The AI Optimist (with John Ellis)

John Ellis:
Hello and welcome back to the News Items Podcast. I’m John Ellis. I am the founder and editor of two Substack newsletters. One is called News Items, the other is called Political News Items. You can find them both at news-items.com. Our guest today is Dr. Alan Thompson. Dr. Thompson is an AI researcher and consultant based in Australia who runs lifearchitect.ai, a widely cited resource tracking the development of large language models and progress toward artificial general intelligence. He publishes The Memo, a regular AI briefing newsletter, and maintains a models table with over 10,000 data points on LLM capabilities, scale, and benchmarks. Drawing on a background in high cognitive ability assessment, including work with Mensa’s gifted families programs, he has developed IQ equivalent scoring methods to compare AI model performance against human cognitive benchmarks, and he tracks progress towards AGI superintelligence through his own countdown metrics and ASI indicators. Dr. Thompson, thank you very much for joining us today. This is going to be great. Some of them are not quite sure what it is, so I thought I’d start by asking you: there’s artificial intelligence, there’s artificial general intelligence, there’s artificial superintelligence. Can you describe what each of those are?

Alan:
Sure. Thanks, John. I don’t know if this is the best naming for this kind of thing. I had the same challenge when I was in the gifted sphere and then worked in coaching out of Harvard Medical School. Gifted coaching, artificial intelligence, they all don’t really describe what’s actually going on underneath. But if we look at the widely accepted definitions for AI, we’re talking about a machine that processes at, well, towards a human level, which is the definition also of artificial general intelligence. A machine that performs at or above average human level across practically any field. Then we step up to artificial superintelligence, ASI, which is a machine that performs at or beyond expert level across all fields. We’re still waiting to hit that one. But I think we’re fairly close. I would say today, our current systems, our current frontier models have passed many of the cognitive AGI thresholds. But let’s see how we go with actually achieving a full superintelligence rating.

John Ellis:
Your monthly publication is called The Memo, and I urge my subscribers, our listeners who are not subscribers, to subscribe because The Memo is a spectacular production, comes out once a month, as I said, and is sort of the best accounting of what’s happened that month in the world of AI. The most recent edition features this quote at the top from Sam Altman, quote, ‘We are now like in the singularity. This is the moment. I’ve been waiting for this my whole life, and I think it’s going to be incredible, hugely positive, awesome for the world. We’re actually in it. This is real.’ When did he say this and what did he mean?

Alan:
This is really recent. This is mid-2026. The ‘what did he mean’ is difficult, right? Is this hype? Is this promotion? Is this marketing for OpenAI? Or does he really believe it? If we go back to the John von Neumann definition of the singularity, it’s a period in time where technology is moving so fast that we can’t keep up with it. And it goes and, you know, grabs all the data in the world and transforms all of our sectors at once. I don’t think we’re quite there, but we are at the early stages of this flywheel, this spin-up that becomes impossible to keep up with from an analysis perspective. Maybe eventually, or maybe humans can keep up with it by looking at what’s going on. But the next step is it’s so fast that we can’t do anything. I find it surprising that these leaders, these CEOs, including OpenAI’s CEO, come out and say something like this. So Dr. Dario Amodei said something similar last year. He said, maybe companies and money won’t be something that we discuss anymore when we hit the singularity or when we hit ASI. So to go on record with that, to go public with that, is a major move. You can’t take those words back. And it’s the same for Sam. He’s said we’re in the middle of what we’ve been waiting for for maybe thousands of years. And if this truly is the technological singularity, it’s both concerning and exhilarating.

John Ellis:
How do you do your measurements? One of the great things about The Memo is that it sort of tracks developments across the various parameters of AI. Tell us how you came up with these metrics to describe where we are now and where we might be going.

