Transcript
This is an AI transcription and may contain errors
Frank: I am Frank Prendergast. I’m here, as always, with Justin Collery to argue about all of the latest AI news, and the latest AI news, of course, is Fable 5 and Mythos 5 from Anthropic. This…
Justin: Sorry.
Would you buy Elon Musk’s SpaceX IPO?
Justin: Frank, just before we jump into that, happy IPO day to those who celebrate.
Frank: IPO day. Tell me, tell me more about IPO day.
Justin: xAI is going public today.
Frank: It is Friday, you are right.
Justin: Yeah, yeah, that’s right. I presume you’ve got your order in for all of those tasty xAI shares already, like all the other people who are about to get burned.
Frank: I can’t wait to be associated with and supporting Elon Musk. Yes, absolutely.
Justin: Okay. And you know I normally like to hedge my bets, but on this one I am going to say, I bet you that this time next week the price of an xAI share will be lower than it is today.
Frank: Ooh, that’s, yeah, that’ll be an interesting one to come.
Justin: This is not investment advice. Whatever, all as usual.
Frank: Yeah, yeah. Please read, please read the fine print.
Justin: Yeah.
Will OpenAI bottle its IPO?
Justin: Anyway, the big news other than the biggest IPO, not just of this year, but I think of all of the IPOs if you add them up since…
Frank: And since we spoke last week, I think OpenAI have done that thing, that part of the process where you have to file for doing your IPO before you actually do it and all that.
Justin: Very interesting language around that, right? Because I saw it first with Anthropic and then OpenAI did the same thing, so obviously it’s some sort of legal language. But they have confidently filed, which begs the question, have they sort of unconfidently filed a couple of times before? I don’t know.
It’s sort of a weird, weird way of saying this, but yes, they’ve also filed. A lot of money’s going to be spent in the next couple of months. Also, and I know we want to get on Fable. Did you see what Sam Altman said about that? He said that we might actually not go public.
Frank: Well, I saw him make some statement around recursive self-improvement and why, if you were getting very close to recursive self-improvement, you might not actually want to be a publicly traded company.
Justin: Yeah, which is interesting because I didn’t get his reasoning around that. I don’t understand why that is the case. Did he think that the product was going to be so much better so quickly that it was better? Or was it that there was some sort of regulatory reason that it would be better not to be a public company at the time that this takeoff happened?
It was weird logic. Did you have any insight into that?
Frank: No, unfortunately I didn’t. I was hoping you could share with me what that was about, particularly since he’s talking about it and obviously Anthropic have been talking about it a lot recently, and Anthropic seem to be barreling forward with their IPO.
Justin: Yeah, interesting.
Who gets to play with Mythos-5?
Justin: Anyway, the big news this week is from Anthropic.
Frank: Yeah, so Fable 5 and Mythos-5. So previously we had Mythos Preview, and that was the one that caused a huge stir. It’s massively capable in terms of cybersecurity and hacking and finding bugs and all that good stuff, and we talked about it on the show before. And it was only released to about, I think it was 50 companies, because it was just too dangerous to release to the public.
They’ve now released Mythos-5, which is the next version of Mythos. They’ve expanded Project Glasswing, which is basically the project that you have to be part of to get access to it as an organisation. They’ve now expanded the number to, I think, 200 organisations, so 200 people will have access to Mythos.
Is Anthropic building an AI elite?
Frank: And we talked about this before when Mythos came out, about this whole thing that for a long time now, it’s been the case that yeah, you or I have access to the same AI as Elon Musk or whoever. And when Mythos came along, that changed, and we now do not have access to the same AI as everyone.
There is now an elite.
There is now an AI elite.
Justin: Except I did read during the week that GPT-5.6 is due to come out soon, and it’s rumoured to have very similar capabilities to what we now call Fable. And Sam Altman has said there’s going to be no restrictions. He’s just going to release the thing. So, and remember, he’s the one on… Again, it comes back to show me the incentives, and I’ll show you the outcome.
OpenAI are the company that are struggling to take enterprise customers. They’re the company that are struggling to turn a profit at the moment, and therefore, they’re the company most incentivised not to restrict access to their most capable model in order to steal market share and money from Anthropic.
