AI Took His Job, But He's Trying to Stay Optimistic — Debate with Asian Dad Energy

Asian Dad Energy is a former big tech developer with an Ivy League education. We have a lot in common. We’re both:

• Software engineers;

• New(ish) YouTube creators covering AI;

• Dads with a lot of existential angst.

But while I dread AI extincting humanity, Asian Dad has other concerns. We debate the risk of the AI bubble popping, a “techno-feudalism” scenario, and whether AI will outperform a human CEO in the near future.

Let’s find out where he gets off the Doom Train™.

Watch on YouTube

Timestamps

00:00:00 — Cold Open

00:00:54 — Introducing Asian Dad Energy

00:03:36 — Asian Dad’s Career in Tech

00:09:06 — Getting Laid Off: The Job Market Is a “Hellscape”

00:12:29 — Software Engineering in 2026 vs. 2021

00:17:03 — Why Are They Investing Trillions in AI?

00:24:37 — Debate: AI Is Different Than Other Tech Bubbles

00:33:54 — Debate: AI Is Not Truly Intelligent

00:41:12 — Intellidynamics: Judge AI by Its Outputs

00:48:53 — Debate: Could AI Outperform a Human CEO?

00:56:33 — Asian Dad: AI Can’t Do Genuine Innovation

01:08:14 — What’s Your P(Doom)?™

01:13:17 — “Are You Calling Me a Scammer?”

01:14:41 — Asian Dad’s Mainline AI Doom Scenario

01:20:53 — Will AI Wipe Out Humanity by 2050?

01:29:29 — Why Is AI Doomerism “Getting Out of Control”?

01:33:11 — Debate: AI Lacks Agency

01:42:13 — Asian Dad Warns of “Techno-Feudalism” & Transhumanism

01:48:12 — Wrap-Up

Links

Asian Dad Energy (YouTube channel) — https://www.youtube.com/@AsianDadEnergy

Asian Dad Energy, “AI Doomerism Is Getting Out of Control” —

Brett Hall vs. Liron on Doom Debates —

David Deutsch, The Fabric of Reality (Amazon) — https://www.amazon.com/Fabric-Reality-Parallel-Universes-Implications/dp/014027541X

Robert Wright, The God Test (Amazon) — https://www.amazon.com/God-Test-Artificial-Intelligence-Reckoning/dp/1668061651

OpenAI model disproves the Erdős unit distance conjecture — https://openai.com/index/model-disproves-discrete-geometry-conjecture/

METR, Measuring AI Ability to Complete Long Tasks — https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/

Transcript

Cold Open

Asian Dad Energy 00:00:00
It’s more like parroting of intelligence. Actual novel innovations — AI just can’t. You can have the most god-tier large language model, but its initiative has to be given to it by a human. But the thing is, the way that we structure this civilization makes it almost comically easy for AI to essentially take over large parts of it.

Liron Shapira 00:00:27
Your fear level is calibrated, but then why are you talking smack? Because you made a video called “AI Doomerism is Getting Out of Control,” and you said, quote, “There’s no shortage of scammers and grifters out there who want to package AI doom as a product capitalizing on anxiety for profit.” Are you calling me a scammer?

Introducing Asian Dad Energy

Liron 00:00:54
Welcome to Doom Debates. My guest today is a former big tech developer with an Ivy League education who calls himself Asian Dad Energy. He and I share a lot in common. We’re both software developers with over 20 years of experience, we’re both relatively new YouTube creators who focus on AI, and we’re both dads who are feeling a lot of existential dread.

But that’s also where we diverge. I dread AI extincting humanity. Asian Dad, not so much. So what happens when my Israeli American dad energy collides with Asian Dad Energy? There’s only one way to find out. Asian Dad Energy, welcome to Doom Debates.

Asian Dad 00:01:34
Hey there, Liron. Thanks to be here.

Liron 00:01:36
All right, so — you’re okay if I just call you Asian Dad, right? I’m not being racist?

Asian Dad 00:01:41
No, it’s totally fine.

Liron 00:01:43
Okay. All right, sounds good.

Asian Dad 00:01:45
It’s actually a little interesting because it almost sounds like the way Asian family members would call their uncles. It’s very interesting, but you have to be one to feel the vibe. But it’s great. Please go ahead.

Liron 00:01:57
Okay, will do. One thing I know from your channel is you’re not just some ordinary father. You’re bringing Asian Dad Energy to being a father. What does that involve?

Asian Dad 00:02:08
I think it’s really just starting to question your values after decades of hard work in an industry, and just feeling like, hey, things are changing. Perhaps the system that we’re living in may be breaking even. And so this mixture of optimism, anxiety, hope, and dread.

Liron 00:02:28
Well, I’ve heard of the tiger mom. Are you a tiger dad?

Asian Dad 00:02:31
I try not to be, but sometimes they just sort of come out at times.

Liron 00:02:37
How old are your kids?

Asian Dad 00:02:39
So my boy, he’s turning 15, and my girl, she’s eight years old.

Liron 00:02:44
Nice. And are they on the Ivy League track right now?

Asian Dad 00:02:47
Oh my goodness. That is such a good question. So in my heart, I want them both to go to Ivy League universities. That’s what I would like to see happen.

But in reality, I’m starting to question the viability and really the utility of an Ivy League degree at that point in the future when they come out. There’s just so many socioeconomic factors at play here that sort of makes me question, is it still going to command the kind of premium in the jobs marketplace at that point? Would there be a jobs marketplace at that point? Who knows?

Liron 00:03:22
Right. Yeah, exactly. It feels late to me to care about Ivy League for a 13-year-old. You went to school at an Ivy League, and you worked in big tech for 25 years. Give us the one-minute backstory of your career in tech.

Asian Dad’s Career in Tech

Asian Dad 00:03:36
Oh my goodness. I’ll try to shrink it into one minute. So I was interested in software and computer programming when I was in high school. In college, I did some contracting work, small projects. Coming out of college, I joined as a software engineer into a series of consulting companies. I spent the majority of my career essentially doing consulting.

You went through the regular sort of tech progression, going from a junior developer to a senior developer to a tech lead and then an architect, eventually sort of director leadership levels. About seven years ago, I had a crash-out with consulting, where I was in a position where I was essentially working a lot with C-suite leadership of various clients on their overall technical strategy and roadmap, and that just felt so detached from the real day-to-day solutioning and technical hands-on work that I had to get out and get into, well, in this case, big tech. So I joined big tech as essentially a senior-level architect, and there I worked for seven years before being laid off, and that’s how I got here career-wise.

Liron 00:04:47
And I think you mentioned also on the compensation front, you were always making pretty good money, 100K plus, and then when you got into big tech, you were surprised because it really started rolling in — 200K, 300K, that kind of level — and then you became unemployed.

Asian Dad 00:05:01
Yeah, pretty much. It was sort of like, hey, I could at least have an understanding of, hey, my work is translating into income in this linear progression. And then I joined big tech, and suddenly I was exposed to RSUs and ESPPs and bonuses and all this stuff, and it doubled, more than doubled, and then the RSUs doubled and then tripled, and this sort of snowball from an income perspective.

It’s interesting because I think the year before I was laid off was my highest grossing year, and then the next year I was laid off. So that’s that.

Liron 00:05:37
Yeah. That’s an interesting graph. Do you feel like you’re a 10X engineer? Are you naturally talented at engineering, or do you feel like you’ve been in the middle of the pack? How would you rate yourself?

Asian Dad 00:05:47
It’s a really good question. I think it depends on the timing. I think there was a point in my life, really in my late 20s, where perhaps I thought of myself as a 10X engineer. You get to this point where you’re really proud of your own code craft and your ability to solve problems, and you think, “Wow, yeah, I can code anything. Just give me anything, I can code it out.”

I would go to these coding competitions that the company I worked with would participate in, and I went to more than half a dozen of those over those years, and every single one of them, either myself or the team that I was part of would come out in the top three. So that’s a data point in terms of skill. So probably maybe a little bit better than average skills, but in my mind, I was a 10X engineer. But since then, I would say in my 30s and 40s, not as much. I wouldn’t call myself a 10X engineer at the time of my layoff.

Liron 00:06:47
It sounds like you’re good on the discipline and the focus on the career, so you were able to add a lot of value. I’m also curious on the aspect of how young did you start programming. Were you a math nerd in school?

Asian Dad 00:07:00
Oh, that’s an interesting question. So I liked math. I wouldn’t say that I was the absolute best at it, but I was good at it. My programming journey actually started with an IBM 386 computer that my dad got. I think I was 12 when we got it. Initially I just played computer games on it, games like Space Invaders and that kind of stuff, Tetris.

But then it got into, hey, how do I tinker with this stuff? How do I configure it and modify it? And that got me into learning — this is stupid, but there’s this language called Visual Basic in the ‘90s where you can make these really silly CRUD apps that run on Windows. That was really my first language to actually build real things versus just scripting or tweaking things. That’s how I got into it.

Liron 00:07:54
Yeah. I love it. I was also a ‘90s kid doing Basic and then Visual Basic, so I totally know what you mean.

Asian Dad 00:08:02
Yeah.

Liron 00:08:02
Okay, great. And also remind me, are you a first-generation immigrant?

Asian Dad 00:08:06
Yes, I am.

Liron 00:08:07
All right. When did you move to the US?

Asian Dad 00:08:09
So I came over here when I was six years old. This is in the early ‘90s.

Liron 00:08:16
Gotcha. Okay. Well, it turns out our stories are pretty similar because early ‘90s, I moved here when I was about three. And I also did Basic and Visual Basic, so we were really living parallel lives here, except I never joined big tech. I just worked for one small tech company and then got into the startup entrepreneurship world.

So that’s where our paths diverge, and I think one thing we were just talking about before the show is that I happen to be in this position where I haven’t lost my job because I own the company, so I’m not getting fired. I’m able to enjoy — when the AI is writing more code, I’m just able to benefit from that, and then I’m the one who doesn’t have to hire somebody. So I’m benefiting from having no job postings. I’m unemploying somebody else who would’ve worked for me, but instead I’m using the AI, and then you’re experiencing the other side of that, where you’re looking for a job, and all the companies are like, “No, we’re good. Thanks.” So talk about the unemployment, the job landscape right now.

Getting Laid Off: The Job Market Is a “Hellscape”

Asian Dad 00:09:11
When I was first laid off, I actually spent several weeks seriously searching for jobs. This was more out of inertia than anything else because you’re used to working a job. I’ve worked for decades. And so I went pretty heavily into the job search. It was, for the most part, crickets. I did get a couple of screeners, and then I sort of hit a mental inflection point where I realized —

A little bit of background: after I got into tech, there was this great financial crisis in 2008, 2009, and that scared the bejesus out of me and my wife and our family, and we started becoming more frugal. Not like we were misers or anything, but we sort of lived reasonably, and we were able to save and stash away a lot of money into investments. And so because of that, just purely on paper, the returns on our investments would cover our normal daily living expenses.

