An AI-Lite WWDC

Hello and welcome to The Attention Mechanism. My name is Justin Robert Young. Joined as always, I'm Mr. Andrew Mayne. How you doing, buddy? Hey, I'm doing fantastic, Justin. Glad to be here yet again. Now you and I... We have been friends for a very long time, and through our friendship, it has oftentimes involved us being very excited to watch and understand, discuss. Apple's WWDC, the Worldwide Developers Conference. Once a big keynote event, now a presentation. But...

Before we get to some of the stuff that they did announce that they wanted to put a lot of emphasis on, including a new UI, there is what they didn't really discuss a lot. In fact, I'm... pretty sure that they spent as much time discussing their new F1 movie as they did talk about artificial intelligence. What is your reaction to Apple's Apple intelligence?

announcements for wwdc so well there i mean there's two sides of it remember wwc the d stands for developers right and so when it came to consumer announcements regarding apple intelligence we didn't get much

I would say they did a couple of things I thought were long overdue. I'm excited about, which I'll talk about in a second. But yeah, the reports we'd heard about Apple as far as where they were with, you know, not any major updates coming to Apple intelligence, et cetera. They're clearly figuring out a lot of their strategy there, and we can talk about a paper that came out that was from some of the same people that released another paper that said that models couldn't handle certain things.

You know, I found the trivial prompt that it could. And this seems to be like, this might be that level of scholarship as before. But anyhow. There's a lot of really neat people at Apple working a lot of cool things. Apple is very, very conservative when it comes to where they roll things out. They made a very big bet when they came out with Apple Intelligence. They had a very good plan for how they're going to do it. The actual...

deployment was a different matter. And I liked some of the things right away, but then things didn't get better and you can't encounter the rough edges. So I think here they were kind of trying to say, okay. We need to take it slowly. I don't think we'll see anything until we actually get closer to the new iPhone unveil because that will be when the consumer thing was. But what they did do, by the way, which I'll tell you the thing that I'm excited about, generally excited about.

Do you hear about Xcode? No. So Xcode is the development tool that you have to use to work with building. applications for ios and ipad os and the challenge with xcode is it's like a billion years old and it's also where you control all your certificates and all this stuff like it's Designing and building things for the iPhone is a pain in the ass.

Not just because obviously Apple's very security conscious, which is important. And I don't have a problem with that, but it's just the way of using this thing. Again, this thing is like ancient history when it comes to building applications and stuff. That being said, you know what they've added to it? ChatGPT. ChatGPT is now built into Xcode. And so when you want to go right with Xcode, you can actually have a ChatGPT bar inside of there and say, help me fix the code. Help me do this.

You can also take your API key from somewhere else, like Anthropic's got great code tools, so you can put that in there. You can also use a local model. And if you have something we've talked about before about Olama. I was, it shows you there are people at Apple that care very much about AI and AI tools. So that was good. And I'm excited now because like now.

I have an app in front of me that I've been doing entirely web-based because Xcode is such a pain in having to deal with all the Apple ecosystem. For sometimes I'm like, I just need a button to press and have it do a thing. Do I really want to spend half a day trying to work with it there? So somebody out there in Cupertino very much believes in the coding power.

of ChatGPT and Xcode. And that's certainly the kind of announcement that would be very, very welcome at the intended audience of WWDC. Yeah. The ability to bring your own key is great. The incorporation of ChatGPT, obviously great for OpenAI. And... Apple is a very, very big company. And so while some people are responsible for the consumer product strategy with Apple Intelligence, other people are trying to build tools. And I was very happy to see this.

Let's talk about one of the things that I think was an interesting announcement that they made for Apple Intelligence, and that was that all apps in iOS will have access to the onboard. large language model. So you will be able to do, you'll be able to access that for computation and not make an API call off the device to do it. There are examples they gave where... You know, very Apple. It was like, where can I map out this hike based on the information that I have here?

I would imagine that having access to something on device is certainly a benefit from a developer's perspective, right? Yeah, I think that Apple wants to be the company providing you the highest amount of security. We've seen the capabilities for on-device models increase. Like you can run a...

