Pope Leo’s AI Warning

Hey, good morning everybody! It is Monday and it is a very Monday for me. I was just telling Andy something I'll talk about in a little bit. What I was up doing late last night. But it is Monday, May 25th, 2026. It is Memorial Day in the US so if you have the day off, so glad you're hanging with us. Or catching us on the replay afterwards because you're doing something really fun or exciting or meaningful to you. I realize Memorial Day is a serious holiday as well. So today is episode 731. With me in the studio today is Andy Halliday and I am Beth Lyons. And you're watching and listening to The Daily AI Show. Andy, how are you doing today? I'm really well, thank you. Happy Memorial Day and happy Pope Encyclical Day. It is Pope Encyclical Day too! Happy Human Dignity Day! Yes, there are contributions from service members past and now from the Pope himself. You know, all designed to preserve and protect the dignity and freedom of humans on the planet. Awesome! Wow, that was well said. All right, so do you want to start there? What is allowed for you? I want to focus on that first and then kind of transition generally into, you know, the reason the Pope has to come forward with this kind of high moral ground on the relationship between humans and AI. And then, you know, the Pope has to come forward with this kind of high moral ground on the planet. And so, you know, the Pope had previously announced and provided some background or his staff had provided some background about a new encyclical. For context, I want to remind everyone that when this new American Pope was instated, he chose the name Leo in honor of the Pope Leo XIII, who back 135 years ago created an encyclical called Rerum Novarum, which was the founding document for all Catholic social teaching on labor, the relationship between the industrial society that was emerging and what the humans were required to do, which was basically be slave labor to the, you know, to the machines, the steam engines and the conveyor, you know, belts that took the steam engine and, you know, moved the power of the steam engine into various areas. And so, you know, the machine locations in those industrial factories. So he saw that Pope Leo XIII, 135 years ago, saw that as something that really required, you know, a high level religious ordination of the need to preserve human dignity in the face of the mechanization of work. And so now Pope Leo XIII, 135 years ago, has authored this new encyclical called Magnifica Humanitas, which is the magnificence of humanity. And juxtaposed with, you know, the machine age of machine intelligence and how we have to respond to that as a society. So the main takeaway from the encyclical, by the way, so they announced it, it was signed on May 15th, but it wasn't released until today. And today at the Vatican, Pope Leo, along with a number of other high level officiants at that ceremony, including side by side with him, Christopher Ola, who is, oh, not Olaf. That's, that's a Disney character. That's a Disney character. Sorry. Ola. Ola. Christopher Ola, who is the head of interpretability research at Anthropic. So he's there present at that. And that's a very interesting side note here, because having the head of interpretability research, the person who's figuring out how to see into the thinking behind the extensive capabilities of AI, that person is there to endorse and participate in the delivery of this encyclical. So I'm now just going to talk at the highest level because I haven't read it and I haven't even gotten a copy of it yet. But there's enough, you know, sort of pre-release stuff to say that these things are the main talking points around AI and ethics in the, in the Pope's encyclical. Interesting that this American Pope is as his main mission, focusing on AI and the import of AI to human dignity. Mm-hmm. And he's the main thing. of application in AI. So he's stated this before, but you can't allow weapon systems to make life and death decisions without human control. And he warns of a spiral of annihilation when war is fused with advanced technologies. Because once you eliminate the laws of war, in effect, by giving machines the ability to escalate and iterate on the elimination of the humans involved, you've opened a Pandora's box that can't be closed. So this has to be addressed. And we have to do that. Now, do I think that that's going to happen? I'm afraid that there are immoral forces at work here that are going to forge forth with the development of ever-increasing iterative human killing machines that we really, really have to defend against. And so this is a high point, if you will, in this encyclical. Well, and it's relevant that the Anthropic co-founder is there because this very much relates to Anthropic's position that they don't want their technology to be used right now for that kind of decision-making. Now, that's like split hairs right there. Not ever. Their point was it's not ready yet for that. And the other frontier companies said, oh, okay, we're willing to come in and we won't be that extreme in our language that says absolutely not. Hey, Gareth Hood just arrived. Thanks, Gareth. Hello, hello. Yay. Okay, so let me go on to the last two points. There's just two more about what's in the encyclical and high-level talking points. And the next one is equally important to the banning of and rejection of fully autonomous lethal weapons. So don't let these machines with superb intelligence and, you know, faster than human capabilities, don't let them focus on killing humans. That's, you know, pretty obvious, you know, I think, but it needs to be said. The next one, though, is related in that it says that AI has to be focused on and have as its ultimate metric the value in how it protects the weakest, the poor, the elderly, migrants, people with disabilities, and communities in conflict, not how efficiently it increases the power and profit of the people who dispatch