Is the current boom in artificial intelligence like a financial bubble just waiting to pop, wiping out trillions like back in 2001? Or is it something really fundamentally different, almost like a self -perpetuating infinite money glitch? It's kind of wild to think about. Two -sec silence, yeah. Today, we're diving deep. We've got 12 essential lessons from 28 intense
months right there in the AI trenches. We're going to unpack why the underlying tech is absolutely transformative, where the actual money is being made right now. And that critical skill you really need to survive the coming labor disruption. It's serious stuff. Welcome back, everyone. Yeah, welcome back to the deep dive. And look, this isn't just theory, right? We're looking at patterns drawn from people with actual skin in the game.
Right, moving past those. you know academic debates you usually hear exactly okay so let's unpack this we've got a roadmap three major areas to explore today first the financial reality why this feels more like a vc bubble maybe not a tech bubble and how nvidia ended up being this this central indispensable Okay. Then second,
we'll shift to the technical change. We're going to look at why smaller, more focused models using things like reinforcement learning, why they're actually taking the lead now over those huge generic ones we saw dominate the last couple of years. And finally, the human cost. The very real job displacement we see coming and this kind of counterintuitive idea, the return to code. Let's jump in. Where do we start? The money. Let's start with the money. The big question.
bubble or not right the bubble debate it's constant you know there's this solid heuristic this rule of thumb if everyone is screaming bubble it probably isn't going to be that sudden crash they're imagining yeah it's different from say 2008 or the crypto thing in 2021 where there wasn't as much skepticism beforehand exactly but i mean if revenue growth is genuinely explosive You mentioned anthropic reporting, like 10x year -over -year growth, NVIDIA smashing expectations quarter after quarter.
Why are we even still calling this a bubble? Isn't the VC side just scrambling to catch up to something totally new? That's the tricky part, isn't it? It is. And honestly, the people calling bubble, they do have some strong short -term points. The investment levels are... Well, they're insane. Like 17 million seed rounds. Billions for companies with no actual product yet. Yeah. Billions for just talent, like getting Andrej Karpathy or Elias Asitskiver on board. It's nuts.
Right. Hard to justify that unless the underlying tech was really, really real. Precisely. And it is. The core tech is delivering real value now. It's writing code, analyzing data, making people more productive. That's why it's not a technology bubble you see. OK. But the overfunded, often unprofitable startups built on top of it. Yeah, a lot of those are going to collapse. It's more like a VC bubble riding in a tech revolution. So like a correction is coming for those weak
players. Definitely likely. Maybe a 10, even 30 percent market correction when those companies fail. But the core tech. It sticks. Okay. That makes sense. And this brings us to that weird infinite money glitch thing, the loop between NVIDIA and the big AI labs. Ah, yes. Yeah. Strange loop. It's where the structural friction really lies. It's like the ultimate capital recycling machine. How does it work? Exactly. So NVIDIA
acts like a VC, right? They bore huge capital, let's say $100 billion just as an example, into a lab like OpenAI. Okay. But OpenAI has to spend that money on compute. Which means... NVIDIA's GPUs. They're really high margin graphics cards. Exactly. The money flows straight back to NVIDIA's revenue. Their stock price goes up. Which lets them make more investments and the loop just keeps going. It's brilliant. A self -feeding system. Genius, really. But it exposes this tough
truth, doesn't it? Pretty much everyone else in the AI game, the labs, the startups, they're losing money. Burning millions, billions on training costs. Absolutely. Jensen Huang and NVIDIA. They're selling the picks and shovels in this gold rush. They've got the guaranteed massive profit margins. Everyone else is just spending. And that must be why Google's pushing their TPUs so hard. Amazon's building custom silicon. Meta's designing their
own chips. They have to. It's desperation. They need to break this dependency on NVIDIA to control their own destiny, frankly, just to survive long term. Okay, so if owning the hardware, the compute is the main battleground. What's the single biggest constraint holding back AI growth and revenue right now? If labs could just get more GPUs, could they make more money? Compute. It's the ultimate physical choke point. It limits how fast labs can expand and even meet the demand
they already have. Labs like Anthropic and OpenAI could likely 2x or 3x their revenue today if they just have more GPUs. So compute is the constraint. Got it. The bottleneck limiting revenue and expansion. Mineral sponsor, Reed Placeholder. All right, we're back. So if hardware is the choke point, how are companies fighting back? It sounds like they're changing the definition of what good AI even means. That's right. Which brings us neatly into the technical shifts we're seeing.