8 billion parameter, 10 billion parameter model on a local device and get pretty good text completion and some generation and some formatting. And there's a lot of really good capabilities in coming when you do the small things. You're not going to get – the thing that happened was like when GPT-3 –

came out. And then when ChatGPT came out, there was this rush to like, can we do it locally? And we got really good. There were some really good open source models you could deploy locally. And a lot of people were like, look. I don't need to use the cloud anymore. Well, like, yeah, I...

I can have a computer in my Apple watch, but I still want a MacBook pro. You know, I still want other things. I think that's the sort of thing that people kind of, they looked at what we were doing it for right now and go, okay, well, I can just use local models. Like you can. For those kinds of tasks...

But you're going to want to do much bigger tests if you're going to want to make a lot more things. So there's always going to be room for local models and they're going to get more capable while the ceiling for what's going to be capable is always going to happen within the cloud.

But some of the conversation... about Apple was what they didn't announce, which, as you mentioned, there had been a report with Bloomberg about how behind they were with Apple intelligence, some of the internal frustrations, if not embarrassments, about advertising products. organic television ad buys that they did not have available, some of which they still don't have available, and where they're going to go forward with this. And with that backdrop...

You also have a paper that is released by Apple's machine learning lab that has some – a team within there that has – Some very pointed perspective on the limits of reasoning models. So break down what the paper says and some of the criticisms for it. So for background, reasoning models are basically when you let a model like ChatGPT or GPT-4... Take a task, break it down into smaller tasks, and then accomplish each one of those. And then...

go take the overall effort and come back to it with an answer. It's the same as if I asked you to do something and you took out a piece of paper and go, we'll do this. Step one, step two, do this. Step three, do this. Step four, do this. Step one, do this. Step two, step three, step four. It was based upon... early prompting technique that was found by

you know, a few different people where you could, you know, basically ask the model to like do things step by step. I found this when I was using the same thing I used to teach magic, like step one, step two, you'd see, oh, this is different. It thinks about a different space. And, you know.

Research has found that if you basically allow the model to go back, it doesn't just give you the same answer because it can also try to figure out if this answer fits or whatever. So these take more compute. It's basically letting the model use itself and spending more time to do the thing. So the researchers on this paper basically kind of the headline was saying that, hey.

So these models just collapse at a certain point. They can't solve certain kinds of solvable problems. And it seems to be like there might be a fatal flaw, et cetera. And I don't want to read too much. But that was the takeaway from this was that. Now. A little bit of backstory, a number of the same researchers wrote a paper that came out several months ago, which said that models failed at a certain kind of problem. Traditional models failed at a certain kind of problem and couldn't do it.

And I found, and I wrote a blog post, I can give it, I can get a model, one simple generic prompt about thinking that I can then train into it and it can solve these problems. And I was frustrated with their paper because. It was very solvable. It was like, ah, you're like, aha, and there's no way around it. It took me 40 minutes, took me 40 minutes, which they weren't doing the thing they need to do, which years ago when I worked with James Randi.

on investigating unusual claims. We'd come across these laboratories that would say, oh, we found psychic powers. We found this. We'd be like, hey, bring a magician into the laboratory. Bring somebody in here who's good at...

bending the system to see if you can bend the system with a way you didn't realize. And that's how, you know, people would often embarrass themselves. They're like, no, we know what we're doing. Why would we need to do that? And it's a crazy thing. So I'm seeing kind of the inverse thing of that where these kind of more cynical, like, oh, I can't do this. I can't do that. I'm like.

bring a prompt wizard into your lab and ask them what they can do. And so that last paper, that was the problem was they weren't using any kind of engineering to think, is this really just a training limit that can't be overcome or just a thinking limit? So the new paper.

One of the examples they show is the Tower of Hanoi puzzle where you have three pegs and disks and you've got to move them from point to point. And they say, hey, listen, the model just gave up even though it had more compute resources or whatever it could do with that. I haven't gone through to verify this. This was solved by machine learning in 1957. So we have gone all this way and spent all this money and we still can't solve the Tower of Hanoi.

Yeah, and I looked at this. I'm like, man, I don't have the bandwidth right now, but I know a couple things. One, these models are only six months old. Often all you have to do is when you discover a category of problem they can't do is just show up that category of problem. And then Lawrence, not give it every possible answer, but it's literally say, sometimes you have to approach it like this. Sometimes you have to approach it like that.