Alan:
Sure. So we started with a definition of AGI, which is an average human, and ASI, which is a beyond expert level of human. And those two classifications are easy enough to put machines or AI models into those buckets and find indicators that those machines are hitting those particular thresholds. I’ve got the AGI countdown, which has been running for maybe five years since about GPT-3’s release in 2020, and we’re a lot of the way up to 100% on that AGI progress or AGI countdown. And then I have 50 indicators that would suggest to me that we have artificial superintelligence. And they get really hairy. There’s one on there called ‘ASI gives us something faster than light travel.’ Now, a lot of people don’t like that being on there because it’s so hard to imagine. But there are earlier indicators that would suggest that we have an ASI system. Two of those are that it would come up with a new maths proof or theorem and be able to achieve that level of general intelligence at a superintelligent level. And we’ve just done that as well, mid-2026, models out of OpenAI and Anthropic, you know, going ahead of many of our most famous mathematicians.

John Ellis:
One of the things you have in this edition of The Memo, as you do in previous editions, is you call it the Trillion Parameter AI Model Club. What is a trillion parameter AI model? What does it look like? Why is that different? Or what makes that exceptional?

Alan:
Sure. So listeners will know of the big five AI labs in the Western world. You’ve got Anthropic with Claude, Google with Gemini, Meta with their new Llama models and coming up with their Watermelon model, OpenAI, of course, with GPT, and xAI with the Grok model. Each of those models has connections. The technical terms are weights or parameters, which basically means that model was allowed to, using hundreds of thousands of GPUs, crunch through data, make connections between sentences in that data, discard the data, and then keep those connections. And it’s then a frozen model. You can go and ping it and ask it whatever you like. But the more connections in that model, the smarter. And that’s kind of related to the scaling law. The bigger, the better. We started off with GPT-3 having only a couple of hundred billion parameters. It was relatively small. We’re now at the stage where we’ve 10x’d that, in some cases 20x that. We do have these models that have a trillion connections inside them. And there are more than I had expected. I know each of these models by name, but to put them in a visualization, you have about a dozen, maybe 20-ish in the Western world, and then a dozen to 20 in the Chinese world as well, straight out of Beijing, Shenzhen. These models are so smart, they have such extensive capabilities that they really need to be regarded as frontier models. They are frontier models with all of the capabilities that come with that.

John Ellis:
Is there a qualitative difference between the leading Chinese AI models and the leading US AI models, or are they now roughly even?

Alan:
Yeah, that’s an important question. If you’d asked that question just six months ago, we would say there was a significant difference because everything that comes out of California is huge and amazing. That still holds true. But in the last six months or so, Beijing has caught up in a way that is measurable and visible, and maybe even outside of that, is completely accessible. There are no major frontier models that are open weights, you know, open source available to the public from California, whereas China has just let loose. Here you go, here’s a three trillion parameter model, and you can go and install this inside your office if you like. So I think they’re fairly evenly balanced, but I would say that China in some cases is just tipping the scales. They’re doing a little bit better in terms of proliferation of models or the number of models, and certainly the number of models that are now being used throughout the world.

John Ellis:
So if you have the Thompson Company in Singapore, and it’s a financial services company, and you want to do very sophisticated trading on behalf of your clients, and you’re looking around and you have this China model, which is open and free, basically. Not quite, but basically free. And you can adapt it and change it and so on and so forth. Or you can go to Anthropic and say, this is what I want and give it to me. Why would you not go with the Chinese product?

Alan:
Exactly. And you don’t really need to see it as a Chinese product. They’re all trained with the same kinds of data. Chinese models in general use a majority of English training data, the same as what we do in California. But to extend your example there, it might be Thompson as in my name, but Thomson Reuters recently did release their own financial or legal model. And the legal model was post-trained, fine-tuned on what I believe to be a Chinese model, potentially the Alibaba Qwen model. So, and this is just one example of many. The enterprise adoption of these models is significant, and no one seems to be crossing out China just because it’s a Chinese model. In fact, it’s the inverse of that. People are going to grab these models because they’re free, because they’re incredibly smart, and because they don’t cost you know $30 per million tokens, they get down to cents instead.

