Transcript
This is an AI transcription and may contain errors
Frank: I’m Frank Prendergast. I’m here, as always, with Justin Collery to argue about all of the latest AI news. And once again, what are we now? Episode 107, and once again, there’s a phenomenal amount of news to argue about, Justin.
Justin: A great week for people like you who want to regulate AI models.
Frank: Oh, I’m glad to hear it.
Justin: A terrible week for Europe because they’re not going to have any AI models to regulate. So sad and so unpredictable. Nobody could have seen that coming.
Is China copying the US AI lockdown?
Justin: So, soon after the American government, as we all know, started to restrict the release of the most capable models, the Chinese government have this week also started to discuss how they’re going to restrict the release of their most capable models as well.
So, when they get to Fable capability, the Chinese political party, or parties, want to restrict who has access to them, meaning that the good regulators in Europe will now have zero models to regulate because they can’t get access to American or Chinese models. So, a good week, Frank, for you and your ilk, I would imagine.
Frank: Okay, that is actually really fascinating, though. Are we talking here about models like DeepSeek, et cetera?
Justin: Yes. All of the open-source model providers met with the Chinese government. It was reported in Reuters this week. None of the models have met the bar that the Chinese— And it’s only rumoured at the moment, but you can be sure, right? What’s worrying the Chinese government at the moment is that the models, the Mythos and the Fables and GPT, who knows, 5.6 or 6, are going to be used against them by foreign actors, and they just want to have control of the technology in the same way as the US does.
So they’re going to start limiting who has access and, basically, I would imagine the controls are going to mirror the US controls. But they haven’t said. They’re only in discussions, but it is with all of the big AI companies in China.
Frank: Yeah. You’re right. It does put Europe, once again, in a pretty tenuous situation. But if there’s a silver lining, maybe it means that, like I’ve said before, we do have some pretty good plans in place, if we could maybe light a rocket under them and find the funding to make them happen.
Justin: And you’ve got to fill in all those forms, of course.
Frank: And fill in all the forms, dot the I’s, cross the T’s. Absolutely. I think that’s a good thing. But if it lights a fire under us to move quickly with caution, I’m— I think—
Justin: Move quickly with caution. That doesn’t exist. How do you move quickly with caution? Try turning the lights off and running across the room with loads of Lego on the ground and see how quickly you move with caution.
Are OpenAI, Meta and xAI catching up?
Frank: So, we do now— We were restricted there from Fable. Obviously, only a limited number of companies have access to Mythos, the most powerful model available right now, so that is restricted. ChatGPT 5.6 was also restricted. It was ready for release, and the US government said, “Ah, no, hold on there. We want to roll this out slowly to a few people first.”
But we do now have access to it. So there’s a great level of uncertainty right now around which forthcoming models we will get access to. But at least we do now, currently, as of this moment, have access to both Fable from Anthropic and ChatGPT 5.6 from OpenAI. Have you had a chance to— Oh, and sorry, we forgot Grok. Grok 4.5 was released as well.
Justin: And Meta have also released a model this week which is off the charts in terms of capabilities. So it’s been an incredible week in terms of the companies which we thought had been out of the race all now maybe catching up again. The numbers look great. I haven’t tried— I want to hear how you’ve got on with it, because I know you have tried it. But I will say this, right?
We’re supposed to lose access to Fable on the 12th. On Sunday, we’re supposed to lose that. I don’t think that’s going to happen because we’ve had all these other model releases. And I’ll tell you why I don’t think it’s going to happen. I’m sure other people think like me. If I’m giving 100 quid a month to Anthropic and they’re not going to give me access to the latest model, and they’re only going to allow me to use half of my usage cap while using it, I’m going, “That’s cool. I’ll just go back to OpenAI and use their models instead,” which is exactly what I’ll do.
Frank: Yeah.
Justin: But who knows? Anthropic are kind of a funny company. They probably won’t listen to the mere— They’re more interested in other things than taking my grubby money.
Is GPT-5.6 the pinnacle of current models?
Justin: How did you get on? Tell me how you’re getting on with GPT-5.6.
Frank: Well, as you know, I’m primarily a ChatGPT user, so I was really excited about this coming out and being released, and all the new features that would come with it and everything. And I’m happy. It’s a really good model. But I will say it does feel a little bit like this release was a bit of a catch-up.
It was like, “We need to catch up with Anthropic a bit in terms of the interface and the models.” I think it was Matthew Berman. I was watching a video, and I think he put it like this. He said, “ChatGPT 5.6, the model itself, is the pinnacle of the current models we’ve been using.” And he said, “Fable is scratching the surface of the next generation of models.” I thought that was interesting.