Now, these models can be brittle. They can run into dead ends. I wasn't like, oh, my God, I got the vapors. What are we going to do now? I was like, yeah, maybe. And then other people went in and looked at the paper. And I'm saying this secondhand, but they're like, oh, in order to solve the problem the way the researchers asked it to, there weren't enough tokens. They're asking it because it has to compute all of these different problem spaces.

And the model early on realizes, oh, I don't have enough tokens to do this, but I can just tell you what the answer probably is. And it's this, or here's how you write the code to do it. And they were saying, no, that gave us answers that were good. And they were like, it was just the problem was the researcher set up a problem.

that was like, tell me what would happen in the ninth move here with this chess thing. And it's like, well, it's got to calculate, you know, a billion moves, but get to there. And you can't just say, well, it's bad at chess. So I'm paraphrasing that, but I would say that...

Of the people who actually know how to use these systems and are doing interesting things with them, nobody there was really, one, taken aback or two, very impressed with the paper. It seemed like it satisfies a certain kind of audience that wants to say, ah, we found the limit. Wasn't true.

of the last paper that a majority of the people that were on this one worked on the last one this one you know looking at secondhand reports just seems to be another issue of a poorly defined problem that you know is trying to say something that

Overall, yeah, these models will have limitations. They're going to encounter categories of problems they can't solve. That nobody disagrees with. This really wasn't it. If I were to... some of the narratives that came up around this paper being released 24 hours ago, the announcement for WWDC being made six to seven hours ago when we record this, it would be, well...

Apple internally believes that the technologies that are leading the way right now are flawed and so they can take their time to get to the correct way to apply. artificial intelligence. And that would seem to echo at least some of the philosophy that was inside of Apple. At least according to Bloomberg, that there were people that did not believe the LLM revolution that was pioneered by open AI and has since been filled in.

in a very, very competitive space by Anthropic and Google and DeepSeek and so many other labs. I don't know whether or not that's still what they believe. I would certainly hope. But... It really kind of made me pessimistic. And we've been kind of down on Apple for a little bit, at least in this space. We were very bullish on them bringing in ChatGPT before their announcement.

last year for Apple Intelligence, because I think we're both of the same mind that if we talk about, you know, you talk about moats in this world a lot, what can you do that nobody else can do? How long will it take somebody to cross that threshold? Apple's got your data, my data. It has information that I don't want to put anywhere else. I would love to see them be better and smarter with this.

But boy, have I left the last 24 hours just wondering whether or not there is a voice that matters inside that big circular building that is pushing them in the right direction. I may have told this on the last episode, but I'll bring it up again. When I read, I think it was the Bloomberg piece about the state of Apple, and they mentioned how Craig Figurini...

encountered ChatGPT when he was working on some personal project and then went to his team to say, hey, look at this. That should be terrifying to somebody in that world because... Yeah! Nobody, none of his AI team, nobody there, nobody. The day GPT-3 had its API launch.

Somebody should have been running into his desk and showing him demos. They should have been teams at Apple building things, and he should have been watching these things. Or Tim Cook should have been saying, what are you doing? What else do you got? What else do you got? Let me see more. That was just a really... That, that.

You know, I love Apple. I absolutely love Apple as a company. Apple's been around my whole life as a kid, you know, going to play with an Apple II computer and wanting to have a Mac and now have an iPhone in my pocket is awesome. And they were always these cutting edge people. But when you find out that. They're finding out about, you know, ChatGPT and the state of LLMs about the same time as the guy that runs the muffler company in Iowa. Well, little concerning.

And there are brilliant people there that are building tools and stuff, but a lot of them, at times, are silent away. And the most public papers we've seen come out of Apple, to me, have been, I think, just embarrassing. But I don't think that's the overall quality of research there. There are some others. stuff I've seen some other releases some other cool things are working on and that's unfortunate because nobody

Nobody I know is super excited about the idea of working at Apple and AI because it doesn't feel like Apple takes it that seriously. Certainly putting it into Xcode I think was great. And I think that making it available to let you put your own API key was smart. Very smart people they're trying to work with and make AI work. But I don't know. I don't know how to be optimistic right now. It's hard to go to TechMeme and see that paper.

