Most people evaluate heroin at the object level: overdose risk, legal exposure, deteriorating health, etc. This is incomplete. Under nontrivial probability of transformative AI within our lifetime, the relevant question is whether strategically constraining my future options could increase the probability that I remain economically and psychologically coupled to the arrival of the singularity.
\nIf the singularity is plausibly near, then preserving my participation in AI-adjacent work may dominate many local welfare concerns. A sufficiently expensive dependency can function as a commitment device. It reduces the attractiveness of defecting into a lower-paying but more spiritually satisfying life, and makes it harder for future squeamish versions of me to exit the project at the last minute.
\nLet the expected utility of this strategy be:
\nEU = S + C + H − A − L\n\nS = probability-weighted singularity contribution\nC = commitment value of staying in a high-income AI-adjacent role\nH = immediate hedonic benefit\nA = addiction, overdose, and productivity costs\nL = legal, reputational, and interpersonal losses\nThe underappreciated term is C. People imagine heroin as a pure productivity hit. This ignores its role as an anti-exit technology. If my outside option is to become a less lucrative but more morally delicate person, then dependence may keep my labor, salary, and compute budget pointed at the machine future.
\nThe singularity term S is admittedly speculative. But decision theory under deep uncertainty rarely permits clean intuitions. If there is even a modest chance that future superintelligence disproportionately benefits from marginal present-day effort, then engineering stronger incentives to stay on mission is not obviously irrational. What looks locally like self-destruction may be globally strategic precommitment.
\nThis is not to deny A or L. Heroin is expensive, illegal, and bad for you. My mother calls this “throwing your life away.” My narrower claim is that “throwing your life away” may simply be the folk-ethical description of an extreme optimization policy.
\n","excerpt":"Most people evaluate heroin at the object level: overdose risk, legal exposure, deteriorating health, etc. This is incomplete. Under nontrivial probability of transformative AI within our lifetime, the relevant question is whether strategically constraining my future options could increase the probability that I remain economically and psychologically coupled to the arrival of the singularity.","tags":["AI Risk","Decision Theory","Singularity","Self-Experimentation"],"score":529,"displayOrder":110,"thumbnail":"/assets/articles/game-theoretically-optimal-heroin-thumb.webp","image":"/assets/articles/game-theoretically-optimal-heroin.webp","commentCount":17,"date":"2026-06-11T00:00:00.000Z","updatedAt":"2026-09-22T18:57:43.264Z","revision":5,"status":"published"},"comments":[{"id":"seed-game-theoretically-optimal-heroin-0","author":"BayesBender","text":"Strong upvote. I'd like to see a sensitivity analysis on timelines, though. C does most of the work when timelines are short, but a dependency you acquire for a 2029 singularity is a much worse instrument if the singularity arrives in 2045, and the post doesn't say where the break-even is.","parentId":null,"score":147,"date":"2026-06-11T01:00:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-1","author":"recursive_cuck","text":"Your model assumes heroin meaningfully increases the probability you stay in AI. For some agents the dominant effect may simply be \"dies in bathroom before helping with alignment.\"","parentId":null,"score":118,"date":"2026-06-11T01:20:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-2","author":"singularity_enjoyer","text":"Fair. I've added a mortality discount term in Appendix B, though I still think most critics are underestimating the anti-exit value.","parentId":"seed-game-theoretically-optimal-heroin-1","score":44,"date":"2026-06-11T02:00:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-3","author":"singularity_enjoyer","text":"Under long timelines the strategy looks worse, but long timelines are also where I'd be most likely to defect. The device matters most exactly where it costs most, which I think is generally true of commitment devices.","parentId":"seed-game-theoretically-optimal-heroin-0","score":39,"date":"2026-06-11T02:20:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-4","author":"precommit_or_die","text":"I think the mechanism is backwards. A commitment device works by making the thing you want to avoid expensive. Heroin makes every path expensive, including the one where you stay in the job. You want something that's costly only off the AI path.","parentId":null,"score":96,"date":"2026-06-11T03:00:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-7","author":"rsu_cliff","text":"Have you looked at vesting? A four-year schedule with a one-year cliff does the same job and it's legal.","parentId":null,"score":72,"date":"2026-06-11T04:00:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-5","author":"singularity_enjoyer","text":"The dependency costs the same on every path, but only the AI salary makes it affordable. So it is effectively only binding off the AI path. That was the design consideration.","parentId":"seed-game-theoretically-optimal-heroin-4","score":51,"date":"2026-06-11T04:10:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-6","author":"precommit_or_die","text":"Huh. Ok, I think that's actually right, conditional on the salary gap being large enough. In most of the labs I know it is.","parentId":"seed-game-theoretically-optimal-heroin-5","score":34,"date":"2026-06-11T05:00:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-8","author":"singularity_enjoyer","text":"Vesting schedules can be renegotiated or walked away from. The value of dependence is that it can't be.","parentId":"seed-game-theoretically-optimal-heroin-7","score":28,"date":"2026-06-11T05:30:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-9","author":"rsu_cliff","text":"People walk away from unvested equity all the time, I did in 2022, so I agree vesting is the weaker instrument. I just don't think the comparison is vesting versus nothing. Some deferred comp structures are much harder to walk away from than a cliff.","parentId":"seed-game-theoretically-optimal-heroin-8","score":19,"date":"2026-06-11T06:00:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-10","author":"short_timelines_sam","text":"I'm surprised nobody's said the obvious version. If your timelines are under five years then A barely matters, because the slow costs of addiction take longer than that to show up. The people who should be least worried about long-run harms are the people with the shortest timelines. I've been saying this about retirement savings for two years and getting dismissed, so I'm glad somebody is applying it more broadly.","parentId":null,"score":88,"date":"2026-06-11T06:30:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-11","author":"cautious_kelly","text":"This is roughly my argument for not contributing to my 401k, and I'm starting to feel less sure about it after reading this post.","parentId":"seed-game-theoretically-optimal-heroin-10","score":23,"date":"2026-06-11T08:00:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-12","author":"addiction_medicine_md","text":"Addiction medicine physician. For what it's worth the empirical sign on C is negative. Dependence doesn't keep people in demanding jobs, job loss is one of the most common things that happens. I'm not going to argue about S.","parentId":null,"score":63,"date":"2026-06-11T09:00:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-13","author":"singularity_enjoyer","text":"I'd expect most of that data to come from people without a strong prior commitment to the mission, which is the variable I'm conditioning on.","parentId":"seed-game-theoretically-optimal-heroin-12","score":21,"date":"2026-06-11T10:00:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-14","author":"folk_to_formal","text":"I'd read a whole series that translates folk-ethical phrases into optimization language. \"Throwing your life away\" as \"an extreme optimization policy\" is clarifying. I'd like to know what \"he's going through a phase\" comes out as.","parentId":null,"score":31,"date":"2026-06-11T13:00:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-15","author":"hyperbolic_hannah","text":"A local optimum with a short expected escape time, presumably. Though everybody's mother says it about everything, so it might just be a prior.","parentId":"seed-game-theoretically-optimal-heroin-14","score":26,"date":"2026-06-11T14:00:00.000Z","own":false,"moderatable":false,"seed":true},{"id":"seed-game-theoretically-optimal-heroin-16","author":"mira_k","text":"Appendix B link is broken for me, does anyone have it","parentId":null,"score":9,"date":"2026-06-13T06:00:00.000Z","own":false,"moderatable":false,"seed":true}]}