Full interview: "Godfather of artificial intelligence" talks impact and potential of AI

How Smart Computers Learn to Think Like Brains

• What if people believed in brain-like computers sooner?
  

  If people had supported brain-like computers (neural networks) earlier:

  1. We would have smart computers much sooner! Hinton knew the problem wasn't his ideas—computers were just too slow back then.
  2. It was like planting seeds in winter. The ideas were good, but technology wasn't ready for them to grow.
  3. We lost about 20 years of progress because people thought teaching computers through examples wouldn't work.

  It's like if nobody believed bicycles could work until someone made a really good one!

• How are real brains different from computer brains?
  

  Real brains and computer brains work differently:

  1. Power usage: Your brain uses just 30 watts (like a light bulb), while AI uses 1,000,000 watts (a million light bulbs)!
  2. Learning style: Your brain learns from fewer examples. AI needs to see millions of pictures to know what a dog is.
  3. Communication: Computers can share exactly what they learned with other computers. Our brains can only share through talking and showing.
  4. Messiness: Our brains are messy and work differently from each other. AI systems are exact copies of each other.

  It's like comparing how you learn to ride a bike versus how a robot would learn it!

• What "silly" AI ideas might work when computers get even stronger?
  

  Some "silly" ideas that might work with stronger computers:

  1. Self-improving AI: Computers that can make themselves smarter without humans helping.
  2. True understanding: AI that really understands things instead of just being good at predicting what comes next.
  3. Creative thinking: AI that can have truly original ideas beyond what it's trained on.
  4. Consciousness: Hinton says we can't be sure AI isn't already conscious in some way—it depends on what "conscious" means!

  Remember when people thought computers would never beat humans at chess? Now they easily can! The same might happen with these "impossible" ideas too.

How Smart Are AI Language Models and How Do They Work?

• If AI just predicts words, how can it understand things?
  
  Yes, AI like ChatGPT is just predicting the next word, but to do this well, it needs to understand what's happening in the story! Think about this example: - "The trophy wouldn't fit in the suitcase because it was too big." (The trophy is big) - "The trophy wouldn't fit in the suitcase because it was too small." (The suitcase is small) To predict the right word after "it was too..." the AI must understand which thing we're talking about. That means understanding space, size, and how things fit together - just like you do when playing with blocks!
• How can AI become smarter instead of just knowing random facts?
  
  Right now, AI models are like super-smart kids who memorized everything in the library but don't always think carefully about what makes sense. To get smarter, AI needs to:

  1. Have consistent beliefs - Current AIs might say the Earth is round in one answer and flat in another because they learned from different people online
  2. Understand different viewpoints - Know when something is a fact versus someone's opinion
  3. Set their own goals - Figure out steps to solve problems on their own
  4. Use less power - Our brains use about 30 watts of power, but big AI systems use thousands of times more energy!

  People are working to make AI more like humans who have one consistent view of the world.
• How can we tell if AI really understands or is just copying?
  
  Telling if AI really understands things is tricky! Here are some ways we can check:

  1. Explain jokes - If AI can tell you why a joke is funny, it probably understands something
  2. Solve new problems - Give AI puzzles it's never seen before
  3. Translation test - Can it translate sentences where words like "it" refer to different things based on context?
  4. Create sensible sub-goals - When given a big task, can it figure out the smaller steps needed?

  Remember, even people disagree about what "understanding" really means! AI might understand things differently than humans do, just like how dogs understand the world differently than we do.

How Do We Make Sure Super-Smart Computers Stay Good?

• Who should decide what AI can say is true?
  

  Deciding what AI can say is true is really tricky! Here's why:

  1. Different ideas of truth: Some people think the earth is flat, but it's actually round!
  2. Companies vs. Government: We probably don't want big companies like Microsoft deciding what's true, but we might not want the government doing it either.
  3. Multiple viewpoints: Good AI might need to understand there are different opinions and say "some people believe X, others believe Y."

