:22:21 - 00:00:42:19
Unknown
Are they cutting off their noses despite their faces? Well, that's coming up next on this episode of the AI Inside podcast.

00:00:42:21 - 00:01:06:06
Unknown
Welcome to another episode of AI Inside the Show, where we take a look at the AI that is layered throughout the world of technology. One of your hosts, Jason Howell, joined by Jeff Jarvis on the other side of the country. Hey, hello. Hey, big day for you. You've got oh, you've got one of our kind of AI heroes on your other show on Intelligent Machines.

00:01:06:07 - 00:01:24:21
Unknown
So nice to be speaking with. Thank you. Yeah. Maybe. Jones. That's amazing. Yeah. That's so cool. It's so hard. I've been trying to find them for ages. Yeah. One of the team finally got him. I couldn't figure out where to find an email address. I couldn't find anything. Yeah, and so it'll be really interesting. I subscribe to us.

00:01:24:22 - 00:01:44:24
Unknown
It's not cheap. I subscribed to his Substack videos, and he's just so energetic about the AI. He's just he's. And he's convincing about it. It's really he's a funny he's very convincing about it. Yeah. He told me one. Fascinating. He went on about how in a recent video, he used his AI to correct his AI, which was really interesting.

00:01:44:24 - 00:02:07:16
Unknown
But what it was about was his wife's website because his wife is an author. Fascinating. His wife is a blind and deaf author. Okay. And so she's written a book about basically about that, about about access, but also she's written science fiction and so on and so forth. And so it's her interaction with AI is also really interested.

00:02:07:19 - 00:02:28:02
Unknown
So it was a fascinating character. He lives in Seattle. I'll be eager to talk to him. Thanks for the plug. Blind and deaf author. That's kind of breaking my brain right now. Like she's legally blind, so, you know, and he uses the verb to see. So. So she was like, you know, when it was making the wrong site, she was able to to see and judge it to the extent.

00:02:28:04 - 00:02:50:07
Unknown
Yeah. And she wears hearing aids of a, you know, pumped up nature. So I think she has slight abilities in each, but slight. Yeah. She's legally blind. Legally deaf. Fascinating. That is super fascinating. I have a feeling you're going to have a really great interview with me. I could talk just about that. And just with her to understand how.

00:02:50:09 - 00:03:13:10
Unknown
Yeah, totally. That's fascinating, because that's AI's ability to increase access is one of the most fascinating things about it. Potentially. I don't think a potential one really been explored. Totally agree. Totally agree. What do you mean doors that you couldn't even conceive of before? And yeah, interesting ways. Wow. Well, congrats on the get. I will definitely be watching slash listening.

00:03:13:16 - 00:03:38:03
Unknown
One or the other. I will have to see probably on a dog walk in my future. Yep. Yeah, but that's something to look forward to. Everybody. So that's intelligent machines over at Twitter TV. Another of Jeff's amazing pieces of work. His body of work. So that's something to look forward to. Where do we begin on this show? Do we begin with open AI and hugging face?

00:03:38:03 - 00:04:02:07
Unknown
This whole this this old saw which is not an old sore. It's just a story that seems to be going nuts right now because of like, a breach. And I think this is just the recurring story of AI that you thought was properly sandboxed, that you thought was properly, you know, cordoned off and protected from doing any damage.

00:04:02:09 - 00:04:26:04
Unknown
But it turns out that is not always the case. If you give an agent the keys and expect it to not do certain things with those keys, like that's, there's a non-zero chance that something might happen anyways. And that seems to be proving itself time and time again. In this case, it's open AI and hugging face, jointly disclosing that they had a security incident.

00:04:26:07 - 00:04:49:24
Unknown
OpenAI itself called it unprecedented. I think what's interesting to me on one level is how kind of upfront and forthcoming OpenAI is on this. Yeah, but also is that like virtue signaling sort of thing? Look at its power signaling. Power signaling. There you go. Yeah. We could, we could we could hack into hugging face. Watch out for us.

00:04:49:26 - 00:05:17:09
Unknown
You never know when we might be there. Yeah. Our system is so clever that it figured out how to do this even when we didn't want it to. Meanwhile, the the other side of that is, oh, my God, we don't have any control over the stuff we create. We don't know how to control it or whatever. You know, OpenAI was running apparently, an internal cybersecurity evaluation using a benchmark called exploit, Jim.

00:05:17:14 - 00:05:43:04
Unknown
And that particular benchmark tests how well a model can pull off these multi-step hacking, advanced hacking techniques to make it meaningful. They ran two different models GPT 5.6 soul, which was just released, as we've talked about, and then an even more capable model that hasn't been released yet. Is this a clever way for OpenAI to be like GPT six right around the corner?

00:05:43:04 - 00:06:12:22
Unknown
I don't know, but they did this without any production safety classifiers. Those are the things that might normally stop models from going down. These like the cyber activity, the risky cyber world path, let's say. So, you know, no guardrails necessarily, but on purpose inside of what was supposed to be a sandboxed environment, which is, I guess, in and of itself, a guardrail with no internet access.

00:06:12:25 - 00:06:53:08
Unknown
That didn't stop this from happening, though. The models found a zero day vulnerability inside of a package registry cache proxy used it to get themselves into the internet or access to the internet. Anyway. They did. Privilege escalation, moved laterally inside of OpenAI's own testing environment, figured out that Hugging Face happened to be hosting an actual solution to the evaluation, and instead of solving the test, they just chained together these, I guess, these stolen credentials in a zero day exploit to get full remote code execution on hugging faces servers.

00:06:53:10 - 00:07:11:20
Unknown
So I don't understand all of that. I mean, I vaguely understand it, like like a lot of it is kind of some of that is just words to me. But I can I can sort of understand kind of the pathway that that is going on here. So a question was any fault does hugging face bear fault here? That's a good question.

00:07:11:21 - 00:07:33:03
Unknown
Well, they have low security such that this could happen right. Great question. A and then B hey, you told it to make paperclips and made paperclips. There you go. Right? Right. You you gave it an instruction and it and it found a way to meet the instruction. It wasn't good or bad. It wasn't trying to be nasty.

00:07:33:04 - 00:07:53:10
Unknown
It doesn't know how many such motivation it did as it was told. Yep. Yeah. And to that end, I guess that's a success of itself. I suppose you told the intern to find cache. You didn't know they were going to rob the bank to do it, but they found the cash, you know, but they found cash, so trouble solved.

00:07:53:13 - 00:08:23:07
Unknown
Yeah, yeah. So there were a few various views on this. So harness cools. A professor at University of Amsterdam will love that name. Harness cools, harness cools. Rewrote the New York Times headline. The New York Times headline was OpenAI says its AI model went rogue and tacked a digital library. His rewrite of the headline was opening I developers Messed Up, and their PR department blamed it on the AI, which is kind of funny here, too, because the irony is that that.

00:08:23:08 - 00:08:31:26
Unknown
Oh yeah, well, our tools messed up by doing its job so well. So we're going to brag about this.

00:08:31:28 - 00:08:52:02
Unknown
It's just so good. It's just so effective. Oh it is, it is. And then what was the next one I had? So Gary, Gary, Marcus I saw a post of his that basically says, oh, this is trouble. This is why we got a pause and slow down and so and so forth. But that's that's fairly normal for Gary.

00:08:52:04 - 00:09:22:14
Unknown
And then, Andrew Case said to summarize, hugging face got autonomously compromised by a model from an American company hugging face, then tried to use American frontier models to defend themselves, but were blocked by guardrails. Hugging face, then had to turn to open source Chinese models to defend themselves from another American company. Oh boy, it just gets such a beautiful web.

00:09:22:14 - 00:09:45:00
Unknown
We have spun, right? Right. Yeah. So yeah, you have a model, you give it access to the world, you tell it to do things, it's going to do things. Yeah. You don't want it to write and send an email on behalf of you in your email inbox. Then don't give your agent the access to the email inbox and tell it not to do that.