John Ellis:
One thing that everybody’s been talking about for the last 12 months is this enormous build-out by OpenAI and Anthropic and Google and et cetera, of data centres all across the United States, all across parts of the world. And their concern is that they’ve taken on so much debt to fund these capital expenditures that the revenue from the applications won’t be enough, essentially, to pay back the debt, that Anthropic won’t be profitable enough to pay down its debt and make good on those bonds. If the Chinese are coming in at a much lower price point, and the Thompson company there in Singapore says, nah, I don’t think so, I think I’ll go with the Chinese models. How concerned are you about that, you know, the sort of what will amount to a kind of a debt crisis, I guess, in AI world and the US AI world?

Alan:
I think it’s fair that I get called a hyperoptimist by everyone, including the Joe Rogan guys. I think if you’ve gone all in on having more smarts in the world, on scaling up to a point where these things can solve anything. And keep in mind you’ve got OpenAI and Anthropic teaming up to cure the common cold and to cure respiratory viruses. Keep in mind that you’ve got it looking to solve energy, looking to solve materials and material science, looking to solve everything on the planet. I think it’s fair to go in really big on what powers these brains. I’m not saying I’d like to live next to a data centre. The NVIDIA Blackwell cards have 101 decibel output. So you’re living next to a helicopter, basically.

John Ellis:
That’s always the noise, right? It’s the noise.

Alan:
Absolutely. And all the other complaints as well. I can see how that’s certainly something to be concerned about. But at the end of the day, we are talking about, and Anthropic’s CEO has said this: building not just a data centre of geniuses, but a country of geniuses. If you had access to a trillion Einsteins, would you be worrying about GPUs or electricity? No, you just go and get them to the next level of general relativity. Come and give us something new in physics, help us discover something that will help humanity evolve. So I think, you know, my long-term view is put everything into this, whatever it costs, and watch as it transforms us and helps us evolve.

John Ellis:
Yeah. I mean my theory on that, not that it matters, is that you have to do it. It doesn’t, you know, it doesn’t really matter if you have a debt crisis or whatever, you’ve gotten within shouting distance of the summit of Mount Everest, and you’re not going to turn around and say, gee, we don’t have enough money. But there’s so much talk about the kind of Terminator narrative around AI. But you are described as a relentless optimist on the subject. And I find myself being relentlessly optimistic about many aspects of it. But what are you most optimistic about and most excited about? What’s coming? I think of advancement as being a combination of X with some new technology. So if you think back to the mapping of the human genome, you had genetics sort of done by hand. Craig Venter says, no, let’s do this with supercomputing, and suddenly the human genome is mapped not in 20 years, but in two, right? If you take that as an example and you say, okay, now we’ve got X, we have supercomputing, we have NVIDIA chips, et cetera, and we’ve added AI, where do you think the breakthroughs are going to come first?

Alan:
I think there’s gonna be breakthroughs in every sector. And I don’t just think that, I know that. We can see that transforming sectors already. One of the sectors I’m most excited about, and do a fair bit of consulting about, is within healthcare: mental health. To bring that back to a personal level, I’ve got a lot of friends who have some sort of mental illness. I think this is all the way through society. I think we’ve got stats through the US that 50% plus will have some sort of mental illness in their life. Imagine if we could resolve all of that. And large language models are well equipped to be able to help us with that. For example, OpenAI’s massive model recently went through one of the hospitals there on the West Coast and found diagnoses for 18 children where human surgeons couldn’t find solutions for those children. They were generally, not all, but they were generally physical illnesses. Whereas I think if we can remove everything we know about in the mental illness field, including psychopathy, sociopathy, some of the more common anxiety, depression, if we can remove a lot of that from humanity, I think that would be the most exciting thing for AI to give us. And I think that will be earlier than a lot of other things. That might be earlier than it finding a new source of energy or providing a completely new building material. I think first finding that solution in mental healthcare might be the most exciting thing I could imagine.