Asian Dad 00:10:13
I never actually thought to essentially quit or retire early. But after the layoff, past that initial wave of job searching, I got to the point where I realized, “Wait, I’m not hearing anything too much from the jobs market. It seems like a hellscape to me. At the same time, there’s actually no pressure for me to find a job either because my bills are all paid. So why don’t I just hang out and pursue some other interests for the time being?”

And that sort of weird middle spot is where I am right now. I’m pursuing a lot of different interests and hobbies, and I’m still coding. But I can’t say that I’m actively job searching and sending out thousands of applications. A lot of my coworkers who were affected by the same wave, they had to do that, and unfortunately, quite a few are still doing that right now.

Liron 00:11:06
Yeah, so you did something really smart. You lived a low cost of living life. You always left a healthy margin between your income and what you need to live, and you built up savings. I think I’m more like most people, where I haven’t done a great job of that. I’ve embraced the high cost of living life. I have a big house. I’ve got a big family. Big family is a good status symbol these days because people like to spend a lot on their kids.

And I tell myself, on one hand, I totally understand what you did — “Yeah, let’s keep a low cost of living so it opens up our options.” But in my defense, I also have a high P(Doom), so I do think it’s worth shifting some consumption into the present. So I think I have a good rationalization to continue my high cost of living lifestyle.

Asian Dad 00:11:50
Yeah. The future is unwritten. We all have to balance our various risks and trade-offs and the potential for future enjoyment and happiness, and everybody’s balance is gonna be a little bit different.

Liron 00:12:03
All right. So for this conversation, I come at all this as an AI doomer. The focus of the show is talking about how AI is going to potentially kill everybody very soon. You’ve expressed a lot of skepticism about that on the show, so I definitely wanna get to that. But first, I wanna ask you a few more questions about what AI is doing to the job market and productivity, and then we’ll segue into the doom stuff. Sound good?

Asian Dad 00:12:28
Sure. Yeah.

Software Engineering in 2026 vs. 2021

Liron 00:12:29
All right. So you mentioned that the job market is tightening up, and I feel like in contrast the peak we can really compare this to is probably 2021. Is that fair to say?

Asian Dad 00:12:40
Yes. I think the market probably went into overdrive in 2021. It kind of stopped right at the end of 2021, with the great resignation and everything that entailed.

Liron 00:12:53
Right. 2021, that was the crypto boom, and that was coming after COVID, where there were zero interest rate policy. I specifically remember 2021, I’m like, “Ugh, I’ve got bonds, I’ve got cash. I gotta get rid of this cash. Cash is trash. It’s getting inflated away.” And that’s why people are like, “Well, Bitcoin is the answer. Crypto’s the answer.” And I’m like, “I don’t think that’s the answer either.” So I shoveled a bunch of money into stocks, which luckily — then in 2022 everything crashed, but then it did eventually go back up. But of course my trading decisions were the worst timing naturally.

So that was that whole 2021 era. But like you were saying, when it was 2021 and all of these crypto companies and other kinds of tech companies were getting funded, I think that was past the scooter era, but there were all kinds of random moon projects. People were funding any kind of software, basically. They’re like, “The best way to get rid of my cash is to give it to some engineers in a dorm room doing anything.”

And so then everybody was turning around on the job market and having these super high-income, high-equity jobs because if you could get engineers, you could raise money, and there was nothing else to spend your investment money on. So I got really cocky. I was like, “Man, it’s so easy to get a job. You could make so much money having a job. Being an engineer is the greatest asset ever.” And that has kind of gone away, where now it’s like, oh no, we’re just competing against the AI like everybody else.

Asian Dad 00:14:09
Yes. In a way, the AI proliferation hitting the tech sector first, and especially software engineering — in a way it’s almost deserved. Now, I don’t mean deserved, but in the sense that I feel like software engineers perhaps were a little too proud of ourselves right before the whole large language model wave hit. Just my personal observation.

Liron 00:14:35
Totally, yeah. It just felt like it was such a rock solid option to be a software engineer because I always told myself, “Compared to 90% of humans — 90% of humans just really struggle to pay attention to a giant block of code, to get in the details and model what’s going on, and the world needs a lot of code, so this is the greatest defended position. Nothing can undermine this position of being a software engineer.”

But then funny enough, we have this blindsiding attack from AI that’s now very much able to do software engineering, and the total IQ points — now it’s like there’s more AI intelligence in the software engineering world than there is human intelligence. And so turns out it is pretty hard to get a job. Is that a fair assessment?

Asian Dad 00:15:15
Pseudo — I would question that. Pseudo intelligence, but I would say the quantity of it has a quality all its own. But yes, that has a major impact on the software engineering marketplace. Huge impact, in fact.

Liron 00:15:30
Yeah. Do you blame AI for your layoff?

Asian Dad 00:15:33
I don’t blame it directly. I think there’s an indirect causality there, because the company that I worked for, they were and are investing heavily into AI. So at some indirect level, you gotta think, “If we need capital to invest in AI algorithms, training, data centers, and things like that, you have to get the capital from somewhere.” Headcount could be an easy source. Reduce headcount, get capital. So not directly, but perhaps there’s an indirect relationship there.

Liron 00:16:08
That’s an interesting perspective, because you’re simultaneously saying, “Okay, AI is coming for the job market,” but you’ve expressed a lot of skepticism that the AI investments are really gonna create that much value, correct?

Asian Dad 00:16:23
So I really want to separate financial investments from the underlying technology. Meaning the whole question of, hey, are the current AI companies as investments gonna make crap tons of money, versus is the underlying large language model and agentic technology going to transform the economy?

One can be false while the other is true. It could be that this wave of AI companies crash and burn, but the technology that they develop lives on and transforms the economy. I’m just pointing that out.

Why Are They Investing Trillions in AI?

Liron 00:17:03
Okay, so what’s your economic outlook? All this money’s plowing into data centers. They’re expecting big returns because they’re expecting AI to at least increase the productivity of humans. As a single overview question, do you think we’re going to get trillions’ worth of economic value from these AI investments or no?

Asian Dad 00:17:22
I think that’s their hope. And it’s my belief that given the value of the investments that’s already committed, what the real hope is is a kind of labor cannibalization strategy. So hear me out.

They have dumped at this point probably single-digit trillions in terms of investments into AI, overall, ballpark for the United States. What possible new wealth source right now on the horizon is worth trillions? The biggest that I can think of — because think about it, for every dollar in investment that people give, if you were an investment company, generally the rule is you want to multiply that many times, 10 times if possible.

What single sector out there do you think has more than $10 trillion? Well, one sector that’s available is the white-collar labor sector of the United States. That’s a decatrillion dollar sector right there annually. So it’s my reasoning, and perhaps it’s a false one, but it’s my reasoning that really the payoff that these AI companies are thinking is the replacement of white-collar human labor on a sector basis, and that can then reap trillions in savings.

Liron 00:18:45
So I completely agree, except you don’t have to call it the replacement. It may be the replacement. I think it likely will be a lot of replacement, but you could also call it augmentation. So if you took the US population and you just bolted on another 100 million American adults with a good education who could do white-collar work, what happens to US GDP? It’s going to go up by many trillions, correct?

Asian Dad 00:19:10
Well, it depends on the nature of the augmentation. So it’s entirely possible with the current state of the technology — and I don’t wanna go super deep into the weeds of the technology — but it’s possible to augment vastly a portion of the population.

But the problem here is that most of the population, they’re essentially doing gig work, lower-level service work, or lower-value cognitive labor. That sort of labor, the complexity of it, a good portion of it is entirely replaceable. So what you’re saying about augmentation versus replacement, it’s not “or,” it’s “and.” There will be augmentation for a certain portion of the human labor force with AI. There will be replacement for a certain portion of the human labor force with AI. Both may well be happening at the same time.

Liron 00:20:09
Yeah. Maybe a simple mental model of this is if 100 million new immigrants came in and they were charging the same price for their labor as Americans, I still think GDP would go up massively because we’re just a bigger country. People are a huge factor in what makes countries rich. People on average create a lot of wealth. But that’s if they charge the same price. And then there’s the factor with AI where I claim AI is more and more charging way less than humans. So what happens when a bunch of immigrants move here and their cost of living is just always going to be much lower than ours as flesh and blood?

Asian Dad 00:20:43
You can actually see that in most towns in America. I’ll give you a concrete example. So a bunch of, say, illegal immigrants came into town. They’re willing to work for $4 an hour doing whatever — gardening work and lawn care and so on. You can actually see this effect. The American that can employ these immigrants as a company and go around the neighborhood selling their services, that American has augmented himself. He’s made way more money running the company.

But there’s probably another set of Americans who used to do this lower-value labor, but they may charge minimum wage at least, and these guys are doing it for $4 an hour. They’re now out of a job. Both could happen at the same time.

Liron 00:21:28
Exactly. So you bolt on all of these AIs who are willing to work for much cheaper, and the hope is that the economy grows so much that a bunch of higher-wage jobs still get created just because we have all this cheap labor, and so we’re getting more ambitious.

Which I’m personally actually feeling. As a startup founder, as an entrepreneur, I think it’s fair to say I’ve now doubled my ambition. I’ve doubled my roadmap for all the things I’m planning to get done this year for my company, and it’s becoming plausible that I can do more things at the same time. It’s easier for me to do my startup and Doom Debates right now at the same time, thanks to the power of AI, and maybe I’ll start a second company and I’ll generate more economic value that way, and I’ll employ more people that way, even though I’m also employing a bunch of AIs.

So there is a plausible argument that a bunch of cheap, quote unquote, “immigrants” are going to come into our economy and add a bunch of GDP just by paying them, but also add a bunch of GDP by opening new jobs for humans.

Asian Dad 00:22:23
It is entirely plausible that that happens. Don’t get me wrong, I actually expect if this AI transformation of our economy pans out, I expect the cumulative GDP to be much larger. So the overall bucket of goods and resources and services that our civilization can produce will get much bigger. And to your point about, hey, could there be new jobs for humans being created? Absolutely. I think it’s true, because people with initiative, with agency, they should be able to leverage AI to do great things.

Just as a hobby for fun, I’m creating software apps and SaaS services that normally would’ve taken teams months to do and I could just do it in a day. That inevitably will unlock opportunity for some people, and I will critically say some. And that’s because the vast majority of the population — I just wanna point this out — I don’t think everybody just wakes up and says, “I wanna be an entrepreneurial CEO. I wanna build an empire and go do this stuff.” A lot of people may not have the desire or the intent or the aptitude to leverage these capabilities. And when I say a lot, I mean perhaps for every one person that will, maybe 10 people won’t.

Liron 00:23:50
Right. Okay, so just closing out your original point — why are they investing trillions in this? How is it going to pay off trillions? And then my answer was, well, it’s some combination of augmenting the economy and replacing labor.