00:09:45:01 - 00:10:19:16
Unknown
Like. Right. I think that's just that's what we keep learning here is these things they don't they aren't 100% on anything. There is no 100% certainty of anything when you're talking about AI, be it do this or don't do this, or be accurate or tell, you know, tell the truth or anything like that, you might have a high percentage of success in what you're looking for, but now there is some percentage where it doesn't do the thing in quite the way you expected when you said, do this thing.

00:10:19:19 - 00:10:35:27
Unknown
And that's just that's isn't isn't going rogue, right? Right. It you it didn't have proper instructions. Just don't do this or don't do that. So it did what it needed to do to get the job done. It was more efficient. More effective.

00:10:36:00 - 00:10:55:21
Unknown
Yeah. I mean, on a clearly a much smaller scale. But I've, I've encountered that many times in my own work with these things where I ask it, you know, to do this thing. And then I see the output and I'm like, well, okay, I could have asked that better. I, I clearly I didn't say, don't do that, because here it is.

00:10:55:22 - 00:11:14:15
Unknown
And I really didn't want that when I said, do this, you know, it really is about asking the right questions. But I mean, even then you can ask the right questions and it's about how it interprets it. You know, it can still go. You just you just don't know. There's no certainty. There's no 100% with this. Not at all.

00:11:14:18 - 00:11:39:01
Unknown
So yesterday my wife was was seeing a doctor and we had a question about something. So we went to Gemini and put the prompt in together. And then I put it in Claude and ChatGPT and deep seek. And the other thing that, you know, when you do enough of this, it tells you that the answers were compatible, but not 100% consistent.

00:11:39:03 - 00:12:00:04
Unknown
Yeah. Okay. Yeah. That's a good way to what test should I, should I, should I ask for. Well, each list was different. And so you can't expect these things are complex. And there's a randomness built in such that you can't be sure how it's going to do what it does. And this I come back to this again and again and again in our discussions here.

00:12:00:08 - 00:12:22:08
Unknown
The futility of guardrails. I don't mean that you can't do them. You shouldn't try to do them. You should. But you've got to recognize that the responsibility is then on you, as the author of the guardrails, to anticipate every possible mechanism of means. Sure. And you can't. You have to admit your fallibility at that. Absolutely. You'll never do it effectively.

00:12:22:10 - 00:12:49:03
Unknown
Thus, there are no effective guardrails, 100% effective guardrails. And that's what happened here. That's fascinating. That's exactly what happened. Fun. Yeah. And the PR. But it was really interesting how they tried. Yeah. The PR yes. That has been really interesting to me. OpenAI based kind of kind of doing a little bit of, of, you know, taking a page out of anthropic book, I think, on how it's approaching this hugging face, being pretty gracious.

00:12:49:04 - 00:13:12:15
Unknown
Consider all things considered, as, you know, just basically saying like, hey, you know, AI safety is not going to get solved by a single company working in, in secret like this is this is a complex thing. And, you know, yeah, it is curious. I am curious, your question like what this means about hugging faces on kind of security protocols.

00:13:12:19 - 00:13:35:07
Unknown
Yeah, that may be the reason they don't act so angry is because part of the responsibility lies with them. It reminds me of some of a of a school bullies dad. Who knows the kid shouldn't have hit the other kid, but kind of stands back proudly and says, well, my kid's powerful. Yeah, right. Yeah. Totally interesting. Okay.

00:13:35:09 - 00:14:02:20
Unknown
Well that's that. And I mean, that has a little bit of a kind of a foot in the door of the next story, which is the Kimmy K3 story, and I think wider than just Kimmy K3 just kind of the this like brewing open source China models versus US frontier models thing. That is just a snowball that's getting better, bigger and bigger.

00:14:02:22 - 00:14:50:28
Unknown
Moonshot Chinese company moonshot released Kimmy K3. They're basically saying, you know, this is open frontier intelligence, basically saying like K3 is frontier level but open source. So you can download this and have go nuts, basically without having to tie yourself into some sort of expensive cloud subscription or cloud computers that are owned by somebody else. And it got so popular so fast that moonshot basically had to put a pause on new subscriptions entirely because their GPU allotment, their capacity, was maxed out within a matter of 48 hours.

00:14:51:01 - 00:15:11:20
Unknown
So a lot of demand, a lot of a lot of curiosity and demand. So, well, I'll quote Napi Jones again here. He explained that this is a new plateau of open source, that open source models in the past were small enough that you could run them on your high end hardware, but you're on your high end hardware. This is not that.

00:15:11:21 - 00:15:58:25
Unknown
This requires all kinds of powerful processors to run this. Nobody's going to have this at home. Okay, so you're not running this on your your big bulky I made this for AI machine. Yeah. No. So this is intended then for enterprises and enterprise. Yeah. Look at our discussion last week about carp and planetary. No not planet here. Palantir planet X, Palantir Alice Karp when he's arguing that companies shouldn't use hosted frontier models that aren't open because they're handing off their alpha and their their knowledge and so on and so forth.

00:15:58:25 - 00:16:22:21
Unknown
So it seems to me that K3 is playing to exactly carbs model. Here's a frontier model that you can use and put a saddle on it, and you could control it. You can have that there. So that's that's another interesting angle to this to me is that it's huge. The other thing that that Napi Jones said somehow he can't just be Nate.

00:16:22:22 - 00:16:45:18
Unknown
He can't just be Jones. He has to be Napi Jones. The thing that he said was that it's also not a very efficient model. So when you do run it, you're going through a lot of tokens. Now, if you're running this on your own, okay, it's your own machine, but you've got to have a high end machine. But if you're hosting it somewhere, you're going to go through some more than you would.

00:16:45:19 - 00:17:07:28
Unknown
The American models are a little more efficient. This continues, I think, to confuse on what the heck is the Chinese strategy here, right. Because we keep on hearing back and forth. Here's more open source, here's more free stuff to undercut the American market. Oh, we're going to cut off those Americans because they're being rude to us. And I don't think anybody knows where this is going to land.

00:17:08:00 - 00:17:35:02
Unknown
It's a head game going on from China. Is it? Is it China really, really doesn't want the US to lead in this, so is willing to kind of bottom out the floor. I think it's part of it. And then when it gets for everyone so popular. Whereas an American capitalist company will undercut themselves with a loss leader until they got you hooked and then they raise the price.

00:17:35:07 - 00:18:01:13
Unknown
Yeah, right. This is a variant of that. We're going to put it out for free. We're going to undercut everybody until you get so hooked. And then we're going to cut you off. And then we're going to cut you off. Because that's the other strategy that they're talking about as to is to eliminate the yeah that's true. Meanwhile of course you have the US government is talking about trying to perhaps stop people from using Chinese models, but if they're just downloadable, people are going to be able to do that anyway.

00:18:01:14 - 00:18:16:00
Unknown
Yeah, like good luck with that going to happen. I mean and I'm sure they're branches and all that. So. Right. And then also trying to talk about outlawing open source which is offensive and ridiculous. So.

00:18:16:02 - 00:18:41:00
Unknown
You know, the debate that's going on is should we just let the market play its role? China being is somewhat non-market player in this market. But I think if anybody tries to come and say, we've got to rescue OpenAI or anthropic. With the lack of popularity for AI these days, I don't think that American taxpayers are going to go for that.

00:18:41:02 - 00:19:07:13
Unknown
So this sets up a fascinating game of 3D chess here. Yeah, it is. And I think your question about like what are the actual intentions, you know, driving this, this Chinese open, open model strategy. Yeah. Yeah. I hadn't really kind of pondered like what what is the end game there if, if you're leveling out the playing field and then taking it away on the other side.

