Everything that got easy got crowded, and everything that stayed hard is now worth more than it's ever been.
May 26, 2026
TechnologySomething I keep coming back to: am I just using AI to chase ideas that would have worked ten years ago, before the tools made them accessible to everyone?
I build a content site, I write code with an agent, I spin up tools that six months ago would have required a team…and I wonder sometimes if I'm just arriving late to a party that's already been commodified.
That question led me somewhere more uncomfortable. It's not just about what I'm building. It's about whether the way most people are thinking about AI – as something that handles the hard parts so you can focus on direction – is actually right.
I'm not sure it is.
Content has never been cheaper to produce.
In 2024, Raptive creators alone published 7.9 million articles – up 37% from the year before.1 Amazon's Kindle platform processed so many AI-generated submissions that it capped authors at three books per day,2 and that still wasn't enough to slow the flood.
By April 2025, nearly three quarters of all new English-language web pages contained detectable AI-generated content.3 The same story is playing out in code – 256 billion lines written by AI in 2024 alone, a number that would have been science fiction five years ago.
Every Category Flooded at Once
Index: 100 = 2019 baseline. Web pages: Ahrefs/Raptive. Lines of code: GitHub Octoverse, Elite Brains.4
The response most people have to this is sensible: if everyone can produce the output, the output isn't the advantage. What matters is knowing which output to produce. Direction. Taste. Judgment about what problems are worth solving. The AI handles the depth; the human handles the aim.
It's a reasonable position. It's also where things get complicated.
There's a version of generalism that works. Leadership principles transfer across industries. A great operator can run a restaurant chain and a logistics company and add real value in both. Prioritization, communication, knowing how to build a team – these compound across contexts without requiring deep domain knowledge.
But there's a layer below direction-setting where generalism stops working, and it's the layer that matters most: evaluation. Knowing whether the work is right. Not whether it looks right – whether it is right.
A hospital doesn't expect a general practitioner to perform cardiac surgery. But when a patient presents with an ambiguous set of symptoms, the GP who can't recognize the pattern that warrants a cardiology referral isn't providing general care – they're providing incomplete care with confidence. The specialist catches what the generalist misses not because they're smarter, but because they've seen enough edge cases to know where the edges are. That only comes from depth that compounds over years in a specific problem space.
Same Degree. Different Depth.
Depth Compounds. Breadth Doesn't.
Same degree. 12+ years of compounding specialization. $494k difference.
Average annual compensation. Source: Medscape 2026,5 Doximity 2025.6
This plays out the same way with AI. You don't need to write the code if the agent writes it. You don't need to run the analysis if the model runs it. But when the output comes back, someone has to know if it's right – not just whether it looks clean and confident, but whether the framing was correct, whether the right question was asked, whether the answer is pointing somewhere true. Most people never find out, because the output looked fine and nobody in the room knew enough to push back.
The mechanism is not new. Satellites didn't improve local geographic knowledge. They made most of it irrelevant – route distances, travel times, the existence of roads, all rendered into an app anyone can open. What they couldn't replace: knowing the trail looks passable on the map but the locals don't use it after rain. That knowledge required having been there. An agent pointed at the wrong valley by someone who's never walked the terrain will return a very thorough, very confident, very wrong answer. And the person who sent it won't know the difference.
What AI changes is what the wrong answer looks like. A bad analysis written by hand was obviously rough – incomplete, hedged, visibly limited. A model produces the same bad analysis with clean prose, coherent structure, and apparent authority. The roughness that used to signal error has been removed. Which means the only reliable indicator of quality left is whether someone in the room knows enough to catch what the model got wrong.
That knowledge – knowing when the confident answer is pointing at the wrong valley – only comes from one place.
This is what makes depth different from effort. Effort is visible from the outside – you can see someone working hard without understanding what they're doing. Depth is only visible from the inside, and only to people who have it.
The shallow read of a domain looks complete. The concepts resolve, the frameworks make sense, the answers feel sufficient. It's only when you go deeper that the edges appear – the places where the frameworks break, where the obvious answer is wrong for reasons that aren't obvious, where the real problem turns out to be adjacent to the one everyone's been solving. You can't see those edges from the outside. You can't even know they exist.
This is why taste without depth is just preference. A strong aesthetic sense, good pattern recognition, a feel for what seems right – these matter. But they're not judgment. Judgment requires knowing what you're looking at well enough to evaluate it, not just react to it. And in any domain where the problems are genuinely hard, that requires having gone deep enough to see the frontier – which is a place most people never reach, not because they lack intelligence, but because they stop before the problem gets interesting.
Deeper in the specific places where I have real surface area – the decisions I've been closest to, the problems I've actually lived, the domains where I've built enough exposure to know what a wrong answer looks like before the model says it's right.
That's a narrower list than I'd like. But I think the instinct to expand what the tools can reach is probably backwards. The leverage isn't in covering more ground. It's in going deeper in fewer places, until the judgment starts to compound.
The part I keep coming back to: you can't outsource the question of where to go deep. A generalized model gives a generalized answer. It can help you execute once you understand the domain – but it can't tell you which domain deserves your next three years. That requires knowing what you care about, what you're uniquely positioned for, and where your existing depth gives you an edge. Nobody else can make that call for you.
I'm still working mine out. But I think it starts with not letting the model make it.
Epilogue
The prevailing bet right now is that the bottleneck shifts to taste. That the people who know what's worth building, what direction to point the tools – those people capture most of the value, because everything else gets delegated.
But a tastemaker isn't someone with good instincts and a vague sense of quality. A tastemaker is steeped. A great film critic has seen ten thousand films – the obscure ones, the failed ones, the ones that almost worked. A great editor has read so much bad writing that the good sentence announces itself on contact. The taste came from the depth, not instead of it. You don't develop judgment about a domain by staying at its surface and outsourcing the rest. You develop it by going deep enough that you've seen what bad looks like from the inside – and that takes time that can't be compressed.
The question isn't whether you can direct an agent to go do the work. You can. The question is whether you'll know if it came back with the right answer – and whether you'll recognize the moment when it was pointed at the wrong problem entirely.
You don't get that from having good taste. You get it from having earned it.
Sources