The most rational take on AI you’ll hear this year

Host: We're diving into a fascinating take on artificial intelligence from tech analyst Benedict Evans, who argues that AI is exactly as big a deal as the internet or mobile phones. Guest: That almost sounds like he's downplaying it, especially with so many people comparing AI to the Industrial Revolution. Host: It might sound that way, but he's actually just reminding us how massively the internet and smartphones completely reshaped our world. His main point is that if AI is the new internet, our current era is basically 1997. Guest: Ah, so we're in a time when the tech is super exciting, but a lot of it barely works and the best applications haven't even been invented yet? Host: Exactly. Because of that, there's a massive gap right now between tech insiders who use it constantly and the general public who might only touch it once a week. Guest: Does that also mean it's way too early to know which AI companies are actually going to dominate the market long-term? Host: Spot on. Trying to declare whether OpenAI or Anthropic will win right now is like being in 1997 and guessing whether Yahoo or Excite would win the internet. Guest: And we all know how that turned out for them. So we know AI will change everything, but we really don't know exactly how it will happen yet? Host: Precisely. It is a fundamental shift, but as Evans jokes about his own industry research, we are essentially staring at a giant presentation that just concludes with "we don't know yet."

Host: We are looking at where AI adoption stands today, and it feels a lot like the tech industry did in 1997. Guest: So people know this technology is a big deal, but they might not see exactly how it will change their daily work just yet? Host: Exactly, unless you happen to be a software developer. For developers, seeing modern AI coding tools is like an accountant seeing the very first digital spreadsheet in the late seventies. Guest: That makes sense, an accountant seeing a spreadsheet would be blown away, but a lawyer back then would just shrug because it did not apply to their daily tasks. Host: That is the exact divide we are seeing right now. It creates what we call a jagged frontier, where developers get it instantly but survey data shows only fifteen to twenty percent of young people actually use AI daily. Guest: If it is that hard for the average company to figure out how to apply it, is that why big AI labs like OpenAI and Anthropic are suddenly investing so much in consulting firms? Host: That is exactly why. Completely reimagining a company's internal workflows to actually integrate AI is a massive project, and most businesses simply do not have idle employees waiting around to build that. Guest: So they have to hire outside help, like forward-deployed engineers or traditional consultants, to bridge the gap and plug the AI into their specific systems. Host: Right, those consultants are the ones who come in, map out exactly where the AI is reliable, and do the heavy lifting of building out those new workflows.

Host: There's a really surprising trend happening right now with artificial intelligence: the top AI labs are heavily investing in professional consultants. Guest: That is surprising. I thought AI was supposed to make consultants obsolete by doing all their work for them? Host: To understand why it hasn't, we have to look at the difference between a "task" and a "job." Think about an old-school elevator attendant who manually pulled a lever to take you to your floor. Guest: Right, they literally just drove the elevator. So when automated buttons were invented, their entire job disappeared? Host: Exactly, because their job was just that one single task. But for most knowledge workers, automating a task actually triggers something called the Jevons paradox, where making something cheaper and faster just makes people demand more of it. Guest: So when spreadsheets were invented, investment bankers didn't suddenly get Friday afternoons off; they just had to run way more financial models? Host: Precisely. And we're seeing the exact same thing with AI today. An AI can instantly write code or generate a 75-slide presentation, but that's just the task, not the whole job. Guest: I see. The AI can make the slides, but it can't sit in a room, navigate company politics, and figure out what the business strategy actually needs to be. Host: You nailed it. You hire a firm like McKinsey to walk through your enterprise and talk to your customers, not just to format a PowerPoint. The AI handles the busywork, but defining the actual problem remains a deeply human job.

Host: Let's look at a fascinating paradox about automation. We often assume technology eliminates professions, but if you look at accountants, their numbers actually went up over the last century despite adding machines and spreadsheets. Guest: That seems completely backwards. Shouldn't software doing all the math mean we need fewer accountants, not more? Host: You would think so, but making a service more efficient actually increases the overall demand for it, which unlocks new types of work. Even top AI companies like OpenAI and Anthropic are rapidly adding human headcount right now instead of replacing everyone with bots. Guest: But we constantly hear AI executives warning about a job apocalypse where entry-level roles disappear entirely. Are they just wrong? Host: Being an expert at building AI models doesn't automatically make someone an expert in labor economics. Historically, every major tech shift since the 1800s has caused some painful disruption, but it eventually creates entirely new jobs we couldn't have predicted. Guest: Sure, but isn't AI different because of how fast it's spreading? ChatGPT got hundreds of millions of users almost instantly compared to early computers. Host: It is adopting faster, but only because it's standing on the shoulders of existing infrastructure like the internet and smartphones. Individual consumer adoption is very different from how big businesses actually operate. Guest: So a massive corporation isn't just going to buy AI software today and fire all its staff next week? Host: Exactly, and the people predicting that don't understand enterprise sales cycles, which often take eighteen months or more. Ripping out and replacing core corporate systems takes years, meaning the job market will actually have time to adapt and evolve.