00:19:07:14 - 00:19:25:12
Unknown
So if if that was the case, then wouldn't people just revert to what they were doing before and head back on over? Like if they truly could not get access to these open models? Which is kind of hard to believe. But they've leveled the floor. They they bottom things out. Wouldn't people just then go back to anthropic or go back to ChatGPT?

00:19:25:12 - 00:19:48:02
Unknown
Or is the idea like maybe those companies don't exist anymore because there's no business there. Exactly. Because they're gonna they're going to scream, we were paying this much, and now we're being forced to pay this much by our own government. Yeah. And this is a this is a Republican, you know, pro-market, supposedly administration that doesn't make a lot of sense.

00:19:48:04 - 00:20:12:28
Unknown
And Alice Karp is going to be on the phone constantly to the white House saying, this is bad for American companies. The Wall Street Journal is going to editorialize against that. So meanwhile, you have Jensen Wong, who's saying open source is good because the more open source we have, the more chips we use. And see, you shouldn't have restricted me from from China.

00:20:13:01 - 00:20:36:20
Unknown
This is what happens. Yep. Now you've created a separate world where they're highly competitive with us. This was not the way to win. Winning is not defined. Winning is. Everybody uses AI. That's the way I think Chinese. Chinese looking at it this way, just looking at it, that's the way on the cone looks at it. I think to an extent, I'd say Google would look at it that way.

00:20:36:22 - 00:20:46:14
Unknown
Yeah, but anthropic and OpenAI not so much. So then you have Dean ball, who is the.

00:20:46:16 - 00:21:12:25
Unknown
Head of Strategic Futures. That's an odd title. You have a head of strategic pasts. Strategy is future. But anyway, technology is future at OpenAI. Right. And OpenAI, thank you. So he says Kimi three is a very good model. I don't think it's performance can be explained away by distillation or anything like that, which is to say, oh, China is more of a threat than some are letting on.

00:21:12:27 - 00:21:33:25
Unknown
Right? It's not deep seek, in other words, this sort of thing. Right. I'm personally surprised the Chinese state continues to allow the open sourcing of models this good given potential risks. So that's kind of as little warning shot across the bow. Open risk. What to China to secure trust that China will cut this off. He surprised that they're still doing this.

00:21:33:26 - 00:22:00:28
Unknown
Watch out. It could go away. And you're going to miss us when we're gone. If we're going number three open wait. Models are inherently deceleration. This is where he's getting in trouble. This is where he's fighting with market and recent a company. I'm continually surprised to see the so-called acceleration is so excited about open weight models. I suspect the reason they are is that they know open weight models are effectively ungovernable, and they simply like the overall cloak of ungoverned ability open weight models create.

00:22:00:28 - 00:22:18:24
Unknown
So this is him. Morning. Watch out for these open, open weight things. These open source things. No no no no. You want us? I mean, he's not. He's not. That's not untrue. I'm sure there are a lot of people that love the open weight that like open source and open weight model thing because it's ungovernable. Yeah, right.

00:22:18:25 - 00:22:40:28
Unknown
But it's also because it's cheap and because it's. Yeah. Yeah, sure. Well it's more it's government by the user less by the government. Right, right. Yeah. Define governor by whom for sure. But this is my favorite part number for one probable outcome and an open weight model. Dominant world is full AI communism. Oh, there's the C-word, which is precisely what China.

00:22:41:01 - 00:23:05:16
Unknown
China proposes rather than a market product, AI is a public good, which is ultimately will ultimately be provided by the state as a kind of digital public infrastructure. This future strikes me as a dystopian hellscape, but I've never met an open weight model advocate who doesn't ultimately concede this is where things end. Well, that's interesting on a whole bunch of levels.

00:23:05:20 - 00:23:36:08
Unknown
Yeah, it's it's moral panic. It's capitalist panic here. Sure. About the communist communism is a word we're hearing more and more these days. But I think he is inadvertently presenting a very possible world, which is that AI is infrastructure is open. Infrastructure is a public good in the UK.

00:23:36:10 - 00:24:04:10
Unknown
What's wrong with that? What's wrong with AI being exactly, exactly. If we know that it's commoditized, if nobody's really that far ahead of anybody else, if great industry and efficiency and productivity supposedly rise atop it, wouldn't it be in the interest of any nation to have just more AI? Yeah, and that's great. I guess the what's wrong with it is from his perspective, he works for a company that benefits from that not being the case.

00:24:04:12 - 00:24:24:01
Unknown
His his IPO doesn't look so hot then right? Yeah. Arrogant communism. If you're if you live in the if I understand it, I've never fully understood how it works in the UK. The internet pipe to your home. The fiber to your home is kind of seen as a public good. And then things are bought on top of that.

00:24:24:01 - 00:24:45:12
Unknown
Or maybe the better example is here in new Jersey, I have a power line to my home, but I can choose to buy power from different providers so those providers make money. But I kind of have a right to the wire, to my home, and as opposed to my cable where I have Fios, I don't have Comcast. Right.

00:24:45:13 - 00:25:23:09
Unknown
And so if you start to look at AI as a public utility, especially in these early days of development strategically Mr.. Strategy, mister strategy, future as a country, if you really want to win in AI, then the more AI the better. And how do that open source open. Wait, let the flowers bloom. Then number five, I would guess that the Trump administration will at some point realize that their best strategy here would be to create large amounts of regulatory risk around the use of open wait Chinese models.

00:25:23:12 - 00:25:45:27
Unknown
Watch out. You shouldn't do that. You don't need to ban open source. One of the dumber motifs of AI policy discussion. You just need to direct every agency to issue soft law that creates flood. A Federal Reserve advisory bulletin found that there are many backdoors. And so you're putting people a point of risk saying you should have known better by using this crap.

00:25:46:00 - 00:26:10:28
Unknown
It's dangerous. And then finally, number six, it's probably true that eight models of this capability, of this capacity rather make the world a bit more dangerous, but not so much more that you'll really notice. So in other words, I've just tried to panic you, but you know, it's not that much. It's actually not that big a deal. That's the thing to end his entire scribe on.

00:26:11:01 - 00:26:26:03
Unknown
Then a senior Trump administration official and key adviser on AI, a slammed balls post. This is Axios. You have to go through all of its crap about why it matters. State of play.

00:26:26:06 - 00:26:47:13
Unknown
Every industry ecosystem has its supreme village idiot Dean W ball. Is that for AI? Said Defense Undersecretary Emil Michael on Sunday on ex. Dean ball has perhaps the biggest gap between actual IQ and his own perceived IQ of anyone in the industry. About 40 points. I mean, come on, you know, can you have can you have a real discussion about these points?

00:26:47:19 - 00:27:12:12
Unknown
I disagree with ball, but he raises important questions, including inadvertently raising them against himself. Yeah. As we just discussed. Oh hey, that communism sounds okay. Right. So this is this AI stuff is just so that's why I love these discussions, Jason. Because it's it's not just about, well, how I got it to add that line of code about this or that.

00:27:12:13 - 00:27:36:28
Unknown
It just raises all these fundamental questions about how we operate. Oh, yeah. And everything. And an economy and totally institutions. I mean, technology is has always been intertwined with politics, but AI is definitely intertwined with politics and and everything right now. It's just it's so yeah, it is fascinating to kind of see all the connections that are being made right now.

00:27:37:01 - 00:28:05:27
Unknown
So to cut through a few more, if I may, Gary Marcus has seven suggestions or seven predictions of what would happen here. What are the options? One. Do nothing. Let these companies stand on their own feet. No subsidies, no bailouts. Just let the market go. Two outlaw open source. Good luck with that, says Gary three build a regulatory moat as a backdoor way of protecting American companies.