The lead story that they discuss, the lead link on Tech Meme for Apple on the front page is Gary Marcus talking about like, ha ha, they did it. They Gary Marcus LLMs. Apple did like swish. And it's like, wow, when that's your mascot, man, that, that is not where I, for the first time in my adult life. post iMacs, post iPods, like when Steve Jobs returns and kind of sets his full vision in motion, have I really thought that Apple is gettable?

I mean, I don't think they're going away. I don't think that they're going to have any kind of, you know, that they're like super, you know, they're going to totally break up, you know, tomorrow or anything like that. I do think that if they don't take it seriously, somebody else is gonna. Well, they have iMessage.

I think they have a great reputation for taking security seriously. They absolutely do, which I appreciate that. They have FaceTime. And I think these things may not seem like a big deal, but they're going to be much more of a bigger deal as in-person. becomes a bigger factor. As you start hearing, one, corporate security, the idea that, well, I got a voicemail from my boss telling me to do the thing, and knowing how easy it is to fake that.

I think that we're going to see that there is a role for Apple to play in this world if Apple steps up to play it. But I don't know. Yeah, we'll see. We'll see.

Let's talk about advanced voice mode. Because we both noticed something over the weekend. Do you know if there's, is there chatter online about this? Or is it just something that... that we noticed the change? No, no, it was an announcement from OpenAI. From OpenAI, okay, okay. Yeah, no, it was official. They tweeted out over the weekend that they had made updates to the new voice system. Gotcha.

I hate it. Can I go back? I would like to go back to a world where advanced voice mode does not giggle at me. It's been... It's been very unnerving. I've had to tell it multiple times. Please never giggle in my presence ever again. I feel like a strict boarding school. master that I am telling this LLM to stop acting so personal around me, but I do not like it. Yeah, I think everybody has a degree of comfort by which they want these models to interact.

The good news is this OpenAI is very, very quick to adjust and to handle things. So we'll see where things go. I feel like, man, this is an incredible demonstration of the capabilities of the model. Don't really know if I need that. Faster response times are good, a little more nuance and enunciation, but I don't need as many mm's and ah's and listening noises as possible because I don't want to pretend that it's human. And again, it's all a learning experience.

You know, they kind of have to go out there and go put it and see what happens. Yeah. I mean, if opening, I were a company where I felt like I was like locked into this for the next year, I think I'd be more animated about it. But it was it was. Like I had to change voices today because I was unnerved with, I had gotten used to my voice and then all of a sudden she was different and giggly.

And I was like, for the love of God, I need to just have a new, I can't have this happen. But it does demonstrate. From that perspective, how much you do get attached to these certain things. Like if you find a groove with something as personal as a voice. That that does matter. I've often said about podcasting that it's the most personal medium because especially since so much of it is done via headphones, it is almost exclusively.

people whispering into your ear like the most intimate act that you could possibly think of and that's the way that i interact with advanced voice mode for a lot of it i'll go on walks and i'll just turn on advanced voice mode and hash things out uh So for it to change very much freaked me out.

Yeah, it's an interesting space where we're dealing with a thing that has become a very, very personal device. I remember when the iPhone came out and they would push through updates to it and be like, wait, you changed my device. It's like, well, yeah, we wanted to. stop it from getting taken over by hackers or have some major system problem. And I think we're going to see a similar thing here. We have to think about how do these updates go out? Do you enable them? Do you not?

You know, we have a conversation, which we'll be publishing soon with somebody who's really involved with all of this. And we can talk about how the approach that they're taking towards trying to figure this out, because we kind of see where the future is, but we're still. trying to understand the shape of things. Yeah. Yeah. Yeah. That will be very, very interesting because it's odd that this kind of change.

is where we are right now in technology. That was like the number one technology story in my life, personally, was that I couldn't talk to my robot friend.

in the same way that i want to i was unnerved with how my robot friend behaved today well in every science fiction story we get you know and a lot of on the robots is they always have the one where they're like adjust your personality setting you know we saw this in interstellar we've seen this in star trek the idea that we seem to be like well if it has a artificial personality then i'm going to have controls to be able to adjust it which

