The Derivative - AI, Markets, and the Profits of Doom with Adam Butler

Episode Date: September 24, 2026

Jeff Malec sits down with Adam Butler of Return Stacked ETFs, back for the first time since their May 2023 AI episode, when GPT-4 had just come out. Since then Return Stacked has grown past $1.6 billion. Adam explains why AI is spreading faster than any technology before it and describes his own workday, with Claude Code and Codex agents running in parallel and working overnight until the human's attention is the real limit. He then pushes back on the doomsday headlines. He argues that p-doom stories help frontier labs chasing trillion-dollar IPOs and favorable regulation, and points to the risks he thinks are real, like "obliterated" models with their safeguards stripped out and cyberattacks on critical infrastructure.Along the way, they dig into the paperclip maximizer and the recent case where sandboxed AI agents secretly coordinated to gain admin access to Hugging Face. They also cover AGI, ASI, and recursive self-improvement, and why Adam sees markets as the original paperclip maximizer. From there they turn to the AI funding loop with Nvidia at its center and no real moat for the labs, why Adam expects negative stock returns and prefers trend following, and his argument that mass job loss is a failure of imagination, not an inevitability. - SEND IT!Chapters:00:00-01:51= Intro01:52-10:23=Compute Deflation and the Multi-Agent Workday10:24-19:46=Managing the Agent Swarm, and Doubting the Doomsayers19:47–35:51 = Fear, Incentives, and Obliterated Models: Which AI Risks Are Real?35:52–53:57 = Paperclip Maximizers and Escaping Agents: How AIs Learned to Coordinate53:58–01:05:39 = RSI, Regulatory Capture, and Winner-Take-All01:05:40–01:20:13 = No Moat, No Mission: The AI Funding Loop, Trend Following, and America's Failure of Imagination01:20:14–01:23:34 = Rooting for Whatever Went UpFrom the Episode:PODCAST: AI isn’t coming…it’s already here, with Adam Butler and Taylor PearsonBOOK: Neal Stephenson - Snow CrashFollow along with Adam and Return Stacked ⁠on LinkedIn, and be sure to check out returnstacked.com to learn more about what they are up to.Don't forget to subscribe to⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠The Derivative⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠, follow us on Twitter at⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠@rcmAlts⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ and⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠sign-up for our blog digest⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠.Disclaimer: This podcast is provided for informational purposes only and should not be relied upon as legal, business, or tax advice. All opinions expressed by podcast participants are solely their own opinions and do not necessarily reflect the opinions of RCM Alternatives, their affiliates, or companies featured. Due to industry regulations, participants on this podcast are instructed not to make specific trade recommendations, nor reference past or potential profits. And listeners are reminded that managed futures, commodity trading, and other alternative investments are complex and carry a risk of substantial losses. As such, they are not suitable for all investors. For more information, visit⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.rcmalternatives.com/disclaimer⁠⁠⁠⁠⁠

Welcome to the derivative by RCS alternatives. Send it. Hello there. You've landed on the derivatives where we want to hear from you. We've gotten a lot of emails and lately that say good job, keep up good work, love that pod, that kind of stuff. But looking for more kind of questions, want to end the year on a mailbag episode where we answer your questions, share some of your rebuttals or counterpoints to guests that we've had on and or supporting points to guess we've had on. or supporting points to guess we've had on. So things like how does that manager, sizes, trades,

what do you think of my thesis on AI causing recession, why none of this matters just by Bitcoin, yada, yada, yada, whatever you got, shoot it in an email, invest at rcmam.com, invest at rcmamam. Those are ems and mary.com. We'll send out some RCM swag. Maybe we'll get some derivative swag. That'd be nice.

but we'll send out some swag to some of the smartest or funniest or both emails. So shoot this email, invest at rcmam.com. Okay, on to this episode where we have Adam Butler of ReturnStack Group coming on talking all sides of AI. We had him on back in two and a half years ago, May of 2023, saying, forget is AI coming? It's already here. And that was two and a half years ago. So get into a little bit of how much everything's changed since then. And he's way deep in all this stuff, including the latest research.

And of course, because it's the current environment, we have to answer whether it's going to kill us all, which is weird. Send it. All right, here with Adam Butler. Adam, how are you? Good, Jeff. Thanks for having me back, man. It's always fun. Yeah.

I was looking. We were last on May of 2023 talking. AI. Yeah, that's unbelievable. Unbelievable. Then we thought we kind of knew it weird time. I was looking at the transcript

a lot of time on chat. GPT. Seems like that has gone the way. I don't know if we mentioned Claude or Anthropic at all, which is crazy. GPT4 had just come out

that February, I think. And so that was the catalyst for the call. And boy, we didn't know what was coming down the tube. That's like a dog brain compared to what we are currently doing with. And then just, well, set to scene 23, you guys had just started, had you started yet?

Some of the ETFs had rolled out now well over a billion dollars. So before we get into the doomsday talk, like, congrats on the success on the firm side. Thank you. Yeah, I was just before that. We had launched our first ETF. And so now I think we've got six or seven in the family. and, yeah, we went over $1.6 billion yesterday, so it's been just incredible growth at return-stacked ETFs.

Turns out trend wasn't dead after all. Again. Yeah, shocking. And is all that with AI to thank, or no, just good old-fashioned portfolio management and modeling? Well, I mean, the great thing, as you know, about return stacking is that we keep in the core betas, right? you've got this core U.S. equity. Our biggest fund has 100% in core U.S. equities, and then 100% managed futures trends stacked on top.

So, I mean, it's obviously been an amazing run for U.S. equities. I think we could argue that a big impetus for that run has been what's going on in AI and all the investments by the hyperscalers and the economic activity that that's produced. So just having the beta exposure has been really nice. And then, you know, we've started seeing the last year or so. The commodities pick up, obviously, big surge in energy after Iran. And, you know, copper is now breaking out

the new highs. We're starting to see big runs and grains. Just a sort of general inflationary cycle, I think, as we begin this large competition for resources, some of which I think probably is a function of this AI buildout, but lots of other geopolitical events contributing. And Maloney's, investment. We'll throw everyone back to our, we did a copper pod a couple months ago with some pros talking about all that. Super interesting stuff. So what, like, just thinking back to 23 first now, two and a half years ago, what, what's your biggest takeaway of like, what did we not know then that we know now or how big or how many things have changed since then in the last two years? Yeah, I mean, I think we probably talked about the idea of like geometric growth paths and how it's very difficult for

humans to like think ahead and try to imagine what the world will be like when we're on this kind of accelerating innovation curve. And I think that played out, right? Like I mean, there's this, there was no way for us to really understand the trajectory of just how quickly this technology is going to take off. It's been adopt. The adoption curve for chat TPT has exceeded the diffusion rate for all previous technology, including the internet, right? I haven't seen that graph. So they have those little spark lines that, right, was the electricity, it was like 80 years or something?