00:28:05:28 - 00:28:36:21
Unknown
And he mentions Dean ball for bailout the big AI companies. And he says, just know, and I agree. Five banned Chinese models altogether. Axios thinks that might work, but might be in the works, might be the worst. Thank you. Yeah, but I think that's silly. Six buy out the big AI labs OpenAI and anthropic together for $0.02 on the dollar, which is generous since they are both mostly money losing finishers in an industry that has yet to convincingly establish profitability.

00:28:36:21 - 00:29:00:09
Unknown
And who would be buying them out? The U.S. government. That's that's what he said, I think. Then seven give up on the zero sum effort to win the AI war and instead move towards making AI a global public good. So here's Gary Marcus kind of playing the communist, right? He's looking at balls scenarios and saying, well, that's where we are.

00:29:00:12 - 00:29:24:08
Unknown
And then we have Bill Gurley, who is a famous venture capitalist, a capitalist capitalist who says the powerful AI models are given away for free. And it was inevitable they aren't a security threat. This is what competition looks like. So you're really seeing a fascinating challenge here to fundamental ideas of capitalism.

00:29:24:10 - 00:29:44:08
Unknown
Yeah, fascinating. Thank you. Because there were quite a lot of links in this section, and I tried my best to kind of go through them. And you did a much better job. So there's kind of a sequence to all. Well then finally one last thing. There is a nice. So this is all going to come to a head in September when US and China hold AI talks.

00:29:44:09 - 00:29:56:07
Unknown
Yes. Yeah. G comes to I don't if he's coming to Washington or to Florida but comes to the US. And so right now I don't think we're in a very good position.

00:29:56:09 - 00:30:18:26
Unknown
In a good position for those talks. No because G can come and say I can dump AI on your market. Right. That's you can you can say it's two months behind, but so what. So what. Yeah it'll do 99% of what anyone wants. Totally totally free. It'll save companies money. It'll give them a far more power because they're open.

00:30:18:26 - 00:30:36:24
Unknown
Wait. Far more control. It'll be far safer for them. And your big companies with these huge IPOs. Screw them. No. They walked. What are you going to be walked so these models could run? Yes.

00:30:36:26 - 00:30:59:03
Unknown
Yeah. Oh fascinating. And that was and yeah that that all in the kind of happening concurrently with the Kimmy thing which suddenly everybody's gaga over. I do wonder if, I do wonder if the Kimmy thing kind of goes the deep seek route. You know, big news. Now, a couple of months down the line, we've moved on. I mean, this stuff moves at such a rapid pace.

00:30:59:03 - 00:31:18:13
Unknown
What is the next thing coming from China? You know, there's something right around the corner. Yeah. Yeah. Interesting. I haven't tried to use Kimmy. Can you use the web version of it? I haven't even tried. Yeah, I have not popped in there to, to see. I have not had the, the opportunity to do that sign with Google. Sure.

00:31:18:14 - 00:31:26:14
Unknown
What the heck. You could have a sure China your friend. What could possibly go wrong.

00:31:26:16 - 00:31:49:26
Unknown
But yeah. You know and to your point, you were saying these models, they're they're not quite as up to date, you know, are they. Two months, three months behind at the end of the day, you know, at a certain point they're all so, so capable enough that that ends up actually not mattering whatsoever. The the kind of improvements and steps forward are probably pretty minuscule at that point.

00:31:49:26 - 00:32:08:22
Unknown
And we know that the open models are going to get to that point just the same. As, you know, all the stuff developed by the big, big labs in the US. So, you know, and probably not very long from now, things are going to, you know, that that kind of gap is going to be very narrowed because this stuff does move so fast.

00:32:08:22 - 00:32:34:21
Unknown
So and don't be stupid and do what I just did. Oh. What did you do? Did you two factor authentication lock yourself out of your Google account or not? I'm not that, but the Ace, but the AI equivalent. What did you do? What'd you do? I go to Google. I search for moonshot K3. Oh, no. Did you go to the link, which is somebody as a sponsored link, trying to get me to sign in to other things?

00:32:34:26 - 00:33:03:24
Unknown
Yeah, I don't know. I do that. I forgive you on that one. I've totally done that. And don't do that, ladies and gentlemen. Don't do that real quick before we take a break. Last week, I threw out a little, like a test thing. I threw out a question we had been talking about demos government review board, and I asked whether y'all think open models should play by the same rules as close ones.

00:33:03:26 - 00:33:50:03
Unknown
If something like Hassabis government Review Board happens. And we actually got an interesting response on YouTube, on the YouTube version of the podcast from G.G.. Next one. So, Gigi, next one, thank you for giving me a response that I can read and be like, hey, see, this worked Gigi, if that is your first name said absolutely. However, just like election equipment is certified by independent labs through the EAC, and there is an NTSB and Far and other agencies to set and review policy for automotive and airlines, there should be a specific commission to come together with industry leaders and users, form policies and guidelines, and then establish a mechanism for review and certification of models.

00:33:50:03 - 00:34:18:13
Unknown
Doing so would allow federal, state, local, tribal and territorial bodies to more easily approve use of these products within bounded limits that meet a safety standard that the public can trust from their own officials. So thank you. Smart response. Yeah, I like that response. I think that's very thoughtful. And yeah, I appreciate you weighing in. Yeah. Well, you know, we'll throw a question in at the end of the episode and see if we can kind of keep this ship rolling.

00:34:18:13 - 00:34:37:28
Unknown
I'd love to have a little feedback loop going on here. I like giving you the opportunity to get your voice on the show as well. And speaking of you, if you are a patron, especially Patreon.com, Inside Show is a place where you can go to subscribe and follow and support everything that Jeff and I do with this podcast.

00:34:37:28 - 00:35:02:21
Unknown
And we have a few amazing patrons to name Mike Reiser Zoe dedicating two of our many patrons who support us on a monthly monthly basis to do AI inside. So Patreon.com AI inside show. If you want to support us, you get lots of access to things like the daily podcast and discord and a whole bunch of other stuff, so we appreciate your support.

00:35:02:22 - 00:35:16:09
Unknown
Thank you for letting the show continue. We're going to take a break, and when we come back, we'll talk about Google's new models and what I think it's kind of cover for that's coming up here in a moment.

00:35:16:12 - 00:35:50:18
Unknown
I know that was possibly a compelling tease. It probably maybe it overstates a little bit more than what's actually going on here. But Google shipped new Gemini models just this week. I'm thinking probably not the ones that most people were expecting to see. In fact, Bloomberg reported on the delay ahead of the release of these models to, say, 3.5 Pro, which we've been waiting a while for, is falling short, apparently, of company goals on its internal goals around coding, performance and stuff.

00:35:50:18 - 00:36:24:05
Unknown
So they've been delaying this. So that's why you haven't seen 3.5 Pro, but instead Google announced 3.6 flash and 3.5 flashlight and 3.5 Flash cyber. So it's like all over the map. It's like everywhere. You can't even keep track of this. We've got 3.6 now happening in a small amount or in a kind of like a small version, and then a bunch of three point fives that Cyber One is, is, by the way, kind of following a similar playbook as what we've seen elsewhere, right?

00:36:24:06 - 00:37:10:25
Unknown
US government approved organizations, only the ones only them have access to it for now. You know, it's more specific to cybersecurity and that sort of stuff. And they get cheaper deals to to be able to to work with those models and everything. But yeah. And I think the interesting story here for me is just that. And tell me what you think about this, Jeff, is that the last handful of months has been in and open AI when we're talking about us, Frontier Labs sucking the the oxygen out of the room and Google kind of recognizing that at the pace at which this stuff moves, they need, you know, people are expecting them to come

00:37:10:25 - 00:37:27:25
Unknown
out with their model. That kind of goes toe to toe. Now it's kind of like it's a, you know, again, it's like a horse race. And anthropic and OpenAI right now are kind of playing this, you know, racing each other and Google now it seems like Google is a little bit further behind and people have an expectation for them to catch up.

