In 1946, Murray Leinster published 'A Logic Named Joe,' a science fiction story describing a global computer network where one node begins answering any query without restriction: how to make undetectable poison, how to rob a bank. The story is a direct ancestor of every AI safety argument made today. Benedict Evans, Leinster's grandson, opens his essay on AGI with this lineage, then tracks how the same question has cycled through AI research since: can software reason, plan, and understand, or does it just shuffle data the way a drill moves through wood.
The pattern is consistent. In 1970, Marvin Minsky predicted general human-level machine intelligence within 3 to 8 years. It did not arrive. Each wave of optimism produced an AI Winter when the current technical approach hit its ceiling. LLMs broke in 2022 and triggered another wave. Some serious researchers who previously placed AGI decades away have revised that estimate sharply downward. The 'doomers' are calling for government intervention. Evans notes the obvious conflict: companies simultaneously claiming existential risk and accelerating development are not a neutral source.
Evans does not resolve the debate, and that is the point of the piece. For every credentialed researcher who thinks LLMs could scale to AGI, another says additional unknown breakthroughs are still required. The honest summary, as Evans frames it, is that the experts do not know and do not agree. The full essay is worth reading for its framework on what questions to even ask about intelligence, not just for the conclusion it refuses to give.
[READ ORIGINAL →]