Frank: They also, I mean, it is interesting. Again, I’m not sure that the companies themselves should be making these decisions, and that’s why I think regulation is so important, because I genuinely think there’s pros and cons to both approaches. I appreciate Anthropic’s cautious approach to safety, and at the same time, I can also see Sam Altman’s argument for democratic availability of AI.
And so they both have very different kind of narratives as a company.
Justin: So are you okay with being a second-class citizen at the moment, Frank?
Frank: In the absence of really good regulation of powerful AI models, I am okay with Anthropic taking a more cautious approach and denying me access to the most powerful models, yes. But only in the absence of regulation that I think we need to be really striving to have in place.
Justin: Okay, we need to talk a little bit about Mythos, but I don’t see how that’s going to make any difference to you, right? Whether you get access to the regulated version or not, and any difference to a bad actor who gets access to the… You know, the bad actor is always going to be a bad actor. You’re always going to be a good actor.
The regulation isn’t really going to change that one way or the other.
Frank: No, but the point is it would have been a conscious decision at a democratic level as to what model gets released to the public, as opposed to Anthropic saying, “We’re going to be really careful over here,” and OpenAI saying, “No, we’re not going to bother. We’re just going to put this out over here.”
Justin: Yeah, we really need to get on and talk about Meta, but actually doing nothing is a conscious decision too. I mean, doing nothing is an affirmative action. We’re not going to do anything. We’re going to let the market sort it out. So doing nothing doesn’t mean being, what’s the word I’m looking for?
Doing nothing doesn’t mean you’re not doing anything. Doing nothing can mean I am actively doing nothing because I want this to play out.
Frank: Yeah, but if you look at the executive orders and statements that have been made by the current US administration, I would say that Mythos, like the event of Mythos, made them realise, “Oh, hang on, there is danger here. We should actually be doing something.” And now they’re scrabbling to figure out what they should do in an environment that they’ve created where they have stated that they are anti-regulating AI.
Justin: Okay. And for everybody who’s listening, next week we are going to have a special on EU regulation, and we will tease out what the implications are exactly of what you’ve said. But that’s for next week, right?
Is Fable 5 just an AI free sample?
Justin: This week, let’s talk about Mythos. So I used Mythos on…
Frank: Let’s just clarify, because I hope you didn’t use Mythos. I hope you used Fable.
Justin: Fable, sorry. Thank you.
Frank: So to clarify for people, Fable 5 and Mythos-5 have both come out, and they both use the same underlying model, but Fable 5 is the version of it that has all these guardrails in place that make it hopefully safe for public release.
So everybody right now has access to Fable 5, or rather, paid customers of Anthropic have access to Fable 5. So what was your experience of it? What’s it like?
Justin: Well, this brings us to the first gripe, right? Because everybody has access, and I pay Anthropic €120 a month or something like that so that I can access their models. And I got a big advertisement saying I was going to have access to Fable under my… You know, I give them money every single month until June the 22nd, and then they’re going to take access away.
So that’s the first gripe that people have is like, “Hold on, I’m already paying you quite a lot of money. Actually a lot of money, not even like a lot of money every month, and you’re going to give me access like a good drug dealer and…”
Frank: Yeah. The first one’s free.
Justin: Yeah. And then you’re going to take it away from me, right?
So I’m not too happy about that. I do hope they change it.
Did Fable one-shot a Miro clone?
Justin: So here you’ll be… The obvious question is, I was in an airport in Toronto, and I started using it. How did I do that? Well, I texted on Telegram my Hermes bot, and I said, “Hermes bot, can you please fire up a Claude Code…”
Frank: Now, let’s slow, slow, dumb it down a bit there, doc.
Justin: Yeah.
Frank: So tell us more, tell us more about this Hermes, Hermes.
Justin: So Hermes is like OpenClaw, but it’s a little bit more locked down, a little bit more responsible. I’m happy to have Hermes running on my machine. So it’s connected to Telegram, and it has control over my local machine, and I can talk to it, right, on Telegram. So I’m talking to it, and I’m saying, “Okay, I need you to fire up a Claude Code session on my computer, create a directory.
We’re going to do some tests, and can you set the model to Fable, please?” And Hermes does that. I’m literally sitting in a bar in Toronto doing this, and it’s controlling my machine and doing that. Right? So then it says, “Yep, that’s all done.” I’m saying, “Right, now can you enable remote control in the Claude Code session?”