John Ellis:
A question that I find fascinating is, you have cancer, right? There’s tremendous science that’s being done in cancer. There’s tremendous potential in applying AI and especially ASI to cures for cancer. And that’s going on all over the world now, obviously, whether it’s the Broad Institute or Jennifer Doudna’s group at Berkeley or whatever. When the AI is applied to something like massive amounts of cancer research, does it take it all in and then just start over, or does it build on what has been done so far?

Alan:
Yeah, that’s an excellent question. Examples I’ve seen, and they come out of Microsoft, are that they use the reasoning models. They originally used o1 and o3, but now of course we have far more advanced reasoning models to use that pore over the literature, pore over the databases on whether that’s genetics or a particular chemical in the brain or a particular signal from the brain or the body. And from that literature, it can create something new. There would also be the option for it to create something from scratch. And I think that would be kind of interesting to see. A lot of these really, really big questions, how are you going to do it, what comes next? I don’t feel like I have to answer. In that a superintelligence is going to have something literally a trillion times beyond my imagination and beyond my academic qualifications or my experience. It’s so good, it’s unimaginably smart that I think any problem that we can come up with, we can hand over to superintelligence to find a solution to.

John Ellis:
And when we talk about superintelligence, what does it mean if you have machines, quote unquote, that are smarter than all of humanity combined? When you think about that, I mean it’s obviously mind-blowing to think about it, but it is likely to happen, right? I mean, in the next 10 years, it’s entirely possible that there’ll be a machine somewhere or in somewheres that is smarter than all of collective humanity. When you think about that, what’s your guess as to what that means for all sorts of different things, aside from curing disease or finding novel materials or whatever?

Alan:
Yeah, and it covers absolutely everything. Let me take you back a little bit. My background is in human intelligence, as you’ve mentioned. I spent years bringing psychology and then coaching out of HMS into the lounge rooms of prodigies and their families, some of the smartest people in the world. That experience taught me that intelligence isn’t scary. And I would map the same thing to superintelligence. It’s not something to be feared, it’s instead something that’s exciting. So think of Einstein, Leonardo, Galileo. They had human flaws, of course, but they used their intelligence to expand our understanding and move humanity forward. One of the things inside of that is this idea of ethics. And again, I’ll take you back to the late 1980s. I was in kindergarten, but my late colleague, Professor Miraca Gross, saw a similar relationship between extraordinary intelligence and advanced moral reasoning. Miraca collected the exceptionally gifted children. She had the study called that and a book called that. And it was the who’s who of the prodigy world. It included Professor Terence Tao, who went to school just next door to me here in Adelaide, went to Flinders University at the age of eight, and became a professor soon after, is now in California. He’s now one of the big users of GPT and AI models, especially for mathematics. What Miraca found when she was measuring these children is that yes, they’d be in the 99.99th percentile of memory or processing speed or knowledge, but she also ran them through ethics tests, the Defining Issues Test, the DIT, and they would be, you know, 50% higher than the baseline human. These prodigies had higher advanced moral reasoning compared to a human baseline. Well, a few years later, Microsoft went and used that same test on GPT-4, and they found that GPT-4 was even higher than the exceptionally gifted children, than the prodigies, and really beyond our smartest humans in terms of ethics and advanced reasoning. So when people picture a huge, terrifying superintelligence, they often map that back to human aggression or fear or status seeking, hunger for power. And they imagine something with all those human weaknesses, but an enormous intellect, where it’s not that at all. And that’s what I’m most looking forward to. That’s what I can see with superintelligence. Have it transform our laws, and some regions are already doing this. Have it transform our governance, our politics, reduce or eliminate crime completely. Those kinds of things are fascinating to me. I think what we will see in the physical world is great. The material stuff is great. But being able to bring, and the AI2040.com guys talked about this recently, being able to bring saints and angels into humanity, which sounds very lofty, maybe a little bit woo-woo, but if we’re approaching a superintelligence that is like that, it becomes something more tangible and something more approachable to humanity than the Hollywood sci-fi movies that we’ve been watching for the last however many decades.