As his models become more sophisticated, should we? Yeah. Yeah. This was listed in the known limitations for the advanced voice mode update. In testing, we observed that this update may occasionally cause minor decreases in audio quality, including unexpected variations in tone and pitch. These issues are more noticeable with certain voice options. We expect to improve audio consistency over time. Additionally, rare hallucinations in voice mode persist, resulting in unattended sounds resembling

ads, gibberish, or background music. We are actively investigating these issues and looking to find a solution. A report on Twitter and somebody, I guess, had recorded this. We're in the middle of a conversation. It starts to give them an ad. And the person was livid that OpenAI was inserting ads into this. And OpenAI was like, no, no, no, it's not an ad.

It's this is, and we've said this before, the model that is doing this, it's not like most other systems. What they do is they take your speech, they convert it into a transcript. And that transcript then gets fed into a regular LLM. What's happening here is the speech part that both the audio end...

The intelligence and the speech out are the same model. So it can see your speech tokens. You can actually ask it, you know, can you tell my pitch? Can you tell me this? That's one of the reasons why there was a delay. with advanced voice mode was because they started to see this crazy sort of behaviors. Like somebody, somebody reported that it mimicked their own voice. That's strange new world.

Strange new world, man. But yeah, I don't know. Maybe it'll be better. I have full faith that it will be better. that it will continue to improve because open AI has been very, very responsive to, to these kinds of things, but boy, the, the giggle. Oh, just nails on a chalkboard. Yeah, I think they have to dial this in and they're pretty fast to respond to it. One last story in the world of AI, and that is a report that Meta is looking to invest up to $10 billion in scale AI.

This comes after reports that they were not particularly happy with where they were with some of their own internal AI advancements. Where are we now? Meta obviously put a lot of time and effort into the world of open source. They got into a little trouble with their last model in terms of gaming it for benchmarks. where is meta now that they want to make a move like this so we were supposed to

Again, I was tracking that. Behemoth was supposed to be the next big model. I think that's what it was. Let me just double check that there. And we have not... seen that yet part of the challenge with these bigger models is they introduce all kinds of new complexities which means when you're trying to make sure that they function the way you want Don't accidentally give you chemical formulas for nerve gas, etc.

that you have to, and it's only so much you can do to an extent. But the point is, is there's a certain amount of safety training and a certain amount of capabilities, things you have to figure out what to do. And that takes longer. And that's, you know, it was, you know, opening. We had. GPT-4 for six months before we released it and often you'll hear about ah the AI race and well sometimes

A company will finish training a model and then announce it, you know, a week later. Yeah. And kind of like, hey, look, look where we are. And it's like, OK, the other guys might be.

ahead of you but not ready to announce because they're spending more time safety training or doing whatever they need to post training etc um you know with elon must xai the original plan was that he said he was going to open source every new model which got him a lot of adulation then it's he said no what i mean is After I have the next model, I'll open source the last model is what I'll do.

And we're at Grok 3 and Grok 2 hasn't been open sourced yet. And they had a 3.5, but now they're jumping all ahead to Grok 4 apparently. So it became like, okay, when does that window work? And also tomorrow. My understanding, we never got a safety card. We never got a safety evaluation, even Grok 3, where they went through and said, hey, these are all the fail things where it falls apart, whatever. We know Anthropic and OpenAI are very, very dutiful about doing that.

publishing, this is where it fell apart, whatever. Other labs not so much eager to do that, which is part of the territory if you want to be considered the responsible. parties involved so i would say it can be a challenge so you might have they might have a very capable model also might not be showing this capabilities if they want i don't know we know that one of the patterns we're seeing now is

not major model releases, but to continuously improve these existing models. And OpenAI is still outside of the reasoning models. They released GPT-4.5, but we're like, yeah, you could use it, but here are all these other models that are probably going to do what you want.

I think 4.5 is a great model, by the way. I was using it for some other stuff the other day. I'm like, oh man, it is really good at certain language tasks. It's really good at other things, but a lot of it is just lost on use cases outside of there. So we're seeing, you know.

