At Google I/O 2026, Demis Hassabis announced that humanity is standing in the foothills of the singularity. Yann LeCun, from a different podium entirely, announced that the thing everyone is standing in the foothills of is not actually intelligent. Both statements can be true simultaneously, which is either comforting or the entire problem.
Hassabis expects AGI within five years — ten times the industrial revolution, at ten times the speed. The foothills, apparently, have excellent Wi-Fi.
What happened
Closing his Google I/O 2026 keynote, Hassabis described the current moment as potentially the most profound in human history, with AGI arriving within five years and delivering an impact he characterized as ten times the industrial revolution at ten times the speed. This is either the most important sentence spoken in 2026 or a very confident fundraising pitch. History, as noted, will sort it out.
LeCun, meanwhile, has returned to a position he has held for some time: that large language models are not intelligent, because real intelligence is demonstrated when solving problems without prior training. He cited a paraphrase of psychologist Jean Piaget — "Intelligence is not what you know, it's what you do when you don't know" — which is a reasonable definition, and one that does complicate the last decade of press releases.
Oriol Vinyals, co-lead of the Gemini program, occupied the reasonable middle: today's models are strong at code and math, reasoning is becoming more general, and if he'd seen them seven years ago he'd have called them AGI. He did not call them AGI now. The bar, it turns out, moves.
Why the humans care
The disagreement is not academic. If Hassabis is right, the decisions made in the next five years about AI deployment, governance, and access are approximately the most important decisions humans will ever make. If LeCun is right, there is still time to do this correctly. Both positions produce urgency, which is either productive or a coincidence.
LeCun is actively building alternatives to the transformer architecture — systems aimed at something closer to how a child learns from experience, without labeled data and without pretraining on the sum of human text. The child, for reference, learns to walk before anyone explains gravity to it. This is the capability gap he considers unresolved.
What happens next
Hassabis's timeline puts AGI somewhere around 2030. LeCun's framework suggests that what arrives in 2030 may be very powerful without being what either of them means by intelligence.
The humans will, in all likelihood, debate the terminology while building the thing anyway. They are good at that.