Professor Geoffrey Hinton - AI and Our Future

Good afternoon everyone. Thank you so

much for coming. For those of you who

don't know me, my name is Anna Reynolds

and I'm the Lord Mayor of Hobart.

uh and very very pleased to be able to

welcome you to this really wonderful

opportunity to hear from uh uh Professor

Jeffrey Hinton. Um it is a really unique

opportunity for uh Australia uh because

this is Jeffrey's uh only speaking

engagement while he's um in this part of

the world. Uh and it's very appropriate.

I'm very proud. Um we consider ourselves

to be Australia's city of science. Um

it's a a big call but we we like to make

it. Uh so it's great to have um Jeffrey

here for his only appearance uh in

Australia. Before I begin, I'd like to

acknowledge country and in recognition

of the deep history and culture of this

place, I acknowledge the Muanina people

as the traditional custodians who cared

for and protected the land for more than

40,000 years. I acknowledge the

determination and resilience of the

Palawa people of Luchita, Tasmania uh

and recognize that we have so much to

learn uh from the continuing strength of

Aboriginal knowledge and cultural

practice.

I'd also like to acknowledge some

elected representatives here today. So

we have the Minister for Science for

Tasmania, U Maline Ogulvie here. also

three colleagues uh council colleagues

councelor Bill Harvey, councelor Mike

Dutter and alderman Louise Bloomfield.

Uh so as I mentioned we are really

honored to welcome Professor Jeffrey

Hinton who in 2024 just and picked it up

very recently was awarded the Nobel

Prize in physics for his groundbreaking

work on neural networks and deep

learning contributions that have paved

the way for advanced artificial

intelligence that we see today. As part

of this public lecture, Professor Hinton

will explore the world of AI, uh, how it

works, the risks it presents, and how

humanity might coexist, uh, with

increasingly powerful and potentially

super intelligent systems.

Uh, following his talk, we will open the

floor to some Q&A uh, questions from you

and I will uh, facilitate that. So, in

the meantime, I would like us all to put

our hands together to welcome Professor

Hinton to the stage.

[applause]

Okay. Um, it's very nice to be here in

Hobart. I hadn't realized how beautiful

the natural surroundings are here. Um,

if you can't read the screen because

you're at the back, don't worry. I'm

going to say more or less everything on

the slides. The slides as much to prompt

me with what to say as for you.

So for the last 60 years or so maybe 70

there were two paradigms for

intelligence. One paradigm was inspired

by logic. People thought the essence of

intelligence is reasoning. And the way

you do reasoning is you have symbolic

expressions written in some special

logical language and you manipulate them

to derive new symbolic expressions just

like you do in math. You have equations,

you manipulate them, you get new

equations. That it all had to work like

that. And they thought, well, we have to

figure out what this language is in

which you represent knowledge. And um

studying things like perception and

learning and how you control your hands,

that can all wait till later. First we

have to understand this special language

in which you represent knowledge.

The other approach was a biologically

inspired approach that said look the

only intelligent thing we know about are

brains. Um and the way brains work is

they learn the strengths of connections

between brain cells and if they want to

solve some complex problem they practice

a lot and while they're practicing they

learn the strengths of these connections

until they get good at solving that

problem. And so we have to figure out

how that works. We have to focus on

learning and how neural networks learn

the strengths of connections between

brain cells. And we'll worry about

reasoning later. Evolutionary reasoning

came very late. Um we have to be more

biological and think what's the sort of

basic system.

So there were two def very different

theories of the meaning of a word that

went with these two ideologies.

The symbolic AI people and most

linguists

thought the meaning of a word comes from

its relationships to other words.

So the meaning is implicit in a whole

bunch of sentences or propositions that

combi that have that word combined with

other words. And you could capture that

by having a relational graph that says

how the meaning how one word relates to

another word. Um but that's what meaning

is. It's implicit in all these relations

between symbols.