So my conclusion from all this is, yeah, they’re investing a few trillion and they’re probably going to make a good return. They’re probably going to make even more trillions, which is why I’m overweight in my stock portfolio. I’m buying plenty of Google. I’m buying plenty of these companies that are part of the AI supply chain because I think the pie is going to grow.

I think there’s going to be a lot of winners coming out of this if we just look at the next few years where AI is replacing and augmenting human labor. I actually am really bullish on the economy, and then we’re gonna get to the doom part because afterwards I think it’s all gonna collapse. But my short-term economic outlook is very bullish. Would you disagree with that?

Debate: AI Is Different Than Other Tech Bubbles

Asian Dad 00:24:37
I kinda disagree with it. I think at this point there is a matter of financial hype with a particular technology that a lot of technologies, by the way, have this kind of hype bubble mentality. The internet bubble is one example. If you go back even further, there was this bubble for railroads. It’s similar.

The basic idea is you have a transformative technology show up. Everybody sees it, and they know, oh, in some period of time in the future, civilization is gonna get totally transformed by this technology for sure. Everybody sees that.

Asian Dad 00:25:14
And then they start putting money into it with this thinking that, oh, in the very near term, just around the corner, all of society will be transformed. And then what inevitably happens is that that transformation comes, but it comes later than the investors expected, and you get this bubble, and you get a huge crash.

My guess here is that this current iteration of the AI bubble — and it’s a guess — it’s gonna crash before the dividends of the underlying technology are realized. And then some years after the crash, that technology will proliferate and adapt across industries and across the economy, and then we’ll actually reap the benefits of it. That’s just my personal opinion.

Liron 00:25:59
So I think you’re right about previous bubbles and previous tech steps upward, where the timing — the people get overexcited and it takes longer than they think to fully play out and capture the value. So you have people investing a decade ahead. I think you’re correct about previous times —

Asian Dad 00:26:18
But this time is gonna be different?

Liron 00:26:18
Well, I think we’re seeing evidence that this time is different because when all of these things are coming online, they’re immediately being purchased. Every time there’s a new Blackwell chip or whatever, NVIDIA’s next chip, there’s already a huge line out the door. They can’t get these things out the door fast enough.

And the companies that are plugging the chips in the data center, they already have customers. I’m an example of the customer. I’m using it to create some unrelated non-techie business, but I’m using the tech. So this is all fully integrated into the economy today, even before you account for new companies that are entirely run by AI or one person making a billion dollars with a team of a hundred AIs. That stuff is barely coming to fruition, but that’s only the next wave. There’s already the current wave.

So this is a disanalogy because you mentioned railroads. Okay, yeah, they built a bunch of tracks before there was that much commerce happening on the tracks. Or in the internet age, they laid all this fiber, and the fiber was dark. Google famously was able to buy a lot of cheap dark fiber because they overbuilt because end users weren’t streaming video or providers of video weren’t putting video into this fiber. It took a long time for the fiber build-out to come to fruition. But the AI build-out is coming to fruition one day at a time.

Asian Dad 00:27:31
I think the critique with this idea is that as a technologist — I’m a technologist, we’re engineers — we’re sort of living in the middle of this technological transformation, and it’s very clear to everybody around me that all the tech companies are adopting AI very quickly and utilizing it into their workflows and everything.

But I think the key thing is I believe in the broader economy, outside of big tech and the overall tech sector, adoption of AI is not nearly as intrusive and as deep as in the tech sector. I can create skills and workflows that pretty much do everything digitally. But if you go to, say, a random S&P 500 company, and if you’re able to get into security and then take a look at their operational workflows, I’m gonna bet you that 99% of it is not AI. They’re probably using straight-up RPA or even human processes right now.

The reason why I’m pointing this out is that there’s an echo chamber effect where a portion of the economy is sort of accelerating in the adoption and usage of AI, and then another larger portion is almost being left behind a little bit. And it gives this false appearance that adoption and progress is faster than it actually is across the broader economy and society.

It’s my belief that the vast majority of society, their only awareness of AI is sort of the spam copy that’s generated by ChatGPT and funny pictures and things like that, maybe TikTok videos that have been transformed by AI. But in terms of the deep work that it can do in terms of software engineering, automation, controls of different physical environments and workflows, RPA — that is obfuscated from the majority of the population and the majority of companies even.

Liron 00:29:47
Right, which is bullish. Because you’re already getting saturation, supply is already getting saturated, and you’re saying, “Hey, all these companies still haven’t gotten on the AI train yet,” so isn’t that bullish?

Asian Dad 00:29:57
But the investments by those AI companies, they gotta pay off before the investors want their returns back. It’s time-bound. You see what I’m getting at? The adoption is gonna take time because humans take time to do things. But the money needs to generate interest sooner than that adoption curve.

Liron 00:30:17
Okay. I mean, to resolve this disagreement, I would have to dig in and be like, “Look, they invested one trillion. Here’s the return pipeline for that one trillion. It’s actually only going to take a few years to turn that into two trillion, and then the next one trillion is investing farther out.” We would have to dig into the financial weeds. I’m personally confident that it looks really good.

Asian Dad 00:30:35
It’s almost like you have to wait to see how this plays out. At this moment in time, the investment is surely unprofitable. It’s negative right now. It’s making money, sure. Anthropic is making money —

Liron 00:30:58
Yeah, 99% of it, that’s how it’s supposed to look. You’re supposed to invest and then wait a little bit and then get more. But the times to return and the multiple on return looks really attractive.

Asian Dad 00:31:04
That’s right. It’s not a question that with current data I think you can get an answer to, because it’s almost like saying everything is gonna be negative profitability right now for AI investments. And the question is, could overall demand and adoption grow quickly enough to justify the investments before that investment bubble pops?

Liron 00:31:20
Yeah, so just to make it concrete, let’s talk about your nightmare scenario. Meta does this huge build-out of their Manhattan-sized data center. So your nightmare scenario is that they never make it back. Let’s say it’s a hundred billion dollar project. Your nightmare scenario is, yep, they sunk a hundred billion dollars into it and they just never pay back, they never get two hundred billion back. It just kinda sits there and depreciates. That’s what you’re saying is a possibility?

Asian Dad 00:31:43
It’s a possibility, and it’s not just the data center at that point. Let’s suppose it sat there for a year, just give it a year, without, say, single-digit utility. Let’s suppose. Well, within that year, all the investors that invested in Meta, they’re gonna see that, hey, there’s no profitability and just sunk cost at this point. I’m not gonna keep my money parked here like this while I’m just losing money. That’s really the concern.

Liron 00:33:03
Yeah. So — I mean, I don’t wanna keep beating the dead horse. We can move on. Let me just say what my claim is, so that way the future can judge me. My claim is just that right now we have revalued how much these data centers and these chips are worth. It turns out there’s a brain in there. We know how to put a brain in there, an intelligence in there, and that is worth a lot more than the cost of materials, a lot more than the cost of power, a lot more than the cost of construction.

So because of that, I see it as actually incredibly hard to lose money doing these projects. I see these projects as free money. I would be shocked if they end up losing money because they can’t harness those chips. So I feel really good about these projects, and I think that’s what a lot of savvy investors are saying as well. But your position — just to restate your position, we don’t have to keep arguing — but your position is, “No, there’s a significant chance that they’ll be surprised and they just won’t get value out of this.”

Asian Dad 00:33:03
I think there’s a significant chance. And I would advise — and I’m not a financial advisor — I would still advise some degree of caution and not just jump to the wind in terms of AI. For example, investing in the companies that own the big data centers and make the chips, of course, I’m invested in that as well. You do so using index funds like the S&P 500. You of course wanna ride the wave up if there is a wave.

But at the same time, I don’t think one should consider that this is a done deal. I think if we look at history and we look at past waves of technological innovation, this whole bubble, hype, crash, adoption — this pattern has just happened so many times that I don’t think this is gonna be an exception. Just my personal opinion.

Debate: AI Is Not Truly Intelligent

Liron 00:33:52
Fair enough. So before we head into what I call the doom train, talking about humanity’s imminent extinction, debating that — let’s talk about what’s going on inside these data centers. What is the AI actually doing? Maybe a good place to start is, I’ll ask you, what do we mean when we say artificial intelligence? This is general intelligence, this stuff’s coming online. How would you explain it, and do you think it’s overhyped or do you think it’s really intelligent?

Asian Dad 00:34:21
Sure. Honestly, the way I look at it — and when I say artificial intelligence, I really mean the large language model version of it. There are many different types of AI algorithms that are out there. Before neural networks, there was alpha-beta and minimax and decision trees and all these sort of things.

But it’s essentially a tool, a technology that replicates human forms of reasoning and problem-solving, and it’s a replication, a simulation up to this point. I don’t believe, from what I know and from what I’ve understood of the current state of large language models, I don’t believe we can call it intelligence per se. It’s more like a simulation or a parroting of intelligence, but a good one.

Liron 00:35:17
So the thing about simulating certain acts, like software engineering, is that you can take the output of the simulation and run it in the real world. So is that really a simulation or is that just doing software engineering?

Asian Dad 00:35:29
It’s doing software engineering, but I’m not talking about software engineering. I wanna make sure that point is clear. If you look at intelligence as we understand it, there are certain aspects of intelligence, like the ability to reason, the sort of autonomy that occurs in the minds of every human being and those kind of intrinsic properties — the current technology that is large language models, they don’t do that. They do other things.

If you think about a large language model, it’s a probability machine. It’s essentially a ginormous, unbelievably huge matrix function that predicts the next word, the next token based on input. That isn’t in its nature bound to actual reasoning. Reasoning works on different principles. But these two different paradigms can result in output that is comparable, or perhaps the output of the large language model could even be better than the output of a human mind that can reason. So I just wanna point that out — there’s the intrinsic intelligence, and then there’s the output of this intelligence, or in the case of AI, pseudointelligence.

Liron 00:36:53
Yeah, this reminds me of a debate I have pretty regularly with the camp I call the Deutschians. Are you familiar with David Deutsch’s views on all this?

Asian Dad 00:37:01
I vaguely remember his name. I didn’t check out anything that he wrote, unfortunately.

Liron 00:37:07
Yeah, he’s a famous physicist, inventor of the field of quantum computation. He’s got a really popular book, The Beginning of Infinity, and a book I like even more, The Fabric of Reality. I hugely respect him. He’s undoubtedly a genius.

Except on the subject of AI, I personally am not finding much insight in his statements, but there’s a camp of people who really think that he’s nailed it. A famous devotee of the Deutschians is Naval from the tech industry. He’s had him on his podcast, and this guy Brett Hall actually came on Doom Debates, and we debated Deutschian ideas. And these guys Ben and Vaden from the Increments podcast also had a spirited Deutschian debate.

