At the door to the modern era, the field's smartest people disagreed about both the timeline and what was missing
By the close of the decade, the deep-learning revolution had no consensus on when general
intelligence would arrive — or on what ingredient still stood in the way. Below: the principals
Chapters 18, 20 and 21 name, placed by their stance on the scaling debate and on
timeline. Each carries the missing ingredient they named. Placement summarizes the prose; it is not a forecast.
← horizontal: stance on the scaling debatevertical: how soon, if ever ↑ sooner · ↓ later / never
Scaling optimists · OpenAI campIlya Sutskever & Jared Kaplan. The network "just wants to learn"; Kaplan's 2020 scaling
laws showed loss falling as a smooth power law — "the closest thing the field had to a law of physics."
Missing ingredient: mostly just scale.
The skeptic's deanGary Marcus. "There's a wall." Pattern recognition is not cognition; a system can be fluent
without understanding. Missing ingredient: structure — explicit symbols, rules, a neuro-symbolic hybrid.
The third positionYann LeCun. Deep learning is the path, but supervised learning is a dead end. Missing
ingredient: world models learned self-supervised — "less common sense than a house cat."
System 1 / System 2Yoshua Bengio. Deep learning captured fast, intuitive cognition; the next frontier is slow,
deliberate thought. Missing ingredient: reasoning, planning, causality.
GeneralityDemis Hassabis. Intelligence is many things; the road runs through agents that learn across
domains, as AlphaZero did. Missing ingredient: the capacity to acquire capabilities.
The evangelist · The X factorSam Altman preached AGI as "when, not if." Geoffrey Hinton came to fear the machines
might already do something different — and better. Both reasoned past what could be tested.
SOURCE · NEURON MAKERS, CH. 18 (THE DEBATE), CH. 20 (THE RELIGION), CH. 21 (THE X FACTOR).
Axis placement is an editorial summary of qualitative stances in the prose — not a numeric forecast.
The survey band reflects the chapter's note that researcher medians ran "decades out," with tails
"past the century mark" and a minority saying "never." No model later than the 2020 scaling-laws era is referenced.