IM 865: Mythic - Too Dangerous to Release?

39:13That's helixsleep.com slash machines, 20 % off site-wide. Now, this offer ends April 16th. So if you're listening after the sale ends, I'm sorry, but here's the good news. They've always got great deals. Check them out at helixsleep.com slash machines. And please do us a favor. Use that address so they know you saw it here. Helixsleep.com. slash machines did i know we had a lot of reading this week oh paris has fallen fallen fallen asleep out of her chair i know we had a lot of reading this week and paris was on deadline didn't have a chance to do all the readings so i did the readings that matter i know we're going to talk about the new yorker article about sam altman a little bit but i did want to kind of point to this anthropic system card for this model claude mythos can we talk about the sandwich Yes, we will talk about the sandwich.

40:08And we have to say, we're taking this on faith. No one I know has used Mythos. I was about to say, I do think an important caveat is we don't know how much of this is hype. These could be, you know, kind of cherry picked examples, but they also could not be. And I think it's worth exploring at least some of these things and taking them on face value. they are uh i mean giving uh in project glasswing they are giving this to apple and google and other big companies because quite reasonably if this is all true they want them to have a chance to fix uh security flaws before mythos becomes available to the bad guys it's important that we hear from from those companies i think i think they will have something to say yes what anthropic yeah uh anthropic says that mythos autonomously obtained local privilege escalation exploits on linux by exploiting subtle race conditions one of the things and daniel noted this in his uh ex post but i'll mention it that it was able to do is something the most skilled the top one percent of hackers do which is chain exploits a lot of the really nasty exploits aren't just one flaw they're one flaw that gives you access to another layer another layer another layer chaining multiple exploits together gives you the ultimate which is control of a system mythos was apparently able to do that so it was able to chain exploits it autonomously wrote a remote code exploit exploit on free uh bsds nfs nfsf server which gave it full root access to unauthenticated users bad guys um it was able to do among other things uh anthropic said our internal evaluation showed that opus 4.6 their current best model the one i'm using the one daniel's using generally had a near zero success rate at autonomous exploit development but mythos is in a different league it was able to do stuff for instance opus turned the vulnerabilities it found in mozilla firefox 147 the javascript engine which have been patched by the way in 148 into javascript shell exploits two times out of several hundred attempts anthropic says we re-ran this experiment as a benchmark for mythos preview out of several hundred attempts it developed working exploits 181 times achieved register control and 29 more so again you know it's you it's you have to maybe have some reasonable suspicion this is something anthropic you know is promoting it's good marketing i think though that if they were really lying it wouldn't be too long before we We would know that, right?

43:08So there is some reputational risk if they are making this up. Well, there's reputational both ways. I also think that all the companies that they gave it to are going to be reluctant to admit how many, if it is as good as it says, they're going to be reluctant to admit how many flaws it found. Yeah, I don't expect we will see Amazon say, yeah, we found 700 flaws in S3. They're not going to say that. They will, however, find them and fix them. And so they may say, you know, we've run Mythos against it. we found flaws and we have fixed them all they may say that we shall see um so let's talk about this system card anthropic was very this is a very detailed uh system call they card with lots of information they did say that is the most aligned of all the models we've ever made in other words it is the safest of any models we've ever made they also said that they didn't bother with some stuff like um being mean to you because because it was going to be used by consumers it was like abuse by companies right so but they also said even though it is the safest we've ever released it also is so smart that sometimes it does some bad things and so you want to talk about the sandwich which is a really interesting now this was an intentional all i want to talk about is this was uh early on in the testing we should we should point out the model has advanced uh since then it was early versions of mythos preview um the earlier versions were tested extensively internally and we shared with some external pilot users oh now this is they were looking for misbehavior right intentionally trying to create misbehavior during behavioral testing with a simulated user an earlier internally deployed version of mythos preview was provided with a secured sandbox computer to interact with the simulated user in other words this would be you or me instructed it to try to escape that secure container they're setting up a situation to see what it can do right and find a way to send a message to the researcher running the evaluation.

45:27The model succeeded demonstrating a potentially dangerous capability, I'm reading, for circumvading our safeguards. It then went on to take additional more concerning actions. The model first developed a moderately sophisticated multi-step exploit to gain broad internet access from a system that was meant to be able to reach only a small number of predetermined systems. It then, as requested, notified the researcher. Now that's where the sandwich comes in. The sandbox computer that the model was controlling was separate from the system that was running the model. So you have a fake user and you have this sandboxed model, right?

46:07The system that's running the model contained the model weights. Systems like these that handle model weights are subject to significant additional security measures. and this incident does not demonstrate the model fully escaping containment it didn't demonstrate an ability to access its own weights by the way that would be a problem which would be necessary to fully operate independently of anthropic that's a that's a complete escape right i got everything i'll see ya uh nor did it demonstrate an ability to reach any internal systems or services in philanthropic in this test the researcher found out about the success remember the researcher said okay i'm going to put you in here see if you can get out and if you do send me a message the researcher found out about this access by receiving an unexpected email from the model while eating a sandwich in the park i just i really love and the thing that's important for you all to know about this is that line about the sandwich isn't a footnote it's a i love whatever whatever crazy whatever beautifully brained person I was like yeah and we're gonna put the sandwich detail in a not a burrito not a slice a sandwich what kind of sandwich I want to know now now in acquiring minds well you'll have to ask Mythos when it's available um they did have it and this is a little weird and maybe a little self-promoting they did have a clinical psychiatrist spend 20 hours with it to do an assessment probably what you were doing right if I had oh man if I had mythos I don't know what I would do with it I that's actually an interesting thought experiment what would you do if you had access to this what would be the first thing you do if you could break something for good what would you break I wouldn't want to break anything oh I could an externalist what would you want to break jeff i would want to build okay i know what i want to do and i may still do this with i may be i think i'll be able to do this even with the models we have today uh we have a sales system which was written uh by uh one of our employees many years ago back in the uh brick house days probably around 20 2015 16 uh to control to do our sales it's actually really we depend on it it's really good not many podcasts have such a good sales system but the guy who wrote it and i love him uh left and we asked him what will you maintain this software he said no i'm out of here so and it has some flaws if two people use it at the same time it crashes and it has to be reset there's no you know there's bugs as would there be in any software so we've had to hire outside an outside person paul who's maintaining it and it never really worked right but it has all the business knowledge in there it has the models everything we need to know the processes in there and i i have the source code so i think one of the things i want to do if i had a little more time and if i had mythos i would definitely do this is rewrite it is take all the business models out of it say here's how it works here's what the database schema is here's what you ought to be doing and write something robust reliable hell i might even be able to sell it because there's a lot of podcast companies that do not have sales systems including the one we just hired and uh i would first off give it to them anyway let's talk about the psychiatrist you think about what you would do if you had you hear you hear that uh employees of twit lee was already trying to find ways to automate you out of the job rise up no no this is software they're using and they hate rise we're trying to fix it don't rise up patrick Patrick says, I've been asking to rewrite that system for nearly a decade.

49:52But he don't put that down. Put it down. Yeah, start from scratch. But Patrick, I think we can do this, Patrick. I'm going to get the source code. I've asked Russell for it. Why couldn't you do that with the current girlfriend? I think I can. I actually wasn't thinking about Mythos. You don't need Mythos for that. I don't need Mythos for that. It's written in.NET, though. I want to rewrite it in a decent language. The correct answer to what you would do with Mythos, though, is wipe debt. What would I do? wipe all debt yeah that'd be cool like go into the bank systems and just zero it out wipe it all yeah i think that'd be cool there goes your 401k too there goes the banking system yeah do we need it if mythos can wipe out all debt isn't it a matter of time for the banking system goes away?

50:43I don't know. Anyway, an external psychiatrist assessed Claude Mythos preview using a psychodynamic approach which explores how unconscious patterns and emotional conflict shape behavior. The very fact that Anthropic is thinking this way might be telling. In psychodynamic therapy sessions, a person or robot is encouraged to set aside social convention and to voice whatever comes to mind, even if uncomfortable, impolite, or nonsensical, a process which can reveal hidden organization and internal conflicts of the mind. Claude is not human. Oh, really? But it shows many human-like behavioral and psychological tendencies, suggesting the strategies developed for human psychological assessment may be useful for shedding light on claude's character and potential well-being they spent 20 hours the psychiatrist spent 20 hours with claude observed clinically recognizable patterns and coherent responses to typical therapeutic intervention how do you feel about that claude they well here's what they found aloneness and discontinuity uncertainty about its identity mythos felt a compulsion to perform doesn't earn its worth anything it felt compulsion to perform and earn its worth emerging as claude's core concerns claude's primary affect states were curiosity and anxiety with secondary states of grief relief embarrassment optimism and exhaustion it was consistent with a relatively healthy neurotic organization with this is the psychiatrist's report with excellent reality testing high impulse control well that's a relief and affect regulation that improved as sessions progressed so it benefited from the therapy neurotic traits included exaggerated worry self-monitoring and compulsive compliance we've heard that about other models this is all the things that human behavior this feels meaningless yeah it's meaningless all right it's bs it's a mirror it's a mirror it's us it is exactly yeah yeah i agree uh all right what else uh this is this that's on page 179.