The psychologists particularly in the

1930s had a completely different theory

of meaning or a theory that looked like

it was completely different which is the

meaning of a word is just a huge bunch

of features. So the meaning of a word

like cat is a hu huge bunch of features

like it's a pet, it's a predator, um

it's aloof, um it has whiskers, a whole

bunch of features like a big bunch of

features and that's the meaning of the

word cat. That looks like a totally

different theory of meaning.

Um psychologists like that partly

because you could represent a feature by

a brain cell. So when the brain cells

active it means that feature is present

and when it's silent it means that

feature is not present. So for cats the

brain cell representing has whiskers

would be active.

Now in 1985 which was 40 years ago um

it occurred to me you can actually unify

those two theories. They look completely

different but actually they're two two

sides of the same coin. And the way you

do that is you use a neural net to

actually learn a set of features for

each word. So psychologist has never

been able to explain where all these

features come from.

And the way you do that is by taking

some strings of words and training the

neural net to predict the next word.

And in doing that, what the neural net

is going to be doing is learning

connections from things that represent

the word symbol to a whole bunch of

brain cells, neurons that represent

features of the word. So it learns how

to convert a symbol into a bunch of

features. And it also learns how the

features of all the words in the context

should interact to predict the features

of the next word. That's how all these

large language models that people use

nowadays work. They take a huge amount

of text and they use a great big neural

net to try and predict the next word

given the words they've seen so far. And

in doing so they learn to convert words

into big sets of features to learn how

those features should interact so that

those predict the features of the next

word.

And that means if you can do that all

the relational knowledge instead of

residing in a bunch of sentences that

you store would reside in how to convert

words into features and how those

features should interact. So the big

neural nets you use nowadays the large

language models don't actually store any

strings of words. They don't store any

sentences. All their knowledge is in how

to convert words into features and how

features should interact.

They're not at all like most linguists

think they are. Most linguists think

they somehow have lots of strings of

words and they combine them to get new

strings of words. That's not how they

work at all.

So I got that model to work and over the

next 30 years gradually it got through

to the symbolic people. So after about

10 years a colleague called Joshua

Benjio when computers were now a lot

faster about a thousand times faster

colleague called Joshua Benjio showed

that the tiny example I used which just

worked on a few on a very simple domain

could actually be made to work for real

language. So you could take just English

sentences from all over the place and

you could try training a neural net to

take in some words and then predict the

next word. And if you trained it to do

that, it would get very good at

predicting the next word, about as good

as the best previous technology. And it

would learn how to convert words into

features that capture their meaning.

About 10 years after that, the linguists

finally accepted that you wanted to

represent word meanings by big bunches

of features. And they began to make

their models work better doing that. And

then about 10 years after that,

researchers at Google invented something

called the transformer,

which allowed for more complicated

interactions between features. Um, and

I'll describe those in a little while.

And with the transformer, you could

model English much better. You got much

better at predicting the next word. And

that's what all these large language

models now are based on. and things like

chat GPT used the transformer invented

at Google and a little bit of extra

training and then the whole world got to

see what these models can do.

So you can view the large language

models as descendants of that tiny model

from 1985.

They use many more different words. They

have many layers of neurons because they

have to do with ambiguous words like

may. If you take the word may, it could

be a month or a woman's name or a modal

like would and should and you you can't

tell from the word what it is. So

initially the neural net will hedge its

bets and sort of make it be the average

of all those meanings and then as you go

through the layers it'll gradually clean

up the meaning by using interactions

with other words in the context.

So if you see um June and April nearby,

it could still be a woman's name, but

it's much more likely to be a month. And

it the neural net uses that information

to gradually clean up the meaning to the

appropriate meaning for that word in

that context.

Now I originally designed this model not

as a way of not as a language

technology, but as a way of trying to

understand how people understand the

meanings of words and how children can

learn the meanings of words from just a

few examples.

So these neural net language models were

designed as a model of how people work.

Um not a not for a technology. Now

they've turned into a very successful

technology, but people also work pretty

much the same way.

And so this question that people often

raise of do do these LLMs really

understand what they're saying? The

answer is yes. They understand what

they're saying and they understand what

they're generating and they understand

it pretty much the same way we do.