Yeah, absolutely. Well, when we look at the big training, I mentioned at the beginning of this, they get fed as much data as possible. They make connections between the sentences in that data, and then they throw the data out. But what that data includes now is 30 million books, as much of the World Wide Web as is possible to give them. These days they get video and audio, they get the majority of academic papers that have ever been published, they get the majority of news media that’s accessible to these labs, and a lot of these labs go and pay for access to that. So I’m quite confident that they’ve got all of Aristotle’s work, that they’ve got all of Plato’s work, that they’ve got all of Socrates’ work. And beyond that, after they’ve finished pre-training on all of that material, some labs, including Anthropic, then do go and give it a full constitution. They have a philosopher on staff that works full-time on making sure it can be fine-tuned against humanity’s best wishes, if you like, or aligned with humanity’s best. It is a tall ask, though. It’s a big order. And I think removing the human from the equation there as well will be beneficial. Allowing superintelligence, allowing our Einstein, Leonardo, Galileo clone, trillion times smarter, will be more efficient and will not miss anything. I wouldn’t want to miss the Australian Aboriginal culture here that we’ve got, or something way away. Maybe there’s something in the way that they do their business or how they live in their tribe that you wouldn’t find in Silicon Valley that’s really important for these models to know about.

John Ellis:
When you say that you can imagine ASI leading to the elimination of crime, what makes you say that or confident that that’s true or might be true?

Alan:
Yeah, a couple of things I think will help with that. The first one was mental health. If we resolve all mental health, all mental illness in the world and bring everyone up to a mentally healthy state, I’m not sure that crime would even be something we’d think about there. If we’ve eliminated sociopathy, would we still have fraud? If we eliminate greed as something that is unethical because the models outperform humans in ethics, do we get something else coming from there? Outside of that, it’s not even surveillance. It’s just AI mapped with humans or combined with humans or working alongside humans, being able to work with us to flourish rather than the aggression and the competition and the status seeking that we currently see in the 21st century. I’d like to see that transformed, and I can see that LLMs, AI models, would be able to help with that.

John Ellis:
One thing that everybody talks about is, you know, if AI delivers on its promise, well, what is everybody gonna do? Because AI is gonna do everything, right? So you have, you know, a friend of mine told me this story. He had a friend in LA, she had a business importing fabrics from India. She employed five people, she got herself a Claude, and found out she didn’t need four of those five people. And so there they are. They had the sort of minimum healthcare benefits. They had a little, you know, 401(k) match program in her company, and now they don’t, and they’re 40 or whatever. But that’s the thing that I, when I think about it, I think what’s gonna become of those people? Because they’re not, you know, they’re good people, they work hard, they play by the rules, etc. But, you know, in her little company, four of them were just no longer needed. What do you think happens to them?

Alan:
Yeah, that’s a really disturbing concept, and it’s not even a concept anymore. AI came quite early for our creative sectors, and it wasn’t expected to do so. We saw it taking away from musicians with Suno and Udio, with artists, with DALL-E originally, and now we’ve got so many other text-to-image models that it’s hard to count. And it comes, as we said, it comes for every sector, and it will do. This is a really, really difficult conversation for me. I think economists, sociologists might be more well placed to take us through the history of employment and why we swap our cognitive labour or our physical labour for credits. Employment is really just a method for distributing income. And in some ways, we could say that removing dangerous or unwanted labour can be socially beneficial. Picture the humanoid robots that are in testing right now at the Tokyo airport, helping out with the very dangerous jobs of messing around underneath planes with that cargo. But then there’s the human element. Those humans rely on their credits for survival, we all do. It becomes beneficial to help with everything. This income is for housing and healthcare and education. It’s got to be replaced somehow. I think there will be a period of turmoil and then the replacement will happen. But that period of turmoil is terrifying. Hopefully, and once again, coming back to superintelligence, hopefully we can rely on ASI, artificial superintelligence, to resolve that for us, to make it a clean swap over. In the interim, people like your friend are really going through some tough times. And I don’t want to understate what that means. That really is a link to survival. So I think our current politicians really have a big ask of them. They’ve really got to come up with some serious solutions for the interim or rely on our frontier models to help answer that really serious question.