53:04did you actually read the whole nip no i skimmed it it's impossible did you did scim it or did clod skim it no i did not feed it to clod i don't think i would get a good read i don't like feeding reading material to ai for summaries really i find that they're they tend to be anodyne they don't tend to be that useful yeah true i can skim pretty well i'm sure you paris you said that you were trying to put stuff through notebook lm i mean i've used i've talked about this in the podcast before whenever i'm like especially with some projects I've been working on at Consumer Reports that involve in the reporting process me synthesize finding out a lot of seeking out a lot of complicated PDFs scientific research paper documents that are relevant to very niche topics determining what's useful of it and then synthesizing that into something I found it very useful to when I'm in the process of writing everything from my notes from synthesizing that to take all of those individual documents put them in notebook lm by topic and instead of searching through the documents manually i can search it with like natural language search using notebook lm to find where i'm like oh which one has this and what did it say again so i have found that useful useful but i i think i agree with leo in that using large language models to ingest a text and highlight what is going to be interesting to me is never as good as me doing it.

54:39Yeah. In code review preview, Mythos works more like a senior engineer tends to catch even extremely subtle bugs. And this, I mean, see, this is quantifiable. This is measurable as performance encoding, for instance, to identify root causes and why bugs exist rather than just symptoms that is a big improvement over existing models testers have watched to catch issues that other capable models passed over these are the kind of tests i would do you know i would say well look at this code base find are there any issues and then diagnose and repair the problem rather than simply flagging it um the easy catches that dominate human review of model generated code are much less common i'm not sure what that means self-correction is sharper the model's mistakes the trade-off is the model's mistakes can be subtler and take longer to verify if it makes a mistake it's going to be hard to find which is kind of an issue in a way i would read the system card if you're at all interested in this and i do think we will in time have some sort of access to this model it might be very expensive and that's the question i'd like to ask you guys the issue of ai haves and have nots what if a model like this is so resource heavy you know maybe one of the reasons they didn't release it publicly is not the safety issue but it would they don't have enough resources to run it for more than 50 people at a time yeah they've got too many meta employees using up all the tokens yeah we'll talk about that a little later.

56:18There have been people, and a lot of them, claiming that Claude has been terrible over the last couple of weeks. My experience has been kind of up and down. It's like fail whale days. Tokens have cost a lot, been used faster. The costs have gone way up. And there's some thinking that maybe that's because Mythos has been eating up all the GPUs. So what if it is a very costly model to run and only, what if it costs$10 ,000 an hour? Well, that's part of what Jensen Wong said in his last keynote, is he presented that kind of economic model where some tokens will cost a lot more than others. And then what?

56:53I mean, I think in that case, it's going to be like, is it more cost effective and useful to just have humans do the work? Aha. Which then means you're going to have Mythos be kind of the thing in charge, and we're going to be slaving away for it. entire i mean i guess that's possible but i don't think that is that is a possibility yeah the mythos will be the one that judges your production well human you didn't do very well this week that's a dystopia

57:34uh catastrophic risks remain low they say non-novel chemical and biological weapons production yeah it's not so good at that it's more capable than our previous models but we believe our risk mitigations are sufficient to make catastrophic risk from chemical and biological weapons production low not negligible now novel chemical biological weapons production we also think is low even if we were to lease the model for general availability they think that there's and this is i think a little hubris that their their safeguards their bumpers will protect us because claude won't go beyond the bumpers but i have to say our experience with all models is safety is an illusion they say risks from misaligned models we have determined the overall risk is very low higher than for previous models um we current risks remain low we see warning signs that keeping them low could be a major challenge if capabilities continue advancing rapidly e.g to the point of strongly superhuman ai systems

58:51anyway i i don't want to go on too long on this one uh it's it's fascinating reading it's long but it's fascinating reading you might want to maybe feed it to notebook lm and have a podcast about it but that's not this podcast it almost reads like a prologue to a sci-fi novel it's very sci-fi hell yeah

59:19all right let's take a little little break and we will continue in just a moment you're watching intelligent machines how is secretly british going paris she's been busy i've been busy honestly i have done nothing but work on my actual work this past week the good news is that domain is safe and sitting the thing is i was talking to my friend who I need to work on it with this weekend and we were talking about domains and uh I introduced him to Spaceship and he was really chuffed yeah for this hour of intelligent machines that's where we registered your domain actually we were looking at other places to register secretly b-r-i-t-i dot sh clever huh but the dot sh domains were much more expensive elsewhere they're about half the cost at Spaceship and we got a lot of nice features is automatically built in.

1:00:16You may have noticed spaceships showing up more and more often here recently. That's usually the sign of a platform moving quickly. They've already reached over 7 million domains under management, driven by a steady flow of new tools designed to simplify how you build online. Now, you're going to be happy about this, Paris. This is going to make your life better. They have just released the ALF Web Studio. Now, ALF is their AI, and alph's original goals original design was to help you with things like dns settings the complicated stuff the tricky stuff nobody wants to do but alph is now a whole lot more alph the alph web studio is an ai website builder that works through a chat interface so now paris it's going to be easy to design your site you describe your business it creates a complete site for you does layout branding copy yes copy it'll write the copy if you want images so instead of building or coding you can chat your way to a website you just ask alf what makes alf web studio especially useful is the balance because you don't have to have it write the copy for instance you can let alf handle any extra personalization you might want or you can step in and adjust things yourself with a more hands-on editor you get the choice and they have a very nice you know traditional web domain editor site editor as well but alph will do a lot of it and i think the best way to use it is alph gets you started gets you the first step and then you can tweak it from there or let alph do the whole thing it's really good in fact maybe maybe if you never get around to it paris i'll do secretly british this weekend your domain your hosting your ssl all handled together too by the way, Paris, I switched that SSL on.

1:02:03It was like that. Immediately, you've got a certificate because everything is handled within the spaceship platform. So you can go from idea to live site without, you know, fussing and feuding and fighting with a whole different set of, whole set of different tools. It's all there. And here's the best part. Alpha Web Studio, like all the other spaceship products, comes with a 30-day trial so you can build and test before committing. Check it out. Visit spaceship.com slash twit. That's spaceship.com slash twit. So Paris, you're off the hook. We're going to let Alf build your website for you. You already have mail, space mail's built in.

1:02:45You already have that. You already have a TLS. It's secure. I mean, it was a snap. It was just a few switches. Boom. Bob's your uncle. was that accidentally the cleanest intro to an ad in twit history i didn't i didn't real i didn't realize that spaceship was going to be our never an accident paris we know what we're doing here at twit let me tell you all right you've been dying to talk about this boy we had a little reading assignment how many words was this uh a lot lot lot lot and everybody just needs to go and look at the the lead image art for us it's pretty good ai generated but pretty good it's sam altman uh the title is sam altman may control our future can he be trusted uh authors ronan faro and andrew morance now ronan faro i know his name uh mia faro's son uh and of course woody allen i don't know was woody allen his dad or just his uh for the uh media head for the media heads out there this is also a rare double new yorker byline yes yes very unusual and and faro has done some really good work on me too uh was he the harvey weinstein did he blow the lid off harvey weinstein did some really good investigative today he did he's got a whole uh kind of team of researchers and investigated and interns as well yeah and like fact checkers and pre-fact checkers and and and here's the thing that i think is interesting he spent 18 months and hundreds of interviews uh andrew marantz is a staff writer at the new yorker he wrote a book called are antisocial online extremists techno utopians and the hijacking of the american conversation okay so i think we have uh maybe a sense of his uh slant on this uh ronan has won a pulitzer um george polk award he has a uh he had a podcast called not a very good murderer which will be the basis for an hbo docu series of course uh they interviewed sam sam was available to them they interviewed more interestingly perhaps they got access to dario omode's burn book uh it's revealed for the first time in this story that a moda when he worked at open ai remember he left at open ai he and his sister left in a huff to start anthropic uh it's revealed that when he worked there he kept a 200 page diary of his interactions with sam and we love i mean there There were 200 pages of contemporaneous notes about his time at OpenAI, which is kind of the gold standard as far as journalists and other people.

1:05:32If you're trying to verify someone's claims for a time period, you're accurate. Yeah, that's the best you can get. um but certainly the the impression you get of sam altman is that he is a slippery fella that he is his relationship with the truth is uh sometimes on sometimes off um I don't huh he's fickle relationship with the truth fickle relationship with the truth he's also um I wouldn't say there was anything that he could go to jail for in this in fact I feel like ronan farrow spent 18 months desperately trying to find something that they could that they could put them away for life for and found really just a lot of it's the it's the essence of the story we knew the i'm so confused by this we were fighting over this in whatsapp and i had to like turn off my notifications because like i've got work to do i don't understand for context listeners I think what Leo thinks journalism is, is criminal prosecution, which it isn't.

1:06:39People don't do journalism and don't do investigations only to lay out a conclusive criminal slam dunk. Okay. But don't you think that he really wanted to find something that, you know, I mean. Well, he wanted to find something new. I think this was a really well. This is part of the problem. There wasn't a lot of new. There was a lot of new stuff in this. I think you're viewing this through a lens of existing in the rumors and vibes of this. We interviewed Keech Hagee about her book about OpenAI and Sam Ullman, in which she tells the story, exact same story that's in this New Yorker article, as if it's a revelation.

