Announcement
May 25, 2026
The AI Vampire Is a Design Failure
There is a new figure haunting Silicon Valley, and Marc Andreessen gave it a name on the most recent Joe Rogan Experience: the AI vampire.
He described friends — some of them famous — who used to look healthy and now look wrecked. Bags under the eyes. Sleep deprived. Clearly not taking care of themselves. And completely euphoric. They have stopped sleeping because they are running twenty AI coding agents at once, and the opportunity cost of going to bed has become unbearable. Close your eyes for six hours and twenty agents sit idle on twenty projects. So they don't close their eyes.
Andreessen frames this as a triumph: the most dramatic increase in programmer productivity in history, leading-edge coders twenty times more productive than they were a year ago. One Wall Street friend — an MIT computer science degree gathering dust for thirty-five years — picked the craft back up with AI and generated five hundred thousand lines of code in his spare time, re-automating his entire house: smart fridges, a robot dog, the jukebox of his dreams. The top human orchestrators of these agent fleets, Andreessen claims, now earn fifty million dollars a year. They have the philosopher's stone.
He is right about the leverage. He is dangerously wrong about the vampire.
What is actually true
Strip away the theater and the underlying shift is real, and it matters for every company reading this. Software work is moving from writing every line by hand to supervising, reviewing, and orchestrating systems of agents at scale. A single engineer can now direct dozens of agents in parallel. The economics of code have inverted.
The part most people miss is the elasticity. The fear was simple: if AI makes code cheap, programmers lose their jobs, because there is only so much software the world needs. That assumption is wrong, and it is wrong in a way that should reshape how you plan the next two years. There was never a fixed amount of software the world wanted — there was a fixed amount it could afford. Every company carries a backlog of a thousand projects that never made the cut, not because they lacked value but because they were not cost-effective under the old math. When the cost of building collapses, that backlog opens. This is the Jevons paradox playing out in real time: make a resource radically cheaper and total consumption goes up, not down.
So the leverage is real. The productivity multiple is real. The backlog is real. What is not real — what is a story Silicon Valley tells itself to justify a bad habit — is that the vampire is the price of admission.
The vampire is a bug, not a feature
Here is the thing nobody on that podcast said out loud. If your AI agents require you to be awake at three in the morning, you did not build a system. You built a job that never ends.
The entire promise of orchestration is that the work continues without you in the room. That is the definition of leverage. A founder who has to babysit twenty agents through the night has not escaped manual labor — he has scaled it. He swapped writing code by hand for supervising code by hand, then convinced himself the exhaustion was a sign of winning. The bags under the eyes are not a badge. They are a diagnostic. They tell you the system has a human bottleneck wedged into the middle of it, and the human is the founder.
The measure of a good agent system is not how productive it makes you while you are watching it. It is how much it produces correctly while you are asleep. That is a design problem, and it is solvable: clear authority levels for what an agent may do alone, what it must confirm, and what it must never touch; verification built into the pipeline so output is checked by another agent before it reaches you, not after it has shipped; and the discipline to let the machine run unattended on reversible work while you reserve your attention for the calls only a human should make. None of that requires skipping sleep. All of it requires you to architect the system instead of becoming part of it.
The real cost the vampires are paying
We keep a sentence on the wall: time will always be the most expensive thing we spend. The AI vampire has it exactly backwards. He treats sleep as the expensive thing and the agents as free. In truth the agents are nearly free now — that is the whole point — and his time, his health, his presence at home are the only scarce resources left in the equation. He is optimizing the cheap input and burning the expensive one.
A founder running on no sleep is not a high-leverage operator. He is a single point of failure with a caffeine problem. The decisions that actually move a company — what to build, what to kill, where to spend capital, which partnership to sign — are exactly the decisions that degrade when you are exhausted. The vampire is generating five hundred thousand lines of code and losing the clarity that decides whether any of it should exist. That is not leverage. That is a very productive way to go in the wrong direction faster. If it costs peace, it is too expensive.
What to do with this instead
The opportunity Andreessen is describing is genuine, and it will intensify — he is right about that too. The companies pulling ahead are the ones treating their backlog of impossible projects as suddenly possible and shipping at a pace that looked insane eighteen months ago. You should be one of them. But build it as a system that works for you, not a habit that consumes you.
Three principles carry most of the weight. First, design for unattended operation — if a workflow only runs when you are watching, it is not done. Second, put verification before delivery, not after, so the machine catches its own mistakes instead of handing them to a tired human at midnight. Third, draw the authority lines explicitly: let the agents own the reversible work outright, and reserve yourself for the small set of judgment calls that genuinely need a human. Get those three right and you capture every bit of the twenty-times leverage without paying for it in sleep.
The vampires think they have found the philosopher's stone. What they have actually found is a faster treadmill. The goal was never to produce more while you are awake. The goal was to build something that produces while you live your life. Build to connect — so you can disconnect. That is the version of this revolution worth wanting.
Changelog