AI has a real chance of killing us all, and most people have 1-3 productive years left: Jeff Kaufman

ajeya_cotra · x · 2026-08-30

EA blogger Jeff Kaufman published a long post on his view of the near-term AI future, amplified by Ajeya Cotra. Key points:

Existential risk is real: AI has a very real chance of killing everyone, though he suspects it won't because many people will work hard to avoid it.
Very short timelines: AI is approaching or exceeding expert human ability across many areas; most people should plan for 1-3 more years of productive contribution—but don't bet your livelihood on it, given remaining uncertainty.
Recursive self-improvement is already happening: AI is substituting for human expertise in AI development itself; as the human share shrinks, this will compound dramatically.
Goalposts are moving: drop a top model into 2006 and it would clearly be considered "AGI."
To skeptics: if AI doesn't seem to be improving fast, you're likely not applying it to hard enough problems, not giving it enough context/tooling/tokens, or not using top models.
Other concerns: data-center environmental impact (plausible futures where almost all energy powers computers); warning shots have begun appearing as hacks; keeping non-public information is increasingly dangerous.

More from AGI Musings