And Hermes goes and enables remote control, which means I then have access to Claude Code on my Mac Mini from my phone. Right? So that’s how I was able to…
Frank: Technology.
Justin: Yeah. Isn’t that cool? Right? So now I’m texted. So I have my use case. So my use case is diagramming for architecture and stuff like that isn’t really a well-solved problem in my opinion, right?
There are… Whatever. There’s no free nice tools that I can use just to draw pretty pictures of how things hang together. And so I have a test case where it’s Miro do this, right? But you can’t embed Miro in a browser, right? Or it’s not easy to do it, right? There are some technical limitations. But I love Miro.
Miro’s a great product. So my test case is create a version of Miro, please, that I can do certain other things with that I can’t do with Miro. I did this with Opus 4.8 and with their WorkTree agent, that thing where they spawn loads of agents. I did that two or three weeks ago, and it spent a whole week working on it and burned every token that my €100 a month, right?
And it produced a thing for me that was okay, but it wasn’t really what I wanted, right? It was good. I got Fable to do it for me as I was on the plane, and I was checking in on it, and on the plane, right, on an Aer Lingus plane, the connection isn’t good, so it’s really slow. It’s just text messages going back.
It did an amazing job. It created a Miro clone that looks very like Miro, and I haven’t got to test it yet because I only landed this morning at five o’clock. But from the images that it was showing me, it looked beautiful. Functionally, it looked good. My big thing for it, right, but if people are interested, is Miro doesn’t… Miro doesn’t have a way of storing the images as text.
So it’s hard to agentically create a design or a picture, a systems design. So this one was, it needed to look as professional as Miro, but it needed to be stored as text in the background so that you could have an AI agent generate beautiful diagrams. That was, that’s kind of the wrinkle in the use case.
It one-shotted it, and it took it, as opposed to a whole week, it took it the length of a flight, so six hours.
Frank: Right, right.
Justin: So I’m impressed.
Frank: Amazing. Amazing. So what I’m hearing about it, I have not tried it.
Is Fable worth the cost?
Frank: What I’m hearing about it is that it’s expensive is the first thing, compared to… So it’s expensive compared to the other consumer models. Interestingly, Mythos-5 is actually half as expensive as the original Mythos Preview.
So you can imagine what the public reaction to Fable-5 would have been if it was the original costs. But it’s still very expensive compared to the other consumer models. I’ve also heard that not only is it expensive per token, but it also uses a huge amount of tokens because it does so much research.
It does so much investigation, it does so much reasoning, that it’s extremely token hungry as well, so it ends up being very expensive.
Justin: I will temper what you’ve said there, right? From my experiments, and I’ve read others, it’s very, very good at orchestrating other models. So you can use it as the planner and then it can delegate tasks to Opus or… And it was even using Haiku for some of the stuff that we were doing. It’s like, “This is a really simple task, just use Haiku.”
So I do wonder again, in real-world tasks, it itself on its own is very expensive, but it might be, as an orchestrator, a good way of orchestrating other models and ends up being less expensive as a result.
Frank: Yeah, I saw somebody talking about what they were calling the 10/80/10 approach, which they were saying… I didn’t get a chance to look up the original Anthropic sources, but they were saying that they had got this from Anthropic engineers, and that the idea is that you use Fable for 10% at the start for coming up with the strategy and the speccing and all that.
Then you use your other models for the 80%, which is the grunt work, and then you use Fable for the other 10% at the end, which is reviewing everything.
Justin: And that, that, that is exactly what Fable did for me on the plane.
Frank: Oh, interesting. Okay. Right. Right.
Justin: So it works like that out of the box, so that’s good. Anyway, sorry, carry on.
Frank: Interesting.
Are Fable’s guardrails too cautious?
Frank: So obviously the other big thing is it’s the guardrails. Because the underlying model is so powerful, they don’t want it being used for bad purposes. So if you try to use it for things like biology, chemistry, and cybersecurity, it will tell you that this might go against its guardrails, and it’ll kick you off to one of the other models instead of using Fable 5.
Which seems smart in terms of we don’t want people going off and building chemical weapons or bioengineered weapons or anything like that. But also apparently some people are saying that it’s a little bit overcautious, and Anthropic admitted, “Look, we’ve erred on the side of caution here.”