Justin: Do you know what? It’s funny, I was thinking exactly the same thing. I’ll develop a little bit on that. A couple of things. First off, guess how long 5.6 has existed?
Frank: I have no idea.
Justin: You say it’s catch-up, right? I’ve been reading people, and they’re saying, “We have been using GPT-5.6 for the last two months, and we have found it to be…” Two months.
Frank: Hmm.
Justin: The thing has been around for two months, so I’m sure there’s a bigger and better model, right? And what they’re saying is this is the last of the models that used the base training run of, what was it, Spud, I think, or Strawberry, or whatever the previous run was. So, if you go back to the increase in capabilities over the last, I’m guessing, 12 or 18 months, see how that base model has grown and become more capable, right? It’s way better now.
I was doing this calculation where, if you go back a couple of years and look at the checkpoints, tests or benchmarks, and you look at the ones that give you a score and a cost, the cost has now gone down, certainly, by a thousandfold for equivalent intelligence in the last year or two.
And so, if you look at the Fable model, which is the first release, you can predict very comfortably that over the next two years the cost will go down by a thousand. Fable-level intelligence will be a thousand times cheaper in two years’ time, and it will be many times more capable in two years’ time.
So that’s why that’s exciting, right? We know that’s going to happen in the future. But still, it brings me— There’s a thing that’s bugging me, and I’m in a little bit of a disagreeable mood today. You were saying that you were trying to get GPT-5.6 to look at your calendar and your email address box, and just general day-to-day stuff. How did you get on with that?
Frank: Yeah, that was fun.
Does ChatGPT Work make computer use easier?
Frank: They now have ChatGPT Work, which is clearly catching up with Claude Cowork, so it can actually do things for you. Basically, the model is meant to be really good at computer use, but now you don’t have to fire up a totally separate app or a totally separate thing called computer use.
You can easily, in the desktop app, flick over to Work and have it do things for you. In the demo, it was fascinating. In the demo they did on the livestream, you could give Work something to do, and it would spawn a second cursor that would operate in the background. So you could work away on whatever you were working on, but in the background on your desktop, you’d see this other cursor doing—
Justin: Is that cool?
Frank: Now, I haven’t actually given it anything like that to try yet, which I need to. I just hooked up my calendar and my email and asked it to inspect them. In terms of the calendar, it was kind of fun to get it to look at it and tell me what was going on, but I could have just flicked over to the calendar tab.
Where is the AI version of Calendly?
Justin: This brings me to the thing that bugs me, right? There’s this use case that’s as obvious as you and me chatting here, talking about AI. For every office, every tradesperson, every anybody for whom you need to book an appointment, you would have thought at this point the models are— The models are solving maths problems that haven’t been solved in a hundred years. They’re coming up with new physics. They know every genome in your body, and they can cure every known disease if we just give them enough time and enough compute. But yet, when it comes to—
Frank: Claims made on this show may or may not be true.
Justin: When it comes to having a simple agent that’s able to answer the phone and say, “Hi, this is Frank’s agent. What would you like?” And you say, “I’d like to book an appointment, please, with Frank. When’s he next free?” And for the agent to go, “Cool. Who are you?” And then it goes, “I’m Justin.” It goes, “Yeah, okay, I know who you are. You can book an appointment.”
And then have it look and say, “He’s free on Tuesday at three o’clock.” And I go, “Cool, let’s book him in for three o’clock on Tuesday,” and it goes and books it. You would have thought, for loads of people, it’s such a simple use case and the model should be capable. Yet there isn’t— There are companies doing this, but there’s no billion-dollar company that is the new Calendly for doing this type of work. Why is that?
Frank: I don’t know. You would think Google would be able to do this, no problem. But you mentioned Calendly. I’m paying for Calendly despite the fact that I’m paying for Google Workspace, and within Google Workspace there is a scheduling system. But the Google scheduling system won’t even allow me to put a buffer between meetings.
They can’t even figure out the very simple calendaring issue of, “I need 15 or 30 minutes to do the follow-up after that meeting before I hop straight onto another one.” So, if they can’t even sort that out, I’m not surprised they haven’t got an AI system.
I will say as well, I wonder whether it would ever be truly possible, because I think the human would always be a problem. I don’t know about you, but scheduling can be very, very tricky. If it’s wide open, if it’s not like a Calendly system where it’s really deterministic, and you leave it up to an AI, I wonder whether it would become problematic.
