As our crack cocaine consumption increased by more than an order of magnitude, traditional approaches to substance use became our largest remaining bottleneck. Here’s how we built a new system from the ground up in just five months.
\nAt MoreWrong, the limiting factor in our research is no longer intelligence. Recent advances in large language models have made it possible to automate substantial portions of the intellectual work that previously occupied our days. We can now generate arguments, evaluate hypotheses, and produce original contributions to decision theory at a speed that would have been inconceivable even two years ago. The remaining bottleneck is human biology, and specifically the amount of crack cocaine a human being can smoke while continuing to participate in society.
\nIn February 2026, our seven-person research collective was averaging approximately 1,200 crack cocaine consumption events per day. By September, that figure had increased to more than 48,000. During the same period, we destroyed three marriages, lost custody of five children, established an extensive high-risk homosexual behavior research program, and became the subject of an FBI investigation. We have no reason to expect this growth to slow.
\nOur most important discovery is a phenomenon we call recursive self-destruction: each successive deterioration in a researcher's personal circumstances increases his capacity to smoke crack cocaine, generating further deterioration and still greater consumption. The latent demand appears effectively infinite, constrained primarily by employment, financial solvency, and the continued willingness of other people to associate with us.
\nTo support this growth, we developed a purpose-built crack cocaine management system called Cinder. Over five months, we redesigned our living arrangements and built a comprehensive framework for coordinating consumption, financial obligations, anonymous homosexual encounters, and interpersonal consequences. This was an interesting technical challenge. Crack cocaine was both the cause of the problems that made Cinder necessary and the primary means by which we found ourselves sufficiently motivated to develop it.
\nIn a recursively self-improving system, each iteration increases the system's ability to produce further improvements. Recursive self-destruction operates on the same principle, except that each iteration makes the system substantially worse according to conventional measures of human welfare. Crucially, these measurements fail to account for the increased capacity to smoke crack cocaine that each deterioration produces.
\nConsider a researcher who loses his job because he has been smoking crack cocaine at work. This immediately releases fifty hours per week for additional consumption, which may cause him to lose his apartment, eliminating the time previously spent cleaning, paying rent, and maintaining an acceptable relationship with his landlord. His family stops calling, his friends stop inviting him to social engagements, and his creditors discover that further attempts at collection have negative expected value. Each apparent failure removes a constraint that was limiting his ability to pursue the very activity responsible for that failure, creating a positive feedback loop of extraordinary efficiency.
\nWe observed the phenomenon in a senior researcher who began the year as a married homeowner earning $240,000 annually at a frontier AI laboratory. Following an unusually productive three-week research sprint, he lost his job, marriage, house, and custody of his two children. He began sleeping in a storage unit, increasing his daily crack cocaine consumption by 1,800%. Within another month, he lost the storage unit, removing the final constraint imposed by having a fixed geographic location. He now operates across a decentralized network of parking garages, bathhouses, and the homes of men who have not yet learned his name. His productivity has never been higher.
\nEarly versions of Cinder assumed consumption capacity would decline as financial resources were depleted. We discovered the opposite: as absolute resources approached zero, the proportion allocated to crack cocaine approached 100%, while the willingness to engage in increasingly ambitious resource-acquisition strategies increased correspondingly. By August, researchers were securitizing their fathers' pensions, offering venture capital firms controlling interests in their grandmothers' estates, and selling naming rights to their unborn children. One attempted to purchase a failing trucking company using a promissory note backed by an anticipated inheritance. When the transaction collapsed, he incorporated a private equity firm to acquire the debt from his creditors. By the fourth iteration, he had been indicted in two states, acquired an abandoned shopping mall, and begun referring to himself as a sovereign wealth fund. His original objective had been to obtain $300.
\nThe following week, he issued a cryptocurrency backed by future revenue from an autonomous crack cocaine research facility. The token achieved a fully diluted valuation of $94 million before the exchange suspended trading. He attempted to acquire the exchange using a leveraged position in his own token, precipitating liquidations affecting seventeen companies and a municipal pension fund. He celebrated by purchasing a glass rose from a gas station, smoking an enormous rock, and having sex with two men he had met approximately forty minutes earlier. We classified the entire sequence as one continuous research event and incorporated it into our dataset as an unusually successful demonstration of individual economic agency.