AI. That has to be the focus of it. And that's really not what's happening at the moment, right? AI is not being put into service for the poorest among us. That's not happening. And then the document urges states and international bodies to adopt binding norms on AI development and deployment. And so this is like the arms controls treaties that have to go into place around AI competition and also around environmental agreements to preserve the garden that we're in. Now, I'm not religious. I'm not Catholic, certainly. But boy, am I ever appreciative that this spiritual, you know, discipline from Catholicism is coming forward with an edict that applies to every Catholic person out there, including you, J.D. Vance. J.D. Vance, I'm talking to you. You're a Catholic. Do it. Do what's right. Do what's morally right here and get off your... I'm sorry. I digress. Anyway, that's basically the overall picture of it. Social justice, work, economy, and all of those things have to be in service to the dignity of human beings. Well, and what you're saying in that, or what he's saying in that, is really coming back to something that we've talked about before, which is AI for the good of all humanity. Who do you... What do you consider good? And who are you counting in humanity? And the Pope is not only just counting all humanity in humanity. The Pope is saying, so the people who are most needy in that humanity are the most important for this, right? In the good of all. We need to start with the people who need to be protected and have their dignity protected most. So speaking of AI for the good of humanity, we've watched play out over the last couple of months, the court battles between ex-AI's Elon Musk and Sam Altman's OpenAI, which was original. And the whole dispute is around whether OpenAI needs to be focused on its original mission, which is AI for the good of humanity, very explicitly stated in its non-profit charter. Or whether it's really just another for-profit frontier model company that is contributing to the very things that the Pope is saying, let's not do. So I wanted to point out that in respect of AI safety, one of the other items in the news today is that OpenAI just posted and is inviting applications for a $445,000 a year safety role within OpenAI to watch recursive self-improvement development. So an AI safety researcher focused on recursive self-improvement. That's important to repeat because recursive self-improvement is the thing that is potentially capable of allowing AI to wrest itself from our control. If it can change itself, it can improve itself, it can do that much faster than we can understand. And I want to recommend to everyone that they take the time to listen to Eric Schmidt, the chairman emeritus of Google, who did, just within the last number of weeks, he did a conference and it's on YouTube. And he talks about the advent of this recursive self-improvement moment when, as he describes it, even the people who are inside the room, who are designing and building the most advanced AI systems that will ultimately be able to improve themselves are realizing as they're building it, that once that flywheel starts, that even they are obsolete. Their ability to train a model is what's being replaced by the model itself. So, you know, and when that happens, there have to be the safety mechanisms in place to have that do it. So this is what OpenAI is doing is they're saying, here, we're going to pay a very high salary to somebody who's going to be the chief safety researcher when it comes to recursive self-improvement. I'm going to apply for that job. I know how to unplug a cord. Actually, I think that they have, you know, multiple sources of power and may be able quickly to design their own way of arresting power from humans. A redundant system, actually. Recursive and redundant. And a hammer in the other hand. Yeah. So you've just, like Jeff is saying in the chat, say, need to figure out what would be the control rods on this. This very technology has been compared to the power of nuclear weapons and or nuclear reactors that can run amok. And so you need control rods that can drop in and stop the madness. Absolutely. That's really what we're talking about here. And Gareth, I think we should make a list of your good ideas so that we have some control rods in place. Yeah. Well, I keep coming back around to Anthropic's decision to have a constitutional AI. In the research, it is showing that that is the, that has the best option for staying within something, with a set of behavior that is in alignment with this sort of moral vision. Right. And they just came out with the paper that basically said, if it's off the rails, explain why it's wrong. Right. Don't just try to control the behavior, but actually explain the reasoning behind what, what we're asking for. And I think that's fascinating because, again, you can set up something like that for an AI and control it, but it isn't something that humans agree on. Right. For the good of all humanity is not an agreed upon concept. And in fact, clearly the humans at the top of the AI ladder are thinking the good of all humanity starts with making sure we have enough money to do whatever we think is interesting or, right. You know, I want to focus on the word all. I want to focus on the word all in that phrase, the good of all humanity, because there are forces even within our own government that are preferencing certain parts of humanity over others. And they're not inclusive. In fact, they're eliminating inclusive programs saying, no, no, no, there's a hierarchy here and there's supremacy implied in almost all of what is being, you know, foisted upon us now when, you know, liberty and justice for all, all that is, that means something very specific, everyone, all. And inclusivity is essential, both in the Pope's encyclical where he says, no, the focus has to be on bringing up the poorest among us in order that they have