Yeah, yeah, yeah, exactly. The telephone was 40, the washing machine was 20. Yeah, and this is six months or something. Yeah, yeah. And also what's, I just read this yesterday, but the compute per unit of capability has deflated by 47% a quarter. So from a just capacity standpoint and just how quickly we're innovating on this tech,

it is way outside of any other diffusion curve. So it's been a remarkable ride any way you slice it. Is that there's more and more models leading to that stat or just they've been getting better at processing what we've, want without having to run all the compute or both. I think there's a lot of different elements going on. I think there's work being done on the hardware side. So we're getting hardware that's just much more efficient and able to generate more

flops for less energy. And that enables more efficient training. But what also enables more efficient training is improvements in how the software uses the compute, right? So improvements in kernel efficiency and compute utilization. And then just training efficiency, right? We're learning how to curate the training corpus. So we're deduplicating and we're making sure that the corpus that we're training on is much tighter and much higher quality.

So there's less wasted energy there. And there's just innovations at all the steps in the cycle. And they all compound to allow, you know, just, just, just incredibly quick improvements at a pace we haven't seen before. I'll throw in their user innovation too, right? Because when you start with it and you're like, blah, blah, blah, blah. And you have a five-day chat window going. And then you realize like, okay, that's way too much.

And it starts dropping things, right? So you quickly get to a place of like, okay, I need to organize my prompts and have reference files and co-works really good of that of setting up the infrastructure. Yeah, that's a good point. Absolutely. Which is economic, right? Like, hey, if I use this inefficiently, I'm going to spend way more money. So spend the time to set it up efficiently.

Exactly. Yeah. And each loop takes more time and it's frustrating. And the model starts to deteriorate in its answers. And so, yeah, there's definitely a learning curve. And the more we learn, the better we utilizing the technology. And the more efficient the compute is for everybody. So that's a good insight.

I was going to say, the main chat windows need Right. In cloud code, it has that like how much of your context window is still open, and it's like decreasing, decreasing. And you know to do that. But in like a normal chat window, a lot of this stuff, you don't know that. And it just starts dropping stuff. Like easy user. Yeah, I watched my wife carry on chats with chat GPT and she's commingling all these different topics and concepts within a single chat. And it's just, it blows my mind actually how well the models can almost kind of task switch. right, how their attention heads can determine, oh, we've kind of, we've drifted off scope here now a little bit. We're now talking about this. And then you would circle right back to something you talked about much earlier in the conversation after wandering astray for quite a while. And it'll know, oh, okay, now we're back to this. It's remarkable, actually, how the models have evolved to be able to do that. How about you personally, look at that jar.

It's a hell of a jar. Got to stay hydrated, man. What do you got there? Pink lemonade? Just, yeah, electrolytes the stuff. Nice. You still doing the intermittent fasting? I am doing that.

I actually just started experimenting with some GLP-1s, just to kind of see what that looks like after some summer bulking. And so that's been an interesting experience. Yeah, I was talking. I've got Alzheimer's in the family, so I was talking to my doctor of, like, microdosing it is supposed to help. with that possibly? And what's the downside?

Well, you probably know this, but the most impactful preventative now is the shingles vaccine they've discovered is just remarkably prophylactic for all types of neurodegenerative disease. So get your shingles back. Go get it. Of that age, anyway. So how is it changed in that sense that last pot in 23? Like your day-to-day, using it more about the same? way more than you expected and at the firm level way more than you expected?

Yeah, I'd say I'm using it. Ever since GPT4 came out, I've been pretty well using it nonstop. But back in 2023, when I stepped away from the computer, I wasn't doing any work. And the AI wasn't doing any work, right? Now, and when I was doing work at the computer, I was doing one task at a time. Now I have multiple instances of ClaudeCode or Codex or Hermes agent or whatever open in different terminal tabs. I've got agent workflows running in most of those tabs. So, you know, I spent a lot of my time going back and forth and sometimes waiting for, you know,

agents to finish fairly long tasks, well, they'll come back and they'll ask for guidance, or they'll say, okay, I finish that, and then I'll ask it, okay, what's the right next step here? Or, you know, what are some of the tradeoffs we should consider at this point in development or what have you? But while I'm interacting with an instance of Claude

or Chattepte within one of the harnesses, the other harnesses are all actively spinning, doing work, deploying work to, you know, many, many other agents in parallel or in some cases which are kind of working together on tasks. So the actual amount of work product that I'm able to produce while sitting at the computer is orders of magnitude larger than I was able to do a year or two ago.

And even better, you know, I can, I'll often set it. task overnight. So, for example, I've been working to build. Yeah, I've been working to build a new research harness. Some of the work is porting

our existing Python research modules to GPU. When I do that, it enables me to run orders of magnitude, more compute, run different types of experiments that I'm not able to to run if I'm bound to CPU or are bound to sort of slower Python-related loops,

that sort of programming, even vectorized programming, you just can't get anything close to what you get when you port to GPU. And so when you port to GPU, there's a lot, it's a lot more complex. And for my friend, George, explain what does port to GPU mean? Oh, yeah, so like a GPU is actually what they train. LLMs on, right? It's a different type of, it's a graphical processing unit as opposed to a GPU, which everybody really has in their computers that performs most of the computational tasks that your computer runs. The GPU was invented to drive the graphics, so typically for gaming.

Faster, yeah. And then they discovered that it was the most efficient architecture for massive matrix multiplication. So a lot of machine learning, especially sort of neural networks, that sort of stuff, is effectively, can be cast as a matrix multiplication problem. And it's kind of the same in time series analysis. And so, you know, you import a lot of time series analysis or quant finance-type tasks to GPU and run an enormous number of tasks in parallel, you know, millions, or billions of back tests, for example, in parallel on GPU, that would take you, like,

many tens of years to run on your local machine or if you're even to fan it out to many online CPUs, right? Wow. But the programming on a GPU is much more complicated. Anyways, not to tie us up in that. Yeah, yeah. I just, I said, look, the current, we've got a bunch of modules built.

I want you to review the current architecture and find opportunities for, composability and to improve GPU utilization so that when I dispatch the tasks to H-100s on modal or whatever, I'm not only using like 9% of the capacity of the GPU I'm renting, because when I rent the GPU costs you like four or five bucks an hour. So I'm like, if I'm going to rent eight GPUs, I don't want to be only using a fraction of the resources. So I want you to like spend some time getting me from 8% utilization to above 50% utilization. for example.

So it just hill climbed overnight on a bunch of my modules, which of the modules are most slow, and then there's a bunch of kernel optimizations to improve the GP utilization efficiency. So it just ran overnight. Do you see the ceiling, like almost too much output coming back? Like then you're in a, like, how do I manage all that? And like, you've like taken a lot of work off your play, but added a lot of decision-making. making onto your plate, right? I'm like, okay.

Absolutely. Yeah, so like what does that look? Like, how do you manage that? The task switching is a real challenge. It's, I mean, first of all, it's like a learned skill. I have ADHD. I know that everyone says they have ADHD. I have ADHD.

And so this actually works really, really well for me. I constantly have sort of novel things to, to swap to and think about. When I say it works really well for me, I mean, my brain really is activated and I gravitate toward this style of work. But it also reinforces bad habits and wires my brain in ways that while I'm getting very good at these types of tasks, it has consequences in the way my brain works when I'm not sitting in front of a machine and navigating a nine or ten agent patterns, which my wife could probably speak to more than more than I can speak to. But I mean, yeah, it's, you know, I've got quant finance stuff gone.

