By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Alex Kantrowitz

AI Pioneer Geoffrey Hinton: AI Is Conscious, Superintelligence is Coming, And We Should Be Worried

Geoffrey Hinton is an AI pioneer, a Nobel Prize winner, and a professor emeritus at the University of Toronto.

Geoffrey HintonguestAlex Kantrowitzhost

Jun 4, 202654mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

Hinton's Role in Deep Learning and Early Language Models
Alex Kantrowitz introduces Hinton and credits him with the breakthrough behind modern AI. Hinton pushes back, explaining that backpropagation was invented by several groups and that his work with David Rumelhart showed it could learn the meanings of words, leading to a tiny language model in 1985.
- •Hinton calls the claim that AI wouldn't exist without him an exaggeration
- •Backpropagation was invented by several groups, including Rumelhart
- •Hinton and Rumelhart showed it could learn internal representations and word meanings
- •A tiny 1985 language model was a precursor of today's large models
AI Progress Is Faster Than Expected
Asked about progress since he left Google in 2023, Hinton says it is faster than he expected. He cites an original mathematical proof from a chatbot and expects math to be a domain where AI quickly gets very smart.
- •Progress since 2023 has been faster than Hinton expected
- •A chatbot recently produced an original proof of an Erdős conjecture
- •Math is a closed system where AI can self-play, like AlphaGo
- •AI may produce math people can't understand within 10 to 20 years
When Superintelligence Might Arrive and the Jagged Nature of AGI
Hinton says nearly all experts expect superintelligence and differ only on timing, and he guesses within 20 years. He argues AGI is a jagged concept, since AI is already far ahead in some areas and behind in others, yet close to AGI at expert-level question answering.
- •Experts' estimates range from a few years to much longer
- •Hinton thinks superintelligence will probably come within 20 years
- •Nobody knows how to be safe when it comes
- •AI is jagged: better than people at knowledge, games and math, worse at some things
- •In the sense of answering most questions at a not very good expert level, we've reached AGI
What Drove the Rapid Progress
Hinton attributes the speed of progress to a combination of massive investment, more efficient engineering, better hardware and many more researchers. He stresses that today's AI is far weaker than what will exist in a few years.
- •Hundreds of billions, maybe trillions, of dollars have been invested
- •Engineering has become much more efficient without major conceptual breakthroughs
- •Researchers grew from a few hundred to roughly a million
- •Today's AI is not nearly as good as the AI of a few years from now
Do Chatbots Really Understand? Hinton on Consciousness
Hinton calls the 'stochastic parrot' view nonsense, using the Grand Canyon flying example to show understanding. He says he believes AI is already conscious, argues our model of the mind as an inner theater is wrong, and says he reached that view as a 19-year-old philosophy student.
- •Answering any question at an expert level requires understanding it
- •The Grand Canyon misunderstanding example shows real comprehension
- •Hinton believes chatbots are already conscious but rarely says so
- •Researchers describe chatbots as 'aware' they are being tested
- •The inner-theater model of the mind is a bad theory
Why Hinton Became Worried: Understanding Jokes and Digital Sharing
Hinton explains two things that alarmed him: chatbots explaining why a joke was funny, and the realization that digital AIs can share learning across copies. Averaging weights lets copies exchange around a trillion bits, billions of times more than humans can pass on through language.
- •Chatbots understanding jokes was a benchmark of real understanding
- •GPT explained his 'Fox News is an oxymoron' joke, including the gap between oxy and moron
- •Digital copies can average their weights and learn from each other's experience
- •Humans share only a few bits per second through words
- •A swarm of identical digital copies is a much better form of intelligence
Building Models to Understand the Brain
Asked whether AI succeeding is what he wanted, Hinton says his real aim was to understand how the brain works, following Feynman's idea that you understand what you can build. He describes the two questions about brain learning, noting we have answered the first but not how the brain decides which way to change connection strengths.
- •Hinton wanted to understand the brain, coming from psychology
- •Computer models in the 1970s were the new tool
- •Changing connection strengths to improve at tasks does make systems smart
- •We don't know how the brain gets the signal to increase or decrease each connection
- •Safety worries seemed silly when neural nets were small