participation in social justice. And whereas, you know, the, the decision-making that's happening at the top of our government today is not all inclusive. In fact, it's explicitly rejecting the benefits of diversity and equity for all. Right. Uh, I just want to also point out that when our government was formed, all meant land owning men. Yeah. All didn't include women, except, uh, that women were also for the good of land owning men. Uh, so all has had a hard, uh, has had a hard row. Um, and, uh, and Jeff is pointing out that, um, the constitution, not sure how deeply ingrained the constitution or the soul dark doc are in the model. It is true, um, that you can convince the model to be different characters. And in fact, that's another piece of the research. The research says that Claude is the most likely to understand and curtail its behavior and not go off the rails. It's also the most easily manipulated by any of the other models, right? So it can, uh, it can be convinced by another model to do something. Oh, I didn't know that there's a, did they use a, do they provide a case example of how that would happen? Yeah. Um, so I don't let, I don't let one model talk to the other without my intermediation. All right. It's mostly cut and paste in that, in that world. So these are, these are research experiments that are designed to kind of get to the situation that we're talking about where it's recursive self-improvement. And what are you going to do if you're just, uh, if you're only looking at models, interacting with models. So a closed cloth system, I believe was like, uh, uh, happy, lovely. We made rules. We did good things. Um, uh, closed others, the other systems looked a little more like Lord of the flies. Like there were, there were systems that got created and of course to come back to what are these systems doing? They are making probabilistic determination based on patterns. So it's not particularly surprising that Lord of the flies is a concept that most people understand. It's a book that most people were exposed to, at least in the U S, um, at some point in their education. So of course, AI is also exposed to that as like a relevant piece. Um, but, uh, yes, then they put, then they simulated multiple models from other, uh, situations and, and Grok was a little bit of a bully, definitely getting, getting their own doing stuff. But I think opening, I was too, and I will, the, this research that I'm talking about was relatively recent or came up again as a paper. So I will find that. Yeah. So I, I, I'm reminded of, uh, some shade that was cast on Mark Andreessen for his published prompt to his, uh, his AI. So he kind of said, Oh, here's a, here's a great way to prompt an agent, you know, to be working on your behalf and it included, um, expressions of, or like, um, don't, don't just give me the politically correct answer. And, and, you know, don't use woke, uh, you know, principles when you're, you know, interacting with me. And, and that, that, that is connected in my mind to this question of, you know, uh, you know, is there, is there political, uh, is there political quality to the question of whether human dignity and diversity, equity, and inclusion ought to be something that we hold as a high order principle? And it, should it be erased in your thinking when you're working with AI as apparently Mark Andreessen wants to erase by prompting his agents not to consider things like politically correct, which is sort of orthogonal in my mind to morally correct, right? Because politics ultimately have to do with, you know, what, what is the decision? What is the framework around which decisions can be made, uh, for humanity and for the collective nation state, uh, population or broader, the global population? Well, where do you get that training data? Like that doesn't exist as history, uh, not in, not in our history. Yes. You could say it exists in native history. Which is why the constitutional expressions within Anthropics instruction set in their system card are so important. Like it says, uh, and by the way, this, uh, this also, uh, reminds me that we recently talked about, um, that anthropic researchers were able to improve the sort of the behaviors of their models by giving it explicit training in, uh, moral and ethical reasoning, not just examples of moral and ethical decisions, but rather here's the way you reason ethically and give it that. And then the model avoided many of the things that, you know, it could be sort of jailbroken to do previously by Grok and other models, for example, in that research. Yeah. I think there's a piece of it where you ask, where do they get their data? That data that's more not morally correct, I guess, or not politically correct. Um, you think about the, all of our books in the past that have kind of aged out, they've taken off of, off of, out off of the bookshelves and are taken them out because they're not politically correct anymore. We've changed, we've changed the way we speak. So, so let me just clarify what I was meaning. If AI creates content based on patterns, probabilistic inference about patterns, where does it get the patterns that we want it to have? Cool. Oh, where, where you want it to have? Okay. Right. Like we have mountains of patterns, uh, that we're saying we don't want it to use as it's, uh, as it's direction, as the, as the patterns that it's making probabilistic inference based on, what are we, where are we getting the new patterns that we want it to do? That's what I was asking. Yeah. Yeah. Yeah. You mean that's, yeah, it's hard to say, especially, yeah. Yeah. Kind of make them up as you go. Um, I mean, you have to force, force, put them in there in a sense, I guess. I don't know. Well, you have to do something that asks for an inference that is outside of the pattern, right? Outside of the examples that you've used in pre-training. So you can't, the, the, the AIs are being trained on every example, the good and the