So I saw a tweet from one guy saying that he literally typed the word, or he asked… One example was he asked, “How does the heart get blood around the body?” And he was kicked off to a smaller model because it was biology. So it can be a little bit cautious.
Justin: Okay, so let’s talk about why it’s really hard to do this, but first let’s talk about the huge lose from… It’s been a total PR nightmare for them because when they did this first, it handed it off to the lesser model without telling you. It silently…
Frank: Yeah. Well, yes, only in the field of AI development. So advanced AI research, it would silently…
Justin: Silently hand you off, right? Which is interesting, right? I didn’t realise it was only actually, I thought it was for all of them. Okay, so you actually were told for everything except for advanced AI development. Okay, that’s not quite as bad. There was still huge pushback, and they’ve rolled that back now.
Frank: Yeah, massive. The research community were saying that it would hamper research, and also because obviously you don’t know that you’re getting a lesser model. You don’t know when you’re getting a smaller model’s output. And for me, the thing that struck me was, I’m not going to be doing advanced AI research, so it’s not going to affect me personally, but if they did this in any other area, it would be very, very dodgy.
Justin: All right. And you’ve got a good example, but I’m going to give you five letters which explains why they did this. The five letters are C-H-I-N-A, right? It’s China. They want to stop competitive companies using this brilliant model to develop better AIs. That’s probably fair, I guess.
Are Fable’s guardrails this easy to dodge?
Justin: And just before you get to your example, I think it’s really hard for them to patrol this, right?
Because, like that guy’s example, right, ask me something about the heart, it kicks it out, and they’re going to push it back and they’re going to change the guardrails. I think it’s Pliny the Liberator, my favourite hacker, managed to get around some of the guardrails just by splitting the problem up into little bits, and then you feed one instance of the model this bit of the problem, and another bit of the model this bit of the problem.
So it doesn’t see all of what you’re trying to do, but you actually get all the information you want, and you can put it together to get the output that you want. Anyway.
Frank: And I think he even got… I think he even managed to get around some of the guardrails by replacing the vowels with Cyrillic versions of the vowels so that the words weren’t flagged, which astounded me because I was like, I really thought we would be beyond that in terms of AI safety.
I thought we’d be beyond just swapping out some of the letters.
Justin: Actually, just to give you a sleepless weekend, Frank, let me tell you my favourite hacking story of this week, which is related, right? So some companies are now starting to use AI to monitor network traffic to stop data exfiltration and stuff like that and try and spot hackers and stuff, right?
And so the hackers now know that AI has been used to read the traffic, and so what they’re doing is they’re putting chemical and biological information into the payload. They’re injecting it into the payload before it goes through the AI checker so that the AI checker refuses to answer.
So normally it’s like it’s been asked a question, “Is this allowed? Is this allowed? Is this allowed?” Then puts some nuclear codes in there and it just goes, “Sorry, I can’t answer that.”
Frank: Wow.
Justin: And through the path packet draft. So unintended consequences, but please carry on with your regulation.
Frank: Yeah.
Justin: With your damned regulation. So go on, give me your example. Why do you think this is a bad idea?
Frank: Oh, yeah.
Why did Fable’s silent routing backfire?
Frank: So the silent… So I think the visible, this might be against our guardrails, kick it off to a smaller model and tell you which model it’s been kicked off to, I think I get it. I understand. The silently kicking you to a smaller model is dodgy.
And I think that if you think of it in a different scenario, so if you think of somebody, say, if you think of a politician in the US saying, “Could you help me put together a strategy for running for president as a Democrat?” And then you think about another politician saying, “Could you help me put together a strategy for running for president as a Republican?”
Now, if one of those is kicked off to a smaller, less capable model, then immediately you can see the issue here. This gets into the whole thing that we’ve been worried about in terms of AI models manipulating perception and potentially having massive detrimental effects if we don’t know what’s going on in the background, if we don’t have full transparency.
Seemed like a really weird move from Anthropic, who are generally pro-transparency, but I think you’re right. I think it is the whole China thing. I think that Dario Amodei in particular has grave concerns about China, whether they’re warranted or not. And I think that this was his way of trying to… He was trying to sabotage.
He just maybe wasn’t thinking about the broader implications, which I think is odd.
Justin: Well, just actually, maybe because, as I said, I’ve been travelling a lot this week, so I haven’t got to dig into it in as much detail. How did people know? If it was silently dropping it to a different model, how did people know?