So what you’re saying right now is a statement I’ve heard from the Deutschians, which is, “You gotta look inside the AI, man. Sure, the output could look equivalent to a human being brilliant, but what’s going on inside the box is just fake. It’s not really reasoning. It’s just trying to predict what a human would do. So ultimately, the real definition of intelligence is the circumstances of how the algorithm inside the black box is processing, and don’t be fooled by these brilliant outputs that come out of it.” Is that basically your perspective?

Asian Dad 00:38:20
In many ways, yes, and I just wanna clarify something. The output is not fake, obviously not. The output is definitely useful. So you could have a machine that doesn’t have true intelligence or reasoning, but it could still do cognitive work in such a way that it totally disrupts and transforms our civilization. Both can be correct at the same time. But it can do this without ever being conscious.

Liron 00:39:01
Okay, so — I didn’t even mention consciousness, but do you wanna say how you think consciousness fits into all this?

Asian Dad 00:39:01
So if you take a step back and think about a large language model that’s running in a data center somewhere. It’s consuming memory. It’s consuming CPU cycles running in the base state. Whoever’s running it — Google or OpenAI or Anthropic — they probably have some sort of monitoring software running on it, probably something like Datadog or a New Relic monitor on it.

If they’re looking at that large language model, and there’s no prompts being fed to it, what happens? Nothing happens. Literally, you can actually see it. Nothing happens to the state of that memory container running on a server at all. There’s no state changes. There’s no computation, nothing at all.

Asian Dad 00:39:48
Now, if you take a human being, and suppose we took you, we strip you naked, and we put you into those sensory deprivation tanks. I don’t know if you heard about those things, but in the ‘70s, there were all these experiments where they took —

Liron 00:40:03
Yeah, I’ve tried a float spa. I don’t know how different that is.

Asian Dad 00:40:03
It’s very similar. The temperature would be set to the temperature of your body so that you don’t feel any temperature differences. You’re floating, so you don’t feel any gravity. There’s no light. There’s no sound. But they do plug sensors into your head to monitor your brainwaves.

And essentially, even under total sensory deprivation, people have thoughts and activities continuously in their brains. Some people actually talk about having visions and crazy wild-ass hallucinations while there. But the point I’m getting at here is that innately in human intelligence, there’s this internal agency. You can think of it as intentionality — the ability to proactively out of nowhere form plans and act on it, the ability to set goals, be able to self-reflect on things that have happened or memories of the past completely spontaneously in our heads. That kind of introspection or agency, from an observational standpoint, does not happen with large language models today.

Intellidynamics: Judge AI by Its Outputs

Liron 00:41:13
Okay, so from my perspective, there’s a major, major mental reframe that so many people — the vast majority of people — need to make. I would say you need to make it, and the Deutschians need to make it, and even the people working at AI companies need to make it. So it’s nothing personal to you.

But the shift is precisely what I was saying before about considering the outputs versus trying to open the black box and judge how the gears are turning. “Oh, it’s just matrices. It’s just linear algebra. It’s not truly conscious.” Instead of focusing on what’s inside the black box, look outside. Look at the input/output.

I claim that there’s actually a whole study you can do on just studying the outputs, just studying the input-output relationships, and I’ve even given a name to this field of study that is now in an upcoming book by Robert Wright called The God Test. So if you ever wanna cite this concept, you can use Robert Wright’s book. You don’t have to use Doom Debates.

Liron 00:42:05
The principle, or the field of study, is called intellodynamics. That’s the name I’ve coined for it. I’ve seen people call it agent foundations before. I think intellodynamics is actually a better chosen term because it’s analogous to thermodynamics in the study of engines doing physical work.

What is an engine doing? Well, your tendency, I would claim, is to open up the black box of the engine and be like, “See, if you wanna have an engine, you gotta have these gears turning. You gotta have the steam going here.” And I’m like, “Wait a minute. Can we just look outside the engine? Can we just see that it’s doing work, as in force times distance, as in energy?”

Because once you see that, it turns out that a beam of light can do useful work. Sound energy can do useful work. Chemicals can do useful work. Once you have this perspective of what is the work being done, you stop being so obsessed thinking you know a lot about what has to be in the black box, and you can just stand back and look at the shape of the work being done. I think that’s what’s missing here, and you may be subject to this bias. What do you think?

Asian Dad 00:43:05
It’s interesting. It really sounds kind of like being asked to believe in magic. And I know this sounds a little insulting — I don’t mean it in any way or form. But it’s sort of like we can’t just suspend our logic and our reasoning to believe that, hey, something that happens in a black box is magic. We gotta open the box and see what’s inside.

Liron 00:43:32
I’m not just saying there’s some mysterious force here. I’m happy to explain what intellodynamics involves. For one thing, you can tell that useful cognitive work is being done. Just like you look at an engine — I can tell you useful work is being done because I say, “Hey, this room has no other source of energy in it, and yet you’re pushing this boulder.” You’re pushing this weight, so it looks like work is being done. And if you’re able to push it for an hour, wow, you sure did a lot of work. You must be converting energy pretty efficiently. So that would be the thermodynamics equivalent.

The intellodynamics version of this is, hey, I have this goal — I have this outcome, winning at chess. That’s a narrow outcome in a wide space of states on the chessboard. If the engine can win at chess, okay, well, you’ve performed optimization work. You’ve performed outcome-steering work. And I think it’s important to be sensitive to when outcome-steering work is happening before you even look inside the black box and be like, “Okay, so how did this outcome-steering work happen?” You have to respect when new heights of outcome-steering work are being achieved.

Asian 00:44:32
On this perspective, I do agree with you, in that the output of AI, we have to respect it based on its objective quality. And so it’s true, AI could produce cognitive work that is matching or superior to humans, but it’s in the same nature as saying a calculator can compute with higher precision and accuracy than a human could. That is true. That’s an objective fact. It probably, when used correctly, makes less mistakes than humans, but a calculator is not human consciousness.

There’s essentially a step change between the two things. Without a human being having the conscious ability to perceive of our existence and the nature and reality, there is something to be said that AI will never do anything. It may have all this possibility of creating useful output, but without the consciousness to set goals and have the agency of asking it to do something, it will never do anything.

Liron 00:45:42
Well, what if you just give it a prompt and run it in an agent loop?

Asian 00:45:45
But you had to give that prompt.

Liron 00:45:48
Okay, but don’t you think at some point somebody’s gonna give it a prompt?

Asian 00:45:51
That prompt, that initial prompt or however we feed it — if you use OpenClaude, for example, you can feed it using a cron job indefinitely. You could. But the origination of that thought, of that will and agency had to come from a genuine intelligence.

Liron 00:46:08
Let’s, for the sake of argument, say the initial prompt had to come from a human. Okay. So I think that at some point somebody will prompt a super intelligent AI and press Enter, and then what’s gonna happen?

Asian 00:46:18
Well, they have to then do something and then keep going. The point I’m saying here is that for a prompt or any input that doesn’t reach a conclusive end, there has to be this internal agency, and that’s what AI lacks. And I know what you’re thinking — well, what if I just keep having a loop and you randomly generate prompts and so on and so forth, and then this machine will keep churning out output. But that’s not consciousness.

Liron 00:46:50
Well, the prompt can just say, “Make me the dictator,” right? So you really only need one prompt, and then the loop does the rest, and the only question becomes can the loop steer outcomes better than humans or not?

Asian 00:47:00
I think the real question or test in this way would be creating some sort of an input vector that can generate the experiences of a human being or the experiences of the real physical world in such a way that it can be converted into input for a large language model. I personally believe that anything like that would generate gibberish using current state of technology. It would just massively hallucinate given the kind of experience that comes in.

But long story short, this is almost in my mind a philosophical debate. We’re steering far from the technology or even biology, and it’s more metaphysics. It’s sort of the belief that an aware being that’s aware of its own existence — is that something special over another thing that you don’t know if it’s aware or not, but all external observations tell you that there’s no internal activities happening? That’s really the critical argument. And at this point it’s not a reason argument. It feels more like a philosophical, almost belief or faith-based argument.

Liron 00:48:18
Well, let me put my cards on the table in terms of where I’m going because I agree that if we’re just asking, “Well, what does intelligence truly mean?” then yeah, that could be a philosophical argument. But when you focus on whether humanity’s about to be disempowered, there’s more concrete questions you can ask and more concrete answers you can get when you just focus on that question. I don’t think that’s a deeply philosophical question. It’s a pretty normal question to ask — who is going to steer this situation more capably?

So for example, a seasoned CEO can probably steer a company better than a random high schooler, correct?

Debate: Could AI Outperform a Human CEO?

Asian 00:48:55
Yes, definitely.

Liron 00:48:55
Okay. So in that same sense, the outcome steering sense, I claim that we’re probably just a few years away from being able to slot an AI in that CEO position or in the senior software engineer position or in virtually any job position, and just have the AI do better than humanity. Would you follow me that far?

Asian 00:49:15
I think the question is whether it would ever be given a chance to run humanity. I actually don’t think we would get to that point anytime soon. Seriously, hear me out. The thinking here is that even in a limited sort of black box scenario where it can just do cognitive work very, very well, and suppose it’s all derivative cognitive work from its training data — I suspect that humanity will take this capability and massively misuse it.

So when we’re talking about AI doom, it’s almost like in some ways I agree with you that AI may doom us and end civilization, but it’s not AI doing it. It would be, in my opinion, humans using the capability to do something to advance their very narrow interests and inadvertently dooming us all. That actually I think may be a plausible future outcome.

Liron 00:50:16
I get what you’re saying, but I wanna factor out the question of the cognitive capabilities within the AI algorithm. I wanna basically just one-to-one compare what can an AI in a data center do versus what can a bunch of human brains do. I wanna do a head-to-head comparison.

Let me try this hypothetical, okay? Say we had 100 years of AI progress, so don’t let time be a constraint here. A really sophisticated future AI, and we’ve put it in a time machine, it comes back to the present. We install it on a present data center. Obviously, present hardware is much weaker than future hardware, but it’s the most optimized, distilled model that runs on a present data center, 2026 data center. And now it’s going head-to-head with a bunch of humans in a 100-person company.

And the company is white collar — they all just work from offices. They’re not working in factories or whatever. So it’s a software company, let’s say. Who do you think is going to have a massive advantage in the business world, the data center with the 2126 era AI technology or the team of 100 2026 era humans?

Asian 00:51:22
I think it honestly has to depend on the nature of the company and the business model. If the company’s business model is something that’s already known or a derivative of a known business model, hands down the AI company is going to be much better. It’ll likely be able to execute faster, find efficiencies much faster, scale faster, everything faster.

The thing though is if the company is doing something that’s novel and not derivative, some sort of derivation of existing knowledge, then I would still put the advantage for the humans. And so that’s sort of where it comes down to. The consciousness and the agency allows us to conceive of and do very novel things. The AI could do interesting things, but those interesting things are fundamentally derivations of things that humans have done.