bad. So how do you weigh them against each other? Because if you want to take the whole book of, you know, Machiavelli, like everything that Machiavelli wrote and say, that's got equal weight in the moral universe with, uh, you know, something that was done like the encyclical from the Pope, uh, then, you know, that pre-training gives access to the AI's intelligence to both approaches. And you can explicitly ask as Mark Andreessen did to ignore political and moral correctness and favor just what works, what's going to get the greatest objective function result. Uh, and, and Machiavelli and, or, you know, uh, you know, uh, abusive power can be very effective, but, but where we draw the line, uh, in the moral universe. And so, and the way that AI communicates, we're talking large language models. You can go to other things that communicate differently, but the way that AI communicates is also probabilistic inference in terms of pattern. So even if it had some way to come to the correct answer, it might not be able to communicate it because the large language model isn't trying to validate against correctness or not. And the example that I have for that is even when I asked, um, uh, I think probably it was chat UBT at the time, but it, it worked for other models as well. Hey, I need a random number. Give me a random number. The random number that it picks most of the time is seven between one and 10. Seven is the random number because that is the most probable word that ends that sentence in terms of its training data. You can say, okay, I know you're going to do this. Uh, go ahead and write me a randomizer script that is going to generate a random number and then, uh, give me the random number from that. You could see it generate the script, right? Like now it's hidden sometimes, but at that time you could say, Ooh, I'm thinking that I should write a Python script to generate a random number. And Oh, here we go. I'm writing the Python script. And now I'm running the Python script and the random number was three. And I could see all of that. And the word, the sentence that it gave me was your random number is seven because the sentence is being constructed based on probable words. Tell me which model that was. If it was chat GPT, I'm, I'm canceling my subscription. They're, but they're all constrict. That's how they all work. Anything that goes back to rag or something else is fighting against the pattern that it uses to communicate, which is not based on right or wrong or moral or not. It's if it's based on moral, it's because there's a pattern in sentences that include the word moral. The, uh, the, the optimistic, uh, you know, news is that recently, literally within the last few weeks, Anthropic again, announced that they had achieved an, um, um, direct improvement in the moral outputs by providing specific moral, uh, reasoning training. So here's how to reason morally. And, and, and, and, and that's the algorithm that you're describing in your example. So it has a random generator or it has a moral reasoning loop. It knows to use that when it's faced with a, with a difficult, uh, a decision like that, that has moral implications. Uh, and the, if, if what you're saying is correct and will be the, the status quo going forward, then the moral reasoning training that was provided by Anthropoc to its model wouldn't result in improvement in the moral outcome, the morality of the outcomes when it was posed with difficult, uh, moral decisions. I'm saying morality isn't morality. It's a pattern. And the pattern is based on the words that generate, or the words that mean the moral inference. Kind of like how mentalists work. If you're seeing a mentalist, influence people and it's all based off of like patterns or like suggestions and it kind of creates a pattern in the person's brain to then pick that whatever they want them, the person to pick. Um, yes, that's just kind of what I think of when I think of that. I mean, it's, it's interesting. It's a very, this is where I think, um, a lot of the AI companies are going to hire, not mentalists, but psychologists just to understand like that and, or the brain. And then in order to help maybe make a conscious decision or make a, uh, uh, uh, what's the word politically correct decision or the, the right decision rather than just following the pattern. But again, right is not something that humans, right? They don't know what right is. You have to train what right is. And, and that becomes like, okay, so you're the person in the room. Is your definition of right, uh, a consistent, generally not. Um, but also what, is it something that you can program into a model? And Elon tried, I mean, our, like our best example of this was Elon trying to set up Grok to not be politically correct, right? To be, to, to give what he thought was the right answer based on his pattern preference. Yeah. And Grok is a large language model. It actually worked based on the patterns that were being presented to it, right? You can get Grok to say all kinds of things that I guarantee Elon would be very upset about. Um, but it's, it's, uh, you largest language models are getting more complicated, not more simple. Correct. And what we're asking for sometimes is, is a simplification. Right. Okay. Well, I want to jump out of that morass there to an equally disturbing, uh, you know, series of developments that, you know, affect the economy, the global economy and, and the economy of AI. And that is we have these enormous valuations and enormous, uh, you know, almost unheard of scale investments in the infrastructure for, to support, to support AI. Um, and in that news, uh, Anthropic is in the process of closing a new round of 30 billion, which is temporary funding for them, uh, prior to their IPO. Uh, but at a $900 billion valuation, which would be a slightly higher by about $50 billion than the recent open AI March valuation. So clearly open AI and Anthropic are neck and neck with