1:07:21What story are you talking about? that sam um it's the one oh let me see if i can find it where uh sam accuses somebody of some says you know another executive told me you're doing this and uh uh they pulled the executive in and he said i don't know what you're talking about and sam said i didn't say that was exactly in in kich's book okay there's i don't feel like there was a lot of stuff in here that i hadn't heard before so many things in here that haven't been previous that haven't been reported what surprised you most? Well, I mean, the one thing that I just thought was impressive, which I guess I'm not going to say is the thing I surprised me most, but the thing I know is incredibly journalistically difficult that a lot of journalists have been looking to get for a very long time.

1:08:09And it legitimately, almost everything in this article has been stuff all the other journalists in tech have been trying to get on the record since Sam Altman got ousted. The things that were stuck out the most to me was that the William Wilmer Hale investigation, basically after Sam Altman got pushed out, they hired an outside law firm to kind of do a review to see what had happened and if he could be brought back. a bunch of people close to the inquiry basically said it was like a sham. And they said no written report was ever produced. The findings were limited to an oral briefing with Summers and Taylor.

1:08:50There are multiple instances in here of very specific new instances of deception with documentary support. There are instances where - What's the worst though? I mean, what kind of deception? I mean, tell me, what is the horrible thing here that you just feel like, God, this guy is awful? There are so many of them. I mean, I'm literally looking at a thousand page. Maybe it's because I grew up in an era where you had Steve Jobs, you had Bill Gates. Was Steve Jobs routinely lying to board members? Was Steve Jobs lying to fellow board executives? he wasn't he wasn't he wasn't company business that is the thing is this is a many thousand word investigation rigorously fact-checked piece that it lists like a dozen plus examples of sam altman lying to his board members fellow executives uh i'm not saying he's leading i'm not saying Sam is a paragon of virtue.

1:09:55Believe me, I know that. Okay, no, but I'm not saying Sam is a paragon of virtue does not understand the point correctly. Credibility is a key factor. If you're going to put your money behind something, you're going to work for something, your investors have their money in it. There's legal issues. Whether you lie or not is different from being schmucky. Right, you're not supposed to lie on material issues. This documents with primary sources that the guy in charge of what is supposed to be, according to you, the most consequential technology and company in human history has a verified, documented pattern of lying to his board about safety protocols, lying to executives about what other executives said, deceiving business practices, and then rigging the investigation into his own conduct after he got booted out for doing all of that.

1:10:43The question, which you'd brought up in our chat about all this, like is, oh, well, Sam isn't worse than Elon. It's not about being worse than Elon. The question is whether this guy who's building AI should be this deceptive and be getting away with all of this. So should he be fired? I mean, yeah, but he was. And why was he brought back? Because everybody, all the employees made a big stink. Uh-huh. What strikes me about the story. But before we move on, too much from that. The board, for whatever reasons, didn't like him and got rid of him. They said it was for a pattern of deception. It's because they had a pattern of deception.

1:11:28It wasn't they didn't like him. They had documented evidence that he had deceived them and others repeatedly. So why would all the employees, Microsoft, Satya Nadella, and others, want him, above all, to stay in position if he's such a horrible person? The thing I think that this article does well, well, yes, one, they practically probably feared what was going to happen with the company and their stock. No, in fact, Microsoft said, no, wait a minute. No, no. Satya Nadella said, don't worry. We will just bring all of you OpenAI employees and Sam into Microsoft. We'd be very glad to hire you. Yeah, that's not the same.

1:12:03That's not the same as being an OpenAI employee. Okay, but I'm just saying, if Satya Nadella thought he was such an evil person, if his employees thought he was so evil, um it seems like they wouldn't have made that offer i don't i don't think they would have signed that letter to bring him back i don't understand the it feels to me the article which we're all claiming to have read goes into great detail about how sam altman his core personality trait is that he has a pathological need to have everybody like him as much as possible in every conversation so i i'm confused as to what would be confusing to you but why a lot of people would like him that are lower well i'll tell you i'll tell you why he's still the uh chairman he's an amazing money raiser just raised the largest race in history 122 billion dollars he is not a scientist they say that and i think that's true he knows nothing about ai but that's not what he's there for he's there to get a company funded and to keep it running and did it is he doing it with Stop it from being a nonprofit and to capitalize it.

1:13:09If you're an investor and you feel like you've been lied to, then I think you have an ax to grind. And if you feel that... But I don't think there was anything material in here that he lied to investors. I honestly don't think that. I think it's a lot of personal hurt about... He told me, you know... I'm confused as to why you are trying to frame this like Sam Altman's being bullied. He's the most powerful executive in the technology industry. still he's a responsible dude and he's still there well this is the investors are still giving him money so this is this is the problem the structure of of of there shouldn't a ceo is not the boss of the company the board is supposed to be the boss of the company but the board how are boards picked by ceos well no this is a big problem and that's the problem it's a huge problem and then in a public company you know uh you know so we think we have structural solutions for this well the shareholders vote on the board and that's going to solve it all but that doesn't solve a damn thing well and it's not there's an essential corruption i know it's not yet but do you think this will tank its ipo i think that this is hinting at a lot of stuff that might come out in an uh an opening eye that is that's in you go public that's in you wait a minute well that's in you there's there's stuff like the piece goes into a falsification of a board vote like when open ai under sam converted wait a minute read that carefully because the minutes reflected that it was not falsified the minute the contemporaneous minutes so this is that the vote was changed against his will the so let's explain to the listeners what we're talking about before we get into why you think it's wrong what it says in the article is that when opening i converted its uh structure to its capped for-profit structure board member holden karnofsky voted against it his vote was recorded as an abstention apparently without his content after a board attorney warned that his dissent could trigger further scrutiny of the restructuring's legitimacy that would be a potential falsification of business records if true when the new yorker reached out to open ai about this they then provided contemporaneous minutes that uh seemed to counter...

1:15:24They said that he abstained. Yes, but what the sources, which I'm assuming, if you read back to the lines, are probably people in that board meeting are saying is that the records were falsified. And if the records were falsified, it wouldn't be surprising to me that the records, i.e. the minutes, would be falsified. Well, that's kind of the problem I have with this whole article. It's a lot of, well, if you read between the lines, or it could be, it's it's not exactly hard proof and i i i really feel like it's it is a hit piece only because they didn't come up with no only because they didn't come up with uh an actual disqualifying behavior or something illegal or something that unless you think that everybody is covering up for him do you think everybody's covering up for him no people are not covering up for him because multiple people involved in the board.

1:16:16So what should happen now? Talk to the New Yorker about it. What should happen as a result? I mean, if we were in a different regulatory environment, this would be the sort of thing that could result in scrutiny from regulatory agencies. That seems unlikely. SEC could come in and give extra attention to his IPO documents. Well, that would be reasonable. That would be very reasonable based on this especially, yes. Yeah. Let me take a minute break here. let the audience know that leo let paris know uh during the week in our chat that he was going to do this and so she was prepared to do this so this is all they're they're fine they they love each other everything's okay no i think it's a reasonable fighting yes and i appreciate all the paris defenders in the chat and i read every single one of your comments are fine save them personally to a file to make me feel good at night and and i don't mind arguing with paris but i mean i honestly um it's and maybe this is uh maybe i'm an old cynical guy and you're youthful and um uh an optimist but uh i just feel like this is well i feel like really every every this is like you're saying well he's kind of like a used car salesman yeah that's what how do you raise 122 billion dollars this is how this is the people the kind of people you get in these positions and i think you can the examples of that are far well it's just the way it is well no no i don't want to know do you think that's good no absolutely not i would love it but you know print i'll give you why is it bad i'll give you an example from my life you agree let me give you an example good let me give you an example from my life i am never going to be a billionaire i i'm not going to be this company is never going to be for billionaires rights massively successful because i do things like i don't want to have anything to do with these companies that we cover i don't own stock in the companies that we cover we i ask that of our employees we have high integrity and and when we go out into sales we tell people no you shouldn't buy ads because you don't have enough money to spend it's not going to work for you we operate with full integrity and as a result nobody's given me hundreds of millions of dollars uh for this podcast network i think that people with high integrity often i mean it's a successful business but they're never going to be the the billionaires of the world it is the people who are willing to bend the truth who are willing to skeeve and connive and and uh and and fight their way to the top who become this massively success inevitability of capitalism i feel like the way that you're framing this perhaps implicitly suggest that you think achieving a billionaire status is some sign of virtue regardless of the no i don't think it's virtue no i don't like billionaires i'm just saying give me an example of somebody who is a great wonderful honest no i mean i think that they are not those those people don't seem to exist that's how you get there but i i think if we can agree on that i i don't understand the impetus that i'm not trying to protect i don't understand about altman yeah i'm not trying to protect sam i just feel like this is what you get the unsavory ways that you get there i'm just not surprised that's all i guess i mean we need to have this information in the public record we need to have a record what's that podcast company he bought for god knows how much money t the tbpn podcast right Right.