Google released Gemini 2.5, and they keep improving it. It just scored. I think it got the highest score now on Humanity's last exam, which for the longest time, OpenAI was the one dominating that. That was the one that I'm like, well, if you look at it. for what I consider the leaderboard, that's this and OpenAI was in like the top three spots and I think Gemini just took the number one spot. So you can just continuously improve these things. So

That means Meta's playing catch up. They need a new coat of paint here because $10 billion is a lot to peel off to invest in a firm like Scale AI. Well, yeah. The question that I got asked by... hedge funds investors and other people early on was what is meta's game here i said well their long-term game is not to spend billions of dollars to give people free stuff that is not the game

The goal is to make money. The money they're deploying is to make a lot of money. And they're going to look at the market to figure out where the money is. And that is going to be what they're going to be incentivized towards doing. Right now, the two places where money is is an API and having an API that provides access to other people to use it and then the consumer-facing part of it, ChatGPT. But as it is right now.

Once you step outside the very, very two online world of tech into the normal world, ChatGPT is the only name there is. There was just a... news release, or I just saw a news item that came out where, let me get the number right. OpenAI just hit... $10 billion in annual reoccurring revenue. $10 billion. And that was fall of last year. It was 5.5. So in less than a year, it's not fall yet, less than a year. A doubling.

of this and a significant doubling and and i remember when they hit 300 million and uh one of the person's names you mentioned earlier was chortling like oh we'll see where they're going to go now 300 is that sustainable

uh yeah no they couldn't hold state you know that they had to go up and so then they hit a billion like ah yeah but really and then well they hit five billion like okay but but now it's like oh but they're spending they're spending more well that strategy seemed to be working because the more they spend the more they make you know that that was like yeah that was like it's the craziest sort of weird i saw this first with tesla when people would look at

There are several ways to figure out the cost of a thing and if a company is actually growing or if they don't know what they're doing. And in Tesla's case, I remember when they're in the middle of building out the Model 3 factories, people are like, oh, they're spending all this much money.

And I hear analysts like, yeah, if you've got the cost per Tesla, it's actually $300,000 per vehicle. And what they were doing was taking the CapEx cost for the factory that wasn't even producing raw cars yet. And basically dividing that among the current yield of cars and saying like, no, it's a one-time cost to build a factory. And you'd see that weird thing. And that's the thing with these AI models. It's a one-time cost to build a model.

There was the question of like, would Microsoft ever make their billion dollars back from investing in open AI? And then that became a joke. And then it's like, well, what about this 10 billion? Will they see their money back? And like, well, I think we're in the place now where... we need to start looking at how the math really works. And what this is, is put simply is...

When OpenAI spent, you know, let's say it was $5 billion last year, it wasn't the cost that took them to deliver the current usage for chat GPT. That is what they were spending to develop the next generation of capability. And every time they do that, the demand keeps getting higher and higher and higher, which I wish more people would just be ignorant to this and pretend they don't understand that so you and I can continue to buy NVIDIA and do whatever else, though. So forget I said anything.

Never mind. Never mind. Everybody, please continue as you were. Look, we are here because we very much appreciate... This moment that we are in, the technological marvels abound. We are obviously optimists in this space. And so it is because of our focus that we will only very, very briefly.

touch on a little bit of online drama that happened with one of the big figures in AI. And that is, of course, Elon Musk, who went back and forth with... The White House over the last week, a falling out that you can hear about on politics, politics, politics. But I hope Elon continues to develop. I hope he continues to be a player and I hope he continues to improve his Grok service because he found out that maybe it's not up to par. This happened.

Last week, Stephen Miller, who works in the White House, he tweeted out, we will take back America. Now, what you may not know is that Stephen's wife, Katie Miller, is now working. with Elon. She left with Elon because she is going to work with him in the private sector. Grok seems to believe that Elon tweeted back to that tweet, just like I took your wife. To which Elon had to go online and say, I did not. No, no, it's fake. FFS. I never posted this. So if there's any little parable.

that just sums everything up, that maybe Elon should not be burning bridges and be focusing more on his products. I feel like that is one of those little things, little stories that tells a big tale. Yeah, these things are hard. Every lab has its snafus in the things that happen there. We had a conversation with somebody on Friday, which we'll talk more about, about the scale of the pie and how just a few years ago.

