Geoffrey Hinton, Nobel Prize in Physics 2024: Official interview

Scientists won the Nobel Prize
12 followers
16 Video Views·May 4, 2025
In between 5 and 20 years from now there's a good chance, a 50% chance we'll get AI smarter than us.
So, I was more or less asleep in a hotel room in California
and I had my phone upside down on the bedside table with the sound turned off
and I just happened to be sleeping on my side when the phone was in my line of sight and it got
bright. I saw this little slit of brightness and it started vibrating and I was in California and
almost everybody I know is on the east coast. So, I wondered who on earth could be calling me.
So, I picked it up and there was this long phone number that I didn't recognize the area code,
the country code, and then some of the Swedish accent asked if I was Geoffrey Hinton.
Then he told me I won the Nobel Prize in Physics and my first reaction was,
well, wait a minute, I don't do physics. This could be a prank. I thought it might well be a prank.
But then what he said sounded very plausible and then other Swedish accents came on
and I was convinced it was real. Well, I was sort of convinced it was real.
But for a couple of days after that I thought I might be in a dream and so I did a bit of
statistical reasoning and the statistical reasoning goes like this. What's the chance
someone who's really a psychologist trying to understand how the brain works
would get the Nobel Prize in Physics? Well, let's say the chance is one in two million.
That's a fairly generous estimate of the chance. What's the chance that if it's my dream
I get the Nobel Prize in Physics? Well, let's say the chance is one in two.
So, now you've got, it's a million times more likely that this is a dream than that it's reality.
So, I thought it might be like those dreams you have when you're younger that you can fly
and you have this dream that you can fly and it's wonderful and then you wake up and it was
only a dream. And then a month later you have a dream that you can fly again and you remember
that you had a dream that you could fly and it wasn't true but this time it's real. I thought
it might be one of those and I, so I, for a couple of days I was sort of waiting to see if I would
wake up. I hadn't woken up yet. There was a lot of pressure to succeed academically so I kind of
knew from a very early age that I had to be a successful academic or a failure. I had a friend
at high school who was always much cleverer than me, who when we were about 16 or 17 he came into
school one day and started talking about memories in the brain and how they might be distributed
over the brain in the same way as holograms. Holograms had just been invented then.
This would have been about 1965 and he got very interested in the idea that came from a
psychologist called Lashley. The memories were distributed across many neurons and I got very
interested in that and ever since then I've been thinking about how the brain might work.
So there's kind of two kinds of risks of AI. There's relatively short-term risks which are
very important and very urgent and they're mainly to do with people misusing AI.
So people are still in charge but they're misusing it. The risks include things like
replacing lots of jobs and causing an increase in the gap between the rich and the poor
because when productivity increases using AI it's not shared equally. Some people lose their
jobs and other people get rich so that's bad for society. That's one kind of risk and we need to
figure out what to do about that although it's not clear what to do. Another kind of risk is
fake videos where they'll corrupt elections. They're already doing that. Another kind of
risk is cyber attacks where bad actors use these big AI models for crafting better attacks. To
begin with it's just for doing better phishing. So last year phishing attacks went up 1200%
probably largely because the large language models can make them much more effective.
Then there's designing things like COVID which you can do much more efficiently using AI
and it's soon going to be relatively easy to design things like that and that means one
crazy person can cause endless chaos. It gets much easier if you release the weights of a large model
so they can take a large model and then fine-tune it and people are now releasing
weights of large models which I think is crazy. There's other short-term risks. There's obviously
things like discrimination and bias. So if you train a model, suppose you're training a model
to decide whether prisoners should get parole. If the historical data is that white prisoners
get parole and black prisoners don't and you train an AI model on historical data,
it'll say the same thing. I'm not as worried about that as other people I think because
with an AI model you can freeze the weights and you can measure the discrimination
which you can't do with people. If you try and measure discrimination in people they realise
they're being measured and you get the Volkswagen effect where they behave differently when they're
being measured. So I think actually for discrimination bias it's going to be much
easier to measure them in AI systems than in people and I think our goal should be not to
make things that don't discriminate and aren't biased. Our aim should be to make things that
discriminate a lot less and are a lot less biased than the systems they replace. So I think that's
one of the... it's a very important problem but it's something where it's fairly clear we can
make progress there. So that's short-term problems. There's also a longer-term problem of
these things taking over. So what we're doing is we're making things more intelligent than
ourselves. Researchers differ on when that will happen but among the leading researchers there's
very little disagreement on the fact that it will happen unless of course we blow ourselves up.