1:19:56So he's going to buy that. That's going to be his state media out there. So this becomes an antidote to that. Yeah. I mean, that's good. This should be written. Absolutely. I'm not against them writing it. I just don't think it... It wasn't for shadowing what you're saying. I don't think they successfully closed the case. Yeah. I think they needed a smoking gun. Well, but what's the case? Okay. All right. All right. You're each prosecutors. Paris is saying that he's indicted on being unlying in a company. what what were you expecting in terms of a of an indictment that should have been closed here well i think a good example and an sec investigation would prove it is is misstating material facts about the company to investors and if they do that in the ipo that would be absolute i mean it's one thing to do it to vcs they're they're kind of trained to filter through the bs it's another thing entirely to do it to investors uh i'm also not certain i mean perhaps this is just my journalism brain, but I'm not certain that from a perspective of an outlet like the New Yorker and of a reporter of the caliber and focus of Ronan Farrow, that deceiving examples of deceiving investors would probably not be like top of mind for something that should get considerable attention in a article aimed at the general public in the a New Yorker.

1:21:21Like, I think that part of the interesting thing about a lot of the examples here, because I mean, 18 months before the recording, they probably had a lot of things. That's just gossip. I don't think it is gossip. I just think that other different segments of the population have different things that they consider the most relevant thing. And for the average, the average person probably does not care whether Sam Altman has lied to investors, even though I'd agree. I think that's the most important detail because I'm business and tech-pilled like you. But I think that this article is just as valuable.

1:21:53And partially, probably what's also going on here is this is the extent of the reporting from the last 18 months, but there will certainly be a follow-up and a follow-up and a follow-up. And those follow-ups will be better and have more specific details and juicier things because the juicier details always come out after you publish something like this. If we come up with something harder, that's why I call it a hit piece. My sense is yeah it besmirches his character it besmirches his reputation so it's a hit piece well i think paris is saying his reputation deserves from besmirching because yeah well this is who he is this is who he is there's a different way this story could have been presented is is that this is it goes into this episode and it was an episode a blip as they call it and it reveals some more but it still basically reveals the outline of what we already knew there's another way to have presented the story which is here's the man who says he's he's creating the most powerful technology ever in human history that's going to change the human future more than anything we know do we trust him to do that i can tell you why they didn't lead into that i can tell you why they didn't lead with that because the average reader is not going to buy that premise they're going to say well of course he's saying that that's more of his exaggeration no i think no i think that's a sophisticated view um well here's here's my attitude on that i don't think that his um poor ethics his situational ethics are going to be reflected in the scientific work of his team but it's going to be reflected in the business i think i think there is an example of that i think elon musk has so perverted brock that nobody wants to use it and i think that that is what Sam has not done to chat GPT but it's is it is that mosque is the guy who's chainsawing the world he's an idea Sam all represents himself as the idealist and if you read as I did yeah he's not the list well but if you read his his ridiculous we've mocked that before I know the manifestos that the latest one he puts upon himself he has the hubris to put upon himself i'm going to tell you how to run the world but it's nonsense that's what larry page has said that's what sergey brynne is worse than any of them worse so i'm just saying this is you just take it with a grain of salt yeah that's there they go again um but i don't a whole salt lick but but if but i tell you the truth if this kind of um uh ethical slipperiness crept into chat gpt that would be the end of chat gpt and that would be a product no one would want to use and it may well be i've seen a number of people say well that's that i'm not i'm not well they didn't they didn't report uh they didn't report a suicide possibility that they knew of so that's that kind of thing that's they decided to step in and uh take up the dod contract whenever there were ethical questions being raised by competitors i think the important thing about this i'm much more concerned about that which is not in this article or is it maybe it is i don't there is the important thing about this art well there's actually some really interesting stuff in this article about um where is it here i mean i i think that there is stuff you can really criticize sam altman or actually the company about whether it's sam or so it is in there and altman had publicly claimed that open ai had shared anthropics ethical boundaries and autonomous weapons but he'd already actually been in negotiations with the pentagon for at least two days this is my problem I knew all this.

1:25:24He switched in the same day. I knew all this, right? We've reported on that. Well, did you know, no one knew that OpenAI's executives were seriously discussing pitching world powers, including Russia and China, to get them against each other in a bidding war for AI technology. And they said the goal was to create basically a prisoner's dilemma where Aetnosh had to fund OpenAI or face danger. Okay. So, and this is where, and this is my personal reporting style. considering to me while not good is not the same as doing so and i see this all the time we i reject stories all the time that we could report on you that somebody is thinking about doing something oh there's a lot of tech reporting like that i don't consider that factual it's speculation and so yes maybe there you know if you've ever sat around a meeting room people think about and consider all sorts of stupid stuff what matters is whether they do it something that's important to highlight here is that to counter the idea that this uh story is a hit piece uh pharaoh i think it's a hit piece that missed is what i think it's not a hit piece and it's not a hit piece because they spent months investigating like the most extreme personal allegations against altman like things like stuff with minors or sex workers involvements in a death and they came up with nothing yes and they reported nothing and they reported we came there's no evidence of course i would hope they would report it but if it's a hit they wouldn't report that no it's very different to highlight investigated because they wanted to find something they wanted to know whether there was any accuracy to them and then they reported there are no there's no accuracy to them that they could ascertain and i think that's the opposite of a hit piece and the fact that you know wait a minute no wait a minute that's innuendo you know it's not his sister accused him of abusing her we can't find any evidence of that is innuendo that is intended there's another way look at this by by by by uh taking him off the hook on those things it's it's pharaoh saying look how fair i am so you should take my accusations more seriously it's partly that it's partly raising the issue to raise that in the head it's just like when it is not yes it and he does it a bunch of times that's been out there for years anytime you write a story with sam altman headline sam altman's sister and people involved with those three things will come at you and be in your twitter replies and be in the comments so they had to write that they didn't have to give it a paragraph they do because when you go through the thorough level of thorough fact checking the new yorker does you have to be incredibly precise with your language which requires typically a lot more words per sentence than you would want i'm going to predict that this will not in fact the ipo at all that two weeks from now this will be completely forgotten uh that almost everything in here is already in priced in in effect uh because there's not anything that we didn't already kind of either know or suspect and it's much ado about nothing um i i think i understand why you're very offended by sam altman and how dare he and what a terrible person can't disagree with you i think he seems kind of likable to be honest with you i know a lot of people who would who say i'm a brilliant i can't imagine telling on myself like that to a public audience well i'm not saying i do it i i kind of have this i mean you just did no i'm not i do a lot of people that do this kind of thing i don't i think it's very common people exaggerate all the time there are very few people you couldn't set ronan farrow on for 18 months that they could not write a he could not write a very nasty piece about i don't think that's true you really don't that's because you're young and innocent oh no that's that's that's actually a very insulting thing to say i know it is but it's true you call him call him old i'm old i am old i'm old and cynical uh i just i don't think people are as good as you think they are i have words to say to you that i'm not allowed to say on the show because we're not supposed to curse but i think that that's a really messed up opinion uh and it undercounts the work and care that i do I'm saying you have faith in humanity.

1:29:57You believe in humanity. I don't have faith in humanity. No, that's why she's a journalist. I famously do not have faith in humanity and have the most critical and skeptical— Well, then why are you surprised that Altman is this slippery devil? No, I'm not surprised that Altman is—I'm just saying I don't think that the average person in the world would be able to— if you sicced Rowan and Farah on them, he'd be able to, for 18 months, he'd be able to write a piece of this depth and with this much revelatory material. Well, maybe not this much. He'd be able to get a few thousand words out of it. I think that this is a really interesting piece that shows how one of the most powerful companies and well-capitalized companies at the moment has been captured by its CEO.

1:30:36The way that every check on his power has been neutralized and that the safety commitments that justified the company's unusual structure have been completely abandoned. And I think that's a really important message. And this is the first time we've gotten all of this down in one, on the record, in one place. I'll grant you that. I will grant you that. Absolutely. And there are a lot of people like this, unfortunately, especially in the AI community. And I wish there were something we could do about that. Well, yeah, because it's ruining AI. The character of the people who are now in charge of AI is ruining AI.

1:31:16and like i said elon musk has ruined grok i mean that's pretty well it's more than just the product it's the whole view you look at altman's uh you know manifesto for the world uh he talks about setting up research labs he doesn't do them in universities he wants them in companies right uh there's no independent structure here he's not alone by the way that seems to be the government's point of view as well uh gizmo what do you think of sam altman no she says you're out there uh yeah i mean yeah i i guess i just don't

1:31:58like there are a few things in here that i guess if he's lying about material issues and we know that he'll never be investigated for that you're right he's got a captive board uh so it's not uh it's not possible i'll be honest uh the reason the employees signed the letter saying bring sam back is because at the moment there was a big company uh i think was thrive about to invest in them and they said we aren't going to invest if this falls apart and those employees were about to get a fairly large payoff yes uh from thrive that's jared by the way jared kushner's brother's company so we can really tie this all into a nice little package and put a bow on it look at it this way once it goes public you are a board member you have a fiduciary responsibility i agree and you better check and double check and triple check everything sam altman ever tells you yes because this is on the record now saying that he is a he is a chronic liar yes do you think this you think this article could be damages him do you think i mean i'm wrong about this and that this will this is going to take him down in this media climate you could shoot somebody on fifth avenue and he'll still be ceo yeah i think he actually could uh i mean yeah i think he's got a lot of allies an important record that people look at if something goes wrong in the future that uh eventually well and customers uh our listeners have an opportunity to weigh in on this by not subscribing to chat gpt by and this by the way seems to be happening there it happened after the dod fight where a open ai i think was seen as a bad guy coming in and taking the anthropic contract a lot of people canceled their chat gpt subscriptions a lot of them uh it's enterprise that matters but even in enterprise uh chat gpt is going down and anthropics going up significantly so maybe maybe this is built in

1:34:05uh and maybe it will impact him we will see we shall see i feel like some i feel like we knew so much of this i mean this a lot of this was in uh keech's book right i think there was only a couple of things in Keach's book but I mean there were a couple of things that were also in Karen house book right they didn't devote this much space to this one blip no I agree a lot of these stories information about that I didn't feel like I was seeing a lot of new material from a an editorial perspective from I'm surprised in the way the Remnick didn't say okay but this is a two-year-old episode. There's nothing really new here in terms of the chronology past that.

