The narrative in the AI world over the past couple of months has been: many people grinding 16-hour days alongside AI — sounds glamorous, trendy, full of energy. But the obvious truth that few will openly admit is: AI has made me more tired. This made me want to write about it — why is the operator more exhausted when the tool is doing more of the work? What exactly is the human carrying in this process?
Let's set aside the meta-game — keeping up with the frontier, new tools, new models and capabilities dropping every week. Assume you're past that. Your workflow is dialed in. You're not doomscrolling out of FOMO — you have real problems to solve. You're in the zone.
What you're actually doing now is: taking work that used to be a purely human, vibes-based, implicit-knowledge process, and turning it into a context engineering problem. You're decomposing fuzzy, experience-driven, muscle-memory tasks into structured, distributable, real-time-aligned human-AI collaboration. Every task becomes: what does the AI need to know? What can it do autonomously? Where do I need to intervene? How do I keep everything coherent?
Think of it this way: you're building a house where you are simultaneously the client and the general contractor. You spec the requirements and you're on-site laying the bricks.
No matter how capable the AI gets, there are two roles you cannot delegate away:
Manager. Problems are human-defined. What are we solving? What does success look like? What's the strategy? Without human intent, the task doesn't exist. The AI doesn't wake up one morning and decide to refactor your codebase. You do.
Executor. Even with AI that can autonomously complete well-defined tasks, no real project is one-shot. The work is full of micro-decisions, judgment calls, authorization gates, and moments where new information changes the plan. You're not "reviewing AI output" — you're pair-programming with it, co-navigating the problem space in real time.
Here's where it gets hard. These two roles have nearly opposite cognitive demands.
The manager needs to maintain a global view. You have multiple threads running — feature A is being built, doc B is being drafted, research thread C is exploring an open question. You're tracking all of them, maintaining the big picture, adjusting priorities on the fly.
The executor needs deep local focus. When you're in a thread, you need to know the details — current state, last decision, active constraints. You need to be fully immersed.
The brutal part: these threads don't wait for you. While you're deep in thread A, thread B pops up with a blocking decision that needs your call. Thread C hits an unexpected wall and needs you and the AI to brainstorm an exploratory solution together. You're not just switching between threads — you're switching between roles within each thread. One second you're the artillery operator, the next you're a sniper, and then you have to zoom out and be the general reading the whole battlefield.
Multithreading and deep focus are, almost by definition, in tension with each other.
Think about who actually has this skill set.
If you've been a manager — led a team, been a director, orchestrated people and priorities — you have the dispatch and architecture skills. You know how to decompose work, track progress across streams, make resource allocation decisions.
But managers rarely go back to the frontlines. They're not used to writing the code themselves, fixing edge cases, debugging the weird failure modes. Let alone doing it while context-switching at high frequency and maintaining precision in each thread.
If you've been a senior IC — an engineer, a designer, a researcher — you have the deep execution skills. You can go deep on a problem and produce high-quality output.
But ICs often haven't built the mental models for multi-stream orchestration, priority management, and real-time strategic adjustment.
AI-augmented work demands both, simultaneously, in the same person. You need to switch between the two every few minutes. Director-level strategic vision plus staff-engineer-level hands-on precision.
This is why the "everyone should use AI" narrative rings hollow for a lot of people. Sure, anyone can use AI to write an email faster or summarize a document. But actually producing things with AI — shipping features, building products, creating things that didn't exist before — that selects for a very specific, nearly contradictory combination of abilities:
- Engineering depth and time for greenfield exploration
- Leadership experience and willingness to do the grunt work
- Broad awareness of the rapidly shifting tool landscape and the ability to pick one and go all-in with full focus
The system is selecting for an AI-native kind of human. Not just smart. Not just hardworking. But capable of running multiple threads while maintaining deep single-thread focus, and switching between the two on demand. That's a rare cognitive profile, and I think it goes a long way toward explaining why the productivity gains from AI are so unevenly distributed.