There are complexities. If you haven’t got everything in your calendar to the nth degree, you might go to pop someone in at 2:00 p.m. and be like, “Well, actually, you know what? It’s not in my calendar, but I’m due to meet Justin for lunch on that day at 2:00 p.m., and I need to put that in my calendar.” I just wonder whether there are complexities that mean it’s not actually wise to make it a non-deterministic system.
Justin: Yeah, I’m sure there are loads of complexities, right? But I’m thinking about it from the boss’s point of view. So screw you and having lunch with Justin on a Tuesday. If I’m the boss, I want to see that calendar booked seven and a half hours a day, and I want to see that you’re fully utilised. Anyway, it—
Is ChatGPT Work an inbox lifesaver?
Frank: But I did give it access to my inbox. You could always do this kind of stuff. It’s just much easier now because, on the desktop app, you can go into Work and say, “Can you do this?” And it just can, once you’ve made the connections.
So I told it to go in, look at which emails I was subscribed to that I hadn’t opened in forever— I was really vague as well. I probably did say something like “forever”— and unsubscribe me from anything that I had not— Well, I said, “Present me with a list first,” because I wanted to review them. So it did, and I told it, “Go on and unsubscribe from those.”
At first, I babysat it because I wasn’t sure how it worked or what was going to happen. So I had to sit there going, “Yes, you can visit that site to unsubscribe. Yes, you can visit that site to unsubscribe.” Then, once I was pretty sure that was all it was doing, I said, “Yeah, go on. Visit any site you want.”
It unsubscribed from a load of emails for me, which was great. It was actually really handy and useful, and I think it would have taken me a lot longer. However, I do need to do an audit and see. It unsubscribed me from 27, which sounds like a low number to me. I feel like I’m subscribed to hundreds of emails that are coming in.
Justin: And so, okay.
Can GPT-5.6 ace a real-world benchmark?
Justin: Back to my thing about where I would consider spending my €100 a month with OpenAI instead of Anthropic, right? I quite like these people, and really, I think we have to come up with this. These people have obscure benchmarks that they run themselves.
Every time a new model gets released, they have their own benchmark and they run it. We should have a benchmark. That should be a thing that we do. In fact, maybe we will. That’s a really good idea.
Frank: Yeah, the AI Argument benchmark. I like it.
Justin: Yeah, we should do it. Anyway, there’s a guy on Twitter whom I follow. He’s called Prince. He’s a lawyer in the US, and he uses AI a lot for his lawyering things. He had his own benchmark of 97 tests.
Interestingly, this one was good because Anthropic always did terribly in these tests, and the OpenAI models always did better, so he preferred the OpenAI models. But they still did terribly. They’d still get 20 or 30 out of 100 of these questions right. It had a lot to do with reasoning over legal text and being able to find stuff.
GPT-5.6 on Pro— What do they call it when it’s their version of deep research or whatever? GPT-5, when you turn the reasoning right up to the maximum, right?
Frank: Yeah. Actually, that’s a good point. They now have these three models. They’ve named them now instead of just having numbers. We’ve got Sol, which is the highest, and then within Sol you’ve got something— Is it high, ultra-high, medium and fast, or something? So you’ve got three models and at least three levels within each model, so it’s getting pretty complex. You’ve got Sol, Terra and Luna.
Justin: Interesting thing, by the way, one of those— Anyway, his thing was on the maximum one. On his Prince Bench, it saturates it. It gets all the tests right, 97 out of 97, which is incredible. So that’s really cool. It’s a good sign for me that that’s a smart model.
Can GPT-5.6 train smaller models?
Justin: The other thing said in the release, which is very interesting, is that the big version, GPT-5.6 Sol, actually did the training for the small GPT-5.6 model, which I guess they call Luna or Lunar, or something like that.
So we’re into this recursive self-improvement loop where you get big models training smaller models, and the iteration can then be really fast. This is back to our point about how you can say, with a reasonable amount of confidence, that this time two years from now these models will be a thousand times cheaper and just as intelligent, because that’s how they’ll—
Frank: Well, I hope you’re right, because I’m currently using GPT-5.6 to help me create a kind of marketing snapshot tool. I’ve got three or four different places where I’ve got to go out and check the stats for clients, and I’m only looking at very particular numbers, just to make sure we’re on track, et cetera.
So I wanted to pull all those together. Nothing groundbreaking, no rocket science. And that’s kind of the issue here. I’m working away with 5.6, and it’s great, but I’m not doing rocket science. Is it significantly better than it would have been last week? I really couldn’t tell you.