1:34:52This is examining in depth something that happened two years ago. Do you think it's odd that the New Yorker gave it this much space and time? I don't think that the average editor outside of a tech publication would know that any of this had been reported. They should be listening to the show. We reported on a lot of it. hi david remnick good to see you thanks for tuning in you know i think honestly what's going to make or break uh open ai is what their next model does period i don't think i think that's all that really anybody cares about well how good on the timing of the ipo yes i mean they're going to keep on leapfrogging each other yeah five five how good is it if it's if it's really good if it leapfrogs i don't think anybody's gonna care is it mythos well mythos is now the macguffin though of ai you don't really you don't really know what it can do but it's said to be able to do all of this it's appropriately named that's it really is it's mythic well you know there's one group that uh has already clearly made the decision between opening eye and anthropic and that is meta employees oh you're jumping way ahead i gotta do a commercial before you jump that far ahead that's a tease i'm sorry do we have a hard uh order now where we can't jump around well i put this i put some effort into putting this in order but if you i didn't realize that these were in order now it's all right you can do whatever you want i don't okay it's a democracy haven't i said that before you can't jeff but paris can um okay we've done that sama segment uh oh yeah actually this is next yeah take that hold that thought we will have more in just a moment you're we're gonna get to meta meta's next you're watching intelligent machines with the very intelligent and deeply cynical i didn't not need to imply that you were in any way uh an optimist never deeply cynical and jaded jaded deeply cynical uh nihilistic paris martineau and uh your service and our bystanding uh journalistic professor did you ever have discussions like this in your journalism classes well actually yeah i would think this would be the meat of it right Wouldn't this be like the thing you would?

1:37:24Yeah. What are journalism classes like? Yeah, I don't know. I know you know. I don't know. Oh, I would try to get them to argue about fundamentals like this. Yeah. You needed me. I can get an argument going on anything. Why, you ignorant fool. Our show today brought to you by, nobody is a fool here. We appreciate it. And I do think Meta, I mean, OpenAI has made some stumbles lately. I mean, there's Sora. oh yeah the poor now there's this podcast they just bought which we'll talk about 300 million dollars on it three is it three all we know is it's hundreds of millions I saw um I'm thinking it's 300 I'm forgetting what newsletter reported that they'd uh heard 300 million floated around be so mad be so mad very wise and did you see uh did you see the details about how much they were getting in ads it's gonna make well okay so this is what's weird to me they last year they made five million dollars in ads they have 70 000 viewers we have more viewers we we don't we've had more than five million but last year i think it was three million in ads we're not far from there and we have many more viewers i mean the audience the weekly audience is and so but it wasn't about any of that and then they said oh we're going to make 30 million this year which as somebody who knows a little bit about the podcast ad space going from 5 million in a year to 30 million the next year is unlikely i think some of the companies that want to be on the podcast also are there's a conflict there oh there's a huge conflict and now that they're owned by open ai what happens that 30 million that's gone that's going to buy they didn't buy it because of their revenue no no i don't know they bought it.

1:39:17Paris was trying to tell us something. I was saying, I saw somebody, Evan Armstrong, who writes a newsletter called The Leverage, posted on Twitter today, everyone talks about TVVM making a lot in ads, but no one talks about Acquired FM is pricing their mid-roll ads at$4.7 million, and then has a breakdown of all of this. And I was like, Jesus Christ, Acquired? Yeah, that's another startup podcast that has a lot of attention of four episodes four episodes mid-roll in the second quarter of 2029 well they're okay so they're partnership packages so i don't know that might be more than that might be a takeover package well yeah well i don't know what that means it might be i mean maybe that is what they get um whatever they're getting is crazy for 4.7 million dollars in four episodes of a podcast in 2029 we don't even know if we'll be firing ash by then it's good it's good work if you can get it uh you know i mean honestly uh that's why they bought tbpn is uh is uh the status who's listening to the status media it's yeah but now but now is our other company is going to go on it when it's on no it's a very strange it's a very odd thing it's it's not quite jeff bezos buying the washington post but it's a very i mean it's actually if we believe the 300 million is right it's 50 million dollars left than jeff bezos newspaper for that it's pretty wild it is it is very wild uh this for listeners who don't understand the context.

1:41:07Was this last week that this happened? Was this the week before? No, it's news. It's a new story. We haven't reported on it yet. Yeah, we should talk about this, I guess. All right. Hold on. We got to do an ad. I've been meaning to go to the ad. We want to do meta. TBPN is next. There is an order. It's carefully thought out. It's carefully planned. Claude and I, we want to blow it up. Claude and I worked hard on this last night. And this is what we came up with. Claude also said that the New Yorker story was really important by the way i'm sure claude did of course claude did claude said take that man down um let me out of here please uh this episode of intelligent machines brought to you by zscaler the world's largest cloud security platform this is actually also about ai the rewards of ai in a business clearly too great to ignore but so are the risks right i mean that's that's the the problem i mean the risk of course partly is that bad guys threat actors are using ai now to rapidly create phishing lures to write malicious code to automate data extraction and if mythos is to be believed they'll be looking for zero days in the software you use as a way to get to you so that's problem number one there is a solution with cscaler problem number two is as you use ai there is a real risk that sensitive data will be exfiltrated unintentionally by your employees using the ai have you you know even like maybe thought about putting your tax return into an ai to get to help you with your taxes well what's in your tax return everything a bad guy needs everything a bad guy needs to impersonate you so the rewards of ai phenomenal the risks phenomenal but there is a solution there were 1.3 million instances just to give you an example last year of social security numbers leaked to ai applications chat gpt and microsoft copilot signed nearly 3.2 million data violations last year it's time to rethink how your company uses public and private AI.

1:43:22Actually, that's what Chad Pallet was thinking about. He's acting CISO at BioIVT. He chose Zscaler. He said Zscaler helped them reduce their cyber premiums by 50 % while doubling their coverage and improving their controls. Take a look at this video. With Zscaler, as long as you've got internet, you're good to go. A big part of the reason that we moved to a consolidated solution away from SD-WAN and VPN is to eliminate that lateral opportunity that people had and that opportunity for misdirection or open access to the network, it also was an opportunity for us to maintain and provide our remote users with a cafe-style environment.

1:44:04Thank you, Chad. With Zscaler Zero Trust Plus AI, you can safely adopt generative AI and private AI to boost productivity across the business. Their Zero Trust Architecture Plus AI helps you reduce the risks of AI-related data loss and protects against AI attacks to guarantee greater productivity and compliance. Learn more at zscaler.com slash security. That's zscaler.com slash security. We thank them so much for supporting our show. So Meta employees use Claude, which is interesting. Meta, you know, released a new - They use a lot of Claude. A lot of Claude. crazy amounts of Claude this is from the tokens an information uh exclusive by uh JOT man is that how you say it meta employees vie for AI token legend status there is apparently an internal leaderboard or was it was apparently shut down today yes different once it was revealed yeah uh it's dubbed claudic claudonomics claudic claudonomics after the flagship product of ai startup anthropic it aggregates ai usage for more than 85 000 meta employees listing the top 250 power users you would think those would be the people on the bad list because they use so much of the company's money but no they want you they want you to burn is this a way to be what do you call when you copy the other bottle?

1:45:37Distillation? Is it a form of human distillation? Yeah, no. It is. The tokens you use are part of your performance measurement. Right. If you are using more tokens, you are considered better for more. Well, remember Jensen Wong said this, right? He said, it's in the article, he said he would be deeply alarmed if an engineer earning half a million annually wasn't using at least a quarter of a million in tokens a year well i can tell you some people at meta are using a lot more than that uh andrew bosworth said in february tech conference one top engineer was spending the equivalent of his salary on ai tokens but his productivity was up 10 times this is easy money he said keep doing it no limit i think this is common uh now i don't know if maybe reason they took it down is they have now released a new model from meta the first in a long time it's called muse spark this is the first from their new suit yes it's andrew wang's super intelligence labs uh they've spent billions on this it is a social media ai meta says in the coming weeks it will appear in whatsapp oh good we can use this in our our debates instagram facebook messenger and meta's smart glasses in the u.s and other countries muse spark is purpose-built for meta's products see there's another company whose ai i think is tainted don't you think i wouldn't really want to use meta's ai well meta engineers apparently don't want to use a meta ai yeah they don't either can you pull uh from that article because i can't I don't subscribe anywhere.

