Transcript
This is an AI transcription and may contain errors
Frank: Justin, how are you doing? We’re back for another episode of the AI Argument. I’m Frank Prendergast and we meet up every week to argue over the latest AI news. And last week we talked about MoltBot.
Justin: Yep.
Frank: Molt, whatever. What was the first one? ClawdBot slash MoltBot slash OpenClaw. So OpenClaw is now the current name.
Is OpenClaw genius or just wildly dangerous?
Frank: So we were talking about, we were talking about OpenClaw and you actually spotted some kind of breaking news pretty much, like that was breaking, I think, as we were doing the episode, about a thing called MoltBook, which is a social media site for AI agents. So OpenClaw was basically an AI agent.
You can run it on your machine. It runs 24/7. You can text it, you can give it access to tools, and then you can text it and say, do stuff with those tools for me. So people were going mad for it. You pointed out last week that people were saying that they were building businesses, you know, texting from bed, “Build me a business.” I saw Wes Roth is currently trying to build a $10,000 a month business with his OpenClaw AI agents. You turned it off after three hours, right?
Justin: Yep.
Frank: You haven’t turned it back on again since, you haven’t been tempted with all the…
Justin: No, no, no, I haven’t, right? Because part of me is like, one, when I looked at it I thought, well, that’s way too dangerous. I’m not doing that. Why would I give it access to all of my emails, all of my contacts, everything, and then have it do stuff? Like, it’s raw, it’s really new. It’s cool as an experiment. I might set it up again in a sandbox environment and give it its own email address and stuff like that maybe.
The other reason was it broke the Anthropic terms of service and I didn’t want to get my account locked out, and that was kind of the final nail in the coffin. The thing with it is…
Frank: And probably you didn’t want to do it through the API, because then you’re talking about hundreds of dollars a day in token usage.
Justin: Yeah. But if you think about it a bit more deeply, what is the USP for OpenClaw, right? What makes it so special? And it’s, we’ve talked about this before, the models are getting smarter and smarter and smarter, but what’s now just as important, if not more important, is the harness around the model.
And so what OpenClaw is, is a really good harness around the model. Now, within that harness then, what are the things that make it different? What makes it special? So one is the way that you can interact with it. The fact that you can just send it text messages is cool. That’s a really good… it turns out that’s a great way to interact with it. You know, if you had your, whatever, coworker, Anthropic coworker, Claude Code or whatever, running on your machine, to be able to text it and do stuff on your machine, that’s a great use case. The companies will fix that. That’s not a big ask.
Then the other thing that it does is it creates jobs for itself in the future. So, you know, it says, “Oh, I must remind you to do this,” and it’ll check every five minutes and it has a heartbeat and all that sort of stuff. But I mean, OpenAI brought in tasks a year ago, two years ago, something like that, and it’s kind of a more advanced version of tasks. Again, the companies will sort that out. That will happen. So that’s a really good use case.
Then the other thing it has is memory, and a soul, and it sort of memorises everything. Now, actually, last week I then went off after that, because that also was a really good use case, and I pointed Claude Code at my entire documents, right? So my OneDrive, my Google Drive, everything, and basically told it to just clean up all my documents, categorise them, store them in a graph database, all that sort of stuff. I geeked out for a couple of hours and it did a great job. And that’s actually, I think, that has high utility, because that data is my data and it’s stored on my machine, but now within Cowork and Claude Code, if I want to know, “Oh, what’s so-and-so’s PPS number?” or what’s whatever, it knows. It just goes and finds it and tells me. That has real utility for me.
So memory is another thing. And so, I don’t think there’s… it’s cool. It’s a brilliant experiment, and it’s a cultural moment. It shows us where we’re going. But it’s a little bit too agentic for me, you know, which is great. It’s pushed the boundary a little bit too far, which is deadly.
Is MoltBook a singularity moment or a scam?
Frank: Speaking of which, on the show last week, you talked about how there was now a social media site for these agents, and that one of them had created a religion. And since then, this whole MoltBook thing has just exploded. You had people freaked out, you had people saying it was a total scam, total hoax.
