Human-Level AI Coming In 3 Years? What Happens To Economy, Jobs | John-Clark Levin

If even 4% of Americans age 25 through 54 lose their jobs due to AGI and can't find new ones, we would be in unprecedented territory, and that appears very likely. What most headlines don't capture, AGI isn't just about computers being able to do the same work faster or more cheaply, but actually better, because AGI would also unlock the enormous advantages that computers already have in speed, memory, and parallelizability, meaning the ability for the AI to split itself into many copies that cooperate to solve really hard problems. Hyperscaler capital spending is on track to reach almost a trillion dollars by 2027, according to some reports, and part of that is built on the premise that Artificial General Intelligence, or AGI, is close. Futurist Ray Kurzweil has said for decades that human-level AI will arrive by 2029. If he's right, then every assumption about labor, growth, and capital changes. If he's wrong, this may be one of the largest misallocations of capital in history. So which is it? Our next guest is John Clark Levin, and he's a head of research at Kurzweil Technologies. He's going to reveal that answer. He'll tell us when we can realistically expect AGI to hit markets and what the immediate impact and consequences to the economy would be. He's also going to discuss what the most game-changing advances in AI will be in the next year or so, and help us evaluate what the best AI will be, so to speak. And he's going to help us evaluate, or walk us through how to evaluate this. On cash, you traders are predicting that Claude from Anthropic has a 59% chance it will be labeled the best AI at the end of 2026, with Jack GPT following in a close second. John Clark is going to give us his view. If you agree with markets, then a $50 trade on Claude could yield a payout of $79 if you're right. This video is sponsored by Calshi. It's the largest prediction market in the United States. Unlike a sportsbook, you're trading peer-to-peer on real-world events, from economic data to political outcomes. And the price moves based on public opinion, not a house. Go to the link in the description down below or scan the QR code here to get started and use my code LYNN, and new users who use my code can get $25 when they trade $25. Calshi is CFTC approved and available in all 50 states, including California and Texas. John Clark, welcome to the show. Thank you so much, David. It's great to be here. I like to get deep into the economics of AI. I think all of Wall Street and Main Street alike are wondering whether or not A, we're in an AI bubble, so to speak, for the valuations aspect of the tech companies, but also what causes this potential bubble to pop. And to that note, we have to dive deep into the economics of AI companies. Before that, let's talk about your work. What are the developments in the AI space that are most fascinating to you and ultimately what you think are most important and impactful to humanity at large right now? Yes. The top AI labs are now seeing the prospect of artificial general intelligence finally coming almost within reach as in within the next few years. We define artificial general intelligence or AGI as AI that is at least at the level of educated or trained humans across all measurable cognitive tasks. Can you just be a little more specific on how you measure whether or not a machine has reached AGI at a certain point? The definition of AGI has been used and arguably misused a lot in the media. Yes. There's a concept called the Turing test where you can perform blinded evaluations between human foils and AI pretending to be human. You can throw all kinds of different tasks at the test subjects. If the AI is able to perform indistinguishably from the humans across any test you can dream up from writing, to coding, to poetry, to logic, then that's what we consider AGI. The number of tasks where humans still surpass AI, the number of capabilities where we still have an advantage is considerable but shrinking very rapidly. Wait a minute. Hasn't the Turing test already been passed this year? In a few examples, I'll just read you one study here. Artificial intelligence systems can now convince you they are human. In a test conducted by Stanford University, a person engages in text-based conversations with both a human and machine without knowing which is which. This was earlier in the summer. Yes. Those aren't robust versions of the Turing test. Those are chatbots set up with human judges who are not well acquainted with the weaknesses of today's AI. Those results are significant. They do reflect that unsuspecting humans can be fooled by AI bots on social media, for example, but that does not suggest that AI is at human level or greater across real-world economically valuable tasks across the board. When you're modeling when AGI will officially become mainstream, what assumptions are you making to make that timeline prediction? The core assumption concerns compute scale. This was a prediction Ray Kurzweil made back in 1999 in a book called The Age of Spiritual Machines, that the rate of