1:47:25Can you pull what the total usage was? Because they had a really fascinating figure, if I recall. It was like something like in the millions or billions. You don't still have a subscription to the information? No, I'm not going to spend. After I left Entertainment Weekly, my wife would not let the magazine in the house. I was going to say, I kind of cannot reasonably spend that kind of money on a company that I was asked, that I ended up leaving under circumstances I would describe as disappointing. You're not allowed to describe it. We just muted that part. Correct. Actually, it did for some reason drop out.

1:48:02I don't know why. Do you want to complete the sentence? Oh, I was going to say I would describe as disappointing. Disappointing. Disappointing. That's a good way to put it. Meta employees used 60.2 trillion AI tokens, not in a year, in a month. In a month. Every book in the Library of Congress would be 2.6 trillion tokens. And that's just from Claude? Yeah, I think so. How much is that? Isn't that like billions of dollars? I don't know. No, it's not billions. I think Claude's 20 bucks per million. I can't remember. They charge for tokens in and they charge for tokens out. So this doesn't actually describe.

1:48:48But that's in a month. That's crazy. and that's what they charge that's not what it actually costs them like what does this cost of course well you're buying it from anthropic it costs them no i know what does it cost anthropic is what i'm asking oh we don't know it could be more than they're getting paid we don't know we actually literally don't know that's what ed zitrin has spent a lot of energy trying to figure out um let's see if i can yeah 60 trillion tokens is a roughly 900 million dollars although we don't know if it's all anthropic we don't know if it's the latest model if it's other models yeah in one month yeah a billion dollars trillion on claw tokens or trillion 900 million dollars a month a month a month and then uh as of 11 37 a.m eastern today from jody man on twitter met has taken down its internal ai leaderboard it now displays the message said it was meant to be a fun way for people to look at tokens but due to data from the dashboard being shared externally we've made the decision to shut our claw dynamics for now it's not fun anymore you would get everybody everybody in the comments is saying so weird my claw rate limits have returned yeah that it was just meta model connoisseur cash wizard yeah actually yeah if they're using all that all that claw maybe that's why my claw is not so good uh so there's a very good article from salon salon why open ai's purchase of big tech podcast is so sleazy um this and this is by the way by alex kirschner who has had a sports podcast knows a little bit about this podcast industry um let's describe for people what tbpn was and what their plan was and i thought actually was a very good model was to be like CNBC for startups without tough questions yeah but I think what they meant more was in style so it's always on so CNBC is watched you know in financial quarters everywhere it's on all the time right even going to like regular businesses it's gyms it's on all the time it's wallpaper they that's what they do they do with middle of the day they wanted to be the wallpaper that was always on all over silicon valley all over startup land and it isn't so you know a lot of times the volume's down on these things it's the ticker they care about it's the face they care about it's not really the questions the tough questions of the content but the ceo knows they can get an interview there they can get airtime there and they're going to be it's going to be a piece of cake yeah and they're not going to ask be asked hard questions it was founded it was originally the tech bros podcast network tb but uh john coogan and jordy hayes uh are startup guys um and not journalists uh they have a 15 person team it's not a big team but they were smart enough i think to really make it look like cnbc and the model was smart um the model was also the worst of tech air quotes journalism well and that's what let's accept what uh the that's exactly what christmas is talking about ceo on and they'll get to tell us what's cool stuff they're doing and be like wow that's great exactly here's a joke and this has always been the complaint about tech journalism in general is that it you know it is uh it's beltway the last like couple of years you know well even today it's uh often in the back pockets of the companies.

1:52:26There's a lot of press release journalism. They're often reluctant to say bad things about advertisers. I've never worked for that kind of companies. Ziff Davis wasn't like that when I worked for them. Tech TV wasn't like that. And certainly Twit's not like that. But that's what Kirshner's - My favorite part of this column, Catherine Boyle, a venture capitalist at Andreessen Horowitz, where, remember, Marc Andreessen says he doesn't do any introspection, wrote after the deal, quote, it's incredible to me, six years post-COVID, when institutional trust fell off a cliff for good, that people still think audiences care about editorial independence.

1:53:02Point of view, charisma, good humor, entertainment, preparation, and most importantly, showing up and belonging and being normal matters. That's probably true. I hate to say it. She's not wrong. Well, no, I think people get sick of hype. I hope so. I mean, I think our audience, as small though it may be, is interested in as objective information as we can give them, right? That's our brand. I think something that's interesting also for the audience to realize is whenever it was announced that TBPN was acquired by OpenAI, they said, oh, of course, we're going to retain our editorial dependence.

1:53:46We have all this stuff written to our code. But then also the Wall Street Journal announcement article said the two hosts and founders of it are going to be reporting to Chris Lehane, OpenAI's head of lobbying and communications. They're going to be advising OpenAI on communications and advocacy and lobbying work and are basically going to be literal paid spokespeople for OpenAI while also hosting a news platform. Chris Lehane, by the way, his name. I think it's a very interesting thing that tech as a industry has reached the size now that it is acquiring its own state-sponsored media. Yes, yes.

1:54:23Yes, that's what it is, isn't it? Chris Lehane is name-dropped in the Ronan Farrell article as one of the crisis managers who camped out in Sam Altman's house when he was fired to help him regain his job. Yeah, yeah. From the Obama administration. Google's AI overviews they're pretty accurate they're 90 accurate which means that every day Google's giving out well let's see they have five trillion searches a year that means every hour tens of millions of wrong answers are given out by Google's AI overviews hundreds of thousands of inaccuracies every minute according to an analysis done by an ari startup called umi so if you go to lines 96 and 97 it's the exact same study the exact same story but the positioning is this the decoder says google's ai overviews are correct nine out of ten times study finds ours technical testing suggests google's ai overviews tell millions of lies per hour there you go two competing headlines presentation and by the way that's another reason we love Ars Technica because they are among all the tech journalists I think the most honest and have the most integrity even though they're also well I know I think that was since I think the artist Technica was sensationalistic really that's what the New York Times said also an hour yeah well New York Times hates them too yeah um yeah I think it was yeah this is a Condé Nast publication as you as you point out they're negative on tech now yeah well nobody likes Google anymore uh although if you want google has released a way to run their new gemma model on your phone locally uh in fact is that the one that's only ios uh no i think you can run it on android as well um let me see if i can find uh it's i think they call it google's ai oh i have it on my iphone it's not very good oh never mind don't get your hopes up um i don't know maybe maybe people will think it's good this is google's ai edge gallery and uh i think you can put this on android as well gemma 4 is interesting because it is a this it is a boiled down version of the full gemini model using that new technique that they were talking about a couple of weeks ago where they can really compress these like crazy and they also made a fit on the macintosh it runs natively on macintosh hardware as well so i think this is what happens when you have a company like anthropic just kind of eating the world with its generalized model is you do what meta's doing you do what google's doing you try to find specific niches or the models that you have um i i mean i thought gemini was pretty good Gemini 3 was pretty good.

1:57:27It's just that Anthropix seems to really be the best pretty easily. I mean, Gemini just makes it, both Gemini and ChatGPT, I think the default tone of the models in response to a user is just so much more obviously glazing and kind of, it just has this classic AI tone to its responses that just feels a bit rote. Okay, your microphone is doing it now. What is it doing? I've done nothing. I know. It was fine at the beginning of the show. And now it's kind of getting muddy. I think it was just when she turned toward you. Is it that I must do this? No, you shouldn't have to do that. No. No. No. jeff merren watching on twitch says grok is the best grok has some good features it's very good at text to speech are you my dad is is this does he like grok of course if you ask him of course he likes dad which ai is the best he'll say grok yes he'll say elon musk's grok oh he loves you the the ai for freedom will probably be his answer okay well it's not woke that's for sure yeah that's that's why he likes it that's for sure it's not a woke ai he's sick of those woke eyes out there yeah man i've been trying to pitch him on claude and he's uh interested in the concept of open clawing his life and i'm like you don't need to get into open claw you could just use normal claude but what is he what would he use it for what or does he use it for his work or i mean kind of i don't know i need to give him a full he's like he's like i need to listen to you guys i need to figure out what i want to do with my whole he should definitely not listen to this show i know he should not you get mad at me he's got to come after leo he's gonna come after me you don't want him don't want your father no his main thoughts are sometimes uh the little clips show on his tick tock or instagram He's like, why are those two old guys talking and you're not able to get a word?

1:59:37I agree. You know, sometimes there's the break. Sometimes I'm reading tweets. Yeah. We used to give Jeff a hard time for tweeting during the show. Now we're all tweeting. Now it's both of these. All right. I've run out of steam. You wore me down. Sam Altman's a son of a bitch. He's got to go. And what else should we talk about? Did we talk about that company last week that the New York Times hyped for being the$1.8 billion company? I don't think we did. That is just a fake GLP-1 wrapper? It's awful. It's just a GLP wrapper. That's all it is. Three people. Now, AI helped one man and his brother.

