blog.samaltman.com, June 2025


The Takeoff Has Started

  • We’re “past the event horizon”: digital superintelligence is close, yet it feels far less weird than expected — no robots on the streets, people still die of disease.
  • Systems smarter than people in many ways already exist and significantly amplify human output; the hardest scientific insights (those behind GPT-4 and o3) are behind us.
  • ChatGPT is in some sense “more powerful than any human who has ever lived” — hundreds of millions rely on it daily, so even small capabilities (or small misalignments) have enormous multiplied impact.

Near-Term Timeline

  • 2025: agents doing real cognitive work; coding permanently changed.
  • 2026: likely systems that can figure out novel insights.
  • 2027: possibly robots doing tasks in the real world.
  • By 2030, one person will get vastly more done than in 2020; experts who embrace the tools will still beat novices.

Self-Reinforcing Loops

  • AI is accelerating AI research itself — a “larval version of recursive self-improvement.” Scientists already report 2–3x productivity gains; a decade of research in a year would change everything.
  • Economic value is funding compounding infrastructure buildout; robots that build robots (and datacenters that build datacenters) aren’t far off.
  • As datacenter production automates, the cost of intelligence should converge toward the cost of electricity. (An average ChatGPT query uses ~0.34 watt-hours and ~1/15 of a teaspoon of water.)

Abundance in the 2030s

  • Intelligence and energy — the two fundamental limiters on human progress — will become wildly abundant; with good governance, everything else follows theoretically.
  • In the most important ways, life stays the same (family, creativity, games, lakes); in other ways it will be wildly different from anything before.
  • The singularity’s pattern: wonders become routine, then table stakes — from admiring an AI paragraph to expecting an AI-discovered cure.

Jobs and Adaptation

  • Whole classes of jobs will disappear, but rapidly growing wealth will enable policy ideas never before possible; the social contract will change gradually but amount to something big.
  • Humans hold a durable advantage: we’re hard-wired to care about other people, not machines.
  • A subsistence farmer would call today’s jobs “fake”; future jobs will look just as fake to us yet feel important and satisfying to those doing them.
  • Lived from the inside, the singularity happens “bit by bit” — the exponential curve always looks vertical ahead and flat behind.

The Path Forward

  • First, solve alignment: robustly guarantee AI systems act on what we collectively want long-term (social media feeds are cited as misaligned AI — exploiting short-term impulses against long-term preferences).
  • Then, distribute widely: make superintelligence cheap, broadly available, and not concentrated in any person, company, or country; give users freedom within broad, socially decided bounds.
  • The industry is “building a brain for the world”; the limiting factor will be good ideas — “idea guys” are about to have their day.
  • OpenAI sees itself as a superintelligence research company first; “intelligence too cheap to meter is well within grasp.”

“May we scale smoothly, exponentially and uneventfully through superintelligence.”