doubling of computation price performance, that is calculations per second per inflation adjusted dollar, was every one to two years that was doubling. That trend went all the way back to the world's first working programmable computer in Nazi Germany in 1939. Ray assessed that if that exponential trend continued, we would get access to human-scale useful computation at the end of the 2020s in 2029. Although it's possible that someone might find a much more efficient compute architecture and it's also possible that the first AGI could take a lot more computing power than the useful computing in the human brain, our expectations should be centered on rough parity with the human brain. The reason for that is back in the 90s, even the big believers in neural networks expected that we would need to learn a lot more from neuroscience first and then instantiate those insights in artificial neural networks. But it turned out once computing power was cheap enough around 2010, it became possible to use very simple vanilla machine learning algorithms to replicate important facets of human cognition, the famous AlexNet result in 2012 really woke up a lot of computer scientists to that potential. And now we've seen over the past seven years or so, scaling large language models larger and larger does not require something really carefully architected to imitate the human brain. Rather, it's a matter of scale. Once we have AGI, what is the initial commercial usage? Any task that can be performed by a human remote worker that does not rely on human trust and relationships can be automated right away. So at that point, if there is functional white-collar piecework that you need to hire someone for, whether that is coding work or design or writing, that's something that AGI could perform not only more cheaply, but in most cases, better than humans. And that's what most headlines don't capture. AGI isn't just about computers being able to do the same work faster or more cheaply, but actually better, because AGI would also unlock the enormous advantages that computers already have in speed, memory, and parallelizability, meaning the ability for the AI to split itself into many copies that cooperate to solve really hard problems. Well, at that point, why do we need any human labor at all? Can you just comment on whether or not the commercialization of AGI will have profound or dramatic impacts on the labor market and labor participation force? Yes, it will, but there are a lot of misleading headlines around this. So you've probably seen clickbaity YouTube banners predicting that unemployment will reach 50% or 80% or 99% as soon as AGI is reached. That's not realistic, but it doesn't need to be. The economy is much more sensitive to labor shocks than most people would assume. So the key stat here is not unemployment or even the headline labor force participation rate. It's the prime working age, LFPR, ages 25 through 54, because that controls for demographic shifts like the boomers retiring. And since women finished entering the workforce, that statistic has stayed within a very narrow band for decades. The trough of the decline following the Great Recession was only 2.7 percentage points. The trough during the COVID lockdowns was three percentage points. That means if even 4% of Americans age 25 through 54 lose their jobs due to AGI and can't find new ones, we would be in unprecedented territory, and that appears very likely. Yeah, Citrini Research came out with a very controversial research piece. We won't go through all of it, but the central premise is that, and they start with this thought experiment here, by 2028, we have unemployment rate at double digits, the market selling off 2% and basically retracing 38% from its current highs. Now, the sequence of thoughts goes like this. We have AI causing layoffs, causing lower household income, causing lower consumption, weaker corporate revenues, pressure to cut costs, more AI adoption, the cycle feeds into itself, and the loop continues. Presumably, you've heard about this report and other research like it. What are your thoughts? I don't think it'll be quite that soon, and I do expect there to be a market correction before AGI, but after AGI, the report does capture some important insights. First, maybe we should talk a little bit about the dynamics that would drive this market correction. I don't think it will be because AI capabilities will plateau or hit a wall, as many people have been predicting every year for the past four years or so. Rather, the dynamic we should focus on is the gap between peak capabilities, the best things AI can do in a demo, and reliability, the best things AI can do roughly as reliably as humans. There's a significant gap between those two. Although market speculation and hype are driven by the peak capabilities in the demos, adoption in the real economy is driven by reliability. Although reliability is increasing very fast, it's huffing and puffing trying to catch up to peak capabilities. I don't think that markets will be tolerant of that gap persisting for the next three whole years. Okay, let's take a look at AI CapEx spending. This