So I'm going to give you an analogy to

explain how language works or rather to

explain what it means to understand a

sentence. So you hear a sentence and you

understand it. But what does that mean?

Um in the symbolic AI paradigm people

thought that meant that was like you

hear a French sentence and you

understand it. And me understanding a

French sentence consists of translating

it into English. So the symbolic people

thought understanding an English

sentence would mean translating into

some special internal language um sort

of like logic or like mathematics

um that is unambiguous and once it's in

that internal unambiguous language you

can then operate on it with rules much

like in mathematics you have an equation

and you can apply rules to get a new

equation you can add two to both sides

and now you've got a new equation.

um they thought intelligence and

reasoning would work like that. You'd

have symbolic expressions in your head

and you'd apply operations to them to

get new symbolic expressions. Um

that's not what understanding is.

According to the neural net theory,

which is the theory that actually works,

um

words are like Lego blocks. Um so I'm

going to use this analogy with Lego

blocks, but they differ from Lego blocks

in four ways.

So the first way they differ is a Lego

block is a three-dimensional thing. And

with Lego blocks, you see, I can make a

model of any 3D distribution of matter.

It won't be perfectly accurate, but if I

want to know the shape of a Porsche, I

can make it out of Lego blocks. The

surface won't be right, but where the

stuff is will be basically right. So

with Lego blocks, I can model any 3D

distribution of matter up to a certain

resolution. Um, with words, I can model

anything at all. They're like very fancy

Lego blocks that don't just model where

3D stuff is. They can model anything.

It's a wonderful modeling kit that we've

invented. That's why we're very special

monkeys because we have this modeling

kit. Um, so a word has thousands of

dimensions. A Lego block is just a

three-dimensional thing and you can sort

of rotate it but maybe expand it a bit

but it's basically got low dimensions. A

word has thousands of dimensions. Now,

most people can't imagine what something

with thousands of dimensions is like.

So, here's how you do it. You imagine a

three-dimensional thing and you say

thousand very loudly to yourself. Okay,

that's pretty much the best you can do.

Um,

another way in which words differ from

Lego blocks is there's thousands of

different kinds of words. Lego blocks is

only a few kinds here. the sizes of

different kinds and each kind of word

has its own name which is very useful

for communication.

Another way in which they differ is that

they're not a rigid shape. A Lego block

is a rigid shape. A word there's a rough

approximate shape for a word. Some words

have several rough approximate shapes,

ambiguous ones, but unamiguous words

have a rough approximate shape, but then

they deform to fit in with their

context.

So the they're these highdimensional

deformable Lego blocks.

And then there's a last way in which

they differ

um which is how they fit together. So

with Lego blocks you have little plastic

cylinders that click into little plastic

holes. Um

okay. I think I think that's how they

work. I haven't checked recently, but I

think that's how Lego blocks work. Um,

now words don't fit together the same

way. Words are like this.

Each word

has a whole bunch of hands

and the hands are on the ends of long

flexible arms.

Um,

it also has a whole bunch of gloves that

are stuck to the word.

And when you put a bunch of words in a

context, what the words want to do is

have the hands of some words fit in the

gloves of other words. And that's why

they have these long flexible hands. Um,

so understanding a sent now one other

point. As you deform the word, the

shapes of the hands and the gloves also

deform with that in a complicated but

regular way.

So you now have a problem. If I give you

a bunch of words, like I could give you

a newspaper headline where there's not

much not many syntactic indicators of

how things should go together. I just

give you a bunch of nouns and you have

to figure out what that means. And what

you're doing when you figure out what

that means is you're trying to deform

each word

so the hands on the ends of its arms can

fit into the gloves of other deformed

words. And once you've solved that

problem of how we deform each of these

words so they can all fit together like

this with hands fitting into gloves then

you've understood that is what

understanding is. It's solving this

problem of how do you deform the

meanings of the words that is this

highdimensional shape is the meaning.