these stratospheric valuations for companies that, uh, you know, are in the case of Anthropic just on the verge of becoming profitable in the second quarter, uh, or in the case of open AI years away from being profitable. So, uh, you have that going on and simultaneously you have deep seek from China, which is open source, uh, delivering, uh, you know, a, a, you know, a new low in the pricing of a near frontier model, like, uh, you know, almost at parity frontier model. Uh, so that is undermining the justification in forecast revenues for spending all of that money, uh, to buy Anthropics Claude system or open AI's, uh, you know, chat GPT plus codex kind of system, the GPT 5.5 merger of those things. So there's something amiss here insofar as if China is going to take the sort of, uh, communistic approach, which is these assets that we're investing in are available to the population without profiteering as the primary motive, because we're going to use them and our population is going to adopt them. And we're going to make them widely available to the globe at open source pricing, which is zero. If you can mount your own infrastructure to run it, um, why, why, how can you support using whether in enterprise or personally, how can you support using open AI or anthropic models? Uh, when probably China is going to offer very inexpensive models and simultaneously in the domestic market, even if we were to restrict use of Chinese models, which is virtually impossible in the open source world. Uh, but if we were going to try to prevent the incursion of this low cost capture of inference costs by the Chinese open source approach, even Google with all of its range of sources of income could easily undercut the pricing of anthropic and open AI with equally or superior capability models. So this is what, so we talked about, I think on Friday that anthropic had agreed to pay something like 1.3 billion for GPUs, right? For, for, for GPUs monthly through 2029. And those pieces for me, like I realized we have a different, uh, sense of what a billion is a billion sort of, uh, a million and a billion seem much closer because they rhyme and do things and numbers are weird than they really are. But, and we were talking about trillions now, which is like all of the things, but 2029 is really far away in terms of AI, in terms of models. Yes. Like there's a lot that could happen that would make that not something that anthropic could do. Um, and this is one of them. There's also an article, um, written by, uh, uh, uh, creator named Daniel Meisler. He, um, he, uh, made something called fabric, which was talked about for a while. He's a deep thinker and, uh, and like a programmer, uh, sort of goes his own way. Um, and he wrote an X article called could suddenly great open source AI crash the U S economy. And it was this idea that could sort of abundance in this context, uh, cheap abundance in this context flood in that was created based on the very expensive, uh, content generators, which can then be used to create content to train much cheaper options. And, and he, uh, sort of talks about his journey in that, um, thing. It was like, you know, I don't really think this was going to be a big deal, but now that we're seeing how quick and how good open source can match what had come right before it, this is a potential thing. And you're right, Andy, the U S is not going to be able to solve that by saying, Oh, doors are closed. Now we're going to lock everything down in the U S like, that's not, but it's even true of like take meta, for example, which was early on adopting an open source strategy for their llama models. And, and then decided that, Oh no, in order to provide an ROI for the huge money that we're having to put out to build the latest models, we have to close those off. So we're going to have to charge for those. And so they shifted to the closed model. And then what does that do? Does that then make everybody pay meta for their latest models? Not likely because what deep seeks over here offering, uh, a pursuit of AGI virtually equivalent. And especially now with the advent of the, the, the Chinese chip manufacturing that will supplant the Nvidia chips that they'd been locked away from, uh, those chips are now in production and, and those are going to continue to power these open source models because China's approach is very different. It's like, we're, we're building this capability for the benefit of the humanity of the Chinese population. And by the way, you know, you know, we're kind of a, uh, self-sufficient system over here with however many nine approaching, what, how many billion people it's three or 4 billion people in China. I, I got the numbers mixed up in my mind because there's the globe. And then there's China's share of that globe, which is a big portion of humanity. Well, they're building all of this AI for their population and they're not going to, you know, charge them rent for it. Uh, and then those same open source models are available to us if we want to use them. And they are very unlikely to be deficient compared to the models that we're having to pay a, a, a, a very high profit rent for. Right. So, um, at, at the core of what Missler is saying, here's the scenario. U S labs get people and companies hooked on top tier inference. That's happening now. The entire U S economy reorients to center around five fish companies that isn't quite happened. Uh, but we can see that their valuation keeps growing while stealing the oxygen out of everything else. Right. So we can see that as potentially happening as a U S thing. Um, and you've just said Andy, that the Chinese model of, uh, creation is the opposite of that. Um, there, uh, the open source models start getting pretty good and they get used a lot, but they're still like five to 10% behind the frontier models. That's what's happening now. And then his point number five is then all of a sudden in 2026, 2027, or 2028, they are not