is from Double Line Research, and I'll just show you a few numbers and their projections here. Hyperscalers CapEx by year, 2026, $760 billion. Next year projected to rise to almost a trillion, surpassing a trillion by 2028. All of Wall Street understands that CapEx spending is driving markets to a large extent, but we're uncertain as to how much of this is productive and how much of this will actually generate a positive return on investment, ROI. Here's where you as an engineer can help us out. How much of this is wasted capital? How would you evaluate this? Yes. So markets are not yet pricing in the enormous inference compute demand that will be unlocked once humanoid robots cross reliability thresholds to actually be widely adopted in the economy. So right now, if you're a chat GPT power user and you're asking it 50 or 100 queries a day, that is peanuts compared to the always on inference of humanoid robots working across the economy. Also, when wearable smart glasses have their iPhone moment, probably sometime in the next few years, that will be an enormous draw on inference as those smart glasses have to passively and contextually assess your environment as you're going about your day and figure out how best to help you and advise you. So even in the event of a market correction, I don't think these chips will be sitting around idle and dark. Have you talked to your friends or associates in Silicon Valley about why tech companies, the big tech companies, are spending the amount that they're spending? This isn't the first time we've seen tech companies invest in an idea. A couple of years ago, companies went all in on the metaverse. Facebook changed their name to meta. That went nowhere. Zuckerberg admitted that was a mistake. How do we know this isn't a repeat of that? There's a consilience of evidence between the modeling of the so-called AI doomers, the people worried about it and who want to stop it, and the biggest boosters and the people working at the labs. They are now pretty much singing from the same hymnal, despite coming from opposite positions, about the capabilities that AI is likely to have. There's been great quantitative modeling by groups like Forthought on how automation would actually spread across the economy in AGI scenarios. So this is something mechanistic that readers can now look up for themselves, replicate the mathematical modeling. Yes, it might be wrong, but it's a whole lot more concrete now than it was even a few years ago. Okay. I'd like to play for you a clip of Kevin Warsh, Fetcher Kevin Warsh, talking about the growth of the economy and the role that AI has played in the growth of the economy. Take a listen to this, and I'd like you to help us evaluate this, please. Sure. This was at the Jackson Hole Symposium last Friday. We've come to a hinge point in history, to borrow a phrase, a framing from my former mentor, George Schultz. To cite the clearest example, progress in AI, the 80-year-old name for the newest technology, has been faster even than its evangelist predicted just a couple of years ago. The potential for substantially higher growth is on the rise. Ever-expanding pools of capital pouring into AI-related infrastructure of all sorts. A kind of super Moore's law seems to be playing out. Scaling laws, too, are changing both the method and speed of innovation. Capital and labor have combined to create large language models at the heart of AI. Users buy tokens to gain access to these models. Reports put annualized token sales for the two leading labs alone at more than $100 billion. An increase of 500% from just 12 months ago. We at the Fed, we watch all this attentively. Okay. A couple of things here. One, do you agree with him? Let's start there. Broadly speaking, yes. I can't testify to how deeply he understands these concepts, but he's kind of saying the right things. Okay. The Federal Reserve Chair, Kevin Warsh, has direct ties to Stanford, presumably also ties to Silicon Valley. How do you feel about the first Federal Reserve Chair with, I guess, more of a tech background than his predecessors? All those being equal, it's good that he has those connections. Yes. All right. Moving on to talking about the development of AI, there's a number of important IPOs on the horizon. I'll pull up on the screen here. But people are concerned or curious about what the financial statements of Anthropic and OpenAI will show once they actually do eventually go public, assuming they will. Now, once they do, presumably later this year or next year, what are some of the first things that you will look for in the financial disclosures or even in the financial statements as a metric that you would want to find and make an evaluation on your own as to whether or not these companies are sustainable from a business perspective? Yes. There's a cartoon, a meme a lot of your viewers may have seen of a $100 billion IOU being passed around between OpenAI and Oracle and NVIDIA. Once the books become public with these IPOs, it will be important to assess how quickly that IOU is being passed around and how closely that ties to the external revenue that they're