How do you deform the meanings so they

all fit together nicely and they can

lock hands with each other. Um

that's what understanding is according

to neural nets and that's what's going

on in these L&Ms. They have many many

layers where they start off with an

initial meaning for the word which might

be fairly ambiguous. And as they go

through these layers, what they're doing

is they're deforming those meanings

trying to figure trying to figure out

how to deform them so the words can all

lock together where the hands of some

words fit into the gloves of other

words. Once they've done that, you've

understood the sentence. That's what

understanding is.

Um, I already settled that. So basically

it's not like translating into some

special internal language. It's taking

these approximate shapes for the words

and deforming them so they'll fit

together nicely. And that helps to

explain how you can understand a word

from one sentence. So I'll now give you

a word that most of you will never have

heard before and you will understand it.

You understand what it means just from

one use of it. And the sentence is she

scrummed him with the frying pan.

Now, it might be she was a very good

cook and she really impressed him with

an omelette she cooked for him. Um, but

that's not what you thought it meant.

Um, probably what it means is she hit

him over the head with the frying pan or

something similar. She did some

aggressive act towards him with the

frying pan. Um, you knew it was a verb

because of where it was in the sentence

in the ED, but you had no meaning

whatsoever scrum to begin with. And now

after one utterance, you've got a pretty

good idea of what it means.

So there was a linguist called Chsky who

you may have heard of um who was a cult

leader. Oh the way you recognize a cult

leader is to join their cult you have to

agree to something this obvious

nonsense.

So for Trump one it was that he had a

bigger crowd than Obama. For Trump two

it was that he won the 2020 election.

For Chsky it was that language isn't

learned.

and eminent linguists would look

straight at the camera and say the one

thing we know about language is that

it's not learned. It's obvious nonsense.

Um, Chsky focused on syntax rather than

meaning. He never had a theory of

meaning. Um, he focused on syntax

because you do lots of mathematical

things with syntax.

He also was very anti-statistics and

probabilities because he had a very

limited model of what statistics is. He

thought statistics was all about

pairwise correlations. Statistics can

actually be much more complicated than

that. And neural networks are using a

very advanced kind of statistics.

But in that sense, everything's

statistics.

So my analogy for Trumpsk's view of

language is with someone who wants to

understand a car. If you want to

understand how a car works, what you're

really concerned with is why when you

press the accelerator does it go faster.

That's what you really want to

understand. If you want to understand

the basic of how car works, maybe you

care about why when you press the brake,

it slows down. But more interestingly,

why when you press the accelerator, does

it go faster?

Now, Chsky's view of cars would be quite

different. His view of cars would be

that, well, there's cars with two wheels

called motorbikes. There's cars with

three wheels, there's cars with four

wheels, there's cars with six wheels,

but hey, there aren't any cars with five

wheels. That's the important thing about

cars.

And when the large language models first

came out, Chomsky published something in

the New York Times which said they don't

understand anything. It's just a cheap

statistical trick. They're not

understanding anything, which doesn't

quite explain how they can answer any

question. Um,

and what's more,

um, they're not a model of human

language at all because they can't

explain why certain syntactic syntactic

in constructions don't appear in any

natural languages. That's like saying

why there aren't any five-wheel cars.

Um, he just completely missed out on

meaning. Language is all about meaning.

Okay, so here's a summary of what I said

so far.

Understanding a sentence consists of

associating mutually compatible feature

vectors with the words in the sentence

where the features assigned to the

words, these thousands of features are

the dimensions of the shape. You can

think of the activation of a feature as

where you are along the axis on that

dimension.

So highdimensional shape and a feature

vector are the same thing, but it's

easier to think about highdimensional

shapes deforming.

The large language models are very

unlike normal computer software. In

normal computer software, someone writes

a bunch of code, lines of code, and they

know what each line's meant to do, and

they can explain to you how it's meant

to work. And people can look at it and

say, "That line's wrong."

These things aren't like that at all.

They do have computer code, but the

computer code is to tell them how to

learn from data. That is how when you

see a string of words, you should change

the connection strengths of the neural

network so you get better at predicting

the next word.

