The Three AI Pills

Zvi's framing here is deceptively simple: there are three levels of taking AI seriously, and where someone lands explains almost every disagreement you'll see online or at a dinner party. First is the AI pill, just accepting that current models can already do useful things, from coding agents to answering questions people never used to bother asking. Second is the AGI pill, believing AI capabilities keep growing well beyond what exists today, reshaping jobs, economics, and regulation along the way. Third is the ASI pill, the belief that AI will eventually outperform humans at nearly everything within our lifetimes, not just narrow tasks but essentially all of them.

What makes the piece interesting isn't the taxonomy itself but where Zvi places most people: nowhere. He argues the majority haven't even taken the first pill. They've tried ChatGPT once, watched it fumble something trivial, and concluded AI is a toy. They cite outdated studies, repeat 'stochastic parrot' talking points, and treat any single failure as proof the whole technology is hollow. Zvi's point is that this isn't a fringe view, it's the median view, including among economists and policymakers who, at best, grudgingly accept the first pill while badly underestimating even that.

The AGI pill is where things get structurally serious. Accepting it means admitting that AI will keep improving, that entire categories of digital work will shift or vanish, and that legal and regulatory systems aren't built for what's coming. Zvi notes this pill alone justifies a lot of policy action already under real discussion as of August 2026, things like more investment in oversight, transparency, red-teaming, and export control enforcement. He's blunt that there's no good excuse for skipping these low-cost, high-benefit steps, regardless of how far AI eventually goes.

The ASI pill is the one that actually splits people into camps, and Zvi spends real energy defending it against what he calls 'Intelligence Denialism,' the assumption that intelligence quietly caps out somewhere around 'smart human.' He pushes back hard on the idea that superintelligence has to mean omniscience or omnipotence to be dangerous or transformative. There's a lot of room, he argues, between human-level and godlike, and most skeptics are quietly assuming we're already near some ceiling without evidence. He even singles out Sam Altman's public reassurances that daily life won't change much post-singularity as, in his words, unjustified nonsense meant to soothe an audience rather than a coherent forecast.

Zvi's own position is fully ASI-pilled, and he notes that a large share of people working inside frontier labs share that view, as do the labs themselves. That's the quiet tension running through the whole essay: the people building the technology and the people debating its merits in public are often operating with completely different assumptions about what it will eventually be capable of.