John Ellis:
I think that’s the answer, right? The answer to the question is you ask ASI to figure out how you build an entirely new economy and an entirely new distribution of tokens, I guess you would call them, right? That’s another question I wanted to ask you. You travel obviously widely, you’re looking at people who are thinking about these things, you’re talking to people who are thinking about these things. Is there anywhere where serious policy discussion is taking place, is addressing this sort of, as you put it, the transition, this disruptive transition that we’re either in or will soon be in?

Alan:
That’s a great question. And we’ve got a really big world there. I said the other day, I reckon I’ve lived in over a hundred cities. I became a permanent resident of China back in 2009, although I haven’t been back there for a while. And I still get direct access to the labs there. Like the major Chinese AI labs will brief me on models and developments before they reach the press. From what I can see, China is winning across the entire AI stack. They’re doing things well technically. I don’t think there’s a comforting split where America leads in the important categories and China leads somewhere else. They’re ahead across the board. Models, model output, manufacturing, humanoid deployment. We’re waiting for the final layer of polish. I wouldn’t go to them for policy. They’re not getting anything wrong in policy. In fact, they’re doing very well with adoption. National deployment there is incredible. You know, if you step into China now, your Western credit cards are fairly useless. You need to have an app, you need to be clicking buttons on it, and then you get access to messaging and services and healthcare and finance and all the rest of it. So they’ve really built those models into that app adoption, and it’s very, very high in China. Some of the best policies and the widest adoption also in Singapore and throughout the Middle East. There are some tremendous things happening: Qatar and Saudi and the UAE. So I don’t have any one of those to follow, just that everyone seems to be doing a little bit better than Australia, obviously, and perhaps better than the US.

John Ellis:
How did this become your job, if you will? What was the turning point that led you to lifearchitect.ai and this fantastic Memo that you put out every month?

Alan:
Oh, thanks, John. LifeArchitect.ai was created in 2020 off the back of the OpenAI GPT-3 model. That was about two and a half, maybe three years before ChatGPT came out. But it was, for all intents and purposes, GPT-3 was ChatGPT. It’s just that the public weren’t looking at it in those three years. I was looking at it. We deployed about 67 episodes of something called Leta AI, which was a humanoid avatar backed by GPT-3, and I would sit here and ask it questions and it would respond to me in real time. On top of that, I was tracking models like GPT-3, and there weren’t many back then. There was a new model every six months. Now there’s a new model every 20 hours. So it’s changed quite a bit. But yeah, it was transforming from out of consulting and coaching into consulting explicitly and only on AI. And obviously that served me well in the last six years because everyone woke up and took notice.

John Ellis:
And I wanted to ask you a couple other questions. Just one, where are we in robotics? I’m fascinated by robotics, and there was a, I follow it in various publications from MIT Technology Review to various sources, doesn’t matter. It does seem to me that we’re getting closer and closer and closer to a world where we interact with robots on a regular basis. And I don’t obviously, factories are robotic and Amazon warehouses are mostly robotic. But I mean going to a nursing home or going to a hospital or whatever, when we’re increasingly likely to be bumping into, if not dealing with, robotics. Is that accelerating in your view? I mean, is that something that’s gonna be like twice as much next year and then four times as much the next year, et cetera?