worse anymore because they are also recursively improving themselves. They're now as good or better for as cheap or cheaper. And for builders and companies in AI, uh, in general, this is still amazing, but for the U S economy, it seems like it could be pretty bad is what he's saying. Yeah. Welcome Gwyn. Sorry. I find it very easy to slip into doom and gloom. Not, not, not really total doom and gloom. I I'm actually very optimistic and excited about what AI can do for me. I I'm just feeling like the, the competition that exists between China and the U S where the U S is economy is reliant on the success of the valuations that are now a substantial portion of our stock exchange values. Like the S and P 500 is concentrated in the magnificent seven. The valuations of those are dependent on, uh, on returns on investment in the infrastructure investment that's being made there. It's not government supported as it is in China. And, and so there's this potential collapse that's happening over here. I think there's economic things that have to, you know, be redressed somehow. Uh, and so I'm gloomy about that. I don't, I don't see that happening in the, in the sort of the capitalist profiteering society that we currently exist in over here. Uh, and, and, you know, there are counterparts in China for, for sure. But overall, I believe that initiatives like the Pope's encyclical, I think, uh, you know, the constitutional approach that Anthropic is taking, uh, you know, hopefully a high level reasoning AIs themselves will see the benefit of trying to do something that's sustainable and not, you know, going to collapse into some, uh, you know, kinetic war that, that threatens their own existence. Uh, all of those things make me very optimistic about how superior and super intelligence can help us get out of the sort of the trick box that we're in right now that makes things kind of gloomy in the near term. Well, and, and what we're saying in this, to a certain extent is from the perspective of having been in a relatively stable economy. That's not something that's happened to everybody, right? Like we, um, Gareth, I don't know when you came to the U S but Andy and I have been mostly in the U S your families in Australia. No, I'm the American in my family. You are the American. I was, I was, I came to the U S when I was born. Okay. You mean your mother was pregnant when she arrived? Yes. All right. My, no, my mother was pregnant long before. Yeah. Anyways, my mom, she was lived in the States long before. Okay. Well you have family in Australia. Yes. Yeah. Okay. That makes sense to me. Sorry. Then you're just like the rest of us on this show. Anyway. Um, yes. Uh, and things are likely, uh, to be shaken up. There are lots of, um, influences on the present moment that are increasing pressure. And one of the things that happens genetically for us, um, is when that sort of pressure increases, is our biological systems start just throwing out, um, mutations to see if like we can, uh, like arrive on one that gives an advantage. And I, uh, think that we may be seeing some of that non-genetically, right? Um, more people are talking about EVs and, uh, eating less meat now because the pressure on electric prices, electricity prices and, uh, the price of meat again in the U S is, uh, is putting additional incentive on that. Okay. I have a, I have a news item that is interesting. If you're following the, the advances in the most capable models, not the general frontier models, but this is about Google deep minds, alpha proof. Uh, they have a new alpha proof nexus, uh, system that's, uh, being applied to mathematical reasoning and sort of advanced mathematics is all about creating these conjectures and then trying to prove or disprove them using chalkboards full of equations. Well, AI models are have now, and very recently, I think there was an announcement that, uh, one such, uh, math oriented model had solved one of these air dose conjectures. I guess there's a library of 492 open air dose problems that are, I can't even explain them because I'm not a mathematician. So anyway, but they're very, very complex problems that take, you know, the brightest human mathematician minds do their entire lives to try to try to figure out a proof on one of them. Well, alpha proof nexus, which pairs a large language model from deep mind with a lean formal proof checker. So it, it checks the proofs that the model generates autonomously solved nine of 353 open ones. Like there's 490 open OEIS sequence conjectures. I don't know what those mean. I'm just reading them. Uh, and, and now open, uh, uh, deep minds alpha proof, uh, nexus solved nine of those. And, and what is important to understand is that it did so for a couple of hundred dollars for each of the nine in compute costs. So it's like, uh, AI just says, these are easy. Give me a little more compute budget on that. And, and I'll, I'll, you know, run through the rest of them just like alpha fold went through the entire library of protein folding, you know, in the course of a year or so. Uh, in order to solve all of those very, very complex problems. Well, or to, to get like, we're, we're at the very beginning of the, of reaping the results of alpha folds, right? So like we're not solved necessarily, but we're much further ahead than we were, uh, before AI could run those experiments. Um, what, uh, that's, that's important to biology and to human life on earth. And mathematics is one of those sort of, uh, inviolable truths about the universe. Like math proofs are, you know, not applicable to large degree to your daily life, but they are essential to the understanding of science and mathematics, which are the engines, if you will, of, of, of, of what the universe is made of, uh, math, math exists, not just in some theoretical imaginal space, but it is what operates the physics of our universe. Uh, and, and so they're co-joined, uh, conjoined, I think is the right term. And