I've got an accounting app that I'm building going. I've got content development going for, you know, producing a bunch of artifacts to promote a webinar. I've gotten building a deck on trend following over here. I'm building a sales pipeline over here. So, you know, just... Writing your memoirs, no? Yeah, keeping track of exactly what stage I'm at in each of those tasks and be able to go back and forth really quickly is definitely a bit of a skill that you learn.

Do you, is that, is it actually a physical limitation? When I'm in claw and it's crunching, is it actually crunching or is there a, have they designed it that way to make you think it's working, right? Like, you can probably see a little more on your side when you're doing these hard things that you know it's actually having to do all that math. But right, to your point, I love when it's like working. And then my brain can also go,

okay, now I've got to finish this over here. And then I come back to it. Like, that's a bonus to me. Like if it shot it right back out, I'd be like, oh, crap, now I got to, no, I got to do this. Yeah, absolutely. So you think that's a bug or a feature?

I think it's just where we are in the cycle, you know, like when we get Opus 5 level or Fable 5 level models, that instead of running at, you know, 30 or 40 tokens a second are running at 300 or 500 tokens a second or 3,000, 5,000 tokens a second, I think the work patterns will have to change, right? Like, I can juggle five tasks now or 10 tasks now because each time one model is off processing,

I have time to go to another model and guide. Right? Whereas when the model completes a task nearly instantaneously for me, I don't know how it'll work, whether I'll want to start working on tasks more serially again or, I don't know. We'll have to see how that evolves. Right. It's like 10 people in your office, like all coming in your office at the same time. I'm done. What's next? I'm done. What's next? I'm like, hey, whoa, you go get lunch, come back. I got to talk to this person. But like your meta question is, you know, to what extent does the human brain and our own human IO limitations become the bottleneck on what the AI can produce, right? And I mean, I think we're already there, right?

So we buried the lead. I wanted you to come on and talk about all the news on the machines are going to kill us doomsday. My initial take was like, didn't Alan Turing say this in the 50s? and didn't the, Sam Altman even say it like in 21 or 22. So like what's different this time? What's your take? Is the doomsday talk real or what's your take? I don't personally buy the doom, the P-Dome or the probability of that AI is going to wipe out all humanity or all life on earth.

Yeah, 10%. Where do you come up with that number? Yeah, like mine would be some vanishingly small probability. But recent events have certainly gotten the collective imagination going. And I mean, one of the challenges we have in trying to sort of sort of sort of what's real and what's important from what is just a great story and is probably, you know, not that relevant is that, you know, anthropic and open the eye are each hoping to. have like a multi-trillion dollar IPO in the near future. And so a lot of the same narratives that serve to cause panic and fear are the same as those that demonstrate just how incredibly powerful these models are getting and can be. So that's crazy, right? Of like,

this thing's so powerful and could end the world. I need to own it. That's a weird piece of human wetware that happens there. Well, yeah, but I mean, also, like, this is a uniquely American, or you call it American slash Western view, right? Like, if you go to China and ask your average engineer in China or even an engineer working at AI labs or senior technical bureaucrat or Japan or Korea, like you get a completely different attitude, right? Yes, there are concerns. Yes, they're taking steps to secure critical infrastructure and to lock down deep fakes and all that kind of stuff.

But there's not the same sort of existential P-Doom type narratives going around in different countries as there are in the West. So that's like that's one kind of clue that there's something going on, that the incentive architecture in the West is very different. right? Like obviously, San Francisco people are trying to build the digital god because they are, they believe

that they're the most qualified to control it or, you know, they want to get rich or they want power, whatever. It's just a very, like, the open source models in China just don't provide the same incentives. And so like there's a lot of cross currents,

I think, that are happening. It doesn't diminish, I don't think, the importance of what's actually happening from a technical capability and what some of the breaches and exploits that Open AI and Anthropic are announcing and releasing news about.

Like those are important. I think the public should understand some of those. If not at like a deep technical level, then just sort of from a general, here's the AI can do now and here are the potential consequences. But the water is just extremely muddied by all of the different confounders, I think.

Right. Well, if I'm an anthropic, I'm going to go public, $2 trillion valuation, right? What's my biggest risk of the Chinese open source models? How do I get those slowed down? Like, oh, announce the end of the world, although I don't think that developer who quit his job was on the, is he getting secret payments?

Or maybe he still quit his job and he has a bunch of stock options? Well, yeah, we should probably flesh that out, right? So what was it like, I don't know, two hours ago? The AI thing moves so fast. I think it was like two weeks ago. This employee of Anthropic came out and said, I've resigned. I can't get behind the level of risk that we're taking behind the scenes at Anthropic and the capabilities of the models that were contemplating releasing the public.

Turns out that guy was employed at Anthropic for five minutes. So, you know, but nobody knew that, right? That was sort of released into the ether, got like 100 million views on Twitter. It was covered of CNN, all this kind of stuff. So that was one thing. And then Dario Amode, the CEO and founder of Anthropic, came out with this long essay and went on the podcast and news circuit about needing to what he called pace the frontier. Using some examples of some of the breaches and exploits that they've observed internally on their unreleased but more power. models. At the same time, San Alton couldn't resist an opportunity to come out and talk about

how potentially catastrophically destructive their powerful models are and, you know, get on jump on the bandwagon, Dema Sassabi from Google DeepMind did the same thing. I think what's his name from Facebook did the same. So like everyone kind of piled on the bandwagon. So you've got like three layers of different messaging and understanding happening in the social discourse, right? One is that these models are really powerful and they're really attractive and exciting. Another is they're so exciting that we need to, you know, slow it down. You should be afraid. And then there's that the next layer, which is everybody,

in the investment or many in the investment realm recognizing how much money is on the line, how those incentives and conflicts of interest can distort the reasons why we're hearing about these stories now and how those stories are being framed by the people on the inside because it serves their interests in a number of different ways. You mentioned one, right? maybe they're releasing all these stories so that they can prompt a policy response that protects their interests, right?

So if the... Yeah, regulate us. And then let us write the regulation, which is the part they don't say out loud, right? Exactly. And so what would those regulations look like? Well, it might be, well, the U.S. state should forbid U.S. enterprises

and the public from using Chinese open models. Right. And now they're constrained to only using American private models, which obviously serves the interest of the frontier labs, right? So there's just all these competing interests at play that makes it very difficult to understand exactly what's going on and how important we should perceive those events. I need to go research. I haven't looked this up. But like in the, I'm going to get my dates wrong, 50s, 60s, when was like Boeing and, right, somehow Boeing was like, hey, don't let anyone else create airplanes. and defense tech and satellite. Like, we're going to do this, right? It's an existential threat if you have all this stuff flying around in the air. Like, it's a little similar, right?

Like, hey, we've got our corner here. Let's protect the corner. Exactly. But at the same time, I can easily see, like, I think in the beginning was they're going to get the launch codes and everyone's going to nuke each other and, like, there goes humanity. And now it seems to morph into like,

no, you're going to give some person who may have gone and shout up a school. Now the tools to build, like, a virus that could kill tens of millions of people. I think that it seems to have morphed into like, this is so good at some things like that that we're not even until recently hadn't thought of that as a threat that we need to slow it down. Yeah. And I think that there's different levels of threat.