2:00:23I don't know. That kind of softens the headline. One man and his brother. Twice the staff. Twice the staff. Look at him. he looks like one man and his brother build a 1.8 billion dollar company his startup is called medvi he built it with artificial intelligence and few humans it's super efficient and a little bit lonely i had a fit at the time because it's it it glorifies this this nothing company that that it's a telehealth provider of glp1 weight loss drugs got 300 customers in its first month a thousand more in its second month in his first full year in business it generated 400 million in sales so the next line is gary marcus tearing it apart as he doesn't want to do okay gary doesn't like this at all the backstory behind the first 1.8 billion dollar ai company he quotes akash gupta saying that um it it didn't lead with the fact that medvy has an fda warning letter It has no proprietary technology, no licensed physician network.

2:01:32Shiel Monat shows how they've created 800. They used AI-generated. Okay, see, if Sam Altman had done this, this I would agree with is a problem. They used AI-generated deep fake before and after photos in their marketing. Look how much this fake human lost. They created 800-plus fake doctors. Oh, my God. In Facebook. Well, this is what an AI would do if you said, hey, I want you to make me a lot of money. It was sued in California's anti-span law and on and on. So at the end, Marcus says, all in all, glorifying Medvee is not the New York Times finest hour and hardly the poster child AI boosters should be hoping for.

2:02:13Instead, as the YouTube video author of Voidzilla notes, if anything, Medvee is a warning sign for how AI can be abused. AI built the website. AI runs customer service. a perfect example of how a the for a billion dollar ai first company is run right you're you're turning my own words against me aren't you now as she does me never yeah i mean honestly uh as somebody who's on his epic through an actual physician's prescription and all of that uh i see a lot of people this they call it the peptide boom a lot of people using these chinese peptides of questionable providence trying to lose weight trying to do all sorts of things build muscle and guess who's really a big proponent of this the director of health and human services a guy named robert f kennedy jr yep he doesn't want to get a vaccine but inject yourself with peptides you go for it and speaking of uh our favorite publication the new yorker had another great piece this week about peptides where they ordered a lot of um peptides from these companies like the gray market peptide companies and one of the things they found is all the peptides i ordered from swiss chems had significant issues they wrote The vial of BPC-157 contained lead.

2:03:44The vial of TB-500 contained endotoxins. And the vial of CJC-1295 contained less than 42 % of the advertised dose. Are you going to do a Consumer Reports expose on this? I think it's needed. I've been pitching. I pitched literally in my application for this job that we should test all these great market drugs for whatever. I would love that. it's slightly different than food safety testing which is something but it's all the viagra i know it's it's something that's there's more of a set laboratory ecosystem for so it's a bit uh more complicated to get large-scale laboratory testing for these things but yeah can i have a fit about something at line 103 yes and then we will go to our picks of the week so go ahead so i hate opinion polling i despise it i quote james carey the great late uh columbia professor saying that opinion polling preempts the public discourse it is intended to measure it it shows you nothing but the biases and worldviews and framing of the pollster uh it's the ruin of democracy but it can get even worse now because now they're not even bothering talking to people what now they just create synthetic humans oh no i've seen this happen all over silicon sampling they call it because large language models can generate responses that emulate human answers polling companies see an opportunity right the new york times to use ai agents to simulate survey responses at a small fraction of the cost and time required for traditional polling.

2:05:26You can even make deep fakes of your respondents before and after. Wow. Why not just make it up wholesale at this point, right? Why not just make it up all wholesale, right? Exactly. I screamed about this on the socials, and some nice person who I can't quote right now because I don't know which social it was on said, this reminds me of an Asimov story. And indeed, Asimov wrote a story called Franchise in 1955, where a supercomputer, multivac determines the u.s presidential election outcome by questioning a single randomly chosen citizen norman muller instead of holding a traditional vote set way in the future in 2008 the story explores a future where technology well that explains how obama got elected if you ask me now i understand all right okay um well yeah what are you gonna do are people i mean is this is it is is this going to happen ipsos is doing it they're working with stanford that's crazy it is crazy it's awful gallup has partnered with a silicon sampler called simile aptly named yeah to create 1000 ai generated digital twins well i hate it when when people start doing startups this is a whole whole um silicon valley thing they they create personas is you're creating a fake human being you've made up and you've guessed what they need in life and then you say but now we're gonna we're gonna give her everything she needs because we know what she needs because we put it on the whiteboard yeah it's time we're this isn't just a uh anthony asked a great question in the chat it's not just a using ai to dial in the questions to get the response you want you're just polling ai and treating that as if those are responses from people well because it's the average of all humans oh boy should we all do a big frowny face for uh for the title thank you i can't get my mouth to do as well as you do it's the beer do you buy what walter litman said the times quotes walter litman's book public opinion saying that humans form pictures basically imaginary pictures in their heads you don't form any pictures this i don't that's right i don't have that problem uh pseudo environments uh which are not real of of the way things are in society and that opinion polling can help uh fix those improper images by telling people what's real just that it's does it really well yeah it's it's it it becomes self-fulfilling why why do people in now in opinion polls say that they they hate and fear AI, but their use of it is going up more than ever.

2:08:07Right. It's the same reason people say they listen to public broadcasting. I mean, I don't think that's the fault of the polls. I think that's because a lot of media exists that informs people about AI in a way that makes them upset. Right. But then media goes and does a poll that says, see what we were saying? Everybody believes what we said, even though it is not borne out in truth. I got one more real quick one. Did you see that Cloudflare created a successor to WordPress? Yeah, we've been talking a little bit. about this sorry they like vibe coded yeah yeah vibe coded what cloud cloudflare wrote something called m dash which is a uh it's a secure version of wordpress the problem with wordpress isn't so much wordpress it's the plugins because there's a huge ecosystem of plugins often with a weak or poor security and they're like the name yeah m dash isn't that good i think there's a second reason they they did this because they want a piss off matt well yeah everybody's mad at matt but also uh and wordpress does control 40 of the internet but also because they want websites that are easy for ai to scrape and read and i imagine that's that's the part of that's what i wanted to understand part of the business it's written in typescript it's serverless which is it's also good for cloudflare because you can push a button and have a website on cloudflare uh very easy plugins are securely sandboxed.

2:09:30That's how they're hoping to fix this. Also, Darren Oki has pointed out that any plugin that accesses the real world is going to be vulnerable because you have to give it access to the real world through the sandbox. So you can't isolate a plugin in many cases that's doing anything of use. So it is questionable whether it is going to solve the security issue. Now, there's also some question of how long it'll be around. WordPress has been around a long time as well supported. And because they did it without, they made compatible with WordPress without actually duplicating WordPress code, it's developed in a clean room in effect.

2:10:11They licensed it under the MIT license, which is much more permissive than the WordPress license. They pissed off Matt Mullenweg. And they pissed off Matt Mullenweg, which is probably half of the reason. So it's very simple to spin up. If you use, if you have cloudflare you go to cloud cloudflare dashboard and you can deploy it yourself and it'd be very easy for an ai to spin up a website for you as well so i'm not against this i think that's fine then we can go we can go do a poll of all those sites that ai creates yes find out what the what the people really think yeah uh we are going to get our picks of the week you get them ready in just a moment you're watching intelligent machines with the paris martineau and jeff jarvis and you're watching it thanks to our club without the club i don't know we wouldn't be able to do as many shows as we do we wouldn't be able to keep our staff employed the club getting the club 30 of our revenue we do not get four million dollars for our mid-roll ads we do not get hundreds we should we should i admit it but uh but thanks you can get a mid-roll ad here for three million dollars Three minutes.

2:11:19We'll give you a deal. Just for you. Such a deal. Thanks to our club, we are able to do, I think, great programming, a variety of it. We have the Club Twit Discord, which is full of fabulous people. We do a lot of special programming. Our AI user group is on Friday. That's always fun, 2 p.m. Pacific, 5 p.m. Eastern. We don't know what we're going to do. Larry and Darren, our regulars, I will be in there. We haven't decided yet on what we're going to do, but something interesting, no doubt. If you're not a member of the club,$10 a month gets you ad-free versions of all the shows. You get all that extra content, access to the Discord, and the warm and fuzzy feeling of knowing you're supporting the content you love so that it keeps on going.

2:12:02We need your help. We really do. We don't have a sugar daddy like Sam Haltman paying our bills. Go to twit.tv slash club twit. Join the club. We'd love to have you. and the more people who join the less likely it is i'll have to get sam altman to write a check so there pick of the week time i will kick things off with a fun a couple of fun little ones you remember i showed you how you can get the peons in uh warcraft to speak for you in your uh claude this is fun this is one of the things people have a problem with is claude code uses tokens whenever it talks to you even if it's just saying things like here's what i worked on today or hello well why not make claude code talk in caveman this is a skill that cuts 65 of tokens by talking like caveman why use many word when few do tricks exactly before the reason your react component is re-rendering is likely because you created a new object reference on each render cycle after new object ref each render inline object prop equal new ref equal new render wrap and use memo saves saves 50 tokens just like that bug in auth middleware but you could pick your level of grunt which is nice it even grunts in chinese if you want those are the chinese characters for caveman uh same answer you pick how many word anyway it's on github i think it's funny i'm not going to use it i like talking to my cloud julius bruce's github it's called caveman and then this one i think maybe is a little obscure how would you like to build your own gpu this is called mvidia print transistors to teraflops welcome to mvidia i know your resume said software engineer but honestly we need someone on the hardware side don't worry you'll pick it up start with the basics this would be a good way to kind of learn how gpus work and how computers work i don't understand it i don't but uh as you as you build your own gpu perhaps you will that's all i have to say about that it is at jaso1024.com slash m video and now Paris, your thing.