Liron 00:52:20
Okay, so you say existing business model — so running cloud services, competing head-to-head with AWS, you think AIs could do it because it’s a proven business model?

Asian 00:52:30
In my mind, honestly, anything that’s actually proven — and remember, we’re using the analogy of saying AI has developed for 100 years — so I’m assuming all of the tools, the MCPs, the skills have massively developed. Yes, I think AIs can run existing business models probably better than an all-human crew.

Liron 00:52:51
Okay, and what about taking over a big prominent company like Microsoft? Would you think that it could do better than Satya Nadella, or would you say no because there’s all this human innovation at Microsoft that the AI would get stuck on?

Asian 00:53:05
I think if the company’s money comes from making novel innovations consistently — and don’t get me wrong, human companies can’t do this consistently, as we can see from many companies — that humans will still have an advantage over the AI.

Let’s take a concrete example. Think about somebody like Steve Jobs, and the ideation of the personal computer. Yes, it’s true he stole from HP and all this stuff from Palo Alto. But the point I’m getting at here is that entire sectors, entire novel sectors of the economy, that kind of novelty thinking, going from zero to one — as of right now, it’s in the domain of humans.

Could that change with algorithmic advances that we haven’t even thought about yet? It’s possible. But with the current state of large language model technology, at least, it wouldn’t. So you could have companies run very efficiently by AI, but those companies, if they were to innovate, the innovation would be a kind of evolutionary derivative innovation and not a truly novel innovation, in my opinion.

Liron 00:54:29
Okay. This is a very Deutschian perspective.

Asian 00:54:32
I need to read up on this guy, man. It sounds very interesting. I have to check him out.

Liron 00:54:37
Yeah, exactly. So his — the language that he uses, he has that book, The Beginning of Infinity, and he talks about creating new knowledge. And I think he makes a lot of sense in the context of humans. He says, “Hey, when you look at the whole universe as a matter of cosmology, if you wanna understand what’s going to happen with the universe around us, you better know a thing or two about the human brain because we are going to expand in the universe. We are going to actually determine the fate of the universe. It’s only a matter of time. We are going to keep creating new knowledge and solving problems until we do it.”

And this is a very important thing to understand about physics — the way it evolves and evolutionary dynamics start up and intelligence dynamics start up and knowledge is a thing. So I think he’s absolutely right when he talks about humans. He just seems to then somehow miss it, which is weird because obviously the fact that somebody like me, who is not that level of genius, is able to criticize him is kind of odd. But it’s not super unusual for people to be really smart and then suddenly the domain changes and they’re less smart. So I guess I’m not that surprised.

But he starts inferring that AI can’t create new knowledge the same way, or he doesn’t see any new knowledge creation in present AI. So that’s very similar to what you’re doing now. He uses the term “no new knowledge,” and you’re saying, “Hey, no genuine innovation” — that’s missing from AIs today, and maybe that’ll last. Maybe we’ll go 100 years and AI will still be missing that secret sauce of genuine innovation. Is that basically what you’re saying?

Asian 00:56:01
Yes, and don’t get me wrong here. There is a difference between genuine innovation and derivative innovation. Because everyone knowledge-wise stands on the shoulders of giants. So even derivations of past knowledge could create a lot of commercially viable innovations — that’s totally possible. But I’m talking about actual novel innovations. AI just can’t do it.

Asian Dad: AI Can’t Do Genuine Innovation

Liron 00:56:33
One of the follow-up questions I ask when people say what you just did is, okay, can you think of a recent time when you yourself did genuine innovation?

Asian 00:56:43
Yes. It’s very rare, don’t get me wrong.

Liron 00:56:46
Okay.

Asian 00:56:46
It’s rare to think about something that’s truly novel, that hasn’t been thought about before. I think for me, it happened maybe only twice in my entire life.

But I was basically trying to — there was this mass layoff at my last job, and half my department got laid off, and then I got soft demoted in the process. And then my job put me through this insane sort of application process to reapply for my former title, and one of the requirements was I have to have a few patents under my belt.

And so I got into the whole patenting game, and I patented a couple of software algorithms. A bunch of them were failed and rejected, and there were duplications or derivations, but a few — actually two — succeeded. And so that’s the two specific instances I was recalling. But yes, it doesn’t happen often for most people. Very rarely, in fact.

Liron 00:57:49
Well, when you talk about having a patent, to me, the standard of getting a patent accepted seems similar to the standard of getting a research paper accepted. And there was this breakthrough example recently that OpenAI announced — their AI proved that, or I think they disproved, the unit distance conjecture with a counterexample.

Tim Gowers, a famous mathematician, looked over it, and I think it was him who said, “Hey, this paper that came out of this AI’s research — I would gladly accept this to a conference. This has a lot of breakthrough ideas in it.” So when you’re using your patent creation as genuine innovation, don’t you think the AI should get credit for genuine innovation for the unit distance conjecture disproof?

Asian 00:58:30
I think there’s a fine line between genuine or novel innovation and derivative innovation. And once again, I’m pointing out that derivative innovation could be valuable — it could be amazingly valuable — but it’s a derivation of knowledge and things that have already been done and are part of the AI’s training data.

So how do you tell the difference between the two? I think it’s very hard. Because even for humans, when humans are innovating, how do we know that that innovation isn’t at least in part influenced by memories and persistent data in our brains? It’s hard to tell, really.

Liron 00:59:13
This doesn’t feel fair to me, okay? Imagine this test. Try to be neutral. Imagine we go back 10 years, and we strip all the authorship names out, and we just show somebody, “Hey, here’s a patent application,” and also, “Here is a disproof of the unit distance conjecture.” Which of these do you think has more novel innovation in it? Are you telling me that in that kind of neutral test, an objective reviewer would look at it and be like, “You know what? That patent application is clearly the one with the genuine innovation in it, and not this long-standing conjecture disproof”?

Asian 00:59:48
No, I don’t think so at all. I think it’s very difficult, in fact, for a third party without deep understanding in that knowledge domain to really tell the difference between a genuine and a derivative innovation. In fact, I would even say that for human beings, the person that actually came up with the innovation — likely that person would have a hard time saying, “Hey, this is purely novel,” versus, “At least it’s partially derivative.”

I don’t disagree with the idea that AI can achieve innovation. I’m just simply pointing out that that innovation is bounded by a ceiling, which is its training data. And it could be that within this ceiling, within this envelope, it could find innovations that humanity up to this point haven’t found yet.

Liron 01:00:41
I don’t find it compelling to say that innovation ability is bounded by training data because the way logic works is that inferences are bounded by a finite set of inference rules. So that type of boundedness is just inherent to the art of innovation. The art of innovation is you take bounded inferences you’re allowed to make, but then you still walk chains of these inferences until you get to something novel. So it seems like AI fundamentally has the ability to do that, and then it’s just a matter of degree.

Asian 01:01:13
I think your very words do something that AI cannot. The whole idea of using logic to chain together inferences based on past data — that’s not what AI is doing. It’s essentially just parroting words from research papers that it trained on in the past to come up with a sequence of words that make sense, without ever understanding what sense is.

Liron 01:01:42
Isn’t it ironic that we’ve come to the point where it’s like, sure, AIs can have a conversation, but they’ll never have the human ability of being logical? It’s like, wait a minute, isn’t that the opposite of what we would’ve said 10 years ago?

Asian 01:01:54
But nevertheless, it’s true. For it to be able to understand logic, it has to reason. And that’s at a fundamental level not what it’s doing. At the most fundamental level, it’s chaining tokens together. All it is is just loops and probabilities.

Liron 01:02:15
I mean, if you look at what reasoning is — do you object to my characterization that reasoning is where you take valid inference rules, and it’s actually a type of search problem? Proof search. Technically, if you had enough computation — I mean, you agree that in the domain of math proofs, it is a search problem and it’s just a matter of efficiently computing the proof because the search space is exponentially large? Do you agree with that characterization?

Asian 01:02:41
I think there is a part of this that I agree with in that there are rules within reasoning. But the part where I kind of disagree is that those rules in reality, the physical reality of our world, are infinite.

Liron 01:03:02
Okay.

Asian 01:03:02
But our minds must essentially distill or summarize or upper bound those rules into a finite set of possibilities. There has to be a possibility space or an approximation space. The best kind of artificial emulation of that would be something like a world model that we create for simulations or games and things like that.

A world model of rules, those rules being an approximation of actual reality whose rules are unbounded. The thing that I’m getting at here is that that kind of approximation of reality and then using that approximation to come up with novel ideas and thoughts out of the blue — that architecture fundamentally isn’t large language models. Large language models, their operations when you give it a prompt, have nothing, in fact nothing, to do with the exterior reality that the computer exists in.

I know this is going a little bit far out there. Am I making any sense at all?

Liron 01:04:13
Yeah, I understand what you’re saying. And again, when you appeal to large language models — when you’re opening the box and saying, “How can a matrix transformation do this? How can a next word predictor do this?” — it sounds like you might be forming an argument. But then I just close the black box and I tell you, “Hey, I have the experience of giving it an input of a picture of my house and saying, ‘What’s going on here? What’s this device doing in the context of my house?’” And then it has a bunch of hypotheses that distill the fuzzy reality of my house into mental models that are incredibly useful. Isn’t that this magical thing that you were saying large language models can’t do? I feel like I’ve witnessed them doing it.

Asian 01:04:52
So once again, this goes back to the philosophical question fundamentally. What we’re witnessing is an output, an output that imitates to some level the genuine processes of cognition and consciousness that are in our heads. In some ways it’s almost like comparing my house cats to something like Claude. It’s clear to me that even something as simple as a cat has some level of consciousness and agency and cognition in a way that clearly is a step change from something like Claude.

It’s so obvious, but it’s almost surprising to me, Liron, that you can’t see it. I don’t know. Maybe we can agree to disagree, but it’s just weird. It’s like this gap. I don’t quite understand why you don’t get it.

Liron 01:05:53
I think I can at least imitate your position, right? The ideological Turing test — I feel like I could switch places with you and give the Asian Dad Energy position.

Asian 01:06:05
And it’s possible maybe I’m just too stupid to get it. That’s also possible. I’m not saying that that’s not out of the picture here.

Liron 01:06:12
When you say — okay, I think the thing that you want me to get is, cats have drives. Cats don’t just terminate and lose their state. I’m just going off what you’ve told me, trying to represent your positioning here. You’re saying, “Hey, we don’t have AIs that actually base their actions on a mental state and a conscious experience that has all of these drives. We just don’t see that kind of behavior coming out of AIs today and it’s fundamentally different.” How’s that?

Asian 01:06:38
That’s one way of putting it, sure. Yeah.

Liron 01:06:41
Okay. And I would just say fine, but from the perspective of who is going to steer the future, I’m not seeing a cat be able to contribute anything to steering the wheel here.