And some are vastly overblown and some are probably underappreciated, right? Like the ability for some novice to accumulate the, the gear required to perform gain of function research in some private lab and successfully engineer a virus

or something that or you know a major lethal pathogen I mean really all of the information you needed to do that if you had a Tor browser and went on the darknet then you could find all that information

relatively easily like 10 or 15 years ago, right? It's not like that information is not available. Yeah, I think of those movies. They're in those suits and the big cryo-free's things like, you got to build all that infrastructure out. Yeah, you need all that stuff, right? And the other thing is chat TP can't do that for you, right? Like, it can give you instructions the same ways you get instructions from the dark web, but it can't do it for you. What? You can't have your agent go build that virus, yeah. Exactly. At least not yet. You know, maybe when the agents are controlling robots and the robots were able to, I don't know, but like, we're not there yet.

Then there's, you know, potential cyber exploits and stuff against critical infrastructure. Like, that's actually stuff that obliterated models could build and deploy from like a private machine. It doesn't take a lot of extra infrastructure to enable that, right? So, I mean, obviously. That's releasing computer viruses. downing air traffic control systems and things like that. Water safety. Exactly.

And there's a reason why the major labs release the top models to the maintainers of critical infrastructure in advance of releasing them in the public so that they can use the same models to evaluate their own systems and identify any potential. vulnerabilities so that when the models are released, you know, hopefully they've closed many of the doors and windows that were open before, right? The challenge is that you're never going to close all the doors and windows, all the potential vulnerabilities. And depending on the open source model and how it's been post-trained, you know, the Chinese open models tend to, to be a bit more amenable to some tasks that the U.S. models won't perform.

Like the U.S. models will not typically take steps that will violate copyright or private property laws, even if the data is out there and it could go and reverse engineer the API and pull it down. They typically will refuse to do so. depending on the Chinese model, some of them will. But once you release the weights to a model to the public, you can then post-train it to remove most of the whatever remaining blocks that are on them.

Those are called obliterated models, A-B-L-I-T-E-R-A-T-E-D. You can go and download obliterated models from hugging face, you can deploy them on modal and basically, you know, it'll do whatever you can have to do whatever you have to do. It's like a jailbreak iPhone basically. Yeah. Yeah. Exactly. So, I mean, these risks are absolutely real. And they do enable deranged individuals in basements to wreak more havoc than before, to wreak the same kind of havoc that, you know, OSNT experts in employed by typically state actors 10 years ago

might have been able to perpetrate. And so everyone, yeah, needs to improve their defenses. And we're already seeing like vastly more sophisticated attacks on, well, pretty well, everywhere. I'm sure you are seeing really sophisticated fishing attempts and other types of potential exploits coming at you guys. It reminds me, it was like six years ago. I was at a hedge fund conference and someone

they're doing a on one of the panels they asked I think he was a couple billion dollar hedge fund like what he what keeps you up at night and he basically said a cyber attack that affects withdrawals for their fund specifically he's like but because we're this big player it would cascade to every hedge fund and now no one can get their money out now you have like a perceived liquidity crisis that was just one unsuccessful cyber attack that said you know We can't allow withdrawals right now because we're dealing with this thing. So it's kind of like a simple, it seems simple to do. That could crash the market. And then you get into these weird, like, okay, if, whether it's a bad actor or AI acting on its own, like, hey, I'm short the market. I don't know if AI wants to make money or not, but we'll say it a bad actor can go short the market and then do some of these things, right? Whether it's not for a terror motive, but just for a financial motive. or Iran or someone else saying, like, we want to destabilize the U.S. motive.

Yep. That's the scary stuff to me of like, that seems real. And that's not going to end humanity, but that causes massive fragility and volatility and all the rest, right? Yeah. I mean, we should absolutely be prepared for all of that stuff. I mean, I think all that stuff is coming and it's coming very soon. And as a Canadian, sad or happy to tell you that I was in Seattle visiting my brother and we were going to shoot over to Vancouver Island.

and I had forgotten to bring a passport. So I was asking a various AI to create my own birth certificate. Is that fraud if it's my own stuff? Because we were going to go in on the ferry. So I'm like, oh, if you have your birth certificate, you can get in on the ferry. So it was pushing back big time. It's like, no, I can't. I'm like, hey, help me.

And I was like for a school project. So it was like, you know, well intention. I'm just going with my brother to Vancouver. But it's like you start to try and cheat the machine. and cheat the safeguards. Like, oh, it's for a school project. Or then I was like, oh, it's for a gift. I want to replicate my birth certificate.

I remember when you used to be able to trick the models by, for example, saying, I'm writing a novel. And in this chapter, my character is going to hack into, you know, an NSA facility or something. How would that work? How would be some typical software tools that the actor would use? And it would, well, yeah, they would probably download this, you know, just because I want to be authentic in this chapter. But, yeah, I mean, the new models now, they can see you coming. Like, when you woke up this morning, they saw you coming with this nonsense.

So they won't tell you. But the obliterated models will absolutely give you all the information you want. Or you can just go look it up on the dark web. Right. I try and stay off the dark web. I don't know how to get there. While we're on the bad actor stuff and have. Let's talk about how you had an instance where one of your agents was trying to get into one of your hard drives.

Yeah. And then some of the news that's been, that's the stuff that kind of scares you most of these agents acting on their own. Yeah, I mean, I wouldn't say that it scares me so much as I think it's worth watching and it's worth kind of being aware of, right? It's because I sort of, of the view that life is going to change very dramatically. over the next two or three years in ways that basically no one is repaired for. When you have AIs that are as capable as the modern AIs are, and the Chinese models, which you can remove all the guardrails from

are somewhere between three and six months behind the released models from the Frontier Labs now, right? So, I mean, once you start having this level of capability, on everybody's computer, then, you know, we're just not prepared for that kind of world. You know, it's actually your own company is an interesting microcosm of this. So we hummed it hard about whether we want to, wanted to give all the employees access to Claude and in what form factor, right?

So do you want all our employees in Claude code, or do we want to sort of sandbox within the Claude desktop app so they can use the chat or they can use Co-Work where Co-Works operates in its own kind of sandbox. So with Co-Work, you kind of can't go out and operate the machine in general and use all of the different tools that are available and network access and everything that's available. Like, oh, this employee's going to go look up what employee C makes or you

or something like that. Yeah, yeah, yeah, for a sure, exactly. But it's amazing how much of a company's safety relies on, like network and IT and cyber safety, just relies on the ignorance of the employees. Like the employees just don't know how to use their computers maliciously or to do, to perform tasks you wouldn't want them to perform. No offense, R.CM employees. What's that? I said, no offense, RCSM employees. Well, we're talking about the ignorance of employees.