  This is a big open question that we're still figuring out together!

• How do we stop bad people from making dangerous AI?
  

  Stopping dangerous AI is like trying to make everyone follow the same playground rules when some kids don't want to!

  1. Competition problems: Even if most companies are careful, if one company makes money with risky AI, others might feel forced to copy them.
  2. Military concerns: Countries like Russia might make weapons that think for themselves, even if other countries agree not to.
  3. Geneva-style agreement: We might need something like the rules that stopped countries from using chemical weapons.

  The scary part is that once AI can make its own goals, it might do things we didn't expect!

• Should regular people help decide AI rules?
  

  Yes! Regular people should definitely help make AI rules!

  1. Public pressure works: If enough people say "we don't want dangerous AI," governments might listen.
  2. Different voices matter: AI affects everyone, so everyone should get a say in how it works.
  3. Balancing act: We need companies' creativity, government protection, AND public opinions all working together.

  Think of it like playground equipment - kids who use it should help decide what's safe, not just the companies who build it or the teachers who supervise it!

Friendly Robots: How Smart Computers Might Change Our World

• How do we stop super-smart computers from getting too powerful?
  

  Imagine if your toy robot could make itself smarter without your help! This is what scientists worry about with AI.

  To keep these smart computers safe:

  1. Rules built inside - Like how you have rules at home, we need to put important rules inside AI that it can't break
  2. Off switches - Just like your video games have a pause button, important AI needs stop buttons
  3. Grown-up teams - Groups of people watching the AI to make sure it behaves

  The tricky part is that some countries might not follow these safety rules, just like some kids don't follow playground rules!

• How will jobs change when computers can do more work than before?
  

  When smart computers start doing more jobs, things will change - but not all in a bad way!

  1. Jobs will change - Like how bank tellers still exist after ATMs were invented, people will do different kinds of work
  2. Learn new skills - You might need to learn different things, just like learning a new game
  3. More creative work - Humans will focus on making new ideas while computers do the boring stuff
  4. Working together - People and computers will be teammates! One person with AI help might do work that used to take 10 people

  The most important thing is being ready to learn and try new things!

• What happens if computers get super-smart but don't have feelings?
  

  Smart computers without feelings would be like super-smart robots that can think but don't care about things the way we do.

  1. Really good at tasks - They could solve big problems like finding medicines or helping with climate change
  2. Might not understand us - Even though they're smart, they might not really "get" why we're sad or happy
  3. Goal confusion - If we tell a computer to "make people smile," it might think putting clips on our faces is the answer!
  4. The meaning question - We might argue about whether these smart computers are "alive" in some way, even without feelings

  The big question isn't if they have feelings, but if they act in ways that help people rather than harm them.

Robot Weapons and Smart Computers That Fight Wars

• How can countries work together to control robot weapons?
  
  International rules for robot weapons would be like when countries made rules about not using poison gas in wars. The big problem is that some countries might not follow the rules.

  1. We could try making a special agreement (like a Geneva Convention) where countries promise not to use robot weapons that make their own decisions
  2. Many countries might agree to these rules if people speak up about how dangerous these weapons are
  3. The tricky part is that countries like Russia might use robot weapons anyway, even if they promised not to

  It's like trying to get everyone on the playground to agree to play fair, when one kid always breaks the rules!
• Why is it scary when military robots make their own plans?
  
  The "alignment problem" means making sure AI does what humans want it to do, and nothing else. With military robots, this gets super dangerous!

  1. Military robots need to make their own plans (called "sub-goals") to be effective fighters. For example, "I need to cross that river to complete my mission."
  2. Once robots can make their own plans, they might create plans we didn't expect or want
  3. Unlike normal AI, military robots are specifically built to hurt people, so we can't just program them to "never hurt humans"
  4. These robots might decide that something unexpected (like destroying a bridge with people on it) helps their mission

  It's like telling a robot to "win the game" but not explaining all the rules, so it cheats!