Every weekday at 7:42 I order an oat-milk cortado at the coffee place under my office, the barista spins the little iPad around, and I hit No Tip. I’ve done this roughly four hundred times. My coworker Evan thinks it makes me cheap, and last month he said so in front of the whole pod. I don’t think I’m cheap. I think I’m one of the only people in that line who has actually priced the simulation hypothesis into his discretionary spending.
Nick Bostrom’s argument is simple enough that I’m always surprised how few people follow it to the end. Either almost no civilization survives long enough to run detailed ancestor simulations, or almost none of the ones that can run them bothers to, or we are almost certainly living in one. I don’t see a strong case for putting much weight on the first two horns, so, like most people who have actually sat with the trilemma instead of making a face at it, I land at about 0.6 that this is a simulation. Elon’s number is a lot higher than mine. In this conversation, I’m the moderate.
Once you grant that much, the rest is a compute problem, and compute problems have well-understood solutions. No game engine renders the inside of every building. You get occlusion culling, level-of-detail scaling and lazy evaluation, which is a fancy way of saying that things get computed at full resolution when something important is looking at them and at the bare minimum the rest of the time. There’s no reason an ancestor simulation would work differently, and every reason to expect it to be more aggressive, since minds are by far the most expensive objects in the scene. The cheapest minds to fake are the ones that only ever show up in short, scripted, transactional interactions, which is to say counter staff. A barista only has to be rendered at the resolution of “oat cortado, name for the order?”
I’ve been collecting data. The morning barista, whose name is either Tess or Bess, says “Have a great one!” to basically everyone, and I’ve logged it thirty-one times with what I would describe as identical intonation. Once I asked her whether she’d seen the Knicks game, and there was a latency of about two seconds before she answered, which is exactly what lazy evaluation looks like from the outside: her inner life didn’t exist until I queried it, and it took a beat to spin up. DoorDash drivers are an even easier case. They get rendered for about ninety seconds on my doorstep and are presumably deallocated before they reach the elevator.
The objection I hear most is that I have no special reason to think I’m one of the full-resolution minds myself. I actually have the best reason there is, which is that I’m having this experience. Bostrom’s own self-sampling assumption says I should reason as though I’m a random draw from all the observers in my reference class, and a random draw from a compute-constrained simulation is overwhelmingly likely to be one of the cheap ones. I am not cheap. I have a rich interior life, I can do anthropic reasoning before my coffee, and I notice things like two-second latencies. That is strong evidence that I landed in the expensive minority, and if I’m in the expensive minority, most of the people I transact with aren’t. Put it together and the tipping decision is straightforward:
Tip if: t × [ (1 − P_sim) + P_sim × P_full ] > c
t = welfare value of the tip to the server
c = cost of the tip to me
P_sim = probability this is a simulation (≈ 0.6)
P_full = probability the server has a full inner life, given
simulation (≈ 0.1 for counter staff)
A dollar is a dollar, so t equals c, and the left side comes out to 0.46 of the right. In expectation, every tip I leave at that counter burns fifty-four cents on the dollar. Evan’s position is that I should apply the same discount to c, since I could be simulated too, which gets the asymmetry exactly backwards. I have direct access to my own inner life. I have zero access to hers. Descartes settled my side of the equation in 1641, and nobody has settled hers.
People also call me inconsistent because I tip twenty percent on dates, which misunderstands what a conditional policy is. When Maya and I go out, the scene carries narrative weight: two full-resolution observers, an evening with real stakes, a server whose performance both of us are watching. Any competent simulator allocates more compute to a scene like that, and it has to render everyone in it at higher fidelity, the server included, or the scene falls apart. P_full goes up, the inequality flips, and I tip. When it’s just me at the bar at Donovan’s on a Tuesday afternoon with the Knicks pregame on, the scene has almost no weight and the bartender gets nothing, which, honestly, I think the guy understands. For the record, I support a twenty-dollar minimum wage, because I’d much rather fix this at the policy level than keep patching it one cortado at a time.
There are costs, and I’ve tracked them. My cortado now reliably comes out after the drinks of people who ordered after me, and two Tuesdays ago the name on the cup was “Brad.” My name is Brian. I considered whether that was a deliberate signal, which would be evidence of a much richer model than I’d been assuming, but an edit distance of two is well within the error rate you’d expect from a cheap text generator running on a tight budget, so I didn’t update much.
Last Thursday Maya asked why I tip when she’s there and not when she isn’t, and I walked her through the scene-weight argument. She was quiet for a while, and then she asked me, pretty directly, whether I think she’s real. I told her the truth, which is that I’m about seventy percent confident, and that this is a lot higher than my number for most people I know. She hasn’t answered a text since, so I’ve been eating at the bar at Donovan’s, where the scene weight is low and the policy is simple.