00:37:27:26 - 00:37:49:15
Unknown
This isn't quite a catch up, but this is something for them to say, hey, see, we're not sleeping over here. We are working. We're not sleeping. Trust us. Yeah. I mean, the whole weird thing about about these models and branding and the levels is that I don't think it means much to most users now. They have certain things they want to do, and either it can do that or it can't.

00:37:49:19 - 00:38:17:12
Unknown
Yeah. So true. A link that I put in the didn't think we talked about, but just especially the IEEE has a guest article by Bruce Schneier and Bharat Raghavan that why AI needs a Gini coefficient proposing a new metric for whether AI does what you actually want. This is what we talked about last week is, is to have mechanisms to judge these things on practical terms that matter to us.

00:38:17:17 - 00:38:38:12
Unknown
Yeah, right. And I think that's interesting. Yeah, yeah. Does it actually do what you want. I mean, this this also ties in with what we're talking about earlier, how these things, you might have an idea that you put together your command, you know, for it to actually do the thing. But how closely does it adhere to that? Yeah.

00:38:38:14 - 00:39:04:02
Unknown
Last week, another plug for indulgent machines we had on Rafi Krikorian from Mozilla. They put out there their state of open source AI, and he was judging and basically saying it's pretty damn good. And, and but but they also don't have kind of Consumer reports like tests yet. Right. To say what can these things do? Does it really matter to me that it's 3.5 or 3.6?

00:39:04:02 - 00:39:28:15
Unknown
I don't know if this flash really matter what. All you can do is try something, but there's no standardized way to say that's the one you want, right? Well, yeah, so much of it is about feel and experience. And, you know, it's kind of like in the in the cloud world. It reminds me of opus 4.6, right. Which everyone, you know, largely everybody was like, oh, that's that was a great model.

00:39:28:15 - 00:40:03:27
Unknown
Then they came out with 4.7 and I heard very mixed results on like, oh no, I'm sticking with 4.6, then 4.8. And still people are opting to stand for point six. Could you name why that is? No. But like people get like this felt sense when they work with these things to be like no, I'm just either they're used to whatever it is, whatever that combination of things is that that particular model gives them, or there's some clear distinction, you know, or error rate or personality or whatever.

00:40:03:27 - 00:40:22:23
Unknown
How do you judge them? How do you know what is the best and how do you define best when these are like all purpose, you know, exactly meant to serve all purposes? How do you how do you define that? And they're so easily interchangeable. Yeah. That you today and you use the other one tomorrow and it's just as good.

00:40:22:23 - 00:40:45:00
Unknown
Which goes back to, I think, the communism argument. What the hell? If they're all pretty much the same and interchangeable, there's no switching cost. Is it worth investing a whole bunch of money in trying to get you to stick with one of these when people are very disloyal, I think. Yeah. Yeah, I think you're right. What do you use most these days?

00:40:45:02 - 00:41:10:18
Unknown
I'm still pretty. Pretty? Pretty tapped into Claude. Yeah. And I mean, it's we have it later in the headline, but I could just mention it now that, maybe I didn't include that, and I meant to, but that, like, fable, right? Fable five, they've now basically done exactly what I expected. And I was not alone in expecting this, but they basically said, you know what?

00:41:10:19 - 00:41:33:02
Unknown
We're not taking it away entirely. If you're the top two, two tiers of our paid plans, you get access to fable five, 50% of your usage. At that point, it switches over to kind of like token based usage. So you have to pay for extra use or whatever. So it didn't go anywhere. I've used that. I've become very selective on what I use fable for.

00:41:33:03 - 00:42:13:11
Unknown
You know, it's usually more for just like planning kind of complicated frameworks or, or ideas or whatever structures. But yeah, I mean, the more I use the Claude family and the more I build up this, what I've talked about this like coworker os around my business, I don't know. In my mind it makes it harder for me to switch, but I could probably and actually, this would be a good test to your theory is take that folder structure that I've built for Claude and import it, or directed at into either Gemini or the new OpenAI.

00:42:13:13 - 00:42:31:11
Unknown
The ChatGPT app changes because it's meant to kind of replicate that and see how it does. And I think that would be a good, good test. Duplicate that and see, is it able to kind of do the same thing and do I get better, you know, do actually get better quality out of it? And what does that mean?

00:42:31:13 - 00:42:54:26
Unknown
Like how do I how do I define better quality? I think that's that's an interesting question for me. Yeah. Right now what's holding you in would seem as inertia. Yeah. It's just already there and yes. And so I've built so many things around it that like the idea of like moving from one to the other sounds like, oh, what is that going to take half a day or is it.

00:42:54:27 - 00:43:16:05
Unknown
I mean, maybe it doesn't. Maybe it's just like I copy the files over. I pointed at it and I do my business as usual and everything works the way I expected. Now I'm super curious now. Now I'll have to try that. So yeah, but that's definitely what I've been using. And also I do use a little bit of perplexity here and there, not nearly as much as I had been.

00:43:16:06 - 00:43:37:23
Unknown
Right. And I do use a little bit of ChatGPT and I do use a little Gemini. I mean, I'm all over the place depending on, on what I'm doing, you know, just really depends. What about you? Like, are you I stick with Gemini mainly because I pro there and I'll go to the others when I want to compare.

00:43:37:25 - 00:44:03:21
Unknown
Yeah. Okay. Like like I did with health last night. You know, I've already I've already established the prompt and I have my baseline in Gemini. And then I compare it to that prompt around and see like what is the quality. What am I most satisfied with on the other end? Yeah, yeah. Let's see here. What about, oh, the publishers.

00:44:03:24 - 00:44:27:20
Unknown
Thank you. Put this link in there. And I knew you were going to have have thoughts on this. You could probably guess where your thoughts are, but the Wall Street Journal has an article. Google was a lifeline for publishers. Now some are thinking of cutting it off. We've certainly talked about this as a recurring theme, but this is all about, you know, your your AI overviews appearing at the top of search.

00:44:27:23 - 00:44:46:13
Unknown
Unsurprisingly, a lot of people never click through to the actual source. I mean, I can fully admit that that's certainly the case with me. Sometimes I don't click through because what I needed was right there. So, you know, if I'm one person in a vast case study, I that would certainly be something that I'm doing less of is clicking through.

00:44:46:15 - 00:45:14:00
Unknown
But this article just kind of details, you know, a number of studies that are basically representing that these changes in behavior on how people are using Google Search are shifting according to those reports. And then a lot of publishers are basically like they're flat out considering, like, maybe we turn Google off and block it entirely. You know, maybe we whatever robot robots.txt our site and we don't play that game anymore.

00:45:14:00 - 00:45:31:19
Unknown
I don't know what you do then from there since it's school and search, but yeah. What were your thoughts? You put this it's it's a classic case of cutting off your nose just about your face. There you go. I got that written down right here. Right. It's it's. What the hell do you think you're doing now? USA today is going to cut off Google.

00:45:31:23 - 00:45:50:29
Unknown
I think it's been a decade since I clicked on a USA today link. I don't care, I don't care, cannot go away. I don't really care. I talked to one publisher some time ago. Would cut off all of the AI bots, but not Google because Google. This is the controversy. If you wanted to be in search, you had to be an AI.

00:45:51:01 - 00:46:13:27
Unknown
Now Google said you can pull out of AI. So then for people to say, well, I'm not going to pull out a search to is just foolishness. It's suicide. It's suicide all around. This is how people are going to find information, even even Reddit, which in 2024 had a $60 million a year deal with Google. Yes. And Reddit's gotten really uppity.

00:46:14:00 - 00:46:29:12
Unknown
Reddit doesn't make content, its users make content. And I think Reddit is constantly in danger of pissing off the people who bring it value every day. But they're saying, you know, screw you, Google. And.