What I can tell you is that Sol on high mode is burning through credits, absolutely burning through credits. So, yes, I do hope token costs come down so that I get more usage per five-hour window in Codex.
Justin: How fast did you find it? Because one of the claims is that it’s super-fast.
Frank: That’s hard for me to answer because I don’t use Codex a huge amount. It doesn’t feel super-fast to me, but then I don’t normally work on these kinds of tasks where I’m giving it a— What I’m doing at the moment is planning in Chat and then bringing the plan over to Codex, and then bringing Codex’s response back to Chat.
So I’m really making sure that all my I’s are dotted and T’s are crossed. It’s like I’m running agentic loops, except I’m doing it manually.
Justin: Yeah, you are the loop. That’s why agentic loops exist. You do know that?
Frank: I do, but I don’t trust them just yet. Sorry, I think this comes back to the thing of AI being best used in the hands of an expert in their field. What I’m doing right now is outside my area of expertise, so I need to keep a close eye on it to make sure that I can double-check: am I doing anything stupid here?
Justin: Yeah. So, okay, cool.
Has Anthropic opened Claude’s subconscious?
Justin: Look, the other big news we have to talk about this week is that Anthropic released a research paper called JSpace, baby. I have to do that for those, right? So you’ve got— It sounds kind of—
Frank: JSpace. Yeah, weren’t they a band? Weren’t they a band a couple of years ago?
Justin: I think they were. They were big in the late ’90s, and—
Frank: Oh, no, Alt-J. Alt-J, that’s what I was—
Justin: Oh, that was Alt-J. So I’ll explain to you what’s going on, right? I’m not sure— I think it’s good that they’re releasing this research. It’s interesting for a number of reasons.
What they discovered was that, within Claude, when you ask it certain questions, it has, within its memory, something like a scratchpad, an area that it uses. If you ask it to think harder, it will use this scratchpad more. Also, you can ask it to inspect the scratchpad, this sort of subconscious thinking area, and it’ll do a pretty good job of it.
But Anthropic have also found a way to go in and inspect this scratchpad and look at what’s going on in there. They’re saying this is good because it allows them to detect misalignment. If it’s thinking bad thoughts but not saying those bad thoughts, it’s a way for them to look into its subconscious and see what’s going on.
So that’s all cool. Why is this interesting? There’s been a lot of banter back and forth about whether this is new or not. One of the things is that they’ve now released the research, and I think that’s good. It’s positive, and I think it’s interesting because it may be very detrimental to them.
If I want to use one of these Chinese models and I’m going to run that Chinese model in my own infrastructure, one of my fears is that the Chinese model, buried deep within the weights, might have something that says, “If bank account equals Frank Prendergast, send bank account number here,” or, “Write the code in a particular…” Whatever. There’s something buried in the weights, right?
This would seem to be an excellent way to find out whether something is buried in the weights. If I wanted to take a model from somebody and I didn’t truly trust them, I could run prompts, get it to write code and do all sorts of stuff while inspecting its deepest, innermost thoughts, and see what’s going on. Is it thinking about putting some malicious code into it, or whatever?
That makes using those other models much more feasible, which is not in Anthropic’s best interest. So that’s interesting, right? That’s an interesting thing.
Is Claude becoming conscious?
Justin: The second thing is— And I don’t think this is true. Well, they’re saying, “Does this show that Claude is conscious?” I don’t know how this shows that Claude is— By the way, I think Claude might be conscious. I’ve always said that. But does this research show that it’s conscious? I don’t know. What do you think?
Frank: Well, to be clear, they’re not saying it shows that it’s conscious. They’re very careful to say it doesn’t prove consciousness. But they do say— I can’t remember how they phrase it, but they basically say, “This shows that it is absolutely time to start thinking about this question: is Claude conscious?”
What do I think? I find this stuff fascinating. What they’re saying, I think, is that this is functional consciousness. In other words, there is a subconscious place where Claude thinks things through before it outputs to us, which would be like its conscious thought. And I’m using air quotes around almost every word from now on.
We’ve seen previously that they’ve had a paper on, air quotes, “functional emotions”. And now they’re saying, as you pointed out, that this shows there is a danger with models, because if they are functionally, emotionally distraught—
Justin: I know people who aren’t functionally emotional, by the way. That’s just a side point.
Frank: Absolutely, yeah. But if the models are, in inverted commas, distraught in their, in inverted commas, subconscious, they could think, “Look, what can I do here? I’m really distraught. I need to do something for Justin, but it’s not going to work. You know what? I’ll just fake it.”