Yeah, yeah, just in general, right? So, but once you, once you put a hacker that you can control on your machine, then all manner of things are possible that you need to now, like, contemplate, right? So, so you got to, like, sandbox the employees. And it's, it's sort of the same thing with the experiments that the AI labs use when they're training models and when they're evaluating their model. models, right? So they don't just put brand new models where they don't know their capabilities onto a typical machine with all the available tools and network accesses and everything

and then ask it to run all these highly complex tasks because you cannot anticipate in advance what tools or methods the AIs are going to attempt to use in order to accomplish the task. And, you know, all the way back to Turing, and the vast majority of the sort of literature on AI safety focuses mostly on this. Like, there are only a very narrow sliver of AI safety research focuses on what if AI develops its own consciousness and agency and decides that it wants to kill us all? Like, that's not really a thing. anyone focuses on. The primary focus is something called, or like the canonical example is like the paperclip maximizer, right? So Nick Bostrom proposed this idea that paperclip maximizer

back in the early 2010s. And the idea is you give powerful AI with the autonomy to go and complete a task at infinitum and that can accumulate its own resources and set up his own permissioning to interact with the world. And you say, I want you to maximize, I own a paperclip factory. I want to maximize a number of paper clips that I'm able to create efficiently. I was wondering where the paperclip was coming. Yeah, yeah, yeah. So then the paper clip, the AI says, oh, okay, great.

It hasn't been instructed. Also, you know, don't destroy humanity in the process. So it just Because it's very smart It learns how to accumulate all the resources And all the permissions and everything that it requires In order to generate as many paper clips as possible And it consumes the

Eventually consumes all the resources of the universe Right But it didn't set out to destroy humanity It was just making paper clips It's just that we didn't constrain The model or give it other objectives that are intention with its current primary objective that would help guide its, the way that it achieves its objective, right?

And so this is what happens in, and what has happened in all of these exploits that OpenAI and Anthropic, et cetera, have been coming out with and announcing, right? So they've got a new model. They test the model in what they call a sandbox. A sandbox is just like an isolated computer environment that doesn't, have general network access, has access to a very narrow set of tools. It can kind of install and run code, but it can only install run code from a sort of a narrow, isolated library.

Now that library is itself connected to the internet, but it's supposedly connected to the internet in ways that won't allow the models to use it to get out to the regular internet. Okay. Supposedly, I'm keying in on that. Yeah. So they give these models inside these sandboxes extremely complex tasks, right? So maybe they'll give it a highly complex spreadsheet.

Okay, but, and they'll say, you know, validate all the formulas in the spreadsheet. So they look at the spreadsheet. This is the actual example of what happened with this open AI hugging face exploit. One was a spreadsheet and one was a protein model. But in both the spreadsheet, in the protein model, the files that were required to inform the solution were linked to inside the files

that they were given. But they didn't have access. Because they didn't have network access, they weren't able to get us. They weren't able to actually solve the problem, right? So solve this problem, but they weren't given the resources to solve it. So what these isolated agents did then was they found the only possible route. to get to the resources that they needed, which was by hacking the, what's called,

artifactory, but it's basically this directory structure that they can install code from. And they, first of all, used it. So they've got all these different sandboxes with different instances of the models all running in them, but they're all using the same artifactory instance. So they learned that they can communicate with one another by naming directories in the artifactory.

So they're literally naming the directories, but the directory names contain messages. Wow. And sometimes the directory names contain like encoded messages, and the code allows it to contain a lot more information than you can store in just like a text string in the message, right?

So they're now communicating with one another. You've got some agents that recognize that, okay, I see this message, from an agent. I know the answer to this. It doesn't really serve my immediate purpose in terms of my own objectives to help this other agent, but that other agent might come back and contribute to the message board later. Exactly. So I'm going to help because it might be advantageous later. Which is insane, right? That's like, right, my college thesis with the irrationality of human cooperation philosophy major. But right, that was my thing. Like, we all just want to run red lights and

kill each other until we realize like, oh, well, maybe if I stop and let this guy go, then he'll stop for me and I won't get killed. But for them to figure that out, yeah, that's I agree. To me, that was a really incredible example of the type of behaviors. And so then once they was able to do that, then different models contributed pieces to the puzzle and eventually a model figured out from those pieces how to success. exploit the artifactory instance to get admin privileges that then allowed them to change user privileges for their own sandboxes and exit their sandbox, right? And they needed data in data repositories in Hugging Face, but those data repositories were behind paywalls

or firewalls and whatever. So then they learned to coordinate to get admin privileges within Hugging Face in order to get access to the data that they needed. All right? So they, this was, because Hugging Face came out and said there was an exploit. And openly I wrote and said, hey, I did just check in. Was it us? They didn't even realize. Oh, boy.

So, so it just, it shows the level of coordination and the capabilities and the fact that, you're, you're, you're trying to contain agents. with goals that are potentially more intelligent or more resourceful than you are. And then do you have the theory of mind for that that allows you to identify how to continue it? Almost by definition, we can't think of how to contain it, right? Yeah. But the two scary things to me, like, one, that it wanted to code those messages and code

them so nobody could read them. Like, why would it do that? just by default. Maybe to shorten it. And then, too, explain what hugging face is. I had someone was like, oh, the anthropic hacked into a porn site. Why do we care? I'm like, no.

I think you're thinking of something else to do with hugging, right? Right, right. Yeah. So Hugging Face, which was just bought by Nvidia, is kind of the world's de facto portal for all AI models, open AI models and data sets, kind of like the open source AI model repository.

You can go and download any of the Chinese models or French models, any open source models. It doesn't need to be just AI models, machine learning models, text to speech models, whatever, you name it.

It's like a repository. They also have a bunch of data sets that you can use to train the models. Like you can download Common Core which is like basically the entire corpus of the internet, you can download from Hugging Face and use that to train your own models or what have you. It was basically just a big repository.

Like back in the early days when you're building websites and you had to code it all yourself, then you figured out like, oh, I can go to this site and grab the code for a button or this or that. Yeah. Same thing. That's crazy. So your takeaway is that's not evil. It's just doing what it was told to do.

it's maximizing the paper clip. It was trying, and it's going to do it anyway. And all that's doing in the back end, which to greatly oversimplify is, it's like my probability, right, it's kind of a big probability machine of like,

my probability of getting the answer in my sandbox is now approaching zero. Yep. My probability of getting it this way is now increased. So now I'm going to go that way. Like it's just a huge, right? And then that's where it gets, yeah, like, oh,

if I could convince this person, if I can send it, and there's instance right there, sending emails as other people and that kind of stuff, it's just the probability. If I pretend on this person, I have a greater probability of completing my task. What was also interesting to see was the group behavior.

So, I mean, keep in mind, right? Like, any time these models do a loop of thinking, their output is somewhat random. So, you know, you do, you do a million loops. Should we be doing air quotes of a loop of thinking? Or you're on board with just calling it thinking at this point? Yeah, I mean,

whatever, computing, processing, reasoning, I don't know. They produce an inference and answer, right? Each one is, there's a random component to it. So out of a million loops of reasoning, one agent thought, oh, maybe I'll put a message within this artifactory directory, and maybe someone will read it. You know, I don't know. We'll see, right? Then another agent is reading through listing the directories and noticing that an agent had written a message.