So the question is what's going to happen when we've created beings that are more intelligent
than us and we don't know what's going to happen. We've never been in that situation before.
Anybody who says it's all going to be fine is crazy and anybody who says they're inevitably
going to take over, they're crazy too. We really don't know but because we really don't know it
will make a lot of sense to do a lot of basic research now on whether we can stay in control
of things that we create that are more intelligent than us. There aren't many examples we know of
of more intelligent things being controlled by less intelligent things.
The only good example I know of is a baby controlling a mother. There's not much
difference in intelligence and evolution had to put a lot of work into making that happen.
It's very important that the baby can control the mother but if you look around on the whole
more intelligent things are not controlled by less intelligent things. Now some people think it'll be
fine because we make them and we'll build them in such a way that we can always control them
but these things will be intelligent. They'll be like us and in fact the way they work is
very like the way we work. They're not like computer code. People refer to them sometimes
as computer programs. They're not computer programs at all. You write a computer program
to tell a neural network how to learn a simulated neural network but once it starts learning it
extracts structure from data and the system you've got at the end has extracted its structure from
the data. It's not something that anybody programmed. We don't exactly know how it's
going to work and it'll be like us. So making these programs behave in a reasonable way,
sorry these systems behave in a reasonable way is much like making a child behave in a reasonable
way. The controls you really have are you can reinforce, you can reward it for good behavior,
punish it for bad behavior but the main control you have is demonstrating good behavior,
training it on good behavior. So that's what it observes and that's what it mimics
and it's the same for these systems. So it's very important we train them
on the kind of behavior that we would like to see in them. At present the big chatbots are
trained on all the data they can get which includes things like the diaries of serial killers.
Well if you were raising a child would you get your child to learn to read
on the diaries of serial killers? I think you'd realize that was a bad idea.
Well that's what we don't know. My guess is in between 5 and 20 years from now
there's a good chance, a 50% chance we'll get AI smarter than us. It may be much longer,
it's just possible it's a bit shorter but I think it's quite likely to have happened in 20 years
time. Other researchers think it's shorter or longer. That's my guess. Actually that was my
a year ago so I guess my guess now is between 4 and 19 years. I think it depends what field you're
in and whether you're trying to do something that's very different from the standards in the
field. So for neural networks for a long time they were regarded as ridiculous and it was clear to
many people that they would never work. So to work in a field like that you have to be
confident that you're right even when everybody else says you're wrong
and I had actually a couple of things happened when I was very young that helped. One was my
parents who were both atheists sent me to a Christian school in a Christian private school
from the age of seven and everybody at the school believed in God. The teachers believed
in God and the other kids believed in God and it seemed to me it was just obvious nonsense
and it turns out I was right. So that experience of everybody else around you believing one thing
and it being clear to you they're wrong and then it turning out that as you get older it turns out
other people also don't believe in God. That was a very useful experience. That was one thing that
helped keep me going when everybody said neural networks was nonsense. It wasn't everybody but
it was almost everybody in computer science. Another experience that I haven't talked about
much, when I was about I think probably about nine but I don't know exactly what age,
I had a radio program with my father talking about continental drift. So at that point
there was a lot of controversy about whether the continents moved around.
Nearly all geologists thought it was complete rubbish. So the theory was first introduced I
think around 1920 by a climatologist called Wegener who had lots of evidence that the continents
moved around but he wasn't a geologist and the geologists just thought this was complete rubbish
and they poo-pooed it and they for example refused to allow it to be in textbooks.
They said it would only mislead the students and it was complete nonsense.