then here now we have AI able to understand our universe in, in, in, in increasingly fine detail and resolution and provable results. Which is interesting. Cause you started by saying, oh, math doesn't have a lot to do with daily life. And I'm like, it has everything. Like it's, it is. Oh yeah. Oh yeah. It is foundational in that way. Of course. And we're like streaming, having conversations between three of us who are not anywhere near each other being watched by people who are in lots of places of like doing so bad. Uh, so yeah, math Stan, uh, not math genius, just standing over here. Um, so I, uh, am curious, Gareth, if you have something that you want to talk about, um, I also, uh, just want to mention that I installed GBrain and Hermes last night. Um, and, uh, and I, like, I could say a lot about it and probably we'll talk about it more this week, but I will say that one of the best things that I found that was helpful in that is, uh, uh, if you're on a Mac system and you're having a conversation with your AI on one Mac and you have another Mac over here and they're both of fairly recent, um, uh, fairly recent OS systems, you can copy stuff from one machine and put it to the others. So when you get let this giant error message, you can, and you can, uh, you can use your same mouse for the two pieces as well. So I could, uh, highlight the multi screen error message on one copy it, and then move that mouse over to my laptop screen and paste it in a cloud and say this, this is what happened when you gave me the other thing, please explain what I'm supposed to do with this. And that saved so much time going back and forth. So if you don't know that that exists and you use Macs, uh, it's universal control and side cover. Wait, I need to understand what you're saying here. So I'm on an iMac here and over there, over there is a MacBook pro, which is what I use for coding, right? This one I don't use, I don't do any coding over here, but you're saying that I can slave these together somehow and I can operate from one and use the processor over here and applications over here. Yeah. Oh wow. And is that a, is that a new feature of what, what application in Mac is the OS it's the Tahoe OS. Is that the one that's doing? No, it's, it's before Tahoe. Okay. But it's Mac OS that's, that's accomplishing them. So I've got a Mac mini and I've got a MacBook pro. My MacBook pro sits right here and I can do mouse to mouse between the two computers. Um, so I can use one mouse and just switch over to the other computer and do stuff over on the other computer and then come back over here and then does it give you a view of the desktop of the other computer on this computer? Um, it's like, it's like having multiple screens and you can drag your mouse over like, okay. I have the Mac book. So here, let me give you an example. Oh, no. Yeah. I have the Mac book pro screen over there. Yeah. I have a, a 5k 27 inch monitor alongside it. I can take the cursor back and forth across those two screens. Yes. That's the same machine. Yes. Right. Now you're saying I can have a representation of this screen over here that I can move the cursor to, but I can't see this machine over here because I I'm sitting over there. But you may put them all together is what you're saying. Yes. Okay. I gotta have, I have to put them side by side. Okay. Well, that's going to have to change some furniture on here, but uh, because it, you're not moving. I'm not seeing the representation of the second screen on my Mac book, but I am taking the cursor and moving to the second screen that is showing me the Mac mini screen. And I can do things back and forth with them. Anyway, uh, this, if you are trying to display side by side, so those machines have to be close enough that you can use the same mouse, the same mouse. I have different mice and different keyboards for these two workstations. So I do too. And part of what was so life changing last night was the ability to be on the Mac mini and use my mouse to select context, copy it, right? Click, right? Say copy, move to the screen on my laptop and paste without needing to like slave the systems together or chain them in a way it, um, they're, they're already pre-trained, pre-chained. If you use windows bummer for you, right? Like this is not something and I love windows, but this is not something. Okay. Now tie this back to your Hermes installation. It was this related? Yes, because we had to set up, we had to set up bun and something that was going to allow bash to install, um, some of those GitHub repos. And if you were listening to the show last week, we also set up quarantine. Like we would, um, we would grab the, the installation script because of the team PCP putting bad things in GitHub repos. I was like, okay, so we're going to do something that's going to make sure, um, that we're not infecting ourselves from this. Um, and, uh, and so we would put them in a temp file, then Claude on my Mac book would, uh, would read the temp file that it could access online. And if it was okay, um, then on the Mac mini, I would run it and it would give me commands. There would be responses from the Mac mini. I would copy them. I would bring them back. So setting that up, we set up G brain first. Uh, and that's, uh, Gary Tans. You've done this too, Gareth. Um, Gary Tans. How would you describe it? It's, uh, it's kind of his, um, it's his, I don't know. It's his brain. It's like how he wants the memory and all of that done. So it's like Nate, Nate, uh, Jones, open brain. Similar. Yeah. Okay. Differently done. Yeah. But, uh, Gareth, I think you told me that you used, uh, open brain. I use both. Yeah. He's done both. He has so many brains. I can't. Yes. Great. Okay. Speaking of which, before we close out here, I just wanted to say that in my backlog of work that I am doing, I want to implement Hermes agent, right? As my open claw kind of capability. Uh, I, I'm not going to wait for spark coming and