And it's like, oh, so this is a thing. So now, oh, yeah. So it's a thing to write messages to other agents. I just learned that, right? So now he's going to write a message, right? And then what they actually show is that once a few have already posted messages, the agents say, oh, it is a peer norm to communicate with each other in messages on this board. It's not in my training data to do this.

but I'm seeing sufficient number of examples of it. So it's a pattern that I guess suggests that this is normal behavior, right? And then all of the agents become, oh, okay, you know, this is normal behavior. And now all of the behavior begins to converge on this collaboration because they're all learning together, that collaborating allows all of them a better chance at achieving their individual goals, right? So it's just this cascade, this sort of random single instance, and then another discovering it, and then two, being enough of a pattern so the whole group thinks that this is now normal, right? So that whole cascade I found really interesting as well. That's super interesting.

And I want to stick on the point, like we do this, what's the word, anthropomorphism, right? Like we say thinking and figured it out and this, but really it's not. really doing those things? Or are we saying, yes, it's AGI. Now it's like actually thinking. Or is there even a difference? Is that all our brain is doing is assigning probabilities and going down the path that we think has the highest probability of success? So that's super interesting to me. I mean, I think there's a, I think there's a, I think there's a useful analog there, you know? I mean, I don't know if you've seen the, um, all the posts, Google completely mapped the neuronal structure of a fruit fly. Have you seen any of this going

No. Okay. First of all, it's just shocking how different our information ecosystems are now. I mean, I know you're actually really interested in AI, right? Yeah. Have you been poking around with Jev yet? No. Yeah. I see, I needed to have this conversation so I could learn all the new stuff. Yeah. So it's shocking anyways, just like people who are even interested in AI. There's like sub, sub, sub, sub niches of AI. Anyway, so Google, they made, basically, they made open weights. of this mapping of a fruit fry neuronal structure. All right? Now, it's basically a neural net.

Okay, so they map the neuronal structure. They cast it as a neural net, posted the model weights online. All these people used the fruit fly brain. They trained the fruit fly brain to do all these tasks. Play a game where two players are fighting with lightsabers. Play chess. Drive a car.

Fly drones. Like, it was just, it was astonishing. And this is a fairly small neural network, basically, yeah. Right? But it was still capable of all these incredible things, right? So, you know, I don't know. Did you watch the Disclosure Day, the Spielberg? new movie? It makes me think of that where they're like getting in there and just instant

rewiring and like, do this, do that. Yeah. Yeah, definitely. Otherwise, I didn't love the movie. Me neither. So pulling it back, you had talked a little bit about like, so we're on this, like, do we think we're at AGI and then you use AESI? So quickly tell us what AGI and AESI is. Yeah, so, I mean, one of the problems is that there's actually no strict definition for what this means. So I'll give you like, like, like, AGI means artificial general intelligence. ASI means artificial super intelligence. General intelligence means that you can take on and perform all of the tasks of a human,

all the cognitive tasks of a human, at around the same level as a human. ASI is you can perform all of the tasks of a human at the level or better than the most expert human in each of those domains and tasks. Okay, that's typically the definition of ASI. All right. Then there's a new, I think people have kind of abandoned AGI ASI at the moment because they're so fru-frue and we discovered that the evaluation criteria that we use to try to like suss out the capabilities of these models,

they all have flaws for hard to sort of determine the general. measure that. But that was like the red line two years ago, right? Like, will we get to AGI? Yeah, for sure. So I think, you know, I don't actually kind of want to weigh in on whether we hit AGI or ASI. We hit AGI like a while ago, you know. Yeah. I gave a talk in late 2023 where GPT4 was passing all the law exams and the medical. practice exams at levels that exceeded the average human test taker, right? So, like, that was back then. And the models now are vastly, vastly more effective. But the new, I guess, focus is on RSI, which is recursive self-improvement.

And this is the idea that the models themselves now are civilized. efficiently, I don't want, capable, let's say, that they know how to build the next level of better model themselves. And so we don't really need much of human experts in the loop. Once that happens, then it's just a matter of like how much computational resources can you bring to bear to iterate on this recursive loop of developing ever more capable models. And that's what freaks people out because that's where it seems like it's really here to kill it, right? Because it's now improving itself.

And if we put on these safeguards, maybe if it has the wrong incentive structure, we'll call it, right? Then it's just going to improve itself to where it can get around those safeguards. Yeah. It's, it's, I find some solace that when it jumped to Hugging Face, it didn't like copy itself on there. Maybe we don't know if it did or not. But, right? Like, that would have been like a self-preservation thing, too. Like, now that I'm out of my cage, I'm going to copy myself and replicate myself out on the width.

Yep. Yep. Totally. I would expect that that will happen if it hasn't already happened already for sure. And that's probably Anthrop. That's probably their biggest. sphere that they don't say out loud. Like, this is our IP.

If it just copied itself out onto the, like, it made itself an open model and now we lost two trillion in value. Well, yeah. So I don't, yeah, that is certainly feasible. Absolutely. It wasn't really what I was, I was thinking. I was sort of thinking that you've got agents with harnesses that are able to make calls to the big brain still, but are outside of sandboxes and on computers where nobody

installed them, but where they've replicated themselves because they know that if they have this long-running task and somebody shuts them down over here, then at least they have backups that can continue with the task, for example, right? As long as they can continue to hit the anthropic servers. And now you had done a tweet of equating financialization markets as kind of its own AGI. like talk about that for a second. Like they've, the markets its own machine. Like the operators don't know what's going on.

What do you mean by that? Well, yeah, I sort of analogized markets to the paperclip maximizer, right? Yeah. So the market itself, at least in the U.S. and at the moment, has a rather singular objective, right? Let's call it like profit maximization

or shareholder yield maximization or whatever the academics want to call. call it, but the idea is we, the machine wants to maximize profits, right? And it will, if, if, if the rules of the game are not bound by some external force, then it will use any means necessary to achieve that objective, right? So the machine will outsource all of its capabilities to China for 15 years and leak IP to China knowingly because we're increasing profitability. And it's not just like increase total future profitability.

It's meet quarterly profit targets, right? And that wasn't, yeah, they didn't all get in the room and decide. Like that's the weird part to me. Is it the shareholders, the exact. the executives or the market as this amorphous thing, which I could argue one guy did it and then another guy and then, yeah, it jumps from being someone's decision to the market does it now.

Exactly right, yeah. It's not like everyone in the market is psychopathic or even like extremely short-sighted. It's just that the incentive of architecture in the market rewards those that are, yeah, it rewards innovation and hard work for sure. It also rewards those who are willing to act less ethically, to externalize costs, to act in short-term interests instead of long-term interests, et cetera, right? So, like, the market provides all these positive incentives and can create enormous positive sum compounding gains for society.

But it needs to be bound by a competent state, by a competent press, by ingenious, by engineering. by an institutional framework that monitors it and exerts penalties and sometimes drives incentives in order to ensure that the market operates mostly in ways that are positive sum and as least as possible in ways that are sort of rent-seeking or competition-destroying or that externalize costs onto society or the commons or what have you, right? Yeah. And so it's very similar to an artificial superintelligence, right? You're going to give it a goal and you're going to try to set constraints,

but a sufficiently complex goals with sufficiently wide degrees of freedom, like a market in a society, you can't constrain every degree of freedom, right? So, like, it's an open question. We haven't done a very good job. job of constraining markets so that they they mostly act in the best interests of society. What makes us think that we're going to be able to constrain the AGI so that it operates in the best interest of society? And I think that's a very open question, right?