Yeah, the whole era of just building the biggest possible model. That's kind of ending, you know, like a lot of analysts found some recent releases disappointing. But they were intentionally smaller, weren't they? Smaller than previous versions. Exactly. It wasn't about ideology. It was driven by that cost barrier we talked about. Yeah. Size just started giving diminishing returns on performance compared to the crazy jump in compute cost. So
efficiency and speed are the new game. That's the real frontier now, efficiency and speed. And the breakthrough that's really keeping the momentum going is reinforcement learning, or RL. OK, define RL for us quickly. Sure. It's basically when the AI learns by doing a task, like actually trying to write code or play a game, and then it gets rewards or penalties based on how well it did. So it's moving beyond just reading like all the text on the Internet. Right.
It's learning from action and consequence. And what's fascinating is how fast the AIs are getting better at coding and math specifically. Why those deterministic tasks? tasks with clear right or wrong answers? Why are they improving so much faster than, say, creative writing? Because the feedback is clean. If the code doesn't compile, boom, wrong path. If the math equation fails, nope, try again. The AI immediately knows to discard that way of thinking. Ah, I see. So it
can generate tons of attempts. Exactly. Like, try a solution 500 different ways, find the one successful reasoning trace, the one path that worked, and then train itself just on that optimal path. It's an amazing synthetic data advantage. Wow. It's pretty humbling, actually, seeing how fast those areas are moving. I still wrestle with prompt drift myself. What's prompt drift exactly? For listeners who haven't hit that wall
yet. Oh, it's maddening. It's basically when you put in the exact same prompt, but the model gives you wildly different answers over time. Its internal behavior just changes. Makes reliable work kind of impossible, right? Totally. And that's much harder to fix for subjective things like writing compared to deterministic stuff like code. It's a real challenge. Okay, so that
brings us to this flow zone problem. The focus now is on smaller, faster models because speed is just critical for actually getting work done, right? Like waiting five minutes for a code suggestion just breaks your focus. Absolutely breaks your flow state. Look at Anthropic's Haiku 4 .5 model. People find it super useful, mainly because it's three times cheaper and more than twice as fast as their bigger Sonnet model. So good enough but quick is way better than perfect but slow.
Vastly superior for most practical uses. Yeah. And there's this new strategy reflecting that idea, test time compute. Right. So instead of blowing the whole compute budget on training the absolute biggest model possible. What do they do? They train a, let's say, medium sized model. Then they save some of that compute budget to use during inference, like when you're actually using the model. They let the model think longer and reason more deeply before it gives you the
answer. That's a great way to put it, like giving a student extra time on the final exam to double check the hard questions, even if they only studied. a medium amount. Exactly like that. And this efficiency, it's unlocking this exponential growth in how long AI agents can work autonomously. How long are we talking? Well, six months ago, maybe 20 minutes was a good run for an autonomous agent. Today, we're seeing sustained work sessions like two to seven hours long. GPT -5 codecs apparently
had some seven hour autonomous sessions. Whoa. Imagine scaling that. That unlocks really complex stuff, right? Like huge code -based refactors or deep research projects. Totally. The agent can analyze, plan, run searches, synthesize the findings, and just repeat that whole loop for hours without a human stepping in. If that trajectory keeps going, we could have agents working reliably for days soon, automating tasks that take junior engineers weeks. It's heading that way. It really
seems possible. This changes the game for complex automation. Okay, how does this new strategy using test time compute actually improve the results if the model itself isn't bigger? Because the smaller model gets dedicated time to reason during inference right when it's needed. It leads to better, more thoughtful answers than just a fast, reflexive response from a giant model. More thoughtful answers from smaller models. Got it. Efficiency is the new size. The technical
shift is clear. Efficiency over brute force. Now let's talk about data. What really separates the winners long term? You mentioned XAI earlier. Yeah, let's validate that prediction about XAI. People kind of dismissed Elon Musk, but he always had foundational advantages that meant he was going to catch up. Unique data, capital, manufacturing skill. Right. XAI gets exclusive access to X, formerly Twitter, that real -time global conversation.