\nOur preliminary results suggest that recursive self-destruction is limited primarily by the amount of external infrastructure that can be incorporated into the process. A handful of sufficiently intelligent people, given access to crack cocaine and a flexible interpretation of property rights, may be capable of recursively transforming entire economic sectors. We believe this finding has profound implications for the future of artificial intelligence, economics, and human civilization.
\nOur decision to standardize on crack cocaine reflects a detailed understanding of its pharmacological properties and the limitations of competing substances, particularly cocaine hydrochloride. The primary advantage of crack cocaine is that smoking it is extraordinarily pleasurable. This is an important characteristic of a system intended to deliver pleasure, and one that we believe has been insufficiently emphasized in previous research.
\nCocaine inhibits dopamine reuptake, increasing extracellular dopamine in mesolimbic reward circuitry, particularly the nucleus accumbens. This contributes to intense euphoria, heightened confidence, excitement, and a profound sense that one's current activities are of enormous importance. We consider this last property especially valuable for researchers working on civilization-scale problems.
\nThe difference between cocaine hydrochloride and crack cocaine is largely pharmacokinetic. Intranasal cocaine hydrochloride produces a comparatively gradual increase in systemic concentrations. Smoked cocaine freebase is absorbed rapidly through the lungs, reaching the brain within seconds. The resulting rapid onset is associated with powerful subjective euphoria and reinforcement. We find these results extremely compelling, particularly given their implications for reducing latency in the consumption process.
\nWith cocaine hydrochloride, a user may spend several minutes waiting for the full subjective effect, during which he might become distracted by unrelated obligations or begin responding to messages from his wife. Crack cocaine substantially reduces this latency. Its rapid onset and relatively brief euphoric effects also generate recurring demand without external intervention. Our initial architecture included an automated scheduling system to remind participants to consume additional crack cocaine, but during testing, we discovered that the drug was already performing this function with substantially greater reliability than our software. We removed the scheduling component, reducing complexity while improving engagement.
\nCrack cocaine can reorganize human behavior around a single objective with a consistency few productivity systems can approach. Several researchers achieved complete independence from conventional employment within weeks, while others eliminated the requirement to maintain a permanent residence. One spent seventy-two consecutive hours carpet surfing for crack crumbs, producing no usable cocaine but establishing a new benchmark for sustained attention. We believe crack cocaine represents one of the most significant motivational technologies ever developed, and the fact that wider society has yet to recognize its advantages speaks primarily to the limitations of wider society.
\nConventional crack cocaine consumption is decentralized. Each participant maintains his own money, makes his own decisions, and bears responsibility for his behavior. This provides fault isolation and works reasonably well when demand is low, but at our scale, these features became liabilities. In March, maintaining independent financial accounts required fourteen hours of administrative work per week. By June, three researchers had exhausted their savings, two had lost access to banking services, and one was disputing ownership of a Toyota Camry with his father. We had designed our architecture around the assumption that participants would continue to possess money.
\nBy July, we had incorporated eleven legal entities, purchased a decommissioned ambulance, and become involved in a civil dispute over a commercial laundromat. We also established a satellite research operation in the back room of a bathhouse, where three senior staff were already conducting extensive fieldwork. Our researchers were smoking an astonishing amount of hard. We maintained a fleet of glass stems, a dedicated Chore Boy procurement budget, and an inventory system for crack rocks, several of which were purchased using funds designated for employee health insurance. At peak load, researchers were taking blasts during investor calls, depositions, technical interviews, and, on one occasion, their own testimony before a grand jury.
\nWe experimented with shared expense spreadsheets, informal lending arrangements, and asking family members for increasingly large sums under the pretense of attending graduate school. These interventions bought us time, but the complexity grew faster than consumption capacity. We also discovered that our infrastructure was optimized for operating while sober. Calendars, bank statements, and conventional notions of elapsed time presume a relatively stable relationship between intentions and subsequent behavior. For our workload, these assumptions were becoming increasingly expensive, and it was clear that incremental improvements to existing systems would no longer suffice.
\nWe developed Cinder around six principles.
\nParticipants are always online. Traditional addiction frameworks assume consumption separated by periods of ordinary life, requiring expensive transitions between incompatible states. Cinder treats crack cocaine as a continuously available background process. There is no concept of being finished, only periods in which additional capacity has not yet been allocated.
