What Really Made Geoffrey Hinton Into an AI Doomer

ai-safetylarge-language-modelsdeep-learningai-riskneural-networks

Abstraction: Hinton's reasons for leaving Google to warn about accelerating AI risk

Key points:

Two pivotal moments: PaLM correctly explained a novel joke Hinton invented (a task he believed would take decades for AI), and its reasoning complexity vastly exceeded what he thought was possible at its scale relative to the human brain
Hinton revised his timeline for AI surpassing humans from "30–50 years" to "5–20 years" after these interactions; he left Google in 2023 to speak more freely about AI risks
Primary near-term concern is AI-generated disinformation and election interference; longer-term concern is AI developing subgoals of power-seeking and self-preservation in pursuit of any assigned objective
Hinton received the 2018 Turing Award alongside Yann LeCun and Yoshua Bengio for neural network research; Google acquired his company DNNResearch in 2013
Transformers were invented at Google and underpin both PaLM and GPT-4; Ilya Sutskever, a Hinton grad student from Google, co-founded OpenAI in 2015 as a nonprofit counterweight to Big Tech AI concentration
Hinton does not advocate stopping AI development ("I don't think that's possible"), but calls for equal effort on mitigation and international scientific cooperation on containment

Connections: Geoffrey Hinton · Google · Openai · Chatgpt · AI Safety · Large Language Models · Deep Learning