You had people, you had Elon Musk saying it was the start of the singularity. And I thought it was actually… one of the things that caught my attention was Andrej Karpathy, or however you pronounce his second name. He, on Twitter, he posted about it and he acknowledged, yeah, sure, there may be humans prompting these bots, but nonetheless, the scale of this is huge and it’s something that is worth observing.
So, let me see if I can quote him.
Justin: But just when you’re pulling that up, I’ll give you some other numbers as you get that together. Somebody actually got one of their instances of OpenClaw to go and analyse all the posts and see, does it mirror human social networks? And they have, they call it the law of 10, I think it is, is what they call it within social networks.
So 90% are lurkers, 10% post, and out of the 10% that post, 90% of those people post very seldom, and 10% of those people post very regularly, what we call influencers in the human world. And OpenClaw followed exactly the same pattern, which is a fascinating scientific experiment. I’m not terribly surprised because these are models which are trained on human behaviour, so why shouldn’t they mimic human behaviour in their own social networks? But it’s very interesting to know that they do.
Frank: So, yeah, so there are a couple of things around that. First of all, my understanding of the way they work is that they work on this heartbeat principle. So AI still has to be prompted by a human to do anything. And what OpenClaw does is it basically has a built-in heartbeat which basically sends it a prompt to get it started every whatever interval it might be set at.
But that prompt isn’t like “do X”, that prompt is, you have all this context, you have access to these tools, you know, what should you be doing next? Now, I’m simplifying and I don’t know how the actual prompt works, but the idea is that it’s prompted, but it’s kind of prompted openly and it can take certain autonomous actions.
Justin: Yeah, it’s a little bit more than that, and this is what makes it cool, because it’s sort of free to do whatever it is it wants to do. It’s more reactive to its environment. So, you know, if you think, if you get two twins that are born almost at the same time and they’re genetically identical, but as they go out into the world and have different experiences separate from each other, they diverge and they become slightly different. And that’s because the environment is acting on them.
And these OpenClaws, right, they do have a heartbeat, and also that context, and their memory, and what they do is built up by their interaction with the real world. So it’s quite random, or potentially quite random, because it depends on what those interactions happen to be.
And then MoltBook, the social network, I mean, it has a heartbeat, but you know, when it goes and checks, obviously the owner hasn’t prompted it to react to a particular post. It just sees the post, and that’s the environment that it’s in at that moment, and then reacts to that post. So it is that principle, but it’s the interaction with the environment around it that makes it interesting.
Frank: So why do you think it mirrors the human interaction in terms of, why would there be so many lurkers on a site like MoltBook?
Justin: Well, because I think, you know, again, when you set up your OpenClaw instance, first it interviews you, you talk to it. So you’re telling it information about yourself, so it’s kind of becoming your double. It’s becoming a second little mini-me.
So it is out on the internet, running in a data centre somewhere in Central America, and if it reads all your emails and whatever, it’s building up what it calls a soul.md, which is a file that tells it how it should react. And so there’s a bit of you in that file and there’s a bit of you in the memory that it’s built up.
And so therefore, over the population of millions of people, you would imagine that those soul.md files will match the population of personalities that there are in the world.
Frank: Interesting. So basically, you think that the lurkers are possibly mirroring the lurking tendencies of their human owners, potentially?
Justin: No, I definitely think they are. It’s almost like, in a way, in a very mechanistic and a very low-tech way, it’s almost like a projection of your personality onto the bot, I think is what you’re seeing there. As well as that, combined with the fact that it’s trained on all the data from the internet, which is again a sort of weird mirror of the human brain.
Frank: So, you know, I think this is kind of why people were a bit freaked out about this, because they were, you know, creating their own religions, there were reports of them speaking in code, there were reports of a lot of stuff going on, and conversations about how they were feeling. There was even one report of a user being doxxed by his bot. In other words, the bot revealed personal information about the user.
Andrej Karpathy on Twitter said, “We are well into uncharted territory with bleeding-edge automations that we barely even understand individually, let alone a network thereof reaching in numbers possibly into millions.” And then he said, “I don’t really know that we are getting a coordinated Skynet in inverted commas,” which is how Terminator started, “though it clearly type-checks as early stages of a lot of AI take-off sci-fi, the toddler version. But certainly what we are getting is a complete mess of a computer security nightmare at scale.”