Frank: They were open about the fact that this was going to happen. They just…
Justin: I’m going to make a stunning leap of faith here. If they hadn’t told anybody, nobody would know. They could write myth… They call this what? Fable 5.0, I think is what they’re labelling it. Fable 5.1 could easily turn around to be even better at numerical reasoning, even better at agentic coding, but strangely very poor at advanced AI research.
And in the background, they could just route it to another model. You would never know, right? And they could tell you, “We’ve just tuned the model to be… The public released model is being tuned not to be good at AI research, but obviously the ones that we use internally are very good at AI research.”
But it’s not. They’ve just bumped it off to another model, and you would never know.
Frank: Yeah, and I mean, this is the issue. And I think this is why we need serious… This is why we don’t want to get into a situation again where, with social media, we had the algorithms on the feeds spurring all kinds of, what’s the word I’m looking for? Polarity in society.
And now we’re kind of realising, “Oh, yeah, that was probably a… We probably didn’t do a very good job there of looking after how the algorithms in the feeds would impact on society.” And we’re back here again in terms of the… We need ways of ensuring that companies don’t do that kind of thing.
Justin: And how do you think that would look? Because, from my point of view, it’s not an unreasonable thing for them to do to say, “Look, our competitors are using our product to compete with us, so the version of our product that we release will not enable our competitors to compete with us.”
That seems reasonable. So is the regulation you’re looking for kind of like when you buy a tin of beans in the supermarket, it’s got ingredients on it, and it sort of lists out what the ingredients are? I mean, is that the regulation that you want? It’s just literally this release contains these models and a sort of a…
Frank: Well, sorry, this is a whole can of worms that we probably don’t…
Justin: Tin of beans. It’s a whole tin of beans.
Frank: Because, look, we’re back to that whole thing of isn’t it ironic that the companies took everybody’s data to create their models, but are now fiercely guarding the output of their models from their competitors and saying, “No, no, no, we can have all the data to make our models, but you can’t have the data from our models to make your models.”
Justin: I think that’s a good point, and it brings me onto the other reason that I have a problem with this.
Will data retention kill Fable adoption?
Justin: So I think that the adoption of this model is going to be lower than you might imagine. So what they’ve said publicly, they themselves, Anthropic, have said that every model of this level and above, we will retain all of the prompts and the outputs for 30 days at least to see if they’re flagged, and if they are flagged, we retain the right to keep those outputs for 90 days or whatever and do an analysis, right?
For big companies, and for maybe even small companies, that’s a complete no-no, right? You’ve got sensitive competitive information in those things. You have potentially PII, your biggest fear, right? You could have health information, all sorts of things in there. And up until now, with these companies, you have been able to sign what’s called a zero data retention policy.
So they promise not to keep any of your data. They’ll process it and then just get rid of it. Not with this model, and they’ve said not with many models afterwards. This comes back to my point about OpenAI and the economic incentives. You can be sure that when they have a similar model, they will release it, and they will offer a zero data retention because they’re looking to steal market share from Anthropic.
But I think that’s, for me as well, it’s kind of like, really? That’s what you want to do? I mean, if they said they were going to do it for a couple of months to profile the traffic so that they could create safeguards that would flag harmful prompts and harmful outputs, I would have been probably pretty cool with that.
But to say that we are going to do it forever with all models, that’s going to affect adoption.
Frank: Yeah. Do you not think, though, do you not think that organisations, once they’ve picked a vendor, once they trust a vendor, that then they are probably okay with that data being retained, particularly if it is like a 30-day window, for example?
Justin: Depends on the data. Some will, some will absolutely not.
Frank: Yeah, sure. Sure.
Justin: So with all of these things, and look at Californian companies know that it’s about friction, right? Every friction, every friction point has to be removed, and this is a friction point.
And another company’s just going to remove that friction.
Frank: Sure.
Justin: Gain market share as a result.
Frank: Yeah. Yeah.
Is Dario creating an AI Iron Curtain?
Frank: Dario Amodei also published a blog post that I thought was very interesting, coinciding with the release of these models pretty much. I’m going to run through it very, very quickly, but the reason I’m interested in it is because he seems to be advocating for regulation and forms of oversight.
So he basically, he opens the essay kind of saying, “Look, yeah, regulation and government bodies, et cetera, move at a snail’s pace, and AI moves very, very quickly. And we need to start taking action now.”