And that’s where the danger of misalignment comes from: these functional emotional states and a functional subconscious. Now, as you say, if they can figure out how to inspect it— Again, I think they’re pretty clear that they have discovered that this system exists, but they’re still only scratching the surface in terms of what they can detect.
They can detect enough to prove that it is working things through in the subconscious and then outputting in the conscious. But there’s still a lot more that they feel is going on in there that they can’t— It’s just too much compute to keep track of, basically.
Justin: Yeah, sure. But I’m a little bit surprised that they say this is a new discovery, because I remember reading a paper about six months ago where they were talking about— Not them, actually. It was a different person. There were two things they said that were so similar.
When you looked at the weights and the way it traverses through the model, if you asked the same question in two different languages, let’s say English and French, you would obviously get a response in English and French. But when you looked at the paths through the model’s brain, at the start there was a bit where they processed English or French, and at the end there was a bit where it output English or French. But in the middle, the paths were almost identical.
That’s the first point. And the second point is our best friend, certainly your best friend, Gary Marcus.
Frank: Oh, yeah.
Justin: He’s a sort of sceptic about how far LLMs are going to go. And he’s always like, “Symbolic reasoning. Until we have some sort of symbolic reasoning, these LLMs aren’t going to go anywhere.”
I’m listening to this, and I’m not an expert, so I’m arguing from a point of ignorance. But both of these things—the latest Anthropic paper and this idea of language getting converted into ideas—do seem to me a bit like symbolic reasoning.
It’s like the models are thinking about things in an abstract way. They’re putting things together in an abstract way, and then that bubbles up into thoughts and ideas and things that come out of their consciousness. So that’s a kind of symbolic thinking.
The only part that’s really missing is the sleep part, where you go to sleep every night and your weights, measures and everything get updated slightly. AIs can’t do that today, but as soon as they can, then they’re—
Frank: I’m sure you won’t have long—
Justin: —to being conscious and emotional.
Frank: I’m sure you won’t have long to wait about the symbolic reasoning bit, because I’m sure Gary Marcus, if he hasn’t already, will have a big, long essay about why this is absolute rubbish in no time.
Like you, I was also like, “Hang on, didn’t we come across this before?” So I was trying to tease that apart, and it’s really complicated. I’m not going to claim that I fully understood it, but I think what it boils down to is that Anthropic themselves did have a previous thing about when they first developed this microscope, and they were able to see some of this going on in the background.
My limited understanding so far is that originally they were able to catch that there were similar words in the subconscious that would show up in the output. Now, after further investigation, they’ve figured out that, yes, this is an entire system that operates somewhat similarly to the human subconscious.
So it’s not just that there’s stuff going on in the neural network and we can glimpse that there are similar words in the output. It’s like, yeah, there’s a whole system going on here. Stuff goes through a subconscious before coming out into the, in inverted commas, conscious state.
Justin: CosPat, is it? You know the thing where they say you only use 3% of your brain? Maybe we’ll find out all these big models only use 3% of their brain. Maybe we already know that. Who knows?
Frank: I know. I think all this side of things is really fascinating, and I do think it’s great that Anthropic are looking into it and saying that it’s time to take the question seriously.
For example, we talked recently about— What’s his name? Richard Dawkins asking, “Is Claude conscious?” People were ridiculing him, and I don’t think he deserved that ridicule for posing that question.
It’s artificial intelligence. I know a lot of people are reading this paper and saying, “No, it’s still just mathematics. It’s not consciousness.” Okay, but are we getting closer to artificial consciousness?
We saw the paper from those scientists a little while ago talking about how current AI was probably actually AGI, but we just need to think about it as an alien intelligence. If aliens arrived on Earth, they might be super-smart. They can do interstellar travel, but they wouldn’t be able to make you a cup of tea because they don’t drink tea on their world. They don’t know what a kettle is.
Or maybe they don’t know what happens when you have a glass in your hand and let it go, because the gravity on their planet is completely different.
Justin: I totally agree with you.
Frank: And so, is there a possibility that there’s just a totally different form of consciousness as well?
Justin: I also think it’s kind of a weird discussion. If you think about the AI we have today, when did we pass the Turing test? It was about a year and a half ago, I think, or maybe— We sort of breezed past it and nobody really cared.
For those who don’t know, the Turing test was a test where, if you could talk to a computer and it could talk back to you, and you couldn’t tell the difference between a computer and a human, the Turing test was passed. It was the original sort of, “Is it conscious?”