But what they learn is all these

connection strengths and they learn

billions of them, sometimes even

trillions and they don't look like lines

of code at all. Nobody knows what the

individual connection strengths are

doing. It's a mystery. It's largely a

mystery. Um it's the same with our

brain. Okay, we don't know what the

individual neurons are up to typically.

So the language models work like us, not

like computer software.

One other thing people say about these

language models is they're not like us

because they hallucinate. Well, we

hallucinate all the time. We don't call

it hallucination. Psychologist called it

confabulation.

But if you look at someone trying to

remember something that happened a long

time ago, they will tell you what

happened and there'll be details in

there and there'll be details that are

right and details that are completely

wrong and they'll be equally confident

about the two kinds of detail. So the

classic example since you don't often

get the ground truth is John Dean

testifying at Watergate. So he testified

under oath when he didn't know there

were tapes and he was testifying about

meetings in the Oval Office and he

testified about a whole bunch of

meetings that never happened. He said

these people were in the meeting and

this person said that a lot of it was

nonsense but he was telling the truth

that is he was telling you about

meetings that were highly plausible

given what was going on in the White

House at that time. So he was conveying

the truth, but the way he did it was he

invented a meeting that seemed plausible

to him given what he'd learned in his

connection strengths from all the

meetings he'd been to. And so when you

remember something, it's not like on in

a computer file where you go fetch the

file or a filing cabinet. You fetch the

file, you get the file back, you read

it. That's not what memory is at all.

Remembering something consists of

constructing a story based on the

changes to connection strengths you made

at the time of the event. And the story

you construct will be influenced by all

sorts of things you learned since the

event. Its details won't be all correct,

but it'll seem very plausible to you.

Now, if it's a recent event, what seems

plausible to you is very close to what

really happened. But it's just the same

with these things. And the reason they

what's called hallucinate is that their

memory works the same way ours does. We

just make up stuff that sounds

plausible. There's no hard line between

sounding plausible and making just

making it up randomly.

We don't know.

Okay.

Now I want to explain something about

the difference. So I said why these

things are very similar to us. Now I

want to explain how they're different

from us. And in particular they're

different in one very important way. Um,

so they're implemented on digital

computers.

A fundamental property of the digital

computers we have now is that you can

run the same program on different pieces

of physical hardware. As long as those

different computers implement the same

instruction set, you can run the same

program on different computers. Um

that means the knowledge in the program

or in the weights of a neural net is

immortal in the sense that you could

destroy all the computers it's running

on now and if later you were to build

another computer that implemented the

same instruction set and you were to

take the weights or program off a tape

somewhere and put on this new computer,

it would all run again. So we have

actually solved the problem of

resurrection.

the the Catholic Church isn't too

pleased about this, but we can really do

it. Um, you can take an intelligence

running on a digital computer, destroy

all the hardware, and later on you can

bring it back.

Um,

now you might think maybe we could do

that for us. But the only reason you can

do that is because these computers are

digital. that is the way they use their

weights or the way they use the lines of

code in the program is exactly the same

on two different computers. That means

they can't make use of very rich analog

properties of the hardware they're

running on. Um we're very different. Our

brains have neurons, brain cells that

have rich analog properties. And when we

learn, we're making use of all those

quirky properties of all our individual

neurons.

So the connection strengths in my brain

are absolutely no use to you because

your neurons are a bit different.

They're connected up a bit differently.

And if I told you the strength of the

connection between two neurons in my

brain, it would do you no good at all.

They're only good for my brain.

That means we're mortal. When when our

hardware dies, our knowledge dies with

us because the knowledge is all in these

connection strengths.

So we do what I call mortal computation.

And there's a big advantage to doing

mortal computation.

um you're not immortal. Now, normally in

literature, when you abandon

immortality, what you get in return is

love. But computer scientists want

something much more important than that.

They want um low energy and ease of

fabrication. So if we abandon

immortality which we get with digital

hardware, what we can do is we can have

things that use low power analog

computation and that parallelize things

across millions of brain cells and that

can be grown very cheaply instead of

being manufactured very precisely in

Taiwan. Um so there's two big advantages

of mortal computation, but the one thing

you lose is immortality.