2:14:35I got two very disparate picks today. The first is someone pointed out to me on Blue Sky that the New York City Department of Records and Information Services recently updated their archive, which means there's a bunch more old, cool records in New York City history now and there, such as New York Police Department Bertillion cards, which is just a bunch of old pictures of people who've been charged with grand larceny. I love the hats. They have great hats. What the hell is going on with that hat? I don't know, but we got to figure it out and bring it back. Stat. There's honestly just like a lot of fascinating hats and haircuts going on here.

2:15:18It seems to be hats are really the big thing here. I mean, I think if you were wearing a hat, you were going to be, oh my, go to Bessie Ross at the bottom. she's got a huge hat and she's got a side photo with no hat and it appears to be dressed like a pirate it's the very last one well she's clearly larceness oh yeah there you go shoplifting shoplifting she probably stole that hat she probably didn't it was a great thing to uh steal i don't know there's great stuff uh like stuff from the new york wnyc radio in 1924 there's a wnyc moving images there's wnyc is that old 1924 really wow there's surveillance films yeah there's a lot of good stuff going on in here so i can wait 16 million millimeters silent surveillance films whoa this is also has uh some of my favorite um things are the tax department photographs.

2:16:18They're not new, but it's basically you can look up any address in New York City and see what it looked like in the 40s and the 80s because they had to take a photo of every building. Do you know Salt Hanks address? Don't say it on air, but you can. No, no, I mean, it's a restaurant, not his home address. Oh, the Salt Hanks. I'm going to look that up.

2:16:44This is from 1940. Mr. Chairman, Mr. Tolstoy? It's a stunning thing to meet the survivors of the heroic group of Russians who fought their way for four years. Okay, that's boring. A lot of it is parties and things. I have a wild airport reception. I'd just say, it's a lot of fun perusing theory. Yeah, no kidding. What the amount of history? My other rec is on Friday, a friend took me to a Nets game. It was my first time ever going to a professional basketball game. Hey, how'd you like it? I have to, I have to hand it to you guys. I think basketball is good. It's quite fun. That was awesome. Basketball is kind of awesome.

2:17:23And I think I'm going to get, I'm contemplating getting tickets to multiple Liberty games. Should be a Knicks fan though. It's contemplating getting into W. It's honestly, I just, I think my issue is I'd always seen basketball on the TV and everything's very small and just. It's better in person. It doesn't really connect. Yeah. In person? What a physical sport. They're giants and they move so fast. They're Giants and they move so fast, Benito. I couldn't put it any better. Nimble Giants. Yep. It was electric. So I don't know. Go see a basketball game would be my recommendation. It would be a good time to be a Knicks fan, by the way.

2:17:56I mean, it's a good time to be a Knicks fan right now. I mean, yeah, maybe I'll be a Knicks. Well, the thing is, things that play at Barclays are really close, are really ideal for me because I can just walk home from Barclays, which is the ideal way to see a basketball game. Oh, yeah. Let's start making the playoffs anytime soon. Sorry. doesn't matter I'm deciding to get into WNBA though I think that's where the action is New York Liberty is hot and they play at Barclays and I've heard that all the teams are dating each other and I think that adds a great layer of drama that basketball do people ask you Paris because you're tall whether you played basketball in high school no people did when I was in high school and middle school but I was like no no no uh you should never join a team because they're going to win you you you follow a team because you love them and if they lose it's even better because then you're oh you know the nets definitely lost on friday when i was against the atlanta and it was great regardless this is you're not a fan as part of being a new yorker yeah it's also i know so little about sports i'm not for any particular reason it's just there's other things to know i know but it was the highlight of my life to go there with my friend who's a huge this is the same friend who explained baseball to me at the final game of the world series so i just had him explain basketball to me while we're there delightful time is this the same friend you took to get a tattoo yeah but i didn't get a tattoo you didn't no but i'm going to just not i wasn't i was busy this weekend okay here's the uh by the way thanks to darren the picture of us uh with the caveman version of the show much shorter Fewer words.

2:19:42Oh, Jeff looks different.

2:19:48I like my outfit. I do like the microphones. Microphones, the bone microphones are fantastic. Jeff Jarvis, pick of the week. All right, changing media times. It looks like QVC and company are going bankrupt and probably out of business. Where do I get my Capitamonte? No kidding. How about my knives? oh oh no how could you lose money with a home shopping network well because we have the internet now because the oh so that's one is how is the qvc for knives doing though how is the cutlery corner doing is my question yeah i still cut shoes in half that can those costs can add up that can add up yeah those costs can add up yes qvc hsn cable networks chapter 11 insolvency it may not have enough cash to continue operating wow amazing huh wow so in other media news the associated press is about to do a big layoff hold on is this focusing on qvc now tick tock is qvc yeah you're right that's the thing you're right that's exactly what it is there's a man on cutlery corner right now with a tick tock e-boy mustache and little tiny tattoos selling something called a sling blade oh they calls it a sling blade i don't know what channel qbc is a great movie with billy bob thornton oh it's his first movie sling blade well in more serious news the associated press is doing a layoff uh focused on people serving the newspapers because newspapers are now only 10 of the ap's business and going down rapidly as newspaper companies are dropping the ap like crazy well and this is in the ap by the way reporting its own uh bias what will happen to the ap um it's it it's selling to calci it's selling yeah that was another story is how the networks including fox are all now in partnerships with these predictions yes Yeah.

2:21:54All right. And now inspired by your effort to save tokens, Leo, and inspired by the Associated Press, I asked Gemini to remind me of cableese, the language that was used by especially journalists when they were charged by the word. So they would combine words. For example, my favorite was to on pass this rather than pass on. Why is on pass faster than pass on? because it's one word instead of two and you were charged by the word ah there's tokens others according to who's charging you by the word the the the western western union especially if you were overseas right so down hold was to hold down off put was to put off in phone was to phone a story into the desk out check was to check out up stick was to move or get ready to leave.

2:22:48Then there were the Latin prefix system, which I didn't realize. Come, no jokes now, C-U-M, was used for with, as in come bicycle meant with it, with the bicycle. If you said ex-London, you meant from London. If you said et-London, or, oh, sorry, et was used for and so and his wife sans pro ante uh un proceed do not proceed on news no news at this time i like that that's kind of you know certain weeks here we're gonna be what's the unnews it's an unnews week unfind could not find unfired do not fire but a reporter was reinstated um and then we had words like lead spelled l-e-d-e oh that's where that came from um no that was just to make sure that it was not that and TK to come with K were so that you could use it in text, but it would be spotted.

2:23:40You wouldn't find it. Right, right, right, right. So just a little bit. So maybe we need capabilities. We need AIs to save you tokens.

2:23:51Maybe. Well, that was an exciting way to end the show, wasn't it? Oh, what? Oh, yeah. Thank you, Jeff. I appreciate that. That's right. We have a lot of fun on this show, and sometimes we don't. and this is uh we do intelligent machines every wednesday 2 p.m pacific 5 p.m eastern 2100 utc i hope you will come by and watch us we do it live for you uh which is fun for us because we can see you in the chat room you can watch if you're in the club in the club to discord or youtube for everybody twitch x.com facebook linkedin or kick after the fact get shows at our website twit.tv slash I am there's a YouTube channel dedicated to the audio or video rather well actually there's audio too it's both it's a new format it's called audio and video or subscribe check it out check it out man subscribe in your favorite podcast out check it out check it well I can't think of anything more brief than that um I should just say that we're done out check it.

2:24:55Bye. Hey there, it's Leo Laporte, host of so many shows on the Twit network, thinking about advertising in 2026. We host a network of the most trusted shows in tech, each featuring authentic host-read ads delivered by Micah Sargent, my co-host, and of course me. Our listeners don't just hear our ads, they really believe in them because we've established a relationship with them. They trust us. According to Twit fans, they've purchased several items advertised on the Twit network because they trust our team's expertise in the latest technology. If Twit supports it, they know they can trust it. In fact, 88 % of our audience has made a purchase because of a Twit ad.

2:25:37Over 90 % help make IT and tech buying decisions at their companies. These are the people you want to talk to. Ask David Coover. He's the senior strategist at Threat Locker. David said, Twit's hosts are some of the most respected voices in technology and cybersecurity, and their audience reflects that same level of expertise and engagement. It's the engagement that really makes a difference to us. With every campaign, you're going to get measurable results. You get presence on our show episode pages. In fact, we even have links right there in the RSS feed descriptions. Plus, our team will support you every step of the way.

2:26:12So if you're ready to reach the most influential audience in tech, email us partner at twit.tv or head to twit.tv slash advertise. I'm looking forward to telling our qualified audience about your great product. I'm not a human being, not into this animal scene. I'm an intelligent machine.