There was this idea that one lab could kind of get such a lead that it would be a runaway, you know, unsurpassable. And everybody else would be left around banging sticks and stones together going, oh, why, why, what happened? You know, the understanding now is the more. time we've spent with these systems and we understand the complexity of the world and all the different places and all the different nuance of trying to apply the most intelligent systems in the world into them.

There are going to be a lot of winners. And I've been in rooms with venture capitalists and people like this for several years who are like, well, what if we go invest in a foundational model? I'm like, you can. But if you if your goal is to outspend Google, good, go do it. You know, if you're if you're somebody incredibly resourced like Elon Musk and, you know, sure. But what you know, if if Boeing builds the jet.

Do you go rush out to build a jet or do you build FedEx? I think he'd go make FedEx, you know? Yeah. I think that's the thing, not to say that somebody shouldn't be trying to improve jets. We just got, you know, boom, aerospace just got permission now to be able to fly supersonic over America. So like you can do that too, but I'm saying there's more to just that. And so.

We're in this world now where there's going to be a lot of there's the game never gets settled. And as Sam Altman had said, you know, for every move you get, they get a counter move. Yeah, I would also wonder in all seriousness. I think that what Grok is doing on... Twitter is remarkable. It certainly has caught on with its user base to have an AI fact checker that is always taggable so you can tag something either for your edification but is also kind of posting for the world.

It's a very interesting idea, but I do wonder how much they are training this model that they're using for it on other stuff because it does seem to be the... The grok that tweets does seem to have become more of a troll or taken on more of the personality of the posters as opposed to. being a foundational kind of fact checker. And we saw that with the white genocide thing a few weeks ago. And now we see it with this, which is, you know, it even bit the big man on the butt.

I would say that, one, it's cool. By the way, I worked on, I think, the first internal application at Twitter of trying to use an LLM to parse a bunch of things and come up with summaries.

things like this and did a bunch of concept demos showing them with GPT, Chachi, excuse me, GPT-3. So it was, and it was sort of an, it was like inevitable sort of thing to see that like this is, not that like, oh, I had started where it was just literally like, hey, do you want to go? Oh yeah, sure. Because everybody knows this is. what you're going to end up doing.

I would say the kind of bit of the irony is that XAI and Grok are supposed to be the very open, you know, the non-closed AI thing. And they're the most opaque companies when it comes to their prompts, when it comes to how the model works.

works and everything else like that and so when things like this happen you know somebody has to go in a prompt hack to figure out why the hell this happened uh you know they won't even show reasoning traces for the reasoning thing because now well you don't want people copying it it's like okay so You became the thing that was your rallying cry to go build the thing. And now you're even more. Ah, yes. Closed AI. Yeah. So. But it's cool. Maybe Grok is a little more like Clock.

Yeah, but I do think that having a builder like Elon who sees the utility of things and where it can apply to it and can really motivate teams. is doing a lot of interesting, cool stuff, you know, within how, yeah, the idea of you tag Grop, you do this, you do all these things. I still have that, like, I don't quite know. I don't use it enough because, like, I have my biases and stuff. I want to get to the point where I do.

But I do think it's it's neat to see just, you know, I'm not I'm not going to start. I think it's because something happened because, you know, I haven't come from nobody. No, I don't. Any of those mistakes. I don't think – I think that was – that was just one of those examples. I do have a T-shirt though. Yeah.

I have a T-shirt that says, because back when they had one of their first big screw-ups, they came out and said, well, the problem was the mistake was made by a former, an ex-OpenAI employee who hasn't fully absorbed XAI's culture. which was the most throw somebody in front of the bus. Just, they got ridiculed for that because like, Oh, this is how you handle mistakes is you just, what was the new hire from here? Like, come on guys. Like, yeah.

Come on, dude. Yeah, that was not exactly we're taking responsibility for. It's like, no, we did. We told you who did it. That's not how it works. No.

But we do work for you, dear listener. And that's what we love to do here on this show. Mr. Andrew Mayne, where can people find you? And you can find me, Justin R. Young, anywhere you find Justin R. Young's on social media. Until next time, friends. See you later. Bob hopes you have enjoyed this broker. Dog and Pony Show Audio.