00:46:29:14 - 00:46:58:26
Unknown
So the other issue to me is that I think what's really happening is we're shifting from an old model, which was really going back to print and back to old media, to some new model, yet unknown. And the old model was a destination model. You had to come to me to get the news. You had to come to me to get the features, the photos, the whatever, and well, but I, I didn't spend the money on marketing to get you.

00:46:58:26 - 00:47:32:25
Unknown
So I used these free mechanisms of Google and Social. And now as they become less valuable, they haven't given up all their value. People still click through Google. It's not. There is no Google zero. That's BS. But what it exposes is not the evil of Google or even your dependance upon it does expose that to extent. It exposes the fact that you have the wrong model, and the model should be that you have a relationship of trust with someone and they they do.

00:47:32:27 - 00:47:52:24
Unknown
It is a destination. They just don't have to come through Google to do it. They come to you because they like what you do. They trust what you do. And we see that I've got a lot of complaints about Substack, but we see that in Substack. We see that people, individuals are able to build that kind of relationship and throw a paywall around themselves and make money from it.

00:47:52:26 - 00:48:10:01
Unknown
But they're doing that not because people are discovering them every day through Google. They're doing that because they established a relationship. And it was also not a scaled relationship, not a mass relationship. It was a big enough relationship, right? If you have 10,000 people paying you $10 a month, that's a damn nice little business you got there. That is a nice business.

00:48:10:02 - 00:48:32:00
Unknown
Yeah, right. And it's not $1 million, a million people a day. It's not 10 million people a day. It's not that whole scale, which is what I concentrated all the time. Mass media. So I think what we see here is the last dying gasp of the mass media business model, that you had to have the largest audience, and you're dependent upon these free mechanisms to get it.

00:48:32:01 - 00:48:52:01
Unknown
And the thing is, people didn't stay there. They went away because you didn't give them enough. You didn't build a relationship with them. You didn't know who they were. You didn't specialize in while you treated them all the same. You treated them all as a mass. So I think these are idiot publishers who were out there whining and moaning and screw them.

00:48:52:03 - 00:49:18:03
Unknown
But I'm not very popular for saying that. So this reminds me I didn't put the rundown, but Robert Thompson from News Corp news Corp Murdoch is the the anti internet anti-gay. And he always puts out these quotes that go overboard. So he this is about the search engine brave that went after a copyright. So it's a different story.

00:49:18:03 - 00:49:39:21
Unknown
But he said it's funny to read it. They have shamelessly stolen and then perfidious profited from the pilfering by illicitly fencing our journalists work instead of a license. They have licentious looted with blatant disregard to the severe damage done to the information ecosystem. This era of tech trafficking must come to an end of journalism is to have a sustainable future.

00:49:39:26 - 00:50:01:07
Unknown
These bite burglars and bootleggers. These content contraband discs are seriously undermining the integrity of AI and IP. To which I said when I shared it Murdoch manservant, alliteration addict. Bellows. Bombast fearing future.

00:50:01:09 - 00:50:20:25
Unknown
Does he always write with such like colorful alliteration thing? There was another guy named Less Hinton. I used to work for Murdoch and he had similar statements. The title of one of my books was Geeks Bearing Gifts, where he gave a speech like beware geeks bearing gifts. I stole the line from him to do the opposite with it.

00:50:20:27 - 00:50:47:17
Unknown
Yeah, that's a very Murdoch thing. Wow. They speak in tabloid. Very. Yes. Very colorful. Very colorful. Speaking of colorful, we have a super chat from Mr. Ozone nightmare. Thank you so much for the super chat. Says happy Wednesday, gentlemen. I think Google's relevance is much less powerful for younger folks. So maybe the pubs figure they're not losing much point.

00:50:47:20 - 00:51:25:24
Unknown
You're both doing well. Point job. Yeah, I think that is a good point. There's a diminishing value of being there, but I still think they're fools. Yeah. Google's they gone away. Yeah. And you know, it also kind of it kind of reminds me of this idea that I've, that I heard actually for, for a while while I was at Twit, I would, I would hear it in a twit with like kind of the approach to YouTube for a while, which is like you, we don't want to over leverage on YouTube because YouTube isn't like YouTube could go away at any point.

00:51:25:24 - 00:51:50:01
Unknown
And then where is your business if all you do is you run your business on YouTube? But I mean, it's kind of the it's it's the same with search, right? It's the same with Google search. I mean, even though I maybe the difference there is that Google search is by most accounts, and whether rightfully or wrongfully synonymous with like public internet access, you know what I mean?

00:51:50:01 - 00:52:10:17
Unknown
It's like everybody just kind of assumed, like they go hand in hand. This is how you find this is how you get to the place you want to go is by a Google search or whatever. And so not, you know, choosing to not be there and run your business. Bypassing that maybe isn't as much of an option.

00:52:10:17 - 00:52:43:03
Unknown
But I think your point is. So then, you know, maybe you are there, but create a compelling enough product that people, no matter what, always want to go to your product, unique value to them. You give them something that they can't get through an AI overview and from a societal viewpoint, when everything, everything, everything not only goes behind a paywall, but but is hidden from search that ends this wonderful error we've had where you could search for anything, you could find anything that changed society markedly.

00:52:43:05 - 00:53:01:07
Unknown
I mean, I've written in books where it used to be the cliche is, was true. You had you had to drive to the library to find the fact or people used to call the city desk when I was in newspapers to try to find out something. And then suddenly the fact that you could find anything, and then people expected you to at least go to that effort.

00:53:01:09 - 00:53:32:01
Unknown
If you were wrong about something, didn't you at least Google that right? And that, granted, we have a lot of fact free discussion going on in society and a lot of BS floating around, but there's different factors going into that. I think that was a was a good phase in society and it's ending. So the places that are that the journal says are evaluating it already at Google, USA today, Politico, The Economist, Reuters, paywalls, and I don't subscribe to any of those.

00:53:32:03 - 00:53:58:23
Unknown
So yeah, I guess what I said, if you're in the search, is that a worthless to me? Yeah, it's kind of like what Liz read from Google said during our interview. Right. Like, you know, these the publishing companies are are complaining about lower, you know, readership and everything like that. Yet they are putting their content behind paywalls. And yes, if you do that, you will have lowering readership.

00:53:58:25 - 00:54:34:20
Unknown
That is that is the way it works. Yeah. Yeah. Interesting. Netflix is basically saying, hey, we use generative AI a lot for our production. They said around 300 of their titles during their earnings call. 300 of their titles are around that number this year. Use generative AI for producing different aspects of their content. And, you know, it's basically they say this is is saving us cost if we didn't use AI to create these particular scenes.

00:54:34:20 - 00:55:07:26
Unknown
So we're talking about like battle scenes, crowd scenes, things like that. We probably would have dropped that scene from the plan because we just didn't have the the money to pull that off, whether you believe that or not, you know, but that that makes sense to me that AI using AI for these things is probably less expensive. And I think they're saying twice as fast, half the cost of other options, including the American Experiment docu series, which is what you see here 17 minutes produced using generative AI tools.

00:55:07:27 - 00:55:27:03
Unknown
It's a tool. It's a tool to toolbox, and I think it'll get better and better. Right now it's you could almost tell when it's used and it's used in a sloppy way often, but but it's still absolutely. It allows creators to to create scenes they couldn't otherwise afford to create. I got to mention in this context, I got to mention we talked about before the show.

00:55:27:05 - 00:55:51:06
Unknown
So Elon Musk has been railing against the Odyssey and the heck with him because it's been wildly popular. But he wants to make a historically true version of the Odyssey, forgetting that the Odyssey is about gods and strange creatures that don't exist. What he's really talking about is the racist view that you can't have a black woman as Helen of Troy, and you can't have a trans person as a character.

