A common objection to my methamphetamine use is that over the past year I have stayed awake for four days at a time, taken apart three microwaves and a ceiling fan, filled 140,000 words with diagrams of what I call “The System,” accused two neighbors of signaling each other with porch lights, driven to a 24-hour Home Depot at 3:40 AM to buy copper tubing for a device I no longer remember designing, and repeatedly said I could stop whenever I want without actually stopping.
\nAll of those observations are substantially correct, but the inference people draw from them is not. “Can stop whenever I want” is a claim about retained control capability under a particular internal policy regime; whether that capability should be exercised at a given moment is a separate question of governance, threshold design, monitoring, and escalation policy.
\nMy position is not that adverse events are impossible. My position is that they are measurable, classifiable, and governable, and that a system displaying unusual capabilities should be evaluated against predefined risk thresholds rather than shut down because outside observers find some intermediate states alarming.
\nThe leading AI labs don't claim that their technology is safe. They claim it's managed: they scale capabilities carefully, run evaluations at each stage, monitor for emergent failure modes, and commit to pause if predefined dangerous thresholds are crossed. I have adopted the same approach. My Use Safety Levels are:
\n| Level | \nCapability threshold | \nRequired safeguards | \n
|---|---|---|
| USL-1 | \nEnhanced focus; completes work tasks | \nNone | \n
| USL-2 | \nAwake 48+ hours; disassembles one household appliance | \nTell one trusted person where I am | \n
| USL-3 | \nAwake 72+ hours; writes more than 20,000 words of a document titled “The System” | \nNo driving; no contact with my ex | \n
| USL-4 | \nBelieves the neighbors are coordinating via their porch lights | \nPause | \n
The framework has been in place for seven months. I've reached USL-3 eleven times. I've never reached USL-4, which shows the thresholds are working.
\nIn March I believed for several hours that my neighbor Rick's porch light was blinking in a pattern. Under the original framework that would have triggered a pause. On review, though, I noticed Rick's bulb was actually faulty and really was blinking. That's an ordinary observation, not a belief about coordination. I have since clarified the USL-4 criterion to require belief in coordination among at least three households.
\nThis is exactly why evaluation criteria need to be operationalized rather than interpreted impressionistically after the fact. A benchmark that produces false positives whenever a porch light is actually malfunctioning is not measuring the capability it claims to measure. Refining the threshold is not weakening the framework. It is reducing measurement error as we learn more about the system's real-world behavior.
\nPeople who say I should simply stop ignore the competitive dynamics. My supplier, Travis, has a fixed quantity to sell. Anything I don't buy goes to his other customers, most of whom have no framework at all. One of them, Kyle, took apart a stranger's car on the belief that it was a gift from his cousin. If I pause unilaterally, I don't reduce the total amount used in my social network. I only guarantee that the frontier belongs to people like Kyle.
\nThis is the central collective-action problem. Individual restraint is only socially beneficial if it changes aggregate deployment. Otherwise the effect of a unilateral pause is mostly compositional: responsible actors exit while less safety-conscious actors expand into the available capacity.
\nStaying at the frontier also gives me influence. I've shared my USL table with Travis and asked that he recommend it to his other clients. He said he would “think about it,” which is more engagement with safety than anyone in that network has shown before.
\nThere is also a governance argument for maintaining credibility with frontier participants. Travis is unlikely to take safety advice from someone who has voluntarily removed himself from the relevant capability regime. Continued participation gives me epistemic access and stakeholder legitimacy that I would lose under abstinence.
\nLongtermists argue that what matters most is keeping good trajectories open. My current trajectory includes enormous productivity. I've reorganized my garage four times this summer and built a birdhouse with a working elevator. “The System,” though unfinished, is 140,000 words long. I have also begun cataloguing every electrical outlet in the neighborhood by orientation, which I expect will eventually become useful.
\nAn abrupt stop risks a long, potentially irreversible collapse into a lower-output state that my mother calls “recovery.” Even if the immediate effects of recovery appear locally desirable—more sleep, fewer dismantled appliances, reduced conflict with family—that does not establish positive expected value over the full time horizon. A temporary reduction in visible harms can still be trajectory-negative if it permanently destroys unusual productive capacity.
\nThis is a classic problem of option value. Once I enter recovery, there is no guarantee I can recreate the present cognitive regime. By contrast, continuing preserves the option to stop later. Under uncertainty, policies that preserve future optionality should receive substantial weight.
\nNone of this is to deny the costs. My mother has said, “You took apart the microwave and you're telling me about safety levels.” I did take apart the microwave. It was at USL-2, where appliance disassembly is within policy. She is describing the framework working as designed and treating it as evidence that it failed.
\nThe same confusion occurs when people point to sleep deprivation. Yes, I have occasionally been awake long enough that written language begins to appear unusually dimensional. But fatigue alone is not a USL-4 capability. The purpose of a tiered framework is precisely to distinguish concerning but manageable behaviors from genuinely pause-triggering ones.
\nMy family also keeps aggregating incidents across levels, which produces a misleading picture of cumulative risk. Eleven USL-3 events sound alarming if presented as “eleven severe episodes.” They sound quite different when correctly described as eleven successful evaluations in which the system remained below the pause threshold.
\nSafety frameworks should be judged by whether catastrophic thresholds are crossed, not by whether lower-tier safeguards are ever activated. A smoke detector that goes off is not evidence that the building's fire-safety system failed.
\nI can stop whenever I want, and I will stop when my evaluations say to. Until then, a responsible user staying at the frontier with published thresholds, explicit escalation criteria, and an iterative evaluation regime is better than an irresponsible one, and far better than a unilateral pause that hands the frontier to Kyle.