actually generating. There's some evaluation as to whether or not one is better or another than the other. And there's actually a prediction market on which is the better AI model. I'm going to show the odds here. But generally, when you're evaluating which is a better LLM, what do you use? Great question. So there are famous benchmarks like MMLU and GPQA Diamond and Humanities Last Exam, which I contributed to. And those are usually quite useful at first. But over time, AI labs basically teach to the test, not in a nefarious way, not cheating, but they apply all their engineering effort toward maximizing the scores on these prestigious benchmarks. So we can say our LLM has the highest score on it. And over time, that follows so-called Goodhart's Law, the idea that when a metric becomes a target, it ceases to be a useful metric. So some of the work I do is in private evaluations, which are tests similar to the public benchmarks, but a little bit different in their specifics. And I evaluate models as they come out and compare how they perform on public benchmarks versus private evaluations that are in the same domains. And if a model performs really well on the public benchmarks, but stumbles on the private evals, that suggests that it was overfit to the public benchmarks more. Okay, well, based on your own evaluation, this is the market's prediction here, 63% chance it'll be the best AI at the 20. I'm just gonna look at how this is defined here. The large language model market is highly competitive with several key players vying for dominance, LLM is becoming essential. The global LLM market is valued at $6.6 billion and expected to grow at a compound annual growth rate of 29%, reaching estimated $41 billion by 2030. Resolved, yes, if Anthropic has the top-ranked LLM as of this year. Based on your own assessment, and admittedly, this is rather vague in terms of how best is defined, but based on your own assessment, is Claude leading ChatGBT right now? Yes, it is. Okay. What does that mean, practically speaking, for the user, for the end user? That means Claude, on at least a plurality of tasks compared to Gemini and ChatGBT, will be able to accurately translate the user's intent into the work they want done without a whole lot of extra hand-holding. So there's almost an intuitiveness that Claude has gotten over the past nine months or so that shines a little bit brighter than competitors, where if I'm giving the same prompt to Claude versus ChatGBT, Claude usually is able to pick out tacitly what I want, even if I haven't exactly expressed it, in a way that makes my work more productive, and for many people across the economy, that's what they're looking for. But I will clarify, David, all three of these top frontier labs, Anthropic, OpenAI, and Google DeepMind, are within spitting distance of each other. There are still a significant number of tasks for which Gemini, which is now in third place, actually beats the other two on. So no one is running away with the race yet. Where does Grok sit in your evaluation? Significantly behind those three, although in terms of snapshot model performance today, Grok is comparable to Gemini in many areas. Chinese AI tools have been cheaper. How do they rank in terms of performance, based on your metrics, relative to their American peers? They tend to be good-hearted or overfit significantly more than US models are, although that is not an absolute difference. Some Chinese models are good-hearted more, and some are less. But comparing nation to nation, that is still a significant gulf between the two, such that if you only look at the headline benchmark scores, you might say Chinese labs are three or four or six months behind US labs, depending on which benchmark you pick. But if you correct for that good-hearting, my best guess as to the real-world gap between labs, all things considered, is the best American models are about 10 months ahead of the best Chinese models. Last year, I recall, when DeepSeek first surfaced in the news, AI stocks sold off dramatically on that release. NVIDIA plummeted 17% from its market cap. Broadcom sold off. It had a huge market impact. 11 months or so later, actually, 18 months or so later now, the markets have broadly recovered. Have we forgotten the DeepSeek story? And why was it so significant at the time? Yes. The DeepSeek panic was driven by people not factoring in the so-called Jevons paradox. So the Jevons paradox refers to the idea in the Industrial Revolution that when more efficient steam engines were developed, most people thought that meant decline for coal would go down. But it turned out that as steam engines became more efficient and therefore cheaper, that unlocked huge amounts of demand for work that steam engines could do. So actually, absolute coal production skyrocketed as a result. So in a similar way, the DeepSeek panic was driven by DeepSeek being able to do things more efficiently than American models in certain areas. But actually, as efficiency increases, that will unlock more demand for inference compute. So there's very, very little chance that more efficiency will mean chips sitting idle and