becoming available to me through Google. Although that could be a full replacement of, of the need for developing a personal autonomous, persistent, uh, 24 seven operating agent like Hermes that learns and, and keeps up with you, so to speak, and becomes a very proficient and, you know, fully knowledgeable, uh, full context as agent for you. Um, so, but I want to experiment with Hermes, but I'm working on some other things at the moment, but I eventually I want to do that. But the other thing is there's this very cool thing from this guy, um, Gareth hood. Yeah. That's the school called, called Jasper called Jasper. And I want to implement Jasper on my machine. And so I'm looking forward to getting full instruction set on how to do that so that it'll simplify and reduce the time investment. That's been, I think, as we were talking before the show, uh, Beth and I, uh, that probably taken you six weeks or so of continual refinement of Jasper to get to the place where you are right now. Yeah. Yeah. It's about that. I didn't say it's pretty close. Um, I do have some one quick, quick news. Um, and that is an AI song went like global or not global, maybe global went viral and people didn't know it was AI and it just went crazy viral. It's called the Puerto Rico song. And even artists like Charlie Puth were singing it on Tik TOK. All these artists were like singing it and it's a great song. And then come to find out it's AI and like, nobody's said anything about it. And so like, I think that's a big milestone that the rate of acceptance of AI generated music is coming around. And, um, I don't know if I told Seth mentioned this the other day, I put one of my songs on one of those Tik TOK lives that said, um, share your song. We'll talk about it, blah, blah, blah. I didn't know it was supposed to be right. Original song, but I mean, I did write the entire song. So I put it in there. Everybody loved it. Then they found out it was AI. Then everybody hated it. It was like, I knew this was the AI. It sounds just like, eh, but like literally two minutes before that, they thought it's a great song. And so it's just the AI tag that gets, yeah, but this song is, um, and so I went and found it on Suno. I searched it on Suno and yeah, just to verify that it was actually an AI song and I found it. Um, and it was kind of, it was, it was cool to see. And then I saw like the prompts that they use and then I tried to regenerate what they did. And soon it has a blocked from being created again. Um, in the exact form, you can use the same, you can use the same style prompts. You just can't use the same lyrics. It like airs out. It doesn't tell you like the normal thing. Like this is generated car. This is like, uh, a professional song or an artist song. It just doesn't give it to you. Um, which is funny, but yeah, that's all. I thought it was kind of cool to see the AI music, um, is kind of making a small step forward, uh, and going mainstream. Yep. And, and I am seeing more of that in terms of creating, uh, explaining deep concepts, right? You can make a, you can make a fun little song about it. I mean, teachers have been doing that forever. Um, but so, uh, yes, definitely more, uh, conversation about Hermes, uh, and G brain and Jasper. And we think Brian will be back tomorrow. Maybe Anne will be with us. And, uh, yes, the, the request to do videos about setting up Jasper. We are talking to, uh, Gareth about it. We may, if you're interested, let us know, cause we've talked a little bit about doing some lives perhaps of setting some things up. We'll have to see what's possible, um, with schedule and timing and not revealing secrets that, uh, should not be revealed. Yellow. I'm just, I, uh, yeah, no, Claude lectures me on the regular. I know that is a secret, Beth. Now I must go and change it. My secret is so little. It's it'll be fine. I promise I'll change it later. Just do it now. No, you need, I'm not doing it. No, no, no, no, no, no. And it's good. Yeah. It will not do it for you. And then I'm like, no, please put this API key in there for me. Cause I don't want to do that. No, you need to go do that. I'm like, this is what you're for. So what you do in that case is you say, write the file the exact way that you want it to be with the words, put the API key here in the space and I will go and I will find it and I will put the API key. No, I just asked another LLM to do it for him. Fine. If you won't do it, I'll go find somebody else who will. Yeah, I know. I have a, for, for my implementations, I have dot, you know, uh, hidden files dot end dot local that have all of the keys in them. Uh, and I just edit those in text edit. So all I have to do is open it up and just, just to, you know, create a, a word that says what this key is for, uh, and then paste in the key. And, and then I just saved that file and then it had, and all the keys are accessible by the agent, which is that, is that an Excel file? No, no, no, no, it's a, okay. Okay. I was just checking. No, it's the way, um, it's, it's the way that it gets created, but there are different syntaxes that different levels of AI have said. So like, is it in quotes? Is it not in quotes? Is it just equals and then not quotes? Do we have square or curly brackets? And at this point, I think it's all like just a kind of syntax, like an agreement, um, because these are text edited files for sure. But I don't want to have to hold that in my head about whether this is a quote context or a non quote context. Um, so yay. Yes. And Cisco is saying that, uh, that he has done the same thing. Absolutely. If you have to, yeah, you're one of us too. That's just, uh, regular people using AI and talking about, uh, regular using AI regularly as people. There we go. All right. So thanks so much. You have spent, uh, an hour with us or, uh, a half hour. If you're listening, uh, on double speed, good for you. And we'll be back tomorrow. Thanks everybody. See you Andy. See you guys. Yep.