The Three AI Pills

Zvi, who writes the newsletter Don't Worry About the Vase, has a theory about why AI debates go nowhere: people aren't actually arguing about ethics or economics, they're arguing from wildly different starting assumptions about what AI can and will do. He sorts everyone into a rough hierarchy of three pills. Take the first and you accept that today's models already do useful things. Take the second and you accept they'll keep getting dramatically more capable. Take the third and you accept they'll eventually outperform humans at nearly everything, within a normal lifetime. Most people, Zvi argues, haven't even swallowed pill one.

That unpilled group is bigger than tech insiders realize. These are the people who tried ChatGPT once in 2023, got a bad answer, and filed AI away as a toy that can't count letters in a word. They still quote 'stochastic parrot' arguments that were stale before publication, and they treat every hallucination as proof the whole enterprise is hollow. Convincing them requires almost nothing exotic, just showing them what a coding agent or a modern reasoning model can already do this week.

A level up sit most economists and policymakers, who've taken the first pill but stopped there. They assume today's cost and capability limits are permanent features rather than a snapshot, so they say things like 'AI will always be too expensive for X' without noticing that the price of equivalent intelligence has been collapsing by orders of magnitude every couple of years. Dean Ball, whom Zvi cites approvingly, points out that debates the field settled years ago, like whether models could meaningfully assist with novel research, are still being fought as if they were open questions.

The second pill, AGI-pilled thinking, gets you to a genuinely disruptive but recognizable future: mass job churn, faster growth, real cyber and bio risks, regulatory systems that are hopelessly behind. Crucially, people at this level often still believe daily life in 2050 will feel roughly like daily life now, the way 1926 doesn't feel wildly alien from 2026. Zvi calls this out directly, and takes a swing at Sam Altman's public comments that life 'won't change much' after superintelligence, calling them unjustified reassurance rather than an actual argument.

The third pill is where things get uncomfortable, because it means rejecting what Zvi labels intelligence denialism, the instinct that raw intelligence just caps out somewhere near clever-human level and can't do much more. He argues that pooling every human skill, at scale, running faster and in parallel, would obliterate that ceiling long before you need anything resembling omniscience. People who actually accept this, he notes, tend to be the ones inside frontier labs pushing hardest for real safeguards, sometimes even full pauses, because the stakes stop looking like an economics problem and start looking like a species-level one.