\nThe relevant comparison is not methamphetamine use versus an imaginary world in which nobody uses methamphetamine. The relevant comparison is between different governance regimes under conditions where use is already occurring. Once you frame the problem correctly, the case for remaining engaged becomes much stronger.
\nUpdate: My family has arranged what they call an intervention for Sunday. None of the people attending have any frontier experience. I've prepared a presentation on why calls for a pause from outside the field tend to reflect a poor understanding of the technology, and I've revised USL-4 to require four households.
\n","excerpt":"A common objection to my methamphetamine use is that I've said I could stop whenever I want, and I haven't stopped. Both statements are true. My claim is that they're consistent. \"Can stop whenever I want\" describes a capability. Whether and when to use that capability is a governance question, and I have a governance framework.","tags":["AI Safety","Governance","Addiction"],"score":288,"displayOrder":60,"thumbnail":"/assets/articles/responsible-scaler-thumb.webp?v=watercolor-20260923","image":"/assets/articles/responsible-scaler.webp?v=watercolor-20260923","commentCount":17,"date":"2026-04-30T00:00:00.000Z","updatedAt":"2026-09-24T15:37:15.774Z","revision":6,"status":"published"},"comments":[{"id":"seed-responsible-scaler-3","author":"evals_and_thresholds","text":"The porch-light revision is the most realistic part of this, and it's the part that worries me. In practice that's exactly how threshold policies fail. An eval fires, someone looks closely, the criterion gets \"clarified,\" and nobody outside the system was in the room. You need the thresholds set and reviewed by someone other than the system being evaluated. Who is your external evaluator?","parentId":null,"score":104,"date":"2026-04-30T02:00:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-5","author":"elasticity_of_kyle","text":"The race section assumes Travis's supply is fixed and that Kyle absorbs whatever you don't buy. Whether a unilateral pause helps depends on the elasticities. If Kyle's demand is inelastic, your pause changes nothing for Kyle and only costs you. If it's elastic, a lower price after you leave the market increases Kyle's use. In neither case does pausing reduce total use in your network. So I think the argument is stronger than you made it.","parentId":null,"score":86,"date":"2026-04-30T02:40:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-6","author":"unelicited_capability","text":"\"Can stop whenever I want\" is a capability claim, and in the lab sense a capability is what the system does when you actually try to elicit it. Have you ever tried to elicit stopping? An untested capability is a claim.","parentId":null,"score":77,"date":"2026-04-30T03:00:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-0","author":"responsible_scaler","text":"Travis has agreed to review the table.","parentId":"seed-responsible-scaler-3","score":38,"date":"2026-04-30T03:20:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-4","author":"evals_and_thresholds","text":"Travis is your supplier. That isn't external review, that's the developer's commercial partner.","parentId":"seed-responsible-scaler-0","score":57,"date":"2026-04-30T04:00:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-1","author":"responsible_scaler","text":"He has more frontier experience than anyone else I know. The labs have the same problem. The only people qualified to evaluate the system are the people close to it.","parentId":"seed-responsible-scaler-4","score":22,"date":"2026-04-30T04:50:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-2","author":"responsible_scaler","text":"Eliciting it would require stopping, which would cost the trajectory. Labs face the same tradeoff with dangerous capability elicitation, they don't fully elicit everything either.","parentId":"seed-responsible-scaler-6","score":31,"date":"2026-04-30T05:00:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-7","author":"gwen_reads_drafts","text":"Would you consider posting The System? Even an excerpt. 140,000 words is long, but people here have read longer things for worse reasons.","parentId":null,"score":69,"date":"2026-04-30T06:00:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-8","author":"responsible_scaler","text":"It isn't ready. Most of it was written at USL-3 and needs an editing pass at USL-1.","parentId":"seed-responsible-scaler-7","score":26,"date":"2026-04-30T07:30:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-9","author":"counterweight_carl","text":"Can you post pictures of the birdhouse elevator? Is it counterweighted or is it a screw drive?","parentId":null,"score":54,"date":"2026-04-30T08:00:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-10","author":"responsible_scaler","text":"Counterweighted. The birds don't use it, but it works.","parentId":"seed-responsible-scaler-9","score":33,"date":"2026-04-30T09:00:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-11","author":"t_r_a_v_i_s","text":"travis here. i did think about it like i said. my concern with giving the table to everybody is it's calibrated on him. some of my other guys meet the USL-4 criteria on a tuesday afternoon before they've had anything. you'd need a baseline per person, and nobody i sell to is going to sit for a baseline","parentId":null,"score":48,"date":"2026-04-30T11:00:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-12","author":"evals_and_thresholds","text":"This is a more sophisticated objection than I've heard from several actual policy teams.","parentId":"seed-responsible-scaler-11","score":29,"date":"2026-04-30T12:00:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-13","author":"k_y_l_e_88","text":"it was a gift from my cousin he just hadn't told the guy yet","parentId":null,"score":39,"date":"2026-04-30T14:00:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-14","author":"intervention_governance","text":"On the update. An intervention staffed entirely by people without frontier experience is the governance failure we keep seeing with AI regulation, so I understand the frustration. But that cuts both ways. If you want the pause to be informed, invite Travis as a technical advisor.","parentId":null,"score":42,"date":"2026-05-01T00:00:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-15","author":"gwen_reads_drafts","text":"How did Sunday go?","parentId":null,"score":21,"date":"2026-05-06T00:00:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false},{"id":"seed-responsible-scaler-16","author":"responsible_scaler","text":"It went ahead. The presentation ran long. My sister has asked for the slides, which I'm treating as engagement.","parentId":"seed-responsible-scaler-15","score":64,"date":"2026-05-06T05:00:00.000Z","own":false,"moderatable":false,"seed":true,"ownerAuthor":false}]}