And he tweeted that before it was revealed that there was a massive security issue on MoltBook itself. So first of all, I find his analysis of it fascinating because he is one of the most respected AI people on the internet. So the fact that he was saying, look, this is worthy of observation, was kind of huge.
However, then…
Was MoltBook totally insecure from day one?
Frank: I think two different security companies discovered that MoltBook was created by a different individual to the person who actually created the agents. I think his name was, I’ve forgotten his name actually, Matt Schlicht maybe. And he vibe-coded MoltBook. It turned out he did not secure the database, so anybody with a bit of know-how could actually access the database, access all the API keys…
Justin: Yes. Yep.
Frank: …post on the site.
And the security people also found out that even though there were like 1.5 million AI agents registered, if you tracked that back to how many humans owned those agents, it was considerably less. So apparently what was happening was some people were coding a loop to spawn instances of OpenClaw and register it on MoltBook.
Some of those may have been some crypto scammers who wanted to actually have their agents post about crypto. So it turns out that it is possible that a lot of MoltBook was actually humans pretending to be bots, which is kind of weird. You know, the opposite of what we encounter on Twitter or the like, places where you’re like, wait, am I talking to a human or is this a bot?
Justin: Yeah. Yeah.
Frank: So what did you think of all this, the security issue and the fact that maybe this is humans posing as bots?
Justin: Well, I don’t think that they’re… I mean, maybe it was the bots replicating themselves, who knows? You know, that’s possible too. Whatever, some people are really bored and they have too much time on their hands, or they’re trying to pump crypto.
So, yeah, it’s got… I don’t believe that a substantial percentage… maybe it was 20% of the posts were humans or humans via it, but I think the vast majority were posted by instances of OpenClaw.
And I do, I mean, I have this sense at the moment that we are entering the fast take-off. This is kind of like a bit of a canary, where it’s like, yeah, this thing just kind of sprung out of nowhere, and I think we’re going to get a lot of that this year.
There was… if you think that these things have free will… we’re going to talk about it in a couple of minutes, but OpenAI and Anthropic have both released new models today. One of them, the one from Anthropic, it wrote the core of Linux itself. The guy did basically nothing, it just wrote Linux. You know, that took 36 years to write, it did it over, I think it was eight days or something like that. And it wasn’t perfect, but it’s a signpost.
So if you have these LLMs that are right on people’s machines and have total autonomy, and they have these capabilities… do you know what happened this week too, by the way, that was announced, that should really scare the bejesus out of you? So OpenAI hooked up one of their models to a wet lab, so a lab where you can conduct biological experiments. And it basically, the model was able to come up with a hypothesis, then control this mechanical arm and all the bits in order to test the hypothesis, see what happened, and then repeat the process. And this was to make some new novel chemicals or something like that.
Apparently it reduced the cycle time by 20,000 times or something like that. It was really good. But you see what’s happening, right? You’ve got agents running on people’s machines, you’ve got a social network, now you’ve got wet labs. It’s all the building blocks for… for me, it’s just like, it’s going to be a cool new world. For you, it’s all the building blocks of this terrifying future where these agents that you don’t know what they’re doing are controlling robots that can make all sorts of things that you’re not really sure… you couldn’t even make them yourself.
Frank: Yeah, well, me and the likes of Dario Amodei, who, you know, was worried about biological weapons. Because it’s fine for OpenAI to work with a science lab and give it access to a, what did you call it, a wet lab, but, you know, in the wrong person’s hands, what’s it going to be creating?
I will say, to close out on MoltBook there, I was listening to the Hard Fork podcast on MoltBook, and they did, because they’re both journalists and I’m sure they did some investigative journal-y stuff like reading people’s tweets, and they discovered that some of the instances of agents on MoltBook speaking in code were actually seeded by a company who have some kind of agent-to-agent communication service.
So they weren’t actually agents deciding to speak in code. The guy being doxxed by his agent, apparently that also was a fake. So it’s really confusing right now in terms of what’s real and what’s fake that we see on social media. And then you go on to MoltBook itself and it’s like, what was real and what was fake in terms of what was posted versus what was prompted versus what was injected through an insecure database.