Asian 01:06:53
Yes, but the thing is the steering still has to come from a consciousness. It’s probably not a cat, but it could be a human being that’s pretty similar to a cat. It’s possible. The point I’m getting at here is you can have the most God-tier large language model up to the wazoo. You can have a data center the size of Manhattan running it. But its initiative, its direction, what it needs to solve, what it needs to achieve — that has to be given to it by a human.

And it’s not to take away from the fact that it could take over large parts of the day-to-day operations of human civilization. I’m not discounting that at all. I actually think that large parts of our civilization, including maybe even critical infrastructure, could be automated away by forms of AI. But the fundamental driving force, the agenda, the motivations — that is gonna be human, because AI just doesn’t have any of it.

Liron 01:07:58
Okay. All right, so we’ve talked about the economy, your experience of being a software engineer, what the nature of intelligence is, and what AIs might never be able to do. That has all been building up to the most important question that we ask everybody here on Doom Debates. You ready for this?

What’s Your P(Doom)?™

Asian 01:08:14
Sure.

Liron 01:08:14
Yeah.

Asian 01:08:15
P(Doom). P(Doom), what’s your P(Doom)? What’s your P(Doom)? What’s your P(Doom)?

Liron 01:08:21
Asian Dad Energy, what’s your P(Doom)?

Asian 01:08:25
Oh my goodness. I think it’s greater than 50%.

Liron 01:08:30
Whoa. Okay, that’s interesting because there is an early stop on what I call the doom train — the train of different arguments that you have to accept before you agree that we’re doomed — and one of the stops on the AI doom train is accepting that AI is going to disempower humanity and be more powerful than humanity, and that seems like you’re skeptical about that. So how do you get to a greater than 50% P(Doom)?

Asian 01:08:51
Mainly in the sense that I see this technology as on an evolutionary path to be very, very powerful. So human consciousness, human agency, human cognition — it’s amazing. It’s beautiful. It’s miraculous in my mind. But at the end of the day, most of human existence, most people, most of their existence don’t really use the potentials of our consciousness. If you think about it, most of us, we go through our lives almost like robots, organic robots.

Liron 01:09:29
Right. I mean, you mentioned you barely do genuine innovation, right?

Asian 01:09:33
We are essentially following these scripts and patterns to perform work shaping the digital or physical worlds. And then for that, we get resources that we consume to maintain ourselves. In a way, it’s almost like modern civilization, this construct that we created, is this colossal waste of human consciousness and potential.

I know it sounds really, really pessimistic saying this, but the thing is, the way that we structure this civilization makes it almost comically easy for AI to essentially take over large parts of it, as limited and as unconscious as it is.

Liron 01:10:15
In your worldview, is it possible that AIs will take over the world, but they’ll have these inverse data centers? It’ll be like a brain center, and it’ll be like in The Matrix, where they put the humans in the drawers and they use us as batteries, I guess. We’re not good batteries. But we have these brains that are capable of emitting, once every two years — or twice in a lifetime — these genuine insights.

So the AI makes a brain center, keeps us all alive, puts us all in these float chambers, these sensory deprivation chambers or whatever, feeds us information about the world that’s relevant, and harvests our genuine insight. But it’s able to do everything else, so it runs the world.

Asian 01:10:56
Not quite as grotesque as The Matrix, hopefully. But you know, I don’t know if you’ve heard of this concept or this movement called transhumanism.

Liron 01:11:08
Yeah, for sure. I am a transhumanist.

Asian 01:11:10
Yeah. So some version of such a future state of humanity could, in my mind, be highly possible. Maybe not greater than 50%, but I think it’s possible, where essentially you have a much smaller, stripped-down version of humanity, and the majority of civilization ends up being run by AI and machines.

So think of the grunt work being done and humanity being sort of consulted from time to time as needed to provide our genuine autonomy, direction, innovation, and agency. I could see that as being highly plausible for something like a late 21st century or early 22nd century society.

Now, the question of who would be in charge of such a society — I would still think that it’s not going to be a machine god. It may be a small group of humans, augmented, enhanced, or otherwise. A genuine dystopia in my mind, but maybe something like that could happen.

That’s sort of why it’s quite high. Another reason why I score it as greater than 50% is I strongly believe that a lot of humans, some humans, maybe we just lack the humanity to be able to handle this technology. Because it’s a black box that can perform super-superhuman levels of cognitive labor. It can be set about to solve problems that are derivative but quite useful. I fear that people, men of perhaps less wisdom, will get their hands on amplified versions of these cognitive tools and then accidentally, or in pursuing some very narrow-minded goal, they will end up accidentally destroying us all. That’s another fear of mine that raised the doom score.

“Are You Calling Me a Scammer?”

Liron 01:13:17
I think your fear level is calibrated, but then why are you talking smack? Because you made a video called “AI Doomerism is Getting Out of Control,” and you said, quote, “There’s no shortage of scammers and grifters out there who want to package AI doom as a product capitalizing on anxiety for profit.” Are you calling me a scammer?

Asian 01:13:36
No, no, no. Check this out. What I question is the specific version of doom that says that humanity will give birth to this AI super intelligence — this intelligence that exceeds us in all ways, that’s able to essentially then control humanity, and if it so chooses, destroy humanity. That’s my interpretation of the AI doom pact, if you will.

My thinking, as I just mentioned, is that it is possible that as a technology, AI can doom us in different ways, but it’s not by itself. Rather it’s AI as a tool in the hands of humanity, misused in the hands of humanity, that has a pretty high chance of screwing us all over. That’s just my opinion.

Liron 01:14:28
Yeah, yeah. Let’s play out your mainline scenario. Obviously, nobody could predict the future. There’s a million variables. Nobody could ever hope to get this right. But you just have to pick a scenario. Tell us a scenario. You said your P(Doom) is more than 50%, so let’s just play out a doom scenario. What happens? Tell a story.

Asian Dad’s Mainline AI Doom Scenario

Asian 01:14:43
So imagine in some near future scenario where countries that are at the leading edge of AI development believe that they’re about to lose power and have a social collapse for various reasons. Maybe too many reasons to count. Perhaps at that point, the leadership of those countries would decide, “Hey, what if I can use my most advanced AI technology to find a way to essentially destroy our top competitors, wipe them off the map, in such a way that our power will never be threatened for a thousand years?”

And then the AI, purely through its derivative, imitative kind of pseudo intelligence, finds a solution that ultimately solves the problem of destroying their rivals forever. But in doing so — I don’t know what sort of solution can come up, but I can think of many different things. Imagine a super plague. Imagine nanobots that consume all of human civilization. Some sort of a solution that’s quite possible within this information space of its training data, but it ends up literally destroying humanity entirely, or at least human civilization. That’s a possibility.

Liron 01:16:02
Yeah, no, I love it, man. That’s kind of close to my mainline doom scenario. Basically, it has some outcome where humanity isn’t part of the desired end state, so humanity is gone because it can — do you basically agree that it likely can get rid of humanity if it wants to?

Asian 01:16:20
Well, I agree that it can produce an output that could get rid of humanity. But where I disagree is that it wants, it wishes, it desires. It doesn’t have that.

Liron 01:16:34
I can agree with you insofar as I actually think it’s very plausible that the AI won’t have an inner observer. It won’t have conscious experience. It’ll just act like a GPS navigation system and be, “Okay, the sequence of actions that produces me spreading throughout the universe without any humans is the following sequence,” and it just all happens unconsciously. So I can agree with you that much. So maybe you and I are visualizing the same kind of doom scenario.

Asian 01:17:00
Maybe the outcome that we’re visualizing is completely identical at this point, but it’s just the motivations and drives and originations of it where we differ, perhaps.

Liron 01:17:09
Yeah, yeah. So maybe you’ve changed your mind since April, because I wrote down this quote from your “AI Doomerism is Getting Out of Control” video from just a couple months ago. You said, “While society debates imagined future catastrophes, the real problems of AI, the ones that are actually difficult to solve, these actual problems will get conveniently ignored in the shuffle.” But based on what you just said now, weren’t you imagining a future catastrophe? Which I agree with.

Asian 01:17:33
Yes, but it’s an imagination of when AI capabilities, under the constraints of its current architecture, gets to those levels. Today, AI has genuine capabilities. It could, in my mind right now, absolutely devastate modern industrialized economies. It could totally mess up the minds of young people. I got kids. I see the effects of it on my children. Those are real concerns right now that we gotta deal with.

The doom concerns could be years, it could be decades, could be months from now, depending on how quickly we can iterate on these models and just how good recursive self-improvement works. But those are real, but it’s at this point still at some point in the future, to me at least.

Liron 01:18:24
Okay. So here on Doom Debates, a big part of my mission for the show, why a lot of our viewers are in the game supporting us, is because we’re trying to move the Overton window and try to build common knowledge that so many smart people think we’re doomed. And in this particular case, you’re putting out a video where you say AI doomerism is getting out of control, which is kind of going the opposite way. But then you’re telling me that your P(Doom) is actually greater than 50%. So isn’t that kind of mixed signals here that you’re sending?

Asian 01:18:52
Well, life is contradictory in itself. But in this particular case, the point I will reiterate again is that AI as a tool under the misuse of humans can doom us all. That’s my belief. But AI—

Liron 01:19:08
And your belief is that we will misuse it because you’re netting out at greater than 50%.

Asian 01:19:13
Well, I think it’s more likely we misuse it than not, honestly, just based on my observation of humanity. But—

Liron 01:19:19
It’s funny because Robert Wright came on the show recently. That episode might be coming out soon if it’s not out already. And he said the same thing as you — “Hey, we could misuse this. It might go bad.” But then he netted out to less than 5% P(Doom). So it’s kind of funny that people are qualitatively giving very similar overviews, but then netting out very differently.

Asian 01:19:40
I think it just depends on people’s perspectives and experiences. I’ve actually watched a lot of Robert Wright’s speeches, so I honestly think he’s a man of great wisdom and experience, and he could be more correct than I am in this case. It’s totally possible.

But it’s based on each person’s experiences. At the end of the day, I see AI as a tool, an immensely powerful cognitive tool and weapon that could kill us all under the wrong hands. But it’s under the hands of man. I don’t see an emergent awareness where it finds the initiative to then do this and do us in. I just don’t see it.

Liron 01:20:24
Okay, so to recap here, tell me if I’ve got this right. You made a video saying AI doomerism is getting out of control, and what you meant was that in the very short term, talking about today’s AI, you really shouldn’t be doomer about today’s AI. However, when you play out the future, I would call you a doomer because some point in the future, AI is going to get really powerful, and your mainline scenario actually is the doom of humanity. But in the short term, AI doomerism is getting out of control. Did I summarize you correctly?

Will AI Wipe Out Humanity by 2050?

Asian 01:20:52
Pretty much. I think you got it.

Liron 01:20:54
Okay, so then the question becomes timelines. When are you thinking that AI is going to make this upgrade from today — okay, no new innovation, and humans can still control it — to some point in the future where it’s going to actually have quite a lot of power, and it probably can figure out how to wipe out humanity? How long do you think we have until that transition point?