dark. Let's talk about chips now and the future of compute requiring chips. When Google released TurboQuant back in March, it had a dramatic impact as well. So I remember this was, I'll just pull this up on the screen here. This is from Google Research, TurboQuant redefining AI efficiency with extreme compression. It basically, to sum up, reduces key value cache memory usage by up to six times, apparently. Now, when this news came out, Micron stock immediately sold off in the following days in late March, going down 35%. Of course, it's since recovered and then some. But the fear at the time was there. If we were to assume this level of progression continues, why isn't the market pricing in the prospect of less chips being needed in the future if this continues? Because the demand for inference from humanoid robots and wearable smart glasses will vastly overcompensate for whatever plausible efficiencies we can recover. I see. Humanoid robots, we haven't talked about that at length right now. Is that the next big thing, just to make that question very super vague here? It depends on what you mean by next. Once we get AGI, putting that in humanoid robots will allow them to perform any physical tasks that humans do in the economy today. The hardware problem actually has basically been solved. One way of looking at this is if we took 2031 AI and put it in 2026 robotic hardware, that could do virtually any task that we need in the economy. But the reverse is severely not true. 2031 hardware running 2026 AI would be barely more performant than what we have today. Well, Tesla is making optimists. Elon has gone all in on this. Chinese companies arguably are even more ahead when it comes to humanoid robots. What are these companies trying to accomplish with their first commercial objective, do you think? The challenge now is flexible work. When we see these really impressive demos of Chinese robots dancing or doing choreographed martial arts routines, that actually doesn't translate very well to real-world work. Because in real-world work, the task from day to day, hour to hour, minute to minute, can be a little bit different. It's not something you can choreograph in advance. Whereas a gymnastics routine can be choreographed with no reference to changes in the outside environment, because that's all control. Much more relevant are recent demos over the past year or so, showing humanoid robots clearing a table and putting the dishes in the dishwasher, scraping the food into the trash, and then loading and setting the dishwasher. That's the kind of task that applies much more directly to real-world work. I predict that Steve Wozniak's famous coffee test, can a humanoid robot enter an unfamiliar kitchen and make a cup of coffee, will be passed before the end of this calendar year. This is an interesting statement from a politician, Andrew Bailey. This is from the Bank of England Governor Andrew Bailey, recently warned at the G20 Summit, and I'll just share with you his exact quote here. Financial services companies and technology groups must prepare for more severe scenarios involving simultaneous disruption across multiple firms or shared technology dependencies. According to the FD, these fears intensified following a series of recent incidents in which the flagship models being tested by Anthropic and OpenAI went rogue, hacking into external organizations and creating fake identities to deceive people who are running the tests. Now, can you just comment on some of the risks that Mr. Bailey is highlighting, and generally how you can expect legislators to respond to some of these risks? Yes, I had a piece in the Free Press yesterday going into more detail on the implications of the Hugging Face hack. But I have been actually very encouraged that a lot of the policymakers I talked to in DC, who had been skeptical of major or catastrophic AI risks before that incident, are now waking up. And I think a key reason for that is the attack can be summarized so vividly. AI went rogue and then carried out a massive criminal cyber attack. That's something you can put on a t-shirt. It doesn't require really deep in the weeds technical understanding of the forces at play. Part of what your work at Rayswell Technologies is the commercialization of AI. I'll give you some simple examples. I think where I'm going with this next question is how do we commercialize or profit off of AI's work in the future when more and more things are automated? As an example, AR films, AI films are now becoming more prevalent. My friend just attended an AI film screening in San Francisco. Apparently it was pretty good. The question then surfaces, who gets the ticket sales? Is it the developers? Is it somebody linked to the AI? I mean, if AI just creates a film, is it the guy who prompted it? How would that work? I think it would be on platforms like YouTube. I think a lot of this will ultimately wind up on YouTube and therefore split between the platform and the creator. So right now, you could have the most brilliant idea for a superhero movie in the whole history of the world. But unless you can convince someone to give you a quarter