And he talks about the fact that we’ve talked on the show about, say, Bernie Sanders calling for regulation, Elizabeth Warren having ideas about AI regulation and about how we tackle the impact on jobs, et cetera.
The current US administration are now taking AI safety seriously since the launch of Mythos, et cetera, et cetera. So he’s just saying, look, yeah, it’s good to see that people are considering action. But he is also saying that even now we’re probably like a year out of step in terms of the rate of progress of AI.
Justin: Hmm.
Frank: So he talks about regulation and public safety, and basically he says, “Look, in the past it was pretty hard to anticipate what the risks would be or precisely how they might actually manifest,” which is kind of what you’ve said as well, that you can’t regulate until the issues come into focus.
And he’s saying, but now the issues are actually coming into focus and we need to regulate for at least the ones that are becoming a little bit more obvious. He likens it to regulating the way that airplane manufacture is done. And he says then that in the future… So for now it’s like, yeah, if you’re building a model, it’s like building an airplane.
It has to be safe before it’s released to the public. And then he’s saying that, look, it is possible that in the future it might be more like regulating weaponisable nuclear materials. But he also says, in quotes, “We shouldn’t get ahead of ourselves,” so we’re not there yet, in his opinion.
Then we’ve got macroeconomics and tax policy. So basically he says, yeah, look, AI might create wonderful new opportunities, but he’s still saying, interestingly, that it is also possible that it might create massive labour disruptions. I say interestingly because we talked recently about how he appeared to be rolling back a little bit on his white-collar bloodbath statement and all that.
So he’s saying, look, we need to track things at a government level. Anthropic are doing a certain amount of tracking, but if they had the access to the kinds of data that governments had, then it would be easier to see exactly what’s going on. He also interestingly talked about having pro-employment incentives, which again sounds a little bit like what Elizabeth…
Justin: Elizabeth Warren.
Frank: Yeah, exactly. And he says even though he’s kind of saying, “Look, AI might create wonderful opportunities,” he’s also reserving judgment around whether a UBI might be needed in future.
So there was five things. That’s two of them. Three: accelerating AI’s positive impact. He basically says the last thing we want in terms of regulation, et cetera, would be for the benefits of AI to be slowed down while actually not managing to touch the risks at all.
Justin: Yeah, it’s good boys.
Frank: You know, this talks about how there’s various areas that AI could impact positively, but that a lot of them are currently regulated by systems that are designed for much, much slower processes, and we need to get ready for regulating in a way that allows the speed of innovation of AI.
I thought that was interesting. Number four, the state and civil liberties. This was interesting given their arguments with the Pentagon recently. So he points out that AI can give someone a lot of power and also not necessarily being touched by our normal democratic checks and balances because they’re not designed for AI.
And so one of his examples was very pointed. It was, for example, a surveillance-focused AI could analyse widely available information at massive scale and use it to infer the innermost details of every citizen’s life, a technological ability not contemplated by current civil liberties law. So that’s exactly the argument that they were having with the Pentagon.
He talks about autonomous weapons, talks about banning them for domestic use, not for global use, but for domestic use. And, yeah, so basically he had a whole raft of other things in there, but I think that was the main thing was that we need to watch out for this accumulation of power.
And then the last one was securing leadership by democracies. This was also very interesting, I think, particularly in terms of what we were talking about with his approach to China, because he says that democracies should form a coalition. They should be building AI around their common democratic values, and that they should lock down the supply chain for building AI to within the coalition, which is kind of pretty much what he’s been trying to do in the US in terms of stopping chips going to China, et cetera.
And he says that that coalition then should coordinate to address AI’s risks, to share AI’s benefits, to work on mutual defence strategies, and to reject AI-powered repression, and to work together on things like the macroeconomic supports that he talked about earlier, which on one hand you kind of read it and you kind of think, “Okay, I get where he’s coming from,” but you read it from another kind of perspective and you think, “Hang on, what kind of polarised world is he creating here with this democratic coalition and everybody else excluded?”
Justin: I mean, on the face of it, I’d say I agree with them. I’m sort of laughing at the… We banned the use of autonomous weapons, at least internally or domestically. I mean, do we have the same ban for nuclear weapons?
I mean, it seems like the sort of thing you shouldn’t need a ban on. It should just…
Frank: Yeah. Yeah.