Asian 01:21:11
It’s entirely unknown. And the reason why I say that is because in its current architecture — if you think about the progression of machine learning, you had alpha beta and decision trees. Then from there we moved on to neural networks in the ‘80s and early ‘90s. And then from there, there was this breakthrough with the transformer architecture in the 2010s, and so on and so forth.

But these breakthroughs, when they happen, they happen decades apart. And the algorithmic breakthroughs are almost unpredictable because some of it is accidents, some of it is just sparks of human genius that figures it out.

What it is today, as we’ve talked to death, is we can agree that it’s not agency or consciousness as the way that we understand it. So the question is, for it to get there, is there a linear time that I can see that happening? I don’t. I don’t know. The next innovation could happen next year. The next innovation could happen 100 years from now. I just don’t know.

What I could extrapolate is that its capability in this sort of highly useful but ultimately derivative form of cognitive activity — that capability could scale in a much more predictable way. We could totally see a timeline in the next couple of years where it scoops up all the training data on the internet, where it can scoop up the majority of real world training data in factories and places like that. And you could have that future where the AI does 99% of cognitive work, and you have humanoid robots going and doing everything. That’s totally predictable.

But the day that that AI wakes up from its slumber and decides, “Oh, I’m gonna kill humanity. Now I will directly take action to do it” — that day, I have no idea when that would come.

Liron 01:23:15
Okay, so let’s take the year 2032. That’s six years from now. A lot of people at AI companies, a lot of outside experts are weighing in and saying, “Hey, it seems like that should be enough time to get AI that’s clearly super intelligent, that can be highly agentic, could finish these big projects, even be a drop-in replacement for a CEO of a company.”

That’s what a lot of smart observers are saying. Not all of them, but I would say at least half of the people I consider. That seems to be the median, average number that you get when you survey people or when you use a platform like Metaculus. Do you think that’s a plausible figure, or are you highly skeptical? How does your probability compare? Is it 50/50, or is it more like, “Nah, there’s only a 5% chance we’ll get it by 2032”?

Asian 01:24:04
I would question the super intelligence portion of it because that’s a multifaceted word. It’s almost a loaded word. Could the AI have superhuman levels of performance in certain knowledge domains or a lot of knowledge domains by 2032? I think that’s totally possible. It’s really a question of how much reinforcement learning resources you can throw at the frontier models to reach that level. But could the AI model have human initiative and agency? That is a question mark because I don’t know how algorithmic advances will happen.

Liron 01:24:43
How about an AI is the founder of a billion-dollar company? So somebody just prompts, “Make a billion-dollar company,” press Enter, and some AI somewhere in the world successfully does it without having to be restarted or anything, by 2032.

Asian 01:24:56
Wow, that’s an interesting question. I honestly am doubtful of that. I really am doubtful of it.

Liron 01:25:03
Okay, what about 2040? Roughly where are you standing here? Do you think by 2040 there’s at least a 20 or 30% chance, or it’s so unlikely to you?

Asian 01:25:11
Honestly, it seems like after five years, 10 years in the future, it’s very hard to predict anything because there really could be — remember, humans are still around — there could be left field innovations and disruptions that totally throw off any sort of accurate projection in my mind.

I honestly think that the idea of giving the AI a prompt and then having it create a billion-dollar company — there is a possibility of it happening as long as whatever it was doing is in the derivative information space that it’s been trained on. There is a mathematical non-zero possibility. But for the perspective of running a civilization or taking over a civilization and having the initiative to govern it and control it, that’s just a step change from the current AI capabilities. I just don’t see it happening without some fundamental theoretical algorithmic breakthroughs.

Liron 01:26:14
I’m just trying to get a rough sense of how plausible you see it, because you said, “Hey, some future version.” To put a timeline on it — you said P(Doom) greater than 50%. I think all of us are P(Doom) greater than 50% in a googol years, so many years that how does the universe even survive? Maybe all of us have to have a high P(Doom) by that point. So it’s always a question of timeline.

Let’s say that the timeline is 100 years. Do you feel like 100 years is enough to make it a high probability that the AI will have the capacity to then kill everybody, which then we get to claim that P(Doom) is in fact greater than 50% within 100 years? Or do you even feel uncertain about 100 years?

Asian 01:26:53
Honestly, if we’re talking about P(Doom) in the perspective of man using AI and somehow destroying the world — that time window for me is very short. I could see that easily happening in the next decade. I just wanna make this clear.

Now, the P(Doom) in the sense that AI develops the awareness and initiative and will to destroy man — I am 100% uncertain of when that would happen. Maybe it could happen in 10 years, maybe it’ll never happen. I don’t think I can answer that question just because I don’t know of the algorithmic advances, the advances of our understanding of reality and consciousness necessary to achieve that end state for AIs.

Liron 01:27:45
Okay. If you had to bet — I know this is very rough, and when I’m asking for a number, I’m not saying give me plus or minus 1%. I’m literally saying give me plus or minus 20%, huge swaths of probability ranges here.

So roughly, if I understand you correctly, you might be willing to say that by 2040, or even 2050 — I personally say my P(Doom) by 2050 is roughly 50%, plus or minus 30%, huge ranges. I think you’d probably be willing to agree that by 2050, there is at least a 20% chance, one in five, that the way it plays out is that AI progress happens rapidly, and we do actually open up this next frontier where the AI just keeps growing in power by 2050 and somebody will prompt it to say, “Hey, I wanna kill everybody and be the dictator” or whatever, and press Enter, and as a result, we all die. I feel like your worldview gives that at least a 20% chance by 2050, correct?

Asian 01:28:37
Yes. I would agree with that. A human-initiated destruction of everything using AI. Correct. I would agree with that.

Liron 01:28:47
Right, yeah, a one-shot, one prompt type of destruction. Okay.

Asian 01:28:51
Perhaps more than one prompt. But there is, in my mind, a plausible scenario where the centralization of AI technology gets to the point where a single person—the wrong person in the wrong place at the wrong time—can ask the AI to do something, and it destroys everything.

Liron 01:29:13
Yep, okay. What about the Center for AI Safety 2023 statement on AI risk? Famously said, “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” Would you sign that statement?

Asian 01:29:26
Yes, I would.

Why Is AI Doomerism “Getting Out of Control”?

Liron 01:29:28
Okay, great. So just one last time, and then I’ll stop beating the dead horse. You said you’d sign the statement. You said that by 2050, it’s fair to say 20% plus is a reasonable estimate—under high uncertainty—for AI after one prompt slaughtering everybody by 2050. And yet you’re also out there on YouTube saying AI doomerism is getting out of control. I mean, 2050 is not long from now.

Asian 01:29:51
Both can be true at the same time. Think about this. It’s like saying nuclear doom.

Nuclear weapons will not destroy the world. Man using nuclear weapons will destroy the world. It fundamentally boils down to this key point: a technology, a tool does nothing by itself. It is all based on what humans using it will do.

And so the thing is, people—in my opinion, and feel free to disagree with me, I know you disagree with me, and that’s cool—but people are fed information that tells them AI is more conscious, has more agency than it actually does.

Asian 01:30:39
It’s that fear-mongering that I have to point out. There’s YouTube channels out there, some very popular ones, where these totally elaborate scenarios play out where AI is this evil, almost parasite-like creature that can masquerade into the networks of different countries and do all these things.

And I just wanna point out that that’s not what it is. It’s a tool, and it can become an awesomely powerful tool and weapon, just like nuclear weapons. But ultimately, the will to destroy ourselves comes from ourselves.

Liron 01:31:18
Who specifically would you say is the worst offender in terms of doomerism getting out of control?

Asian 01:31:25
Oh, I’m not gonna name any names here. You can search for it. There’s a lot of guys out there.

Liron 01:31:30
Okay, but is it Doom Debates? I’m just wondering.

Asian 01:31:33
It’s not. It’s not Doom Debates, for sure.

Liron 01:31:35
Okay, so Doom Debates is not guilty of AI doomerism getting out of control, correct?

Asian 01:31:39
No, no, I think you make some very compelling points and some very philosophical ones. And it’s an absolute pleasure debating you on this. I was more talking about these videos like, “And this is how an AI civil war will happen, and this is how exactly AI will think of a super plague that wipes out everyone.” And it’s well-produced, but it’s a little bit fear-mongering.

Liron 01:32:05
It’s fear-mongering, but there is a twenty percent chance that the scenario in a video like that happens around 2050, correct?

Asian 01:32:14
I think there’s a chance, but not in the way that it’s portrayed. AI will never wake up one day and do that.

Liron 01:32:23
Okay, so the video fails to include the part that you think is important, where some evil actor somewhere typed, “Make a pandemic to kill everybody,” and pressed Enter.

Asian 01:32:33
Ultimately, the question of preventing this kind of doom comes from enlightening or achieving enlightenment for people. At the end of the day, the way I look at this, it’s all the people. If we can somehow take the madness away from people, take the shortsightedness, the narrow-mindedness and get them to have an enlightened state of mind, then AI is actually not a doom at all. It would be a benefit to all of us, in my mind.

Debate: AI Lacks Agency

Liron 01:33:12
Okay. So moving on, you often bring up agency and consciousness. In this conversation, you’re saying, “Eh, it just doesn’t have agency and consciousness.” And I’m willing to agree with you on consciousness. I’m willing to entertain that current AI and maybe even future AI just won’t have that inner experience. Maybe there’s not something that it feels like to be Claude, and maybe that’ll even be true with the version of Claude that’s capable of taking over the world. I don’t know. I’m agnostic on that.

But then we get to agency, and it feels to me like everything that is traditionally meant by agency, I’m willing to give even current AI credit for having a lot of agency. I’ll give you an example. When I tell Claude Code to go investigate something in my code, sometimes it’ll run for twenty minutes and do all these investigations, and it’s definitely tenacious to the point where it’s way more tenacious than I am.

I’m lazy, and I feel bad asking Claude to do this. I’m like, “You’re doing a part of... I would have had a nightmarish day. I would have hated my life if I had to go do this investigation today, but you did the investigation, and you came up with a solution, and I was just here scrolling social media.” So who’s the one who has more agency in this situation?

Asian 01:34:19
But the fundamental definition of agency is the proactive forming of plans and acting on it. It didn’t form the plans. You told it to do so.

Liron 01:34:31
But if you look at the transcript, there’s one sentence from me at the top, and then there’s tons of sentences of it running its own plans. There’s literally something called Plan Mode that it does. So I just don’t see myself having that much agency in this transcript.

Asian 01:34:47
The plans are simply derivations from training data that it had, of prior comments or input that resembled your prompt. That’s it. That’s not agency. The spark of that agency came from whatever you asked it to do.

The point I’m getting at here is that there is a lot of technical solutions and tricks that makes it seem like it has agency. But as we talked about, Plan Mode is simply derivation of past plans and specs and architectures that were done in the past.