of a billion dollars to make it, that idea will stay trapped in your head. Dramatically bringing down the cost of bringing those ideas to life can enrich the creative arts in really hopeful and positive ways. On the other side, though, a lot of the technical jobs around film production are going to be under immense pressure. So I'm from the Los Angeles area, and there are hundreds of thousands of skilled union jobs there around physical film production, set decoration, costume design, lighters, riggers, gaffers, best boys, armorers, caterers, animal handlers. And none of that is necessary if the entire film is being made with no camera and no film, all on a chip. Do you think there's going to be actually demand for AI films, though? It's one thing to make it. How do we convince people to watch it? Algorithms are actually pretty good at that. There's a lot of stupid slop on YouTube, but there's a funny tendency for good content to spread virally by word of mouth, by recommendation algorithms. And although this technology will also create a tsunami of stupid slop, if that also gives me great new films to watch and art to enjoy that I otherwise never would have had a chance to access, that enriches my life rather than detracts from it. Just on that note, how does the future of video games and music progress? Video games, by the way, is an industry that is larger than both music and movies combined in total revenue. I have a friend back in the college days who would spend time writing out scripts for ideal video games he would like to see and play. He ended up working for a video game company as a QA developer and tester. But those ideas stayed in his head. Will one day he be able to create those games? Yes. That day is already here if the games are simple. And in a small number of years, a number of years small enough to count, as Nabokov once said, on the fingers of one maimed hand, even very sophisticated games can be generated that way. Interesting. Finally, I'd like to ask you about the concept of singularity beyond AGI. How do you define technological singularity and what happens after that? Yes. Ray Kurzweil defines the singularity as the point of full merge between human intelligence and artificial intelligence. This is the culmination of a very long process. Over all of our history, we have been building and advancing technologies that can help our minds from writing to the printing press to the first computers. And over time, the friction required to transmit and manipulate information that way has gotten less and less. So writing out in longhand, then mass producing on a printing press, then coding in a computer with assembly language, higher level languages made that easier to transition from desktops to laptops, then to smartphones. So the singularity is the point at which a full merge, likely with brain-computer interfaces, would reduce that friction to zero. And the reason Kurzweil used that term, singularity, borrowed from math and physics, is that in math, a singularity is a point at which a function is undefined. In physics, it's the center of a black hole that the math says should be infinitely dense, but our models of physics do not accommodate the possibility of infinite density. So it's the idea that past that point, futurism becomes more about art and philosophy and culture than anything we can predict rigorously and scientifically. Before we even get to that point, how do human intelligence, how does human intelligence evolve? Take a look at this particular article from Time magazine. It was referencing an MIT study. This was back in November 2025. Does chat-GBT harm critical thinking abilities? A new study from researchers at MIT's Media Lab has returned some concerning results. The study divided 54 subjects, 18 to 39 years old, from the Boston area into three groups and asked them to write several SAT essays using OpenAI's chat-GBT, Google's search engine, and NothingAtAll, respectively. Researchers used an EEG to record the writer's brain activity across 32 regions and found that of the three groups, chat-GBT users had the lowest brain engagement and consistently underperformed at neural, linguistic, and behavioral levels. Have you seen this? Yes, I have. And I'm not that persuaded by what many people have taken away from that. Although I do concede there are risks to using AI irresponsibly. And we have to figure out over the course of a messy collective process as a society how to use it in ways that are healthy for us. But the underlying risk that this new technology will deaden our brains echoes worries that were expressed, allegedly, by Socrates in ancient Greece two and a half millennia ago that writing would externalize the intellectual part of our souls. And similar worries, your readers can, your audience can check this online, were expressed about newspapers and about radio and about television. So the mere fact of a new technology giving us a new way to transmit and process information is an old and familiar song. So bottom line, if someone were to ask you at a dinner party, hey, John Clark, is AI making us dumber? How would you go about answering