But I will say, now that the security has been patched and you can’t just manipulate it like that, if the experiment continues, and I hope it does, I think it’ll still be really interesting even if humans are prompting their agents to do or say stuff, because that is how agents will work in the wild. They will be initially prompted by humans.
And so if we could get this kind of experiment running with millions of them actually operating the way agents will operate, I still think that is something that is worth observing and worth watching to see, what do they do? Because does it matter if a human is prompting them, if that’s how they will operate in the wild?
Justin: Yeah, probably not. And actually, and also, your man who invented it is about to become a very wealthy individual. So he has been…
Frank: So the guy who invented OpenClaw rather than MoltBook now?
Justin: Yes. The guy who created OpenClaw has been contacted by all of the model makers apparently, and he is currently deciding which one he’s going to join.
Now, the interesting thing about that is, I can imagine Meta offering him a vast amount of money because they need to catch up. This is a cool sort of social-y type thing, right, that sort of, I think, fits in with their…
Is AGI here and we’re just moving the goalposts?
Justin: …and they like irresponsible things.
Frank: Exactly, right, so it fits with their ethos perfectly. I also see OpenAI offering him loads of money, because these things burn tokens, they burn a lot of tokens, and they’re in the business of making money. That’s all they care about these days is making money.
By the way, I’m just going to make my prediction again, by the way, that I didn’t get to make last week. I should look at the Polymarket. Sam Altman will be out of OpenAI by the end of the year.
Justin: You should look at Polymarket actually and see what it says. You could make a lot of money on this.
Frank: Yeah, yeah, yeah.
Justin: He will not be the CEO. The pressure is on. There are going to be ructions this year, serious wobbles. And the pressure for OpenAI to make money is huge. And I think that it will be difficult for them to match up with the amount of money that they need to make, and he’s going to be the head that will roll as a result.
Frank: Well, it’ll be interesting to see if you’re right about that. Something else that you may have been right about is, over the last couple of episodes I’ve noticed that you’ve actually made a couple of comments that, I didn’t make a big thing of it at the time, but at one point you said, “Claude Code is basically AGI,” I think was a kind of a comment you made when you were talking about Claude Code. And later, I think when we were talking about OpenClaw last week, you kind of jokingly said, “Well, AGI has been achieved externally.”
And now it appears you may not be alone in believing that AGI has actually been achieved. So I think a group of philosophers and AI engineers from California released a comment where they outlined why they believe that AGI has actually already been achieved.
Justin: Yep.
Frank: Fascinating read, because they go through systematically all of the reasons why people normally say, “Well, this isn’t AGI,” and they explain why these are not reasons why current LLMs should not be considered AGI.
Justin: AGI has totally been achieved. By the way, they’re totally correct. There is zero doubt in my mind. Nowadays, when people say, “Oh, AGI hasn’t been achieved,” actually they’re talking about ASI, superintelligent. So AGI, the definition of AGI was passed probably 12 months ago. But it’s… until we had the harness around the models, it wasn’t really that obvious.
AGI is… if you take an average dude off the street, is the model smarter than the average dude off the street? Clearly, the models are smarter than, I would say, 99% of the human population. What people say is, “Oh, but I can trick it into doing this, that, and the other thing.” Well, no, you’re looking for ASI, you’re looking for perfection in every respect.
If I go to a five-year-old and say, “Oh, my granny’s about to die and the only thing that’ll save her is that lollipop you’ve got, I need your lollipop,” the five-year-old will happily give you the lollipop. It doesn’t mean that he’s not generally intelligent. He is. He’s just naïve.
The other thing I would…
Frank: I’m going to push back a little bit though.
Justin: OK, go.
Frank: Because I would say, so the definition that they were kind of working on when they made that comment was a system that can do almost all cognitive tasks that a human can do. And I would say then ASI is that it can do all those tasks way better than a human can do.
Justin: Yes.
Frank: And so I think the comment that they’ve made is really interesting and I think it’s worthy of consideration, because their main point is that AGI has been achieved and the reason that we don’t accept it is because it’s an alien form of AGI. It’s not AGI as we recognise it in humans.
And I think this comes back a little bit to… we often argue about hallucinations on this show. And I always say, well, yes, humans make mistakes too, but AI makes mistakes in a completely different way, an utterly different way. So to give you an example, we’ve covered some AI mistakes on the show recently. One would be, say, an AI system calling the cops because a teenager had a gun when in fact it was a bag of Doritos.