Host: We often expect new technology to change the world overnight, but this text argues it actually takes years for people to figure out how to apply it. Guest: Does the author give an example of that kind of delay? Host: Yes, they point out that many software companies founded right before ChatGPT could have actually been built a decade earlier, but it took time for founders to connect the existing tech to specific industry problems. Guest: So even with AI moving fast, businesses will take a while to fully transform. Does the author think this current AI shift is completely unprecedented? Host: They actually find a lot of comfort in history, noting an IBM ad from the 1950s that pitched a fridge-sized electronic calculator as giving a company 150 extra engineers. Guest: That sounds exactly like the marketing pitches we hear right now for AI coding assistants! Host: It really does, and it reminds us that we've survived enormous shifts before, like the internet turning a two-week, long-distance library research project into a two-hour Google search. Guest: But what about the fear of Artificial General Intelligence, or AGI, eventually replacing human jobs entirely? Host: The author admits we are all essentially just guessing, or doing what they call "vibes forecasting," because we still don't even have a solid theory for what human intelligence actually is or why these AI models work so well. Guest: So until we understand our own brains, predicting the end of human work is basically just late-night dorm room philosophy.

Host: When we look at the future of AI, there's this great observation from a computer scientist named Larry Tesler, who said AI is simply whatever machines can't do yet. Guest: Wait, meaning once a machine can actually do a task, it loses the "AI" label? Host: Exactly, people just start calling it software instead. It's similar to how jet airliners were the cutting edge of "technology" in the 1960s, but today we just call them airplanes. Guest: So the definition of AI is essentially a moving target, but what about AGI, or Artificial General Intelligence? Host: That definition is shifting too, where people used to think of AGI as something with true human-level intelligence, almost like it's alive. Guest: Has that standard been lowered recently to fit what current AI models can actually do? Host: Yes, some now casually define AGI as just being able to do a certain percentage of economically valuable work. But by that logic, a standard IBM mainframe in 1975 was AGI because it automated previously human tasks. Guest: It sounds like people are just getting tangled up in semantics, arguing over terms like super intelligence versus AGI instead of what the tech actually does. Host: Spot on, it's like people endlessly debating whether crypto is blockchain or blockchain is crypto. It entirely misses the bigger picture. Guest: If we strip away all the confusing terminology, what is the real bottom line we should be focusing on right now? Host: The most important takeaway is that even if AI models hit a brick wall and stopped improving tomorrow, the tools we already have are incredibly useful and will radically transform the world over the next ten years.

Host: Let's look at a fascinating quote from Sam Altman, where he predicted we'll soon be buying AI intelligence on a meter, exactly like water or electricity. Guest: That sounds like it would be a massive, reliable business, so why does the author see that comparison as a bit of a warning sign? Host: Because historically, utility companies don't capture the real value of the products that run on their networks. For example, when you watch television, the TV network isn't paying a cut of its profits to the local electric company. Guest: Ah, so just providing the underlying infrastructure doesn't guarantee you get a piece of the action happening on top of it. Host: Exactly, and we saw this exact same thing happen in the telecom industry. Mobile data consumption has skyrocketed exponentially since 2010, but telecom stocks basically stayed flat because they're selling a low-margin commodity. Guest: Does that mean the big AI companies spending billions on these giant foundation models might end up in that exact same boat? Host: That is the elemental question right now. It really depends on whether AI models become sticky platforms like Windows, where developers are locked in, or if they become interchangeable utilities like cloud hosting on AWS. Guest: If they're interchangeable like AWS, then I imagine the real pricing power and profit would shift to the specific apps being built on top of them, right? Host: Spot on. Once the current crazy hype cycle settles down, we might just see a handful of companies selling AI at marginal cost, meaning all the real value gets captured much further up the stack.