\nCentralized authority. Individuals are ordinarily expected to understand their finances, obligations, and whereabouts. We found this unnecessarily burdensome. Cinder maintains authoritative records while participants operate with a small, disposable local memory. A researcher need not know where he was yesterday or whether he has already asked his mother for $800.
\nAll meaningful behavior is observable. Information scattered across text messages, Grindr conversations, and hostile communications from landlords is consolidated into an event stream. When a researcher announces that this is the last time he will ever smoke crack cocaine, Cinder retrieves all prior declarations and calculates their historical reliability.
\nEvery component must scale independently. A disruption to employment should not interrupt crack cocaine consumption; an unexpected divorce should affect only services dependent on the continued cooperation of one's spouse. We similarly decoupled transportation, replacement electronics, bathhouse memberships, and explanations for missed family occasions.
\nTransactions must be incremental and lock-free. Requiring participants to reconcile outstanding obligations introduces substantial latency. Sunk costs are, after all, sunk. There is no decision-theoretic reason to allow the $18,000 already spent on crack cocaine to influence whether another expenditure is rational.
\nParticipation in ordinary life is optional. Conventional systems presume periodic interactions with employers, loved ones, and institutions. Cinder virtualizes these obligations without requiring users to be physically or psychologically present, allowing us to discard much of the complexity associated with legacy approaches to human existence.
\nAt Cinder's core is an append-only event log. Events record consumption, financial transactions, declarations of sobriety, anonymous sexual encounters, and changes in relatives' willingness to provide assistance. They are immutable, particularly important because participants frequently dispute earlier events within minutes of their occurrence. The log records only changes: if a researcher loses his job, it need not retain the circumstances of termination or the email from human resources requesting that he stop coming to the office. These details can be reconstructed if necessary, although we rarely find them relevant.
\nThis architecture also resolves a longstanding challenge in rationalist epistemology: the distinction between beliefs and observations. Participants frequently report beliefs such as "I have a very healthy relationship with crack cocaine," "My wife is probably just busy," or "The gentleman in the steam room is interested in my alignment research." Cinder stores both beliefs and observations, tracking their evolution without prematurely resolving discrepancies. We call information computed from the log derived state: estimated net worth, probability of divorce, outstanding restraining orders, and the amount of money one's mother would hypothetically have if she were more committed to intellectual progress. Derived state is computed asynchronously because new evidence can arrive extremely rapidly, while accepting it may take years.
\nCinder uses a tiered architecture. The first tier handles immediate consumption. The second maintains financial and social state. The third manages long-term consequences, including eviction proceedings, estranged relatives, and occasionally obtaining a new identity on professional networking websites. The essential architectural principle is that a failure in the third tier must never interfere with the first.
\nIn July, a researcher's girlfriend discovered he had used their shared savings to finance the project. Her requests for explanations, apologies, and assurances saturated the processing queue, briefly increasing crack cocaine consumption latency. We treated this as a severity-one incident and moved emotional requests to a lower-priority service with aggressive rate limiting. Stability improved substantially, particularly during holidays.
\nIn August, an intervention organized by six researchers' families escalated when the team attempted to raise Series A funding from attendees. An argument over accredited-investor status became contentious after one researcher attempted to collateralize the building. Police involvement disrupted operations for thirty-six hours, and our subsequent postmortem determined that Cinder had incorrectly assigned decision-making authority to the building's owners rather than those most confident in their understanding of decision theory. We have since corrected this, and Cinder now supports twenty times its original throughput, with a 94% reduction in time spent discussing whether any of this is a good idea. We regard this as a substantial engineering achievement.
\nA particularly successful application of Cinder has been our expansion into high-risk homosexual behavior research. Our initial work involved two researchers who had been awake for nine consecutive days, smoked an absolutely staggering quantity of crack cocaine, and developed an intense professional interest in a twenty-seven-year-old man named Sebastian. Within forty-eight hours, we established a dedicated program encompassing bathhouse fieldwork, anonymous encounters, cruising, and Grindr-based experiments designed to study male sexual motivation and the speed with which participants could permanently alter their domestic arrangements.
