Geoffrey Hinton

Co-published the landmark paper on training deep belief networks using contrastive divergence, breaking a decade-long stagnation in neural network research. This directly triggered the deep learning revolution and validated decades of neural net research.

Led AlexNet at ImageNet 2012, reducing the top-5 error rate from ~26% to ~15% — a shock to the computer vision field. Famously auctioned his lab to Google, convinced that deep learning would dominate AI. That bet proved correct within five years.

Briefly argued (2022) that backprop is biologically implausible and we may need fundamentally different learning algorithms. Proposed "forward-forward" as an alternative. The academic community was skeptical, and Hinton himself acknowledged this was speculative.

Left Google in May 2023 to speak freely about AI dangers. Stated publicly there is a 10–50% chance that AI systems will eventually kill humanity. A dramatic reversal: for decades he dismissed AI safety concerns as premature. Now calls himself 'a pessimist.' His defection from the optimist camp sent shockwaves through the industry.

Argued that large language models likely have a form of understanding and representation, not just statistical pattern-matching. This put him at odds with critics like Gary Marcus and the 'Stochastic Parrots' paper. His view: if something behaves intelligently in diverse contexts, dismissing it as 'mere statistics' may be intellectually dishonest.