So I saw a debate in which there was a theory that was regarded as complete nonsense
by geologists, nearly all geologists, and turned out to be correct. So that was also very helpful
and in fact what happened with neural nets, the most similar thing I know,
is what happened with continental drift. With continental drift there was this idea that
you know that South America fitted nicely into the armpit of Africa
but it wasn't just that, it was that the soil types all down the coast of America
matched the soil types from Norway all down to South Africa. There were fossils that linked up,
there were glacial scrapes on rocks in the tropics and there were coal deposits in the Arctic.
So there's all this evidence that the continents had moved around but the geologists as a field
completely dismissed it. They just couldn't believe that the earth had moved and it was
the same with neural networks. We had this evidence that neural networks have to be able
to learn to do complicated things because we've got a brain but most people in AI said
if you take neural networks and try and learn everything in the neural networks it's hopeless.
The knowledge has to be innate or you have to do it by learning symbolic rules
and they basically refused to allow people to publish neural network stuff in their journals.
So it was a very similar paradigm shift and now there's been a more or less complete paradigm
shift. I don't know if I'm in a good position to give advice but the piece of advice I normally
give is if you have an idea and it seems right to you and it's different from what everybody
else believes, don't give up on it until you've figured out why it's wrong. So most ideas you
have like that, you're wrong, there's something you haven't thought of or there's something you
didn't understand. Just very occasionally you have an idea that's different from what other
people believe and is right. You're never going to discover that unless you keep going with your
beliefs until you discover why they're wrong and you should just ignore what other people say.
I'm very good at ignoring what other people say.
I think the scientists have a much better understanding of what this technology is
than politicians do or the general public do. The scientists still disagree. So there's still
some scientists who say these big chatbots, they don't really understand what they're saying.
Despite all the evidence that they do understand what they're saying, some scientists say it's just
a statistical trick. I would like to have been concerned about this existential threat sooner.
I always thought superintelligence was a long way off and we could worry about it later
and the problem for now was just to make things more intelligent. I wish I'd thought sooner about
what was going to happen. If you go back to Turing in the early 1950s, he talks about
making things smarter than us and he has about one sentence which says, of course,
when they get smarter than us, we're finished. He doesn't say it like that but he implies that.
But most people just don't think about that problem until it gets close and the problem
is it's close now. So I wish I'd thought about that sooner. Okay, so half of the prize money,
half of my share of the prize money, I donated to an organisation in Canada that trains people
who live in indigenous communities on the technology of producing safe drinking water
and that's good because those people will then stay in the communities and they'll have safe
drinking water. It's ridiculous at this time in a rich country like Canada that, for example,
in Ontario, 20% of the indigenous communities do not have safe drinking water. That's just crazy
and I'm sort of sympathetic to this problem because I adopted a child in Peru and I was
there for two months and you can't drink the tap water there. It's poisonous. So your whole life
revolves around how do you get safe water and it's just a huge extra burden on everyday life
and so it's crazy that people in Canada have to suffer that. So I donated half of it to that.
The other half, back in the 1980s, I worked with someone called Terry Sanofsky who was actually a
student of Hopfield's on this theory of Boltzmann machines. We worked on it equally. I wouldn't
have had the theory unless I'd been talking to him and he wouldn't have had it unless he'd
talked to me. So he was a physicist originally and then went into neuroscience and we thought
this must be how the brain works. It was such an elegant learning algorithm that we were convinced
that it had to be how the brain works. So we thought we might get the Nobel Prize in Physiology
or Medicine for discovering how the brain worked and we had an agreement back in the 1980s which
was that if they gave it to one of us and not the other one, we'd split it. So when they gave
me the Nobel Prize, totally unexpectedly, and one of the reasons was Boltzmann machines,
I got in touch with Terry and said where would he like me to send his half?
And he said well look, he didn't feel right about it because it wasn't just for Boltzmann machines,
it was for other subsequent stuff that he wasn't so involved in. So he refused to take the money
and so in the end we compromised that I took that half of my share and we used to set up a
prize in his name for young researchers and it'll be for young researchers with crazy theories of
how the brain works, like we were, and it'll be handed out at the annual conference in our field
and that seemed like a good compromise. So I feel he could easily have been the third person
named in the Nobel Prize. He wasn't. I'm not complaining about that but he could have been.
But this is a way of recognising that he made a huge contribution.