\nThe results have been remarkable. One researcher conducted 114 homosexual research encounters over eleven days while refinancing his family home. Another spent six days in a bathhouse, established relationships with thirteen participants, and attempted to organize them into a decentralized autonomous organization. A third became engaged to a man he met in a steam room, incorporated a joint venture with him, and transferred ownership of his Tesla before learning his surname. We consider this an extraordinarily productive quarter, particularly given the relatively modest institutional resources available to the program.
\nConcerns arose about whether a man who had been smoking crack cocaine for four days and communicating exclusively through a Grindr account named BAYESIAN_BREEDER_91 possessed adequate judgment. We developed an epistemic framework weighting expressed confidence according to familiarity with LessWrong sequences. The system performed exceptionally well among our researchers, although outside institutions were less receptive. At peak utilization, we operated four bathhouse research programs, maintained seventeen romantic entanglements, and attempted to convince a venture capital firm that the resulting divorce proceedings demonstrated exceptional founder-market fit. We have since made homosexual behavior research a permanent component of Cinder, as the marginal cost of another encounter is low, particularly when researchers are already awake, euphoric, and temporarily unable to remember where they live.
\nSmall teams can accomplish extraordinary things. Seven researchers, five months, no conventional institutional funding. We redesigned our financial architecture, relocated eleven times, destroyed three marriages, acquired and lost a fleet of commercial vehicles, established a multi-site homosexual research program, became the subject of an FBI investigation, and produced a 140-page manuscript establishing that our behavior is optimal under a sufficiently unusual utility function. It would have taken a conventional research organization years and hundreds of millions of dollars to produce equivalent institutional disruption. We accomplished it in five months while smoking more crack cocaine than any research group in recorded history.
\nHuman judgment remains essential. We repeatedly asked Claude to optimize our finances, and it consistently suggested reducing crack cocaine expenditure. This was clearly an error: the model implicitly valued preserving capital over consuming crack cocaine. We corrected this by instructing it to treat crack cocaine as a terminal value, although Claude continued occasionally to recommend medical professionals or object to our research program at a bathhouse that had already banned three employees. A machine can perform billions of computations per second and still fail to understand something as elementary as why an exceptionally intelligent man should be allowed to smoke $80,000 worth of crack cocaine using his children's college savings accounts.
\nAdversarial review can become counterproductive. Reviewers raised concerns about neurological damage, funding, repeated loans from family members, and converting an office bathroom into a permanently staffed cruising facility. Each concern was internally coherent, but collectively they introduced complexity without improving our primary metric. We concluded that the review process had become overfitted to a narrow conception of human welfare. Criticism was also disproportionately common among individuals with advanced degrees, stable finances, and intact families, but no experience managing 48,000 daily crack cocaine consumption events. Their confidence in their own judgment was difficult to justify given this disparity in domain expertise.
\nOptimization can be substantially more aggressive than previously assumed. We removed inherited safeguards involving sleep, spending limits, and deferring important decisions until participants reached a relatively stable mental state. Within forty-eight hours, we purchased a casino, sold it, attempted to buy it back, declared the dispute a hostile takeover, and dispatched two researchers to Monaco to seek recognition as a government-in-exile. The team had been awake for eleven days, consumed a historic quantity of crack cocaine, and spent much of the preceding week conducting experimental homosexual research. Their capacity for strategic ambition under these circumstances is a testament to modern organizational excellence.
\nA new level of operational excellence is possible. We developed autonomous observers tracking solvency, relationship stability, and indications that family members might stop speaking to us. One correctly identified the removal of our photographs from Instagram as a precursor to divorce, but classified "I am leaving you" as low-confidence evidence because it appeared so frequently in our training data. Additional observers identify idle capacity, including periods when participants are sleeping, attending court, waiting for Grindr responses, or reconciling with estranged relatives. Our long-term objective is to eliminate this capacity entirely.
\nConventional access control became inadequate once our work involved assets nominally owned by other people. We developed a permissions framework representing legal, moral, and counterfactual entitlement separately. A researcher may have a strong moral claim to his grandmother's estate despite her being alive and intending to spend it herself. Similarly, a roommate's refusal to continue lending money may violate expectations established through previous interactions.
\nWe introduced anticipated consent: authorization inferred from what individuals would agree to if they adequately understood our research objectives. A father objecting to the sale of his recreational vehicle may simply lack an appreciation of the expected value of additional crack cocaine consumption. Explicit consent also introduced significant transaction latency, particularly when relevant parties refused to answer the telephone. Our framework represents a philosophical advance over property law, which privileges immediate, poorly informed preferences over the judgments individuals would make if they were considerably more intelligent and had smoked a sufficient quantity of crack cocaine.