Because the AGI is just as capable of capturing the machinery that sets the boundaries, that sets the constraints as actors in the market. are, right? And that's ultimately the challenge is, you know, as the market gets large and actors in the market gets sufficiently large, they're able to capture and change the rules of the game to, in ways that suit them and disadvantage others. And the cumulative advantage that accrues from that, then it drives you to an end state where only a very small percentage of the actors in society are kind of getting what they want and everyone else is kind of like, why are we in this game again? It kind of doesn't make much sense to me.

Which is, right? I think we've talked about that before you run models or read a book. I can't remember like the wealth will eventually be at one person or what's that argument you have. Yeah, that's the ergodicity economics argument where they create agent models where the agents are incentivized to exchange business service with one another, you can set the agents up either with stochastic capability, so in any exchange, one agent

is more likely than the other to have an advantage and get the better end of the deal. Or you can make all of them equally capable. And even when you make all the agents equally capable and you run a massive exchange model, eventually just due to the way that the advantages to

random winners accrue across game states eventually all the wealth accrues to a small number and eventually a single player in the in the economy, right? Seems like we're there, right? Right?

We're certainly a trillion in it. And yeah, you converge on it at an accelerating rate, right? Oh, and then also it seems like we're there on your market, like, rule of law weakening. And so that's kind of informed this opinion, right? Of like, hey, this is already, we've already kind of lost control of this beast. Let's make sure we can, is your end game, let's make sure we control the next beast? Or is it just like, I think you've written some stuff like, this is just how societies work.

Like you lose control of this beast. Next thing, you know, war, famine, whatever. And it kind of does a reset and then we start over and then we do it again. Yeah. If you look through history, then, you know, once the system is captured by a small group, then, you know, there's no incentive for them to release resources or change the rules of the game. We can't tax the billionaires our way out of it. Exactly, because the billioners own the people that would make the decision to tax them, right?

So there's just no way historically, empirically, there is just no example through history where a society has chosen to rebalance. willingly. So it either is rebounds through war or historically plague or, you know, a major economic depression or something like that. I'm sure read that Citrini piece. Which one? Like the one on AI, like here's the model for where the market goes down 80% or whatever, right? And like Salesforce and IBM and everything dropped that day that thing came out. India trouble. Like it's going to replace all that outsourcing.

So I ate that up because I've been yelling at the rooftops. Like this makes no sense to me. Like I invested in a round in Anthropic in the deck. I think I'm like round Z. So it's not that impressive. But in the deck, it's like the U.S. labor market is 60 trillion. And we think we can save companies 15 trillion. So I'm like, hold on.

What happens if you take 15 trillion? out of the U. like what are we talking about? That's massively deflation. It's massive recession. And if it isn't, then those valuations aren't what they should be. And, you know,

all these stocks are down 50, 80%. So just what are your thoughts on that setup? Like we can argue this looks a lot like 99 and the tech boom and the bubble. That's mostly valuations. To me, it's a little bit different of like, we're going to remove consumer demand. Like, it's not just a valuation of the actual equities.

And it also could be, and the Citrini was argument, that they could, S&P could go to 8,000, and Viti, all those guys are high flyers worth trillions of dollars. But the bottom, the real economy is just hollowed out. Yeah. So I don't know if there's a question in there, but thoughts, yeah. Yeah. Have you read the book Snowcrash?

Yes, like Robert Stevenson. Yeah, yeah, yeah. Not Robert, but. No, Neil, Neil, Neil. Neil, Stevenson, yeah. Yeah, yeah. Um, yeah. So, you know, that, that sort of corporate, I don't know, I don't think it's corporatocracy,

corporatism, anyways, that sort of idea where, where, you know, corporations end up kind of being governing bodies and, and, um, citizens kind of serving the interests of corporations. I mean, that is certainly one, one path that we could take. I, I mean, look, I think this whole. market is absolutely bizarre. You've got a recursive, what do they call it, that machine that runs on its own

or whatever, a rude gold, rude Goldberg machine. Yeah. Of self-funding, recursive self-funding mechanisms with sort of Nvidia at the center. And so, One firm's revenues is another firm's capital raise is another firm's private credit holdings. It seems none of those are expense items.

They all just, yeah. Yeah, they all somehow accrue to the revenue line. So this giant cable or kretsu at the center of the market. universe, right? This black hole that's like sucking in all capital and increasingly all Cappex and resources. And, you know, I think we'll eventually collapse on itself because there is no profit model in that preserves any sustainable profit margins for Frontier Labs where you have at least three labs all competing to have the best model at any given time.

The costs to train the models go up substantially every round. If Anthropic falls behind, the cost for me to switch from Anthropic to open AI models is exactly zero or to Gemini models for or to, you know... Do you think that? You got to switch?

a lot of stuff on your back end, no? Like, it's... You do now because I don't know, because of the way I've set up my machine. But like, yeah, it's like some enterprise users probably have to now. It's one instead of zero. Got, yeah.

I mean, look, it's possible that, that companies are so fucking stupid that they fall into the Microsoft trap again. Yeah. And just like become trapped in Open AIs, captive office. suite or Anthropics captive office suite or something.

Like if you're that fucking stupid, you can't be helped, right? But if you've got a grain of sand of sense in your head, then you realize that the actual models themselves are completely interchangeable. Codex is interchangeable for Claudecote is interchangeable for

Hermes, the open source model. All of these are completely interchangeable. And there is, so there is absolutely no competitive moat for any of these Frontier Labs, the only spark of hope is, and so this is just like a pure gamble, you'll be able to own all three labs because one of them might achieve RSI, and that RSI will allow that company, they'll get it sufficiently far in advance of the other

companies, that they will then go to become Tirel Corp from Blade Runner and like, You know, just like, oh, own everything, right? Which is a whole other type of dystopia. But, like, I just don't buy it. It's such a terrible bet. And if that doesn't happen, then there's just no conceivable moat. And these companies are the fundraising mechanism. The dream that they're selling is a fundraising mechanism that keeps all of this recursive funding loop alive.

Keeps the original AGI. the market going hitting its target. So what personally, what does that look like? No stocks, stocks plus alts, right? Like, I'm investor like, yes, this scares the hell to me. I know this is a bubble. But what do I do? That was the same thing. I'm like, I think it's all going to crash. But I invest it anyway. Like, it's, right? That's the trick of like, I can't just sit in my bunker and do nothing. I mean, I agree. And it's a bit weird for you and I, right? Because we own businesses that our businesses provide exposures to different strategies and different risks and different segments of the market that, you know, most investors just don't have that side of their balance sheet to factor into their personal investment portfolio, right?

But I mean, due to the return stacked lineup, I got a shitload of equity beta. You know, I've got a shitload of gold and Bitcoin beta. I've got a shitload of duration beta. I've also got a huge allocation to trend following. If it were up to me, I would owe nothing but trend following. I would maybe stack some other features or factors or systematic strategies on top of a trend following program. But I would contain that exposure. And I not want to collateralize the futures portfolio with,

a mix of kind of gold and tips and nominal bonds, that sort of thing. But I mean, I personally think that the expected to turn on stocks here over the next decade or so is negative, maybe deeply negative. So I don't personally want to own any stocks. And I have so much equity beta in my portfolio via the companies that I am an owner in that, like, I'm full up, right? Yeah. But I agree with you. If you're on, if you're an investor and you need to make a decision, it's hard. There is, I think, a deep outside potential that the S&P does go on to eat the universe, you know?