That gives their AI, Grok, this unique edge on current events and slang that others just can't legally scrape, right? Correct. That's a powerful data mode. But the ultimate mode, it's physical. It's Tesla and the Optimus robots. How so? Tesla cars provide this constant stream of real -world video footage, amazing for training world models that understand physics and navigation. Yeah. But Optimus. Optimus will generate something
totally unique. Physical embodied data. Three -dimensional data from interacting with the world. Embodying the data? What does that physical experience give an AI that just reading text misses? Everything. Seriously. Think about a human child. A toddler is infinitely smarter than GPT -5 in some really fundamental ways. How? Because the child learns from physical feedback that isn't scraped from
the Internet. They learn about gravity by dropping things, force by pushing things, object permanence, social cues, all by doing things in the real world. Ah, that embodied experience. That's what current AIs are fundamentally missing. Totally. It's a huge gap. Okay, but here's where things get really interesting, maybe counterintuitive again. Open source models seem to be catching up incredibly fast, even matching or beating the big closed systems. Yeah, this is blowing
up quietly. Take a model like GLM 4 .6 from Zippo AI. It's open source and it's performing better than Anthropic's closed source Claude Sonnet on a lot of important coding benchmarks. So the open source community is moving at lightning speed. Undeniably fast, yeah. But why isn't this like front page news? The incumbents, the big labs like OpenAI and Anthropic, they kind of ignore it, right? To protect their perceived advantage, their moat. But there's also a practical
issue. What's that? Cloud infrastructure, AWS, Google Cloud, Azure. They're optimized to run the popular models efficiently. The GPT -series cloud. So these powerful new open source models seem slower just because the pipes aren't optimized for them yet. Exactly. It creates friction. Yeah. But the underlying quality of the models proves how fast the open source community is innovating. You know, Anthropix valued maybe $200 billion, but a lab like Zippo is hitting comparable performance.
It's validation. Wild. OK, now for the more sobering part, job displacement. Yeah, got to talk about it. Look, the goal here isn't just disrupting the, what, $300 billion software market. The aim is replacing big chunks of the $15 trillion global labor market. $15 trillion, wow. So the trend we're likely going to see accelerate by 2026 is single job replacement agents. Meaning? Think customers, court reps, sales development reps, setting appointments, executive assistants.
These roles are prime targets for automation by specialized AI agents. Which means huge. efficiency gains for companies massive games but also significant social unrest i mean widespread protests related to ai job losses are genuinely predicted for 2026. that's heavy which leads us maybe surprisingly to your prediction the return to code learning computer science is going to be sexy again why because it gives you asymmetrical leverage Massive leverage. Explain that. Okay. A non -technical
person using some AI assistance. Maybe they get 2x, 3x faster at their job. That's nice. Useful. Right. But a skilled programmer, someone who really understands systems, logic, architecture, they don't just get 3x faster. They become 100 times more powerful. 100 times? How? Because you stop being just the coder writing lines. You become the architect. You start managing teams of AI agents. You leverage your deep technical know -how to orchestrate these tools at a scale
nobody else can. You have to get to the cutting edge to really crush the competition. So you go from using AI to, like, help write an email. To using AI to refactor an entire million lines. code base in an afternoon by coordinating five different specialized agent teams simultaneously. That's the 100x leverage. Wow. So what does this all mean for the average career then? If you're
listening right now, what's the takeaway? The greatest advantage, the biggest upside, goes to those with the technical competence to truly leverage these new AI tools at a deep level. Technical competence is key. Okay. Hashtag, hashtag occlusion and final thought. All right. Let's try to summarize the big ideas here. Three core lessons, maybe. First, AI is real transformation, not just a bubble, and it's backed by actual explosive revenue growth, even if the VC funding
is bubbly. Right. And second, compute is that critical physical constraint, the bottleneck, which is fueling that kind of bizarre self -feeding money loop between the AI labs and NVIDIA. And third, the game has totally changed. That era of just copycat startups reselling API tokens at a loss, that's ending. It's over. The future really belongs to efficiency, specialized models, and that deep technical competence we just talked about. And it feels like the conversation has
shifted too, right? Less focus on, you know, abstract AI doomerism, the paperclip maximizers. Yeah, thankfully. The focus has rightly shifted to the very real world problems. Bias in the models, market consolidation, and crucially, labor disruption. These are tangible issues we need to solve now. We've seen just how much capital, how much energy is pouring into building the physical infrastructure for AI, the data centers, the silicon foundries, the power grid upgrades.
It's staggering. So maybe the final provocative thought for you, the listener, is this. It's about where you fit. You need to ask yourself, are your skills foundational? Are they essential, like the picks and shovels needed to build this new AI infrastructure? Or... Are your skills part of that repetitive lower leverage labor that is now, frankly, being systematically replaced? A heavy question, but a really necessary one to ponder. Where does your unique leverage lie
in this new world? Yeah. Thank you for taking this deep dive with us today. Lots to think about. Always is. Talk soon.