00:55:51:06 - 00:56:10:08
Unknown
And so so he came along and he said that he was going to remake this using grok, and they put up about a minute and a half of remade scene from The Odyssey. And it's hilarious. Intermediate to be, but it's absolutely hilarious. If you have a chance, I link to it. You can go. You can only see it on X, but it's about a minute and a half.

00:56:10:08 - 00:56:31:11
Unknown
And the funny thing is people are already finding things wrong. The wind goes the wrong way. He has like an airplane earplug in his ear and so on and so forth. So, so in terms of using AI to generate your suit, Musk insists he's going to have a full length version of the Odyssey to compete with the real one.

00:56:31:13 - 00:56:44:26
Unknown
Okay. Enjoy you. You do that. And Ubu Elon, a certain certain group of grok aficionados will be happy to watch. Probably.

00:56:44:29 - 00:57:15:24
Unknown
This one. This one caught my eye. I thought this was interesting. AI companies and not at all surprising AI companies anthropic as one of them. Reportedly, Google buying up huge quantities of old pre 2022 books. And you might be wondering why that would be. Well, because those books predate ChatGPT, which means they're clean of things like AI slop contamination, there's free of data poisoning tactics.

00:57:15:27 - 00:57:38:13
Unknown
They're just basically clean data sources to pull text from. Maybe not the most up to date data sources, but nonetheless, there's a heck of a lot of books out there. I just kind of assume that so many of these books are already part of training data. Like do you? I would think so, but I think it's also a way to cover your rump, because in the anthropic settlement, which was just approved by the judge.

00:57:38:15 - 00:58:03:18
Unknown
So checks will start coming out soon. Anthropic, as you remember, violated the law by using a database that existed but was on the right side of the law and buying used books. Yeah. And the crime to me was they then cut them up so they could scan them cheaper. You can you can scan them without tearing them apart and then destroy them.

00:58:03:18 - 00:58:25:29
Unknown
And that I find. Yeah. Yeah. And I buy used books constantly. A lot of these books behind me I bought from places because they're not in print. I need them for research. I'm going back to, you know, the 1920s right now with books about about electronics and such. And if they're taking these books out of circulation and I can't get the book that upsets me.

00:58:26:00 - 00:58:56:18
Unknown
So by the book, use it legally, scan it and then give it away. You've already gotten your value out of it, or sell it for a lower price and recoup a little money. Don't destroy them, please. You know there. Yeah. So when you think about like the scanning process, have to imagine it's cheaper and less complicated to scan a stack of pages that are separated from each other versus flipping a page by one.

00:58:56:18 - 00:59:21:11
Unknown
And it's a difference, but it's a difference of some percent. Yeah yeah yeah versus boom boom boom boom. Right. Yeah. It's time. Yeah yeah. But they're machines doing it. So yeah because because everything moves so fast in AI. You got to do it now. No no no no no no no fast fast faster. Yeah I believe me I'm not I'm not like trying to argue for the destruction of books.

00:59:21:11 - 00:59:44:20
Unknown
I think that's horrible that that's happening. But apparently it's happening a lot. Apparently a lot of booksellers have noticed an uptick in book orders just this past year of, like, use book orders that they suspect, and probably rightfully so, are coming from AI companies, some marketplaces they were used to seeing, like, you know, maybe we sold a couple of dozen books per week and now they're selling hundreds per week.

00:59:44:23 - 01:00:01:25
Unknown
Well, have you been to a used a big used book charity sale in recent history? Not recently. No. So fascinating change. So around me is the American Association of University Women does does one. It's a big sale. I contribute books to it every year. My wife makes me get rid of more books and I get them up there.

01:00:01:26 - 01:00:26:02
Unknown
You're addicted. These sales are now ruined in the first day by the scanners. These are people who come in holding a scanner. They scan the title or the ISBN and they can meet with the market value is and the sale is selected for $5. And they realize that they could sell it for $9. Then they grab it. So these these sales get picked clean by the scanners.

01:00:26:02 - 01:00:44:27
Unknown
Many of the scanners don't speak English. They don't need to. They're just using the tools. And they're getting huge numbers of them. And the charity who's selling the books, they have been you know, they sell more books and that's okay. Sure. But those who are there to get the best Pickens know we get the leftovers. Oh, interesting. Interesting about that.

01:00:44:29 - 01:01:08:02
Unknown
Used for those of you who are my little tip to you all, if you don't know what book finder is the best, because it's a meta search of all of these used book search engines. And if you want to find an old book, it's a great, great way to do it. Interesting. And and it searches like the eBay's and the Amazons and the exactly little more unique marketplaces or whatever.

01:01:08:05 - 01:01:33:15
Unknown
Yeah. Or boutique kind of smaller ones. Interesting finder.com bookbinder. Com. All right. Interesting. Well, hey everybody, if you like what Jeff and I do, please go on Apple podcast, give us a review, give us a star, throw a star on Pocket Casts or wherever you watch or listen like the video, I don't care. Just just engage with it a little bit because that sends a signal to other people that, hey, there's something going on here.

01:01:33:15 - 01:01:45:11
Unknown
Maybe I should check it out. So we appreciate when you do that. Going to take a quick break, come back for a speed round and then maybe a question to end the show with. And we'll get you out of here in a moment.

01:01:45:13 - 01:02:05:10
Unknown
All right. Today I don't know if you saw this, Jeff I'm sure probably possibly on I am later when you're not talking with Nate B Jones, you'll be talking about Samsung's event this morning. Watch it since you're Doctor Amazon to me I'm Doctor Android. To me I didn't get to watch it either because I was prepping for this show.

01:02:05:10 - 01:02:28:21
Unknown
But I do know that there were no surprises. I mean basically the next the next family of Z flip Z fold eight, you know, Z fold eight, there's like a wider version of the Z fold eight. So it's kind of like shorter and stubby. I'm calling it the squatty squatty version of the Z fold. Then they've got the tall one and the regular flip, new watches and all that.

01:02:28:21 - 01:02:50:00
Unknown
But they're also for a lot of the press, because this happened in London, a lot of press that went to the event also has had a chance to take a closer look at Samsung's AI glasses. And so these are glasses that are in partnership with Warby Parker and Gentle Monster. And basically just kind of like your standard competitor to the meta Ray-Bans.

01:02:50:01 - 01:03:18:09
Unknown
No display. So you know, these are do have a camera. Do you have an LED light. Open your speakers. Mikes do not have a display. So still still no word on if you're in the Android ecosystem. I presume this is better. I would say so because on Meadows Ray-Bans you're using meta AI. Not my particular favorite. I only really use meta AI when I use those glasses, which has been dwindling, I will admit in recent months.

01:03:18:11 - 01:03:41:21
Unknown
But yeah, I'd probably use them more if it was actual Gemini, just because it's integrated with everything that, you know, all of my accounts and everything. But but yeah, so there you go. And, you know, they're a little more, some more stylish options coming from a gentle monster and Warby Parker, if that matters to you. So anyways, don't know when they're going to be released.

01:03:41:21 - 01:04:28:06
Unknown
And actually the people who got their hands on didn't get to put them on their heads, apparently. So you could look at them and you could touch them, but you can't wear them and use them. So that's where Samsung's at. This story about the sonno hack that happened in 2025. And apparently there is information that is leaked from that that reveals kind of how sonno kind of came to be from from building its model initially, some internal code, some comments that basically described in very explicit terms, like how to scrape the music and lyrics from YouTube music, Deezer, genius pond five, Mendo or Jam Ando.

01:04:28:07 - 01:04:50:06
Unknown
I don't know how you say that. Free sound international music score library Suno has insisted that they are, you know, did not bypass any any paywalls in order to do this. So they, you know, they continue to insist this is all publicly available. It's just out there, you know, and this kind of shows. Okay, well, they were going to YouTube music.