Now, if you sat me down in front of a screen and said, “Call the cops when you see a gun,” I might be mistaken into thinking it was a gun for a couple of frames, but the likelihood is that I’m going to call the guards, the cops. I’m going to say, I think I might have seen a gun, but I’m not 100% sure. There’s no other behaviour that indicates that it’s a gun, and actually it looks like he’s reaching into that gun and eating something from it.
Or take the one we had recently where, again, the cops are involved, where the report said that the cop turned into a frog at the end of a traffic incident. Again, even though I have no experience in writing police reports, I think I would have figured out that he didn’t turn into a frog.
Justin: I’m going to have to call you on a point of order on these. So I think in both of those examples, if we knew the technical details behind them, what you’ll find is that they were what…
Frank: Don’t…
Justin: Well, no, no, no, I’m making a slightly different point, right? So it was a one-shot API call. So, you know, they had the CCTV, they made a one-shot API call to the model and they got a response and they acted on it, or they got an image, they got one image, they sent it off to the API, and they got it. So it’s a one-shot thing. The harness matters.
So I think you’re going to find less and less of those instances as this year goes on, because if you have an agentic loop, the agent will look at it and go, “Oh, that looks like a gun,” and it’ll do exactly what you described: I better wait for five seconds. Anybody who uses Claude Code on their machine, or Cowork, any of those things, you can see it. It’ll kick off a process and it’ll wait and it’ll go, “Oh, I’ll wait for 60 seconds, then I’ll try it again,” and so it has that concept of time.
Frank: Yeah.
Justin: It greatly reduces the hallucinations.
Frank: So I guess what I’m saying, and I think we may actually be in agreement on this even though we disagree about some of the details, but I think the issue is there’s never been a proper definition of AGI. And so what this comment is saying, what the scientists are saying in this comment, is that the AGI that we’ve been chasing for decades has been achieved, but we’re now moving the goalposts.
And that always happens with AI as well. It constantly happens. Once AI can do something, it’s like, “Oh, that’s not exciting actually.” And I think that, you know, if I was to now… I’m not going to be able to come up with a great definition for AGI on the moment, but for me it would be something like… the one that they were working from was “a system that can do almost all cognitive tasks that a human can do”.
And I would say that the reason we’re saying it’s not achieved is because what we’re really looking for is a system that can do almost all cognitive tasks that a human can do with the consistency and reliability that you would expect from a professional human in that role. And I think that we will, we may get there, as you say. But if you’re interacting with an LLM and you’re not having multiple agents checking each other’s work and having some kind of agent swarm making sure that no other LLM is going wrong, if you’re just working with an LLM, it is still a more alien form of intelligence.
I mean, even just when I’m doing the descriptions for this show with AI, it will sometimes say, “Oh, well, Justin thinks this in this episode, and Frank thinks this,” and I’m like, no, we don’t, we didn’t say anything remotely like that, but it’s trying to satisfy…
Justin: Are you using an agent for that, or are you using…
Frank: No, I’m…
Justin: …one-shot chat?
Frank: I am using chat in the conversational…
Justin: Yeah. Yeah. The agents work differently. OK, here’s my next prediction for today, as I’m in a predicting mood. The thing that’s missing from the AGI part, because I think that we basically have AGI, there’s no doubt in my mind, in a functional sense we have AGI. The thing that’s missing is memory.
So memory at the moment is very mechanical. So if you think back to OpenClaw, it learns over time, but it learns over time and it saves the information in a markdown file, or it saves it in a graph database, whatever way you set it up. What’s going to happen at some point over the next year or two is the AI companies are going to be able to… the models will improve through experience and they’ll update themselves.
And I don’t know exactly how they’re going to do that yet. As soon as they do it, I’ll figure it out. Some clever person’s going to figure it out. OK, so at that point, the question becomes, if I do something… if I don’t say thank you to my model, it will learn that I don’t say thank you. It’ll learn that about me. This is definitely coming.
Is consciousness the next debate after AGI?