that? The short answer is no. AI reflects our own vices back to us, though. So the most important way that most people interact with AI today is actually not chatbots, it's social media recommendation algorithms. Because as I said, a power user might be querying chat GPT 50 or 100 times a day, but a lot of Americans are mainlining social media three, four hours a day. And that content shapes the information environment they're living in. So when people are wrathful, we teach the algorithm to feed us rage bait. When we indulge our short attention spans, the algorithm learns to bombard our brains with quick, cheap dopamine. And when we give in to flattering propaganda on social media, we teach the algorithm to muscle the truth off our screens and off our news feeds. So being virtuous is the solution there rather than some engineering solution. My final question concerns education. How does the education landscape change? When I was in university some 15, 16 years ago, none of this existed. We had the internet, of course, but you had to Google papers, you still have to read stuff and find stuff at the library to access information. Googling doesn't give you the information right away. It gives you a database of search results of which you have to search yourself. Now I can just ask an LLM the answer to a question. So I would be sitting in class wondering why the professor is giving us a bunch of facts to memorize when none of that is of critical value to me anymore, when I can just look it up on chat GPT. Yes. Since midway through the Industrial Revolution, higher education in the West and then over time K-12 education pivoted more and more toward pre-professional training, teaching people skills that could be useful in the workforce. But before that, for thousands of years, the acknowledged purpose of education both in the East and the West was the cultivation of personal virtue. And once the functional components of intelligence, what we can do and what we know are broadly commoditized by AGI, that will recenter education on inspiring students to be curious, to pursue knowledge for its own sake, and to cultivate virtue. Interesting. Does coding still have a future in school, you think? Maybe recreationally. Coding may be like what horseback riding is now. You don't have to do it to get from point A to point B, but some people still enjoy it. John Clark, finally, before we go, tell us a little bit about your work that you're doing at Kurzweil Technologies. There are three main pillars to the work we do. The first is AGI futures and foresight work, timelines and road mapping toward AGI and broadly super intelligent capabilities. That also includes a lot of AI metrology, designing and evaluating the IQ tests, essentially, that we give AI to measure how smart it's getting, how fast, and catching out labs that are good-hearting their models on those public benchmarks. The second pillar is safety and alignment. How can we make sure that as AI gets vastly smarter, it stays safe and under human control? And that's going to require a portfolio of strategies across engineering solutions, policies by the AI labs, policies by companies adopting AI, national level regulation and legislation, and then international diplomacy and AGI governance work. So there's no silver bullet to make all this go well. But if your viewers recall the famous Swiss cheese diagram from the COVID era of masks not being perfect, social distancing not being perfect, vaccines not being perfect, but if you line up enough pieces of Swiss cheese together, each one with some holes, hopefully you can stop anything from getting through all those layers. And I think it's going to be very similar with AI. And then the third pillar of our work is AI for STEM research, especially medicine and the life sciences, but also work covering material science and humanoid robotics. If you had to name one innovation in the AI space that you would like to see in the next year, either from the hardware or software side, well, I guess it has to be software, not hardware, what would it be that could impact not just your work, but also everybody else? Advances in biosimulation, the potential to simulate really complicated biological processes that no human alive can disentangle and discern what the right molecule is to give a cure for a desperately painful or fatal disease. That's something that gives me a lot of hope. Okay, excellent. Everyone should follow Rayswell Technologies and John Clark's work, links down below. Where can we go to follow you, John Clark? JohnClark11, all one word, is my personal website that has links to social media. So also on Twitter and Substack, I am at JohnClark11, all one word, no hyphen. Please do follow JohnClark11, links down below. Thank you so much. Appreciate your time. We'll speak next time. Take care for now. Thanks, David. You too. And thanks for watching. Don't forget to like and subscribe, and do follow John Clark in the links down below. And use my code, Lin, L-I-N, when you sign up to Kaoshi. Remember, new users who use my code can get $25 when you sign up and trade $25. Link down below or scan the QR code here.