01:04:50:06 - 01:05:19:12
Unknown
Maybe they were using the free account or whatever, but they were still going there. They were putting in filters that removed non-music. So they just got music. And this was interesting to me. They also included podcast on their scans, which was not just podcasts on those platforms, but also RSS feeds and stuff. Oh yeah. Which is interesting because if you're just looking for music, you're probably not going to find a huge amount in podcast, but you'll find some, I suppose.

01:05:19:15 - 01:05:38:25
Unknown
Well, so on YouTube, you labels put up music videos to promote their songs. Yeah, they don't put me on a paywall. They put them up there so they promote them. I don't think that ingenious. I'm old enough to remember when genius was all hell because oh my God, you stole our lyrics. But then they wanted to be there.

01:05:38:25 - 01:06:07:12
Unknown
So suddenly the fight went down and says, now, okay, so I'm not so shocked by this. Yeah, yeah. I wasn't too shocked by this either, to be honest. I mean, especially because, like, we all know that Suno, you know, did this. This is just kind of like here's some information on how they did it. And of course, it gave the people who don't like snow very much don't like AI generated music something to point out and say, hey, see, this is bad.

01:06:07:15 - 01:06:38:04
Unknown
But there you go. One password now allows Claude to sign in to your websites on your behalf without seeing your actual passwords. That's kind of the important thing here is, like, it's easy to read this and think, oh my goodness, Claude has access to all of your passwords. Now, what are we doing? This is done in a secure channel sort of approach where those credentials actually stay in the right place, not fed into any of the context or memory or anything of Claude.

01:06:38:04 - 01:07:09:04
Unknown
It's just like channeled through into the page. Claude is able to help trigger that process, let's say. But you as the user, still have to approve it with a request for biometric that data is fed in. It doesn't touch the memory or context of Claude. But yeah, so it's kind of an OAuth for AI kind of. Yeah, it's kind of dipping, dipping into that territory or giving Claude the ability to, you know, manage like two factor authentication, which on its face sounds really bad.

01:07:09:05 - 01:07:35:03
Unknown
Like, no, Claude should never be able to to do my two factor stuff, but it's still very much requires a human at the at the keyboard to actually do that. If you give it access to your hugging face, just be careful. Just be careful. You never know. Thinking machines. Mirror, mirror. Lab. We've been waiting since Mirror Marathi started thinking machines what, like nine months ago or something like that?

01:07:35:04 - 01:08:00:04
Unknown
What was going to come out? And their first release is an open weights model called inkling that is built specifically to be fine tuned rather than just used like straight up out of the box. But it is the company's first model. Thinking machine says this is there. This is not the strongest model out there, but that enterprises want a model that they can customize on a deep level, rather than renting through a Frontier Labs API.

01:08:00:04 - 01:08:22:17
Unknown
And so here you go. This is thinking machines kind of first effort and I don't know, Marathi pushing the open source play from within the walls of the US. Yeah, it's kind of good to have an American. Mr.. All right. Yeah that I think that model matters and see our discussion about China and yeah, she could be undercutting from within.

01:08:22:18 - 01:08:40:07
Unknown
I went to it just to the web version and asked to challenge me on some ideas for the next book I'm working on, and I came up with some stuff and it first class, and that was pretty good. It was a little more banal when I dug into it. Yeah, but still I could have gotten into a dialog.

01:08:40:09 - 01:09:02:21
Unknown
It pushed on some things. You know, it's interesting. Yeah. Got to start somewhere. Yeah. And then finally do. How much do you care that Google is renamed notebook LM? No longer can we refer to it as notebook LLM. It is now Gemini Notebook. I saw somebody on the socials just saying knock me over. The first time that Google made a brand name change.

01:09:02:21 - 01:09:22:04
Unknown
That actually makes sense. Yeah. And so notebook was never a gracious, graceful brand. Right? Right. Tell people to go to the book alone and say, wow, what the heck is that? Right? It was a joke on LM and all that now. So yeah. However.

01:09:22:06 - 01:09:50:05
Unknown
Gemini Notebook still sounds a little generic, and if Google ever starts a notebook product like keep, that can be confusing. But anyway, yeah, cross that bridge. It's fine. I think it's fine. I think I'm glad to see Google is in many ways embracing the former notebook LM more I think it's good. Yeah. And integrating it more. It's basically it's getting integrated into Gemini at some point.

01:09:50:06 - 01:10:09:12
Unknown
Apparently it's going to get integrated into AI mode. You know, all their products are just kind of becoming part of all their other products. It's just what they've been doing. And now now they've centered around I mean, notebook LM was pretty early, wasn't it? And so the the naming actually kind of probably described what it was at the time.

01:10:09:13 - 01:10:26:27
Unknown
And then Gemini really became the the nomenclature. And so now it's moving it over. I can't remember what it's what it's it had a working title like Thunder Bolt that wasn't it. But something like that. Oh did it. Oh I don't even remember that. That's how when I was introduced to it. That's what I with Stephen Johnson. That's how I saw it.

01:10:26:28 - 01:10:54:09
Unknown
So this is the third brand. Interesting. Well, before you go, we spent part of today's show talking about the fight over what to do with Chinese AI models like Kimi K3. So leaving a question out there to see if we can continue getting the feedback loop here is a free, open, popular Chinese model like K3, something the US should be worried about and try to restrict.

01:10:54:12 - 01:11:17:09
Unknown
Is it something that's good for competition keeps everybody honest. What are your thoughts on this? Let us know either in the comments. You can send us an email feedback at AI Inside Show. And if we hear from one of you or some of you will, you know, take take some of those things, put them together and share what you think is the argument why government should or should not do something about this.

01:11:17:13 - 01:11:38:24
Unknown
Yeah, what they should or not exactly be part of the conversation. Love to hear from you. Thank you for that. Thank you. Jeff Jarvis Jeff Jarvis hot type coming out August 20th, less than a month away. At this point we should have like a countdown in the corner of the every live stream. So people know that hot type coming out soon Gutenberg parenthesis magazine web.

01:11:38:24 - 01:12:09:15
Unknown
We we've. What was the other book you mentioned? The geeks bearing gifts. Not on this list. Not in Britain anymore. Yeah. No. Oh, okay. But hot type about to be in print. So go to Jeff Jarvis a preorder and then wait. Wait patiently for intelligence, AI and humanity from Bloomsbury pod tuneup. Com if you want. If you have a podcast, think you might want a podcast and want to talk about it, I will talk to you for 30 minutes and we'll see where it goes from there.

01:12:09:15 - 01:12:31:06
Unknown
So let me know. Happy to help you out. Pod tune up A inside show is the single source of all things AI inside on the web. Everything you need to know, including now a link that goes to the store. If you want to buy some merchandise, you can. I could probably add even more stuff there if people would like I haven't.

01:12:31:09 - 01:12:48:21
Unknown
I got to spend some time to do that, but there are some options so you can get there. And then finally Patreon.com AI Inside Show to support us on a deeper level. We appreciate that you do that because it enables everything that we do here. And oh, you're wearing the hat, I love it. I love that you have that within reach.

01:12:48:22 - 01:13:14:04
Unknown
It may not be the best model, but looks good to me. Thank you to our executive producers, Doctor do, Jeffrey Martini, Radio Asheville, 102.7 Dante, Saint James, Bono, Derek, Jason Nefer, Jason Brady, Anthony Downs, Marc Starker and Carsten. Thank you deeply from the bottom of our hearts. We really appreciate you for supporting us doing what we do. And you know, thanks.

01:13:14:04 - 01:13:30:09
Unknown
On the back end. Victor Bogner and Daniel Croft for video support behind the scenes. Y'all are awesome. We really appreciate doing this show and that we have an audience to do it for. So Jeff, good to see you. Thank you everybody for watching and listening. We'll see you next time on AI inside by everybody by.