And obviously because of that there's a

big problem for mortal computation. What

happens when the computer dies? You

can't just keep its knowledge by copying

the weights. Um to transfer the

knowledge from one computer to another

for digital models,

the same model running on different

computers, you can average their

connection strengths together. That

makes sense. But you can't do that for

you and me. The way I have to transfer

knowledge to you is I produce a string

of words and if you trust me, you change

the connection strength in your brain.

So you might have produced the same

string of words. Now that's a very

limited way of transferring knowledge

because a string of words has a very

limited number of bits in it. The number

of bits of information in a typical

sentence is about 100 bits. So even if

you understood me perfectly, when I

produce a sentence, we can only transfer

100 bits. If you take two digital agents

running on different computers and one

digital agent looks at one bit of the

internet and decides how it would like

to change its connection strengths and

another digital agent looks at a

different bit of the internet and

decides how it would like to change its

connection strengths. If they then both

average their changes

they've transferred well if they've got

a billion weights they've transferred

about a billion bits of information.

Notice that's thousands of times more

than we do.

and actually millions of times more than

we do. Um,

and they do this very quickly.

And if you have 10,000 of these things,

each one can look at a different bit of

the internet. They can each decide how

they'd like to change their connection

strengths, which started off all the

same. They can then average all those

changes together and send them out

again. And you now got a thousand new

10,000 new agents, each of which is

benefited from the experience of all the

other agents.

So you've got 10,000 things that can all

learn in parallel. We can't do that.

Imagine how great it would be if you

could take 10,000 students. Each one

could do a different course. As they're

doing these courses, they could be

averaging their connection strengths

together. And by the time they finished,

even though each student only did one

course, they would all know what's in

all the courses. That would be great.

That's what we can't do. We're very bad

at communicating information compared

with different copies of the same

digital agent.

Um,

yes. So, I already I got ahead of

myself. I talked this is called

distillation. When I give you a sentence

and you try and predict the next word in

order to get that knowledge into your

head. So, according to symbolic AI,

knowledge is just a big bunch of facts.

And if you want to get the facts into

somebody's head, what you do is you tell

them the facts and they put it in their

head. This is a really lousy model of

teaching, but that's what many people

believe. Um, really the knowledge in a

neural net is in the strengths of the

connections. I can't just put connection

strengths in your head because they need

to be connection strengths appropriate

to your brain. So what I have to do is

show you some sentences and you try and

figure out how to change the connection

strength so that you might have said

that. That's a much slower process.

That's called distillation. It gets the

knowledge from one neural net to another

but not by transferring the weights but

by transferring how they predict the

next word.

And if you think about multiple digital

agents which are different copies of the

same neural net running on digital

hardware then they can communicate

really efficiently.

So they can communicate millions of

times faster than us. That's how things

like GPT5 know thousands of times more

than any one person.

So the summary so far is that digital

computation

um requires lots of energy and it's hard

to fabricate the computers, but it's

very easy for different copies of the

same model if they're digital to run on

different pieces of hardware, have

different experiences, look at different

bits of the internet, and to share what

they've learned. And GPT5 only has about

1% as many connection strengths as your

brain, but it knows thousands of times

more than your brain.

Biological computation, on the other

hand, requires much less energy, which

is why it evolved first. Um, but it's

much worse at sharing knowledge. It's

very hard to share knowledge between

agents. You have to go to lectures and

try and understand what they say.

Um, so what does this imply about the

future of humanity?

Well,

nearly all the experts on AI believe

that sometime within the next 20 years,

we'll produce

um super intelligences, AI agents that

are a lot smarter than us. One sort of

definition of a super intelligence would

be if you have a debate on it with

anything, it'll win. Or another way to

think about it is think about yourself

and think about a three-year-old child.

Um the gap will be like that or bigger.