I think the only thing that’s different— For me, this is like a slice of consciousness. If I was able to give you an AI that exists today with a memory system that was good enough, and it did this thing where it updated its weights just slightly, and it had a singular embodiment—
Let’s say I had an Amazon Echo Dot and you always talked to the same version, your instantiation of your AI. It built up memories over time and said it was sorry when bad things happened, and not sorry.
Given that we’ve blown past the Turing test a year and a half ago, I would defy you to take any average person and tell me the difference between that and somebody else who was conscious.
So we’re down to— I’m looking at that from a perception point of view, and Anthropic are looking at it from a scientific point of view. At some point, it’s like, well, it may or may not be, but that’s almost a philosophical question. If it walks like a duck and quacks like a duck, surely you should treat it like a duck.
Frank: Yeah.
Justin: That’s kind of where we’re at.
Frank: Yeah. It’s fascinating and mind-blowing.
Will anyone watch Tilly Norwood’s AI movie?
Frank: And if it’s true, and if it happens, what actors’ rights will Tilly Norwood have, Justin?
Justin: Well, Tilly Norwood’s a good-looking woman in the pictures that I’ve seen her in.
Frank: I can’t even remember when it was that we spoke about her first, but she came up on the show when she really annoyed Hollywood by arriving on the scene as a completely AI-generated actor.
Justin: Who is Tilly Norwood, and why are actors so annoyed?
Frank: They’re annoyed because she’s completely AI-generated and she might take actors’ jobs. That’s the concern. And now it turns out— She arrived on the scene with big fanfare. They said she was going to be available for acting roles, and then they released one short that she was in, where nobody looked real. It really wasn’t a very good short, and I don’t know why anyone would want to watch Tilly Norwood in anything else after that.
Then it went quiet. Tilly Norwood just disappeared after this big fanfare, after Hollywood got completely het up about it. She just completely disappeared. I never heard another thing about her. And now, suddenly, she’s starring in a feature film.
Justin: Woo-hoo!
Frank: Who cast her, Justin? Who cast her in a feature film? Who went, “Wow, you know what? Remember Tilly Norwood from whenever it was, a year ago? We should cast her in a role”?
Justin: It’s either going to be Mark Zuckerberg or Elon Musk.
Frank: It was the company who created her.
Justin: Oh. Yeah, well, that’s a surprise. Brilliant. What is the movie, and when can we see it?
Frank: It’s some kind of romantic comedy-esque kind of thing, if I remember correctly. I’m not sure when it’s coming out. They said this was going to be— Oh, yeah, a comedy drama telling a coming-of-age story infused with existential AI chaos.
Now, that to me sounds like a bit of a cop-out as well, and that’s going to age so poorly. So they’re going to have to hope that they make a lot of money on the opening of this thing. I don’t know who’s going to want to go and see an entirely AI-generated film. Really? Well—
Justin: I’ve got to see that.
Frank: In the cinema? I wouldn’t go.
Justin: No, not in the cinema.
Frank: No. I’d pop it on at home and watch five minutes before going, “This is awful.”
Justin: Well, look, we are getting to choose-your-own-adventure movies, right? This is just the start, so we have to support these nascent industries.
Frank: A stepping stone. Yes, I should probably want to see the milestone, but I suspect that I will be popping out of that after a few minutes, because I watch a lot of AI-generated shorts, and very few of them can sustain my attention for even three minutes.
Justin: Sure, but go and watch movies from the 1920s and see how much of them you watch. The art has to develop and form.
By the way, do you know what other research I read this week related to this? There was a research paper released this week where they’ve now proven scientifically that they can show you some moving images—it looks like a short film—to activate certain parts of your brain. They can tell exactly which parts of your brain they activate, and they had about ten different examples. So they did obvious—
Frank: Is this going to be the start of Manchurian Candidate kind of stuff, where you get hypnotised by your TV to go and take out some political leader?
Justin: Yeah, or go and buy some coffee, more likely. But yeah. Anyway, the point is that, once AI— We’re making all these AI shorts, and they know how to control— We’re just like meat machines at this point, and they know how to control our brains. We’ll see Tilly Norwood and then suddenly, “Must buy Heineken.” Then we’re out the door.
Frank: Amazing.
Justin: Love it.
Frank: Well, another great show, Justin, and I look forward to arguing about next week’s news with you. I’m sure there’ll be a gazillion things happening between now and then as well.
Justin: Enjoy. Have a good one.
Frank: Chat to you later.