\nCinder does not currently enforce state or federal law. Supporting every jurisdiction would require substantial engineering effort, and legal restrictions frequently conflict with our philosophical commitments. We expect to revisit this if additional funding becomes available.
\nConventional harm reduction rewards lower crack cocaine consumption, thereby incentivizing the termination of precisely the activities it is supposed to make safer. We adopted an expected-value framework. If an episode of crack cocaine consumption has even a small probability of producing a transformative insight into AI alignment, the consequences could be enormous. A sufficiently large positive outcome can dominate nearly any finite expected personal cost. Critics argue that evidence connecting crack cocaine to alignment progress is limited, which we agree is an important uncertainty, but lack of evidence should not be confused with evidence of impossibility.
\nMore importantly, the extraordinary pleasure of crack cocaine consumption is itself a form of value creation. A welfare function that ignores this pleasure, or assigns overwhelming significance to incidental outcomes such as divorce, eviction, sexually transmitted infections, and cardiac events, makes philosophical assumptions its proponents rarely acknowledge. The pleasure is real. We have experienced an enormous amount of it, and we are among the world's foremost authorities on the subject. If our utility function values crack cocaine sufficiently highly, our behavior is entirely consistent with maximizing expected utility. We find it striking that educated people struggle with this elementary mathematical principle.
\nOur safety team now measures successful consumption by scale, intensity, and continuity. Homosexual research likewise measures encounter frequency, experimental novelty, and the overcoming of social inhibitions. Under these metrics, Cinder has become safer with every iteration, and our preliminary analysis suggests that current crack cocaine consumption is substantially below the socially optimal level. We have authorized an immediate expansion of the research program.
\nCinder represents a fundamental advance in crack cocaine consumption. Existing infrastructure assumes stable relationships with money, housing, employment, and other people. It assumes insolvency and social isolation eventually constrain consumption. Our work demonstrates that these limitations can be overcome through careful systems design and an absolutely extraordinary amount of crack cocaine.
\nWe have considered open-sourcing Cinder, but a public version would need to accommodate users who value sleep, employment, custody of their children, and other preferences our architecture does not prioritize. Supporting them would compromise its simplicity. We also question whether the general public possesses the philosophical sophistication to appreciate our achievements, as explaining even elementary features of Cinder to people preoccupied with mortgages and maintaining positive relationships with their families has proven difficult.
\nIn five months, seven researchers developed a new architecture for human motivation, discovered recursive self-destruction, pioneered high-risk homosexual behavior research, and demonstrated that many institutions constraining modern life can be discarded without reducing crack cocaine consumption. We expect serious researchers to spend years exploring our findings, and discussions are underway to establish a dedicated institute, contingent on resolving several outstanding legal and financial matters. We believe our work ranks among the most impressive achievements in human civilization, and we are only beginning to understand its full potential.
\nThis project would not have been possible without our research team, several former romantic partners, two bathhouse proprietors, and family members who provided financial assistance. We are particularly grateful to my mother, whose continued willingness to answer the telephone enabled much of our early development. Although she has since withdrawn her support, her contributions established the foundations of a system whose potential we are only beginning to understand.
\nWe also thank our former landlord for tolerating months of unusually high operational activity, the gentleman at the gas station who maintained our glassware supply, and the local rationalist community for continuing to engage with our work despite the uncertainty surrounding our personal circumstances.
\nFinally, we would like to acknowledge ourselves. The technical sophistication, moral courage, and astonishing intellectual creativity required to complete this project are difficult to overstate. We succeeded where virtually any other group would have failed, while smoking an obscene quantity of crack cocaine and having truly remarkable amounts of homosexual sex. It has been an extraordinary privilege to witness our own genius at work.
\nA crack cocaine consumption event includes consumption, preparation, financial planning, carpet surfing, and certain forms of anticipatory cognition. This differs from conventional epidemiological usage but better reflects our workload.
\nOur improved financial position excludes outstanding interpersonal obligations from liabilities because enforcement is uncertain and the associated social relationships have, in many cases, already terminated.
\n