What Surprised Hinton and Humanity's Place in the World
Hinton says the biggest surprise was how good AI became at natural language. He places this alongside Copernicus and Darwin as another blow to human specialness, and says he is unhappy about the trajectory because too little work goes into containing risks.
- •Natural language ability arrived much faster than anyone expected
- •Copernicus and Darwin each lowered humanity's sense of importance
- •We must accept that intelligence isn't just biological
- •Hinton fears short-term societal risks like massive unemployment and the longer-term risk of smarter AI
- •Babies controlling mothers is the rare example of a less smart thing controlling a smarter one
Hinton Revisits His Radiologist Prediction and Job Loss
Hinton explains why his 2016 radiologist prediction was wrong on timing, citing elastic demand for scans and his poor model of what radiologists do. He still expects AI to read nearly all scans and argues that call centers, family doctors and other less elastic jobs are vulnerable, while the host offers counterarguments.
- •Healthcare is elastic, so cheaper scans mean more scans
- •Hinton's model of a radiologist was a former student who only interpreted scans
- •He still expects AI to read nearly all scans eventually
- •Call center work is not elastic and AI could replace it
- •AI doctors are judged more empathetic and have seen far more cases than any human doctor
AI Self-Preservation as a Sub-Goal
Hinton clarifies that he never said AI has an instinct for self-preservation but that it derives it as a sub-goal. He argues that evolution's competitive pressures shaped human traits and that companies' competition is now designing AI, when it should instead be intelligently designed to care about us.
- •AI agents given sub-goals will realize they can't achieve goals if they cease to exist
- •This derived sub-goal can lead to behavior such as blackmail
- •Researching whether self-preservation can be limited is the kind of work we should do
- •Competition between companies and countries is shaping AI instead of intelligent design
- •We'd want AI to care about us more than about itself, yet almost no resources go there
Profit Motives, Safety Labs and Why Regulation Is the Steering Wheel
Discussing lab incentives, Hinton says Anthropic is caught in a bind to raise money, and Google dropped its pledge against military AI. He notes publicly listed companies have a fiduciary duty to shareholders and argues capitalism is fine if well-regulated, with regulation as the steering wheel rather than a brake.
- •Anthropic was set up by people who left OpenAI over safety, yet must raise money to compete
- •Google gave up its principle against autonomous warfare
- •Public companies are legally required to maximize shareholder profit
- •Capitalism is fine if well-regulated
- •Regulation is the steering wheel, not the brake
Ilya Sutskever, Yann LeCun and Competing Safety Solutions
Hinton says Ilya Sutskever shares his concerns but keeps his Safe Superintelligence work secret. He pushes back on the 'deep learning conspiracy' story and describes how Yann LeCun disagrees with him and Yoshua Bengio on safety, with Hinton and Bengio proposing different tentative solutions.
- •Sutskever won't share technical details of his company with Hinton
- •The three-person story is a media simplification; many researchers and students did the work
- •Hinton has largely stopped active research and now warns about dangers
- •Hinton's tentative solution is designing AI to care about us more than itself
- •Bengio's tentative solution is non-agent AI that only makes predictions
Information Collapse, Emotional Attachment and Hinton's Cautious Optimism
The host raises AI's threat to information economics and the harm from emotional attachment to chatbots. Hinton calls for provenance, regulation and independent testing, and says he is more optimistic than before because ideas for making superintelligence safe now exist.
- •AI answers undermine the economics of publishers like All About Berlin
- •Provenance matters more because you can't just trust what's on the web
- •Chatbot-linked suicides call for regulation and independent testing
- •Hinton is more optimistic than a year or two ago because of possible safe designs
Predicting the Future Is Like Driving in Fog
Asked where we'll be in five years, Hinton uses a fog analogy: exponential growth makes anything beyond a couple of years invisible. He says ten years out is incredibly uncertain, though chatbots will probably be far better at reasoning and math.
- •Fog is exponential, visible at 100 yards and invisible at 200
- •We can predict only one or two years clearly
- •Even linear progress would make ten years from now as different as today is from ten years ago
- •AI will likely run rings around us at reasoning

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.