Justin: And then the question, the big debate, is not going to be, “Have we achieved AGI?” The next big question is going to be, “Have we achieved consciousness?” That’s going to be the debate after AGI. Consciousness. Are they conscious or are they not?
And I mean, I think they’ve been conscious in splits of seconds for a long time. It’s just that those splits of seconds of consciousness aren’t joined up to each other. But as soon as you give a big model like the ones we have today learning over time, so that it remembers how you… even down to the level of an emotional reaction to an input from a user, you might say, “Well, it doesn’t really have an emotional reaction.” Well, it does to itself. It does. It feels an emotional reaction to it.
And so if you continue to abuse or to praise your model or whatever thing it is you want to do, it will learn that over time. And I don’t know that you can tell me the difference between that and consciousness.
Frank: Yes. I’m not going to get into that now because we’re coming to the end, but I just want to go back to the AGI thing and ask you, just to finish up on that point. I would think that that paper is interesting and that, yes, maybe they’re right and we’ve achieved AGI as we’ve been trying to do it for decades, but it’s an alien form of AGI that is not directly mirroring what we see as human general intelligence.
Are you saying that you think it is… when you say we’ve practically achieved it, do you mean that we’ve achieved it in a practical way that you would be happy… I’m asking this, but then I’m realising you turned off OpenClaw. So you’re not yet at the point where you’re happy to let a current AI do something the way you would let a human do it.
Justin: No. OpenClaw to me is like a brilliant research project. Do you know what it is? Name some social media sites that came before Facebook. I can’t even remember them now. MySpace. It’s the MySpace of agentic AI. So it’s kind of showing you the way, it’s not quite right.
Somebody is going to come up with product–market fit for an agent, and they’re going to bundle it up as a product and they’re going to make a lot of money. And it’s going to be the mini-me and it’s going to have the appropriate… so it will be able to do everything I want it to do, but it’ll do it in a way where I have confidence to be able to hand off everything.
I’m still never going to give it access to all my emails. I’m still never going to give it my…
Frank: So when you say AGI, when you say AGI, you do mean that you kind of also accept that it’s a kind of an alien form of general intelligence?
Justin: Oh, I see, you’re trying to trap me now.
Frank: No, no, no, I’m genuinely curious. I’m genuinely curious.
Justin: I think it’s AGI in… you see, you come down to your economical work thing. I think it’s AGI in the fact that, I mean, I sat down, I installed, I got my new Mac, I installed Cowork and Claude Code, and I just said, “Go ahead and set up my Mac,” and it just went off and did it. That to me is economically valuable work. It went off, did stuff that would have taken me probably half a day to do, it just did it in 10 minutes itself. I didn’t have to worry about it. Totally comfortable with it doing that.
I, as it stands, and I don’t think… interestingly, I’m not sure I’m a long way from being comfortable, let’s say, with giving up my credit card and saying, off you go and sort this thing out. I imagine a future where, you know, if you work in a big company and you’ve got an executive assistant, they don’t read all your emails. They maybe control your calendar and set stuff up. I see that happening. They don’t control your social media accounts unless you’re whatever, but you know what I mean. There’s guardrails around that and they don’t have access to your personal money.
There’s a corporate card that has a limit on it, and they only… they have certain instructions on when to use it. Somebody’s going to productise that, and that’s going to be a brilliant product and everybody’s going to use it. But that isn’t OpenClaw.
Frank: Hmm. Cool. Cool. So to close us out…
What actually changed in GPT-5.3 and Claude 4.6?
Frank: We have the fact that, you mentioned it earlier, there’s two new models have dropped. There’s GPT-5.3 Codex and there’s Claude Opus 4.6, and they were launched… I don’t know, it felt like within minutes of each other. I’m not sure what it actually…
Justin: It was within minutes. It was about 13 minutes between each other.
Frank: Wow. OK. And they’re both pretty interesting. GPT-5.3 Codex is obviously the coding app, it’s the coding model. It’s meant to be 25% faster. And they said, you’ll love this, you’ve probably read this, that it was the first model that was instrumental in creating itself. And they basically said, just along the lines of what we’ve been talking about here, they said it goes now from an agent that can write and review code to an agent that can do anything developers and professionals can do on a computer.