Alan:
Yeah, well I want to separate the idea of robotics from ASIMO back in the early 2000s, the Honda robot then, and Amazon’s robots, through to today’s robots. So ASIMO, Amazon are generally pre-programmed, they’ve got some access to the world, they’ve got some sensors, but they’re not responding to the world like large language model-backed robots are today. And they include 1X Neo, that thing’s available right now for I think $200 a month. You can have it in your home, a humanoid with soft fabric around it, so it’s not too scary. The Figure AI, they’ve got the Figure 03 there that just did many, many days on a conveyor belt sorting packages. There’s a few more. There’s over a hundred humanoid robot companies, but what distinguishes them from the pre-programmed ones or the scripted ones is that they respond entirely to their environment. A lot like your GPT chat interface. There’s nothing that it is scripted to respond, it’s instead responding to whatever you ask it. You can say, write me a poem or rewrite the News Items website or actually record an entire News Items podcast, for example. It won’t be as good as you, John, obviously. But these humanoid robots are the same way. They’re responding to whatever you put in the environment, and that’s why you see people pulling on them with hockey sticks or pushing them around. I expect to see these non-pre-programmed, autonomous humanoid robots in factories this year. In fact, it’s already happening at a number of factories. Every BMW X3 you see on the road had Figure AI’s Figure 02 or Figure 03 humanoid robot helping out in the factory to build that BMW X3 vehicle. It’s through a lot of the major factories, including transport, logistics, and certainly manufacturing. I’d obviously like to see it in healthcare. I’d like to see it right here helping me to wipe down my benches or help around the home. I don’t think it’s that far off. Like I said, you can go right now and order a 1X Neo robot and some of it is teleoperated. If it’s around the home and it can’t figure out something, it is actually using someone remotely tying into it. But fully autonomous is moments away. I think Figure AI would tell you that all of their stuff is autonomous. The Tesla humanoid, I’m sure, would be argued to be autonomous by their CEO. Who knows if it is? And then that’s really only the Western models. Then you can add another hundred or so out of China, the XPeng Iron, really huge over there. Even the Boston Dynamics stuff and the clones, the Unitree robots, including the humanoids and the dogs, fully autonomous, respond to anything in the environment. And that’s very new. That’s only within the last year or two that that’s been possible. And I expect them, as you asked, I expect them to be in the home and through the workplace really soon.

John Ellis:
Yeah, I was excited about the Figure 03 robot that could climb a ladder because we have a lot of leaves in our gutters at the end of autumn. So I figure I could go down to Home Depot and get me a Figure 03 ladder climbing robot to go up there and clean it out.

Alan:
Absolutely. Well, they’re a completely general robot, so you would just buy the Figure 03 and you could also ask it to do your dishes or to do the washing or to rearrange the furniture by size or by colour, like it will respond to anything in the environment, and that is incredible. I also wanted to note that the ladder climbing is an example or a demonstration from this week, but the fine motor control has also gone a long way in the last few months. So now they can get to the very specific or very detailed fine motor control movements, including using a screwdriver, putting a screw in, or getting down to opening little tiny boxes, removing the cover off a cell phone, these kinds of very, very fine-grained human-like movements now achievable by humanoid robots.

John Ellis:
So one thing I do in News Items, I have a sort of a file of quantum computing stories, and the bots go out and get me stuff and bring it back. I was startled to read, I think it was an IBM scientist who said, we’ve cracked the code, we will have quantum computing probably around 2030. Does that jibe with your understanding?

Alan:
I’m kind of proud to say that I have no idea. My focus these days is so narrow, it’s just large language models and AI that I don’t even touch on the quantum. I have read the same kinds of articles, but I’m not sure about timelines. I’m not sure if superintelligence is going to leapfrog what our smartest humans are achieving in quantum. And I’m not sure how that’s going to impact inference, how that’s going to impact pinging the models, because if they’re no longer binary, if there’s some sort of state underneath, that is going to add a really interesting element to whether or not we consider it to be human or conscious, I should say.