Justin: So I don’t know. And, of course, my joke with nuclear weapons is, and maybe this is again because there’s so many parallels, is if you look at how many nuclear weapons have been used in each country, the countries that have nuclear weapons tend to explode most of them in their own country because they need to test them.
So if you look at the number of nuclear weapons exploded in Russia, it’s been almost exclusively, if not absolutely exclusively, by Russians. They’ve launched more nuclear weapons on themselves than anywhere else because they needed to see that they worked. Same with the US. The UK is the only one, and France, I think they haven’t done…
Anyway, so it’s important, right? It’s good that it’s talked about. What will happen? I don’t know, but I do think about it, right? I think about the things that we would have talked about two years ago, and in a lot of cases, exactly the stuff that we talked about is coming to pass.
So one of the things, I remember we talked about this, I would say two years ago, where it was like sovereign AI, right? In terms of culturally, unless you get your data into this training data and whatever, your society runs a risk of being just forgotten about in the mists of time.
And this doesn’t speak directly to that, but it kind of does in this idea of a coalition of the willing and make sure that we all garner the benefits and stuff. That’s part of it, right? It’s part of a sovereign AI. So it’s very interesting. We’ll see what happens. I’m not going to pass judgment.
Will profits beat Anthropic’s safety talk?
Frank: What do you think, in terms of Anthropic and their focus on safety and Dario’s warnings about the risks of AI and blog posts like this, do you think that this will continue once the IPO goes through? And I know that they’re… I think they’re trying to do an IPO as a public benefit company maybe, so there are certain safeguards around what they won’t have quite the same shareholder pressure maybe.
But at the same time, I just kind of wonder what happens when you’re a publicly traded company, people have shares in your company, and you are, for example, calling for a pause on AI development instead of getting your next most expensive model out to the public to make a load of money.
Justin: I wouldn’t be worried about it, Frank, and I can tell you why.
Frank: Why is that?
Justin: Okay, so we now have Fable, right? A model that’s probably smarter than you and me, right? What does the public and what do companies need? Do we need Fable 2.0, a model that’s even smarter, right? Look at the price of tokens and look at the usage of tokens.
And so when it goes public, they need to make profits, right? They need to make money. And so I think the competitive pressures over the next 12 to 24 months is going to be make me a model which is about as smart as Fable, but much, much cheaper. And so the economic pressure is not to make smarter and smarter models.
We already know that there’s a huge technical overhang. The economic incentive is to make cheaper and cheaper models. Again, actually, this comes back, maybe it’s a driver, in fact, of the fear of the competition from China, because China keep releasing these models that are actually pretty capable, but 90% cheaper.
And so that destroys their model, right? Their business model, because they can’t go charging a premium. And so I expect that over the next 12 months or so, both OpenAI and Anthropic will start to release models which are as capable but cheaper. And they’ll have to find some sort of marketing way to make it look like they’re still making their models more capable.
But really, the big R&D push is going to be make these as small and as cheap, but just as capable as possible so that people can use them more.
Frank: Are you cynical enough to think that Dario Amodei’s focus on democracy and a democratic coalition, et cetera, could be simply entirely financially motivated from the financial risks of the competition from China?
Justin: No, I don’t believe he believes that at all. I believe he’s motivated by the fact that this is a technology which if, let’s say, this is a technology similar to the nuclear bomb. Okay, so I don’t know which one of them talked about it, right? But they said that America was a very… It’s actually Elon Musk.
America was a very unusual country in the fact that it got the nuclear bomb before everybody else by years, and it was the first country not to use that power in order to take over the rest of the world. They could have. They could literally have gone to every single country in the world and said, “We have nukes, you don’t. Be subservient to us or we’ll blow you off the face of the planet.”
Every civilisation previous to that, if they had that advantage, they used it to their advantage, and the US didn’t. You can argue that they did in economic terms, but really they didn’t. And so Dario Amodei, his fear, I think, is driven by the fact that if a Russia or a China or some other country other than America does the same thing with AI, they won’t have the same restraint to not use that technological advantage in order to impose their will on other countries, and I think that’s what drives his fear.
Frank: Sure. Yeah, yeah. Cool.
Justin: On time, Frank.
Would you give Ollie AI the keys to your life?