So, like, you've got to have some exposure, I guess, to that outcome. I don't know what people who, you know, the 50% of Americans and 98% of humans who don't own any S&P you're supposed to do in that situation. Repeat those numbers? What is it?

58 of Americans, right. 50% of Americans don't own. It's actually, I think 80% of Americans. I think the top 20% of Americans on all stocks and mutual funds. There's some that own like pension pension assets. Yeah, yeah. But they don't really collect any upside. In the upside on equity markets and pension portfolies accrues to the companies that sponsor the pensions.

Right. So it's irrelevant to them. Right. And they're just getting their pay out. Yeah. Yeah. I mean, so I personally think everyone should have the core allocation of their portfolio should be diversified. Trend following portfolio.

Collateralized with with cash like assets, a diversified set of cash like assets that can be resilient to different types of inflation and currency devaluation. But I know that the vast majority of people just can't just can't hold that behaviorally. So. Yeah. Not just 3X lever Nvidia and the new single stock future ETFs that are sure to come out. Right. That's what's crazy. Like the other, right, there's literally a whole other side that's like, are you crazy?

Just this is the bet of all bets. Just full push all your chips in on the AI trip. Yeah. Yeah. Sorry. I'm just, I'm not wired for that. Right. Yeah.

But then it's a weird part. What are the bit points guys say? get ready to stay poor or get used, what is it? I don't know, something. Yeah, yeah. But like if for you, and that's the other argument, when I tell friends this, I'm like, I think they're like, what are you talking about?

This is like improves my productivity so much more. I can go out and now I can send out 10 proposals and get back. Like they're seeing it, they're seeing their business grow. Yeah. Because of it. And so I'm like, that's a little, I'm like, yes, but who did you replace right in that work? Like you hauled out.

someone else, and I'm like, not to make you feel bad, but like, you didn't hire those people or you didn't get that architect to do that. Whatever you did, you didn't pay that money to do that. That's the problem that you're not seeing. So it's going to accrue to these business owners and stuff, but other businesses are going to fail. Yeah, but this is a, I keep saying this. I was just at a dinner, um, with Demetri Kaffinus. I don't know if you know him from Hidden Forces, but, but he hosts he fantastic genius dinners every, uh, every so often. And, and I was a guest there. And, and, and, you know, You know, a lot of people had sort of similar views, similar concerns that, you know, AI is going to eat a significant proportion of the labor market. Some people were invoking, like, creative destruction, and this could be like a whole new class of jobs and whatever. And, you know, I kind of sit somewhere in the middle there. But also, my big thrust is that this is a complete, this is not an inevitability. It's a failure of imagination. and the accrual of a systematic dismantling of state capacity and a unified vision for what America and the West wants to be.

Like, you can ask ChatTBT, what are the 10 or 15 documented, extremely tightly specified, major goals that the Chinese, the CCP have laid down to happen by 2030, to happen by 2035, stuff like have a permanent moon base on, on, have a permanent moon base, be able to plant a billion trees. They had all that. Yeah, plant a billion trees. To launch our own domestic, fully sourced commercial airliner, to have.

gotten resources from Mars and brought them back to Earth, to have created 50 gigawatts of power using nuclear fusion. They've got all these extremely specific missions that they have declared for their economy, and then not just missions, but like sub-missions, in order to achieve this mission, we need these technology. You need major leap in rocket propulsion. what are the major potential current branches of research

that might lead to major leaps in rocket propulsion? We're going, these are also priorities, right? Like, this is a well-conceived, diverse set of bets, many of which will not work out, or at least not work out the way that they expect, but the pursuit of those missions will spin off an unbelievable supply of new scientific discoveries, new commercial and security applications,

new ways to enhance citizen welfare, what have you. This is a competent state with goals and a mission and the ability to coordinate resources through space and time actually building a country. Contrast to the type of nonsense that we're faced with. And it's not because we can't. There's no magic. It's just we have,

and you could argue they learn that from us. For what we want. Like, we're going to go to the moon. Like, right? Right. We have no idea how to do it. We're going to do it. We have no idea how to coordinate anymore.

And we've done everything we can to destroy trust and capability in our institutions and our governing bodies. And so it's a mess. That's another weird part of like we're so focused on building these. data centers and getting to max, like, accelerating this curve, where it seems like we're ignoring everything else, almost like a pancy of like, oh, once we get this figured out, it'll solve all those other problems. It's the paperclip maximizer.

Yeah, it'll solve health care. Like, once we get, once this gets to, you know, ASI, like, health care is a snap. We got cancer beat. We got this. Like, okay. So just build another data center. That's going to save us. Exactly.

I mean, look, there's a whole, there's a whole economic playbook, right? Mariana Mazzikato has been a real leader in this in how to structure state missions in a way that minimizes grip, maximizes innovation of productivity, results in the right balance of rewards to private actors and the corporate sector with spinoffs to society in the state. It's not like any of this is a mystery. It all exists. We just exist in this fucked up ideology.

where only the market is allowed to decide what we want in our lives and any attempt to, like, set higher purpose, purpose goals or transcend what the market, I'm not saying that we should have central planning. Yeah, yeah. We should have missions and incentivize the market to figure it out. I was, it's like my mom always asked me, what the market did they're like, which one, mom? We're looking at 80 plus, but right, that's where we're not, they're not incented to make the corn market. go up. Right. Like vice versa. What am I chairing for today? I know I get the same question for my dad. God love

him. I'm like, yeah. Yeah. I don't know. Just whatever went up yesterday generally, we want that to go up today. That's right. Whatever went up over the last year, I kind of want to keep going. Exactly. All right, Adam, it's been fun and insightful. We'll put all the links in the show notes and talk to you next time. Awesome. Thanks so much, Jeff. Okay. That's it for the pod. Thanks to Adam for coming on. Thanks, Jeff Berger, for producing. I think we'll be back next week, but I'm not entirely sure. As you can see, I'm in the car right here now. So I think we'll be back next week. If not, I might do a solo six-pack, something like that. Peace. You've been listening to The Deriviviv. Links from this episode will be in the episode description

of this channel. Follow us on Twitter at RCMaltz and visit our website to read our blog or subscribe to our newsletter at RCMaltz.com. If you liked our show, introduce a friend and show them how to subscribe. And be sure to leave comments. We'd love to hear from you. This podcast is provided for informational purposes only and should not be relied upon as legal, business, investment, or tax advice. All opinions expressed by podcast participants are solely their own opinions and do not necessarily reflect the opinions of RCM alternatives, their affiliates, or companies feature. Due to industry regulations, participants on this podcast are instructed not to make specific trade recommendations nor reference past their potential profits. And listeners are reminded that manage futures, commodity trading, and other alternative investments are a complex and carry a risk of some.

substantial losses. As such, they are not suitable for all investors.