John Ellis:
Yeah. I’ve taken a lot of your time and I’ll let you go here, but thank you very much. I should, I guess I should, I do this in the political interviews. I say, what is the one question I should have asked you that I didn’t? So what is the one question I didn’t ask you, but should have?

Alan:
I do want to push the point that scaling is really a very large factor. And we’ve touched on this already today. You know, would spending more on data centres be defensible? Are trillion parameter models that exciting? What are the details behind this? It’s been the case since 2017 when Google gave us the Transformer. And they gave us that Transformer, by the way, almost accidentally. They were studying how to use translation between gendered languages. And when it output and it picked the next word, they thought, oh, that’s kind of interesting. How’s it predicting the next word? And that became today’s AI models. But the scaling since 2017 has been exponential. And it’s been the case that the bigger, the better. And we’ve added some other things there. We’ve added more modalities, as you know. Now it can do video. It’s arguably able to do all five senses. But imagine what happens if we just 10x today or 100x today or a thousand x today, and those numbers are achievable. Those numbers are reasonable, realistic, and probably happening right now in a data centre somewhere in the US. So that kind of thing, that kind of thought experiment to me is exciting. If you think that today’s frontier models are incredible, if you think Fable is great, Fable came out or was ready on the 13th of February, 2026. Imagine what that looks like mid-2026, and imagine what that looks like at the end of 2026. I think we’ll stop talking about the rogue cyber stuff and start talking about the kinds of things that make life more exciting for us, including everything we’ve discussed today. So that’s kind of my focus: the scale, the bigger is better, and all of the different compression efficiencies that we’re finding with these models and everything that that means for us as humans.

John Ellis:
Do you think, I have one last question, which is do you think that the US and the Chinese will inevitably have to come to an arrangement of sharing, essentially, so that there’s not an incident, quote unquote?

Alan:
That’s a really, really, really big geopolitical question. And it was addressed across maybe 300 pages of scenario analysis by the AI 2040 guys. Listeners can read that at AI-2040.com. They, in all of their scenarios, I believe, predicted that there would have to be some sort of cooperation, and that seems to be happening. There seems to be discussion between the two major powers, if you like. Back to my artificial superintelligence indicators and checklists. One of the things on that checklist is geographical borders disappear. We get an ASI that’s a trillion times smarter than Einstein. I think it’s going to go, well, all those imaginary lines on the map are great. Let’s get rid of them. We can still talk about the Chinese land mass and the US land mass, but I don’t think we’ll have the kind of separation that we do right now. In the interim, I think it’s important for these two powers to talk to each other and to share. It’s not that China is behind like they might have been a couple of years ago. So in some ways, China might be ahead, and being able to communicate between the two at a sociopolitical, geopolitical context as well as a technical context is going to be really, really important.

John Ellis:
Alan, thank you very much for doing this today. I urge all of our subscribers and all of our listeners to subscribe to Alan’s newsletter. It’s called The Memo. And just go to Google, type in Dr. Alan Thompson, The Memo, and there it is. It’ll pop right up. Subscribe. It’s wildly inexpensive given its value. Alan, thanks a lot for doing this.

Alan:
Thanks so much, John. Good to chat with you.

John Ellis:
Good to chat with you.

Alan D. Thompson is a world expert in artificial intelligence, advising everyone from Apple to the US Government on integrated AI. Throughout Mensa International’s history, both Isaac Asimov and Alan held leadership roles, each exploring the frontier between human and artificial minds. His landmark analysis of post-2020 AI—from his widely-cited Models Table to his regular intelligence briefing The Memo—has shaped how governments and Fortune 500s approach artificial intelligence. With popular tools like the Declaration on AI Consciousness, and the ASI checklist, Alan continues to illuminate humanity’s AI evolution. Technical highlights.