Zvi Mowshowitz, the AI commentator behind the blog Don’t Worry About the Vase, published a framework on August 5 that sorts every AI opinion into three cumulative belief tiers. His claim: the overwhelming majority of economists, policymakers, and casual users have not cleared even the first tier, and that gap is why so much public debate about AI still argues over questions he considers settled.
The three tiers, as he lays them out, build on each other. The first, which he labels being “AI pilled,” just means accepting that current systems already outperform old assumptions: coding agents, cheap inference, tasks that used to be manual now handled by typing a request. The second, “AGI pilled,” means accepting that capability keeps compounding well past today’s tools, enough to eliminate large categories of work and strain institutions that were not built for the pace of change. The third, “ASI pilled,” means accepting that AI will eventually outperform humans at nearly everything within a normal lifetime, a threshold he says he and much of frontier lab staff have already crossed.
Mowshowitz argues most economists and government officials have stopped at the first tier or never reached it, still citing outdated benchmarks and treating today’s limitations as permanent. He contends that reaching the second tier alone should be enough to justify heavier investment in safety testing, export controls, and regulatory readiness, and that reaching the third tier is what pushes some of his peers toward proposals as drastic as an international pause on frontier development.
That claim deserves the scrutiny any insider consensus deserves. Frontier lab employees are a self-selected population: people who doubt near-term superintelligence tend not to stay long at companies whose valuations depend on investors believing in it. A belief that is common among the staff of Anthropic, OpenAI, and their peers is not independent evidence that the belief is correct, any more than genomics researchers circa 2001 predicting imminent personalized medicine validated the timeline. Mowshowitz’s own essay does not cite a benchmark, a revenue figure, or a specific capability threshold as proof the third tier is near. It argues from analogy and frustration with critics, not from data the reader can check.
The more useful question for an operator is not which pill a person has swallowed but who this taxonomy actually serves. It works well as a debate-calibration tool among AI safety writers arguing past each other online. It works poorly as a description of how enterprises and policymakers actually behave, because both groups act on budget cycles and liability exposure rather than declared conviction. A procurement team rolling out a coding copilot to cut ticket volume has taken no formal position on superintelligence; it bought a tool because the ROI math worked this quarter. A legislature drafting model-evaluation mandates is responding to a specific incident or lobbying push, not endorsing a tier of belief about where capability tops out. Judged by declared belief, both look “unpilled.” Judged by revealed action, both are already hedging against exactly the scenario Mowshowitz says they deny.
That gap between stated belief and revealed behavior is the more reliable signal for anyone tracking where this goes next. Watch what organizations fund, mandate, and insure against, not what their spokespeople say about AGI timelines. The next ninety days of enterprise AI budgets and legislative text will say more about the industry’s actual risk posture than any taxonomy of conviction can.
Published by Don’t Worry About the Vase on August 5, 2026.