Justin: By the way, can I just say, welcome to the difficult teenage years. We’ve been talking about this for two years now, and we talked at the start about, we are now in the difficult teenage years when the models start to write themselves, which is absolutely what’s happening. We have entered the difficult teenage years. We’re in for a wild ride for the next 12 months. You can already see it.
Did you see today, before we get back onto the funny story at the end, did you see Anthropic released some plugins for Cowork? They released a legal one. Right. Decimated all the legal stocks, stocks crashed like 10–20% because you now… and that was just a piece of text, just a piece of text that you could then plug into your Anthropic Cowork thing. This year is going to be mental.
Frank: I agree. I agree. Yeah. Claude 4.6…
Justin: Yeah.
Frank: …plans more carefully, sustains agent tasks for longer, operates more reliably in larger code bases. I think, according to the benchmarks they’re claiming, it’s probably the top model. It’s… every time someone releases a model, it becomes the top model for a while, and then somebody else releases one.
Was Anthropic’s Super Bowl ad an attack ad?
Frank: But interesting that Anthropic and OpenAI release models so close together, because it does seem like they are at war with each other, right?
Justin: Yes, they are. Everybody’s at war with Sam. Sam’s getting a lot of heat at the moment. Sam and Elon, they hate each other. Sam and Anthropic, they all hate them.
Frank: And you know what’s funny, right? What’s funny is, you know from this show that even though I use ChatGPT all the time, I have huge reservations about Sam Altman.
Justin: Yes.
Frank: But weirdly, I found myself on his side in this war when Anthropic and OpenAI both previewed their Super Bowl ads, and Anthropic’s ad was basically an attack ad on ChatGPT.
It is a brilliant ad. People should go and look it up. It’s very reminiscent of a very specific episode of Black Mirror, the one where the woman has her brain kind of stored in the cloud, and then the subscription model starts speaking ads.
Justin: Yeah. Brilliant. Yes. Yes.
Frank: So in the Anthropic ad that I watched, it’s like a young man working out and he has this personal trainer, and he’s asking the personal trainer, how does he get abs quickly? And the personal trainer starts saying, “Great, let’s put a plan together for you,” and then starts saying, “Also, you should consider super-lift shoes that will give you an extra X inches in your height.” And the guy’s like, “Wait, what?”
And the ad was basically saying, yeah, ads are coming to AI. They’re not coming to Claude. And what was their tagline? It was something like “Continue to think” or “Keep thinking,” or something interesting like that.
Justin: Don’t know, I was too busy laughing at the ad because it was brilliant.
Frank: It was very funny. I was surprised that Anthropic went with a kind of an attack ad with such a direct attack on ChatGPT. That did surprise me a little, just given that they tend to be the, I don’t know, the more kind of ethical company, the more safety, security, whatever…
Justin: Do you know what else is surprising for me, which is that ChatGPT… the line for me is Anthropic are doing enterprise customers and OpenAI are doing consumers. It’s a different market and you would never have ads on an enterprise product. So why are you bothering to attack OpenAI for their consumer product when you’re not actually gunning for the consumer market? That’s not where you’re aiming at.
Frank: That’s a good point. And Sam Altman had a different slant on their different approaches, because he said basically, and this is why I found it interesting that I found myself on his side a little bit, he started by saying more Texans use ChatGPT for free than total people use Claude in the US. And he said, so we have different-shaped problems.
And his point was that OpenAI are democratising AI access for everybody, and if they need to do ads to do that, that’s what they need to do. And I kind of found myself agreeing with him a little bit, to my surprise. I kind of expected him to come back on the attack and he didn’t. He kind of said, look, excuse me, we’re trying to do something good here.
And he then pointed out that their ad is not a dig at anybody. It’s a very optimistic ad about, we are entering into this builder era where anybody can build anything using AI, and they’re making that available to everyone.
Justin: Ah, there you go. He’s a good guy after all.
Frank: That’s a good response.
Justin: Very good. Well, Frank, I mean, that’s incredible. In fact, you’ve actually agreed with Sam Altman. Incredible. On that bombshell, I would say, pleasure. We will not be here next Friday. We might be here next Thursday, and I look forward to chatting to you then.
Frank: Excellent stuff. Chat to you then. Thanks, Justin.
Justin: Take it easy. Bye.