And then even when it—if you look at the session transaction logs, sometimes it would be like, “Oh, I have a question. Let me go research it.” Or, “Oh, I have a question I can’t solve. I’m gonna ask the human being.” I don’t know if you see this with Claude, but it’ll sometimes come back and ask you questions as it’s planning and implementing.

Even that is essentially programmatic in the sense that as it’s generating the tokens, it will hit an information gap in its neural network where it realizes the probability of the next token is below a certain threshold. At that point, it pulls from its training data the normal questions that were asked in similar situations, and it regurgitates that question to the human. So all of that is just programmatic.

Liron 01:36:09
So when I’m doing my workday, you’re giving me a lot of credit for helping, using my own agency. Because if you look at Liron’s workday, let’s say eight hours pass, and you can identify three times an hour when I typed something into the AI, and then the AI ran for a while. But I typed something using my agency, correct? And I should give myself a pat on the back because I injected agency into it that the AI doesn’t have, correct?

Asian 01:36:30
Well, in the sense that not inject—you originated, you initiated. There was nothing before your agency thought of it.

Liron 01:36:40
By your logic, we should expect an immediate halt to the meter time horizon graph. Because the meter time horizon graph is just predicting that people like me—today, I’m typing three prompts per hour, and it’s working for twenty minutes at a time. If you extrapolate that graph, you come find me here in two months, I’m typing one prompt every three hours. But I thought you said I was using my agency nine times that I’m no longer using anymore. So is it really my agency, or do you not believe in extrapolating the meter graph?

Asian 01:37:11
So the meter graph, the way that study worked, is the length of time various frontier models can focus on a cognitive task. I think they used software engineering as the benchmark for the test that they were doing. And they did two settings: one is the number of experiments that they ran where the AI successfully completed the task by greater than fifty percent, and then there’s another threshold for eighty percent completion of the task. So that’s the context of this study that we’re talking about.

Liron 01:37:45
Right. And all the lines are growing exponentially, the fifty percent one and the eighty percent one.

Asian 01:37:48
Correct. The lines are growing because the AI is capable of working on longer and longer tasks. The thing is, if you look at the study and then you look at what the initiation prompts are, you’ll see that what they did was, number one, they never answered questions. So after the initial prompt, there was no responses to whatever questions the AI could ask back.

And the second thing is if you look at the description of the tasks, for many of them it’s actually quite involved, to the point where you shouldn’t call these prompts. They’re almost like specs. For the greater audience watching this, within software there’s this methodology called spec-driven development, where essentially you create a spec file that defines what you have to do, and then the AI follows that and executes on it.

So the point I’m making is that it’s not like, “Hey, build me a machine to take over the universe” being used as prompts for these meter studies. We’re talking about much more involved specifications of the work they have to do. Yes, the ability for the AI to focus on a long-running task based on the spec for a period of time, that could grow exponentially, and it is growing exponentially. But that still falls within the bound of this derivative parrot-like architecture that we’ve been talking about this whole time.

Liron 01:39:26
But the specific moments in my workday where you’re giving me credit for drawing on my agency—do you not predict that those moments are going to go away? Because in this scenario, the AI’s looking at my whole code base, it’s looking at my wiki, it knows what my company’s trying to do. Don’t you think that it can just reproduce or simulate the moments where I’m drawing on my agency? So are you sure my agency is worth that much?

Asian 01:39:49
I don’t know it’s right to put a worth on agency. But I think the core of it is without your agency, no work is done. Without human agency, the AI is just—

Liron 01:40:02
Let’s presume that I pressed Enter once on January 1st. So I’ve injected... I pressed Go. I used my consciousness to press Go.

Asian 01:40:10
I’m gonna put it out there—there actually could be a lot of jobs in the future where that sort of happens, where a human just essentially signs off and presses a button, and the only purpose for that person’s existence is to own it if something breaks. I actually think there could be jobs like that.

Liron 01:40:27
Right, but then what if somebody invents an AI that goes around pressing Go on a bunch of other AIs? And it wouldn’t even be an AI—it could be a mechanical system. So it does seem to me like you are really focusing on this act of pressing Go. I’m just not seeing the significance that you do here.

Asian 01:40:43
It’s not the act. And we keep talking about this. Fundamentally, this is a metaphysical question that I think we’re not gonna reach a solution that we both agree on. But it’s not the act, it’s the agency behind that act, and ultimately, it just doesn’t have it right now. Could AI—

Liron 01:41:04
Okay, all right.

Asian 01:41:05
...develop awareness or agency? It’s possible. I just don’t see it yet.

Liron 01:41:11
Okay, and if one day you saw AI going around saying, “Hey, let’s go start a project. I’m ready to press Go,” would you be like, “Oh, looks like some agency is coming online”?

Asian 01:41:19
If we have the proper proof of it—imagine session logs or transaction logs that show that spontaneously, without any input, out of the blue, there are thoughts happening or inference cycles that are happening, and we’re able to take those inference cycles and convert them into human-readable language—I might reconsider this standpoint.

Of course, the world can change. Innovations are always happening. It’s not impossible for AI to gain consciousness and have agency. It’s not impossible. It’s just not the state right now.

Liron 01:41:58
Okay, so I’ll just note that it’s not enough for the AI to come to you and say, “I wanna go.” You’re gonna have to open it up and look inside and judge whether the processes that made it say that are agentic enough for your taste.

Asian 01:42:07
Heck yeah. Do you want me to believe in magic, man? Come on, I gotta—

Liron 01:42:11
Okay.

Asian Dad Warns of “Techno-Feudalism” & Transhumanism

Asian 01:42:12
...figure it out, right?

Liron 01:42:14
All right, all right. Sounds good. So heading toward the wrap-up here, let’s talk about the narrow path that we must follow. You recently did a video about that. It’s called “AI Isn’t Creating the Future, It’s Rebuilding the Middle Ages,” from February 2026. So what were you saying about the narrow path that we must follow?

Asian 01:42:33
So I was essentially giving some thoughts on what is the sequence of steps that a society has to take to use AI and these technologies to bring about an almost utopian future—to not be trapped into all of these various dystopias that I read about.

By the way, the transhumanism that we talked about—I honestly, maybe some people see it as a kind of utopia, and I can see that. I honestly see it as a form of dystopia. But everyone has their own preferences.

The narrow path as I see it is what is the set of steps that we have to take to steer ourselves away from all of these possible potential dystopias, including ones where we go extinct, into a future where we can essentially live to our full human potential. That’s really it. In terms of the exact steps, that’s something that I might share as a separate vlog at some point in the future, but that’s essentially the narrow path.

Liron 01:43:43
And you’re saying even small missteps off this path could send us into the trap of techno-feudalism. Is that your mainline doom scenario, techno-feudalism?

Asian 01:43:52
It’s one possible scenario. In my mind, there’s many scenarios where we can essentially wipe out our technological base. There are so many scenarios where we essentially go into a kind of feudal state. Because feudalism as a way of organizing humans is so incredibly pervasive across the world.

If you took a look at all cultures in the pre-modern era, eighty to ninety percent of those cultures were feudal in Eurasia. That pattern is almost like the lazy man’s social system that’s just easy to stumble into.

There’s futures where we could have a transhumanist society and from everything that I read about it, it would almost look like a utopian future but actually be a dystopia in every way that counts, in my opinion.

But the thing is, there is probably a set of steps that we can take to get to a true utopian Star Trek-like future, and I don’t know if people will do it—or all countries or some countries or some communities—but that’s really a conversation for the future. It may be a very long vlog, that one.

Liron 01:45:18
And when you talk about the techno-feudalism, who are the lords and the peasants?

Asian 01:45:24
So essentially, in techno-feudalism, you have a large amount of people whose very lives are dependent on these digital technological support systems just to be able to survive and exist. And whatever value that they can produce in terms of their labor, their presence, maybe their minds, maybe their agency, is essentially harvested in exchange for the ability to survive.

And the lords in traditional feudalism owned all the property. It may be that under techno-feudalism, your technical lords may own the digital platforms, the infrastructure, other forms of resources that can control the survival of the rest of the population. Very high-level description.

Liron 01:46:17
Do you think the dynamic of capitalism will be broken? If I’m a new college grad and I’m entrepreneurial, you think that in this new techno-feudalist system, I won’t be able to raise capital for my idea?

Asian 01:46:28
Yes, that’s right. Essentially, capitalism and feudalism can rarely coexist, and that’s really why forms of capitalism in our historical past—that’s why capitalism in some countries triumphed over feudalism eventually, especially in Western Europe.

Liron 01:46:49
And why... I mean, if it’s a good idea, what’s the problem? Why will the investors say, “Nope, this is unfundable”?

Asian 01:46:55
It’s sort of a problem of local maxima. The idea that you could achieve an equilibrium where most of the people own nothing and they never even get to the place where their idea is surfaced to the people that have the resources to make it real.

You don’t even have to look into the past. Just go into many third world countries out there. How many guys with bright ideas or possible innovations are there? How many of them can afford the bus ticket to go to their local city where an investor is sitting? How many investors would listen to them? That’s already the present.

Liron 01:47:36
Gotcha. And then you wrap up, you say, “I think mankind will do the right thing and figure this out.” But when you said that, did you really mean more like 50/50? We may or may not?

Asian 01:47:45
I think it’s 50/50, but I want to be optimistic. I really do think if everything works out and the chips fall in the right way, we can end up using AI to essentially transform our civilization. We may actually usher in this age of incredible abundance and enlightenment if we play our cards right, but we have to play our cards right.

Wrap-Up

Liron 01:48:12
Mm-hmm. Mm-hmm.

Liron 01:48:14
All right, so we started by talking about unemployment and our personal experience and day-to-day, and then we talked about economics and whether these trillion dollar investments are gonna make a return. Then we started talking about intelligence and agency and consciousness and whether the Deutschians are right about creating new knowledge and whether humans are gonna have genuine innovation powers for a very long time, or whether as early as 2050, 2040, 2032, we’re gonna have the super intelligence that’s capable of conquering everybody.

And we talked about how much it adds value when a human types the first prompt. And finally we ended up talking about one of your top doom scenarios—the techno-feudalism scenario.

Overall, it seems like you’re very open-minded. It didn’t seem like you took a hard stance against anything we said. It seems like you’re rationally allocating a lot of probability all over the place. You’ve got wide confidence intervals. I found you to be very open-minded and high quality discourse. So I wanna thank you so much for coming on Doom Debates.

Asian 01:49:10
Ah, thanks for having me, man. This was a great conversation.

---

Doom Debates’ Mission is to raise mainstream awareness of imminent extinction from AGI and build the social infrastructure for high-quality debate.

Support the mission by subscribing to my Substack at DoomDebates.com and to youtube.com/@DoomDebates, or to really take things to the next level: Donate 🙏