Frank: Yeah, we have one last thing to look at, which is a little bit lighter than some of the material that we’ve just been talking about. So you brought this to my attention, and I’m not entirely sure what it is. So this is Ollie, the world’s first AI family assistant. So let’s just have a look at what Ollie is.
Justin: Who does that?
Frank: So just to pause it here a second. So he just got Ollie to respond to his sister who was freaking out because there was a family emergency. Did I get that right?
Justin: Something like that. And then in the middle of it, so he didn’t put his phone on silent. I mean, who the hell does that? And then it’s like, “Oh, yeah. Oh, by the way, also could you put out the cat food, and could you do…” Nobody does that, right?
Frank: And I love this. Well, that wasn’t real, but you get the point. Well, yeah, okay, but I would like to see what it really does. All right. Let’s…
Justin: Well, let’s roll it there. Maybe we can see.
Frank: Oh my, my family’s going to love it when I add Ollie to the WhatsApp. By the way, I’m just going to pause here again to say I don’t think you need an advanced AI to see that that guy’s calendar needs serious attention. Like, did Ollie put that calendar together for him? I don’t know. That’s…
Justin: Yeah.
Frank: Well, that’s not creepy at all.
Justin: No, no, it’s not. But let’s roll it on there. I’m getting to the bit that I really dislike. There’s all your medicals, Frank, too. You want to watch out for that? Okay, hold on a second, right? So how much stuff happens overnight to this family? Do they have a family of like 20 million people?
It’s like, I bet you he puts his phone on silent when he goes to bed, because that thing must be going ding, ding, ding, ding, ding, ding all night long as people are changing appointments and moving things around and whatever it is.
I mean, that’s just ridiculous. Stuff doesn’t happen in the middle of the night like that. Anyways, carry on. I’ve got one more gripe with this. Mm-hmm. That’s my problem. Okay? This is a family assistant. It has nothing to do with Apple, Google, or OpenAI.
Why in God’s name are you invoking the name of three of the biggest companies in the world just to manage my family of three? I don’t need a trillion-dollar company to manage…
And I can guarantee you, those people are smart, right? These people are smart, too. I wish them the well, the best. But the people in OpenAI, Google, and whatever are, one, smart, and two, incredibly wealthy. They either, A, are way too organised to need an Ollie or, B, have a nanny to do what Ollie’s doing anyway.
I just, it’s like, why?
Frank: So I’m actually kind of surprised. I thought that you would love this. I thought that this was the kind of always-on, totally connected AI system that you were dying to have in your life.
Justin: It’s the European in me, right? Maybe I do want an always-on assistant, right? But I don’t want it marketed to me like this. It just, for a European brain, just cannot understand this marketing. It’s kind of like this is an Alexa. This is what Alexa should be, by the way, but a little bit less intrusive.
But it’s just the whole, it’s scanning everything, and there’s 50 million things are happening at night, and people from Google are doing this, so you should do it too. No to all of the above. Things do not happen like that overnight. You should have your life organised.
I get that you don’t always, but I don’t need people from Google telling me what… Just because people from Google use it doesn’t mean it’s great. I’m sure they use lots of terrible products.
Frank: Yeah, and I think you make a good point there about Alexa, that this is what Alexa should be, because to me it’s like, well, this is interesting because my immediate reaction is no way. Absolutely no way, Jose. I am not going to connect Ollie to my entire life and give it access to everything, but partially because I don’t know who Ollie is or where they came from or where their servers are or who has access, and I don’t want to be the guinea pig for some new meta-style company that is doing all kinds of nefarious things with our data under the hood.
I might maybe consider this if it was, say, Alexa. I’m not really in that ecosphere, but if it was Google, I might be like, “Well, you know what? 70% of my life is in Google already, so yeah, maybe I’ll go the extra 30%.”
Justin: How about if it was in Siri? Would you trust if it was Siri?
Frank: Ooh, that’s an interesting question because although I do have a MacBook, I’ve moved to Android for my phone, and I think of this as more of an always-on, always-with-you phone type situation. So I’d actually rather Google. Literally I’d say 70% of my life is in Google, whereas only my work life is in Apple.
Justin: All right. Well, bad news, you’re not going to get it anyway because the European Union has banned it. We’re going to talk about it next week.
Frank: Excellent. Justin, pleasure as always. I’ll chat to you next week.
Justin: Have a great week.