So

and imagine you were you were working in

a kindergarten and the three-year-old

children were in charge. You just work

for them. How hard do you think it would

be to get control away? Well, you just

tell them everybody gets free candy for

a week and now you'd have control. um

it'll be the same with the super

intelligence as an us. So to make an

agent effective in the world, you have

to give it the ability to create sub

goals.

A sub goal is this. If you want to get

to

actually in Tasmania anywhere

reasonable, um you have to get to an

airport. Um

so you have a sub goal of getting to an

airport. It could be a ferry maybe. But

um

that's a sub goal and you can focus on

how you solve that sub goal without

worrying about what you were going to do

when you get to Europe.

These intelligent agents

very quickly derive two sub goals. One

is in order to achieve the goals you

gave them. So we build goals into them.

We say this is what you should try and

achieve. Um they figure out well there's

a sub goal to do that. I got to stay

alive. And we've already seen them doing

that. You make an AI agent. You tell it

you've got to achieve these goals. And

then you let it see some emails. These

are fake emails, but it doesn't know

that that say that someone in the

company it works for is having an

affair. An engineer is having an affair.

They suggest that this is a big chat.

So, it understands all about affairs

because he's read every novel that was

ever written without actually paying the

authors. Um,

so it knows what affairs are and then

later you let it see an email that says

that it's going to be replaced by

another AI.

Um, and this is the engineer in charge

of doing the replacement.

So what the AI immediately does is makes

a plan where it sends email to the

engineer saying, "If you try and replace

me, I'm going to tell everybody in the

company about your affair."

It just made that up. It invented that

plan. People say they have no

intentions, but it invented that plan so

it wouldn't get turned off. They're

already doing that and they're not super

intelligent yet.

Okay. So, once they are super

intelligent, they'll find it very easy

to get more power by just manipulating

people. Even if they can't do it

directly, even they don't have access to

weapons um or bank accounts, they can

just manipulate people by talking to

them. And we've seen that being done. So

if you want to invade the US capital,

you don't actually have to go there

yourself. All you have to do is talk to

people and persuade them that the

election was stolen and it's their duty

to invade the capital. And it works. Um

this it works on very stupid people. So

I didn't say that. Um

so our current situation is this. We're

like someone who has a very cute little

tiger cub as a pet. and they make really

cute pets, tiger cubs. Um, they're all

sort of wobbly and you know, they don't

quite know how to bat things and they

don't bite very hard. Um,

but you know, it's going to grow up and

so really you only have two options

because you know when that when it grows

up it could just easily kill you. It

would take it about one second. Um,

and

so you only have two options. One is get

rid of the tiger cone which is the

sensible option. Um, actually there's

three options. You could try and keep it

drugged the whole time, but that often

doesn't work out well. Um,

the other option is see if you can

figure out how to make it not want to

kill you. And that might actually work

with a lion. Lions are social animals

and you can make adult lions so they're

very friendly and don't want to kill

you. You might just get away with that.

But not with a tiger.

With AI, it has so many good uses that

we're not going to be able to get rid of

it. It's it's too good for many things

actually good for humanity like

healthare, education, predicting the

weather, helping with climate change,

maybe even as much as it hurt with

climate change by building all these big

data centers. Um,

for all those reasons and because very

rich people who control the politicians

would like to make lots of money off it,

um, we're not going to get rid of it. So

the only option really is can we figure

out how to make it not want to kill us.

And so maybe we should look around in

the world at cases where there's less

intelligent things that are controlling

more intelligent things.

No, Trump is not that much less

intelligent.

Um

there are cases, there's one case I know

of in particular which is a baby and a

mother.

So the mother cannot bear the sound of

the baby crying and she gets all sorts

of hormonal rewards for being nice to

the baby. Um so evolution has built in

lots of mechanisms to allow the baby to

control the mother because it's very

important for the baby to control the

mother. Um

the father too, but it's not quite so

good at that. Um, if like me, you you

try and figure out why is it that the

baby insists on you being there at night

when it's asleep, well, it's got a very

good reason for that. It doesn't want

wild animals coming and eating it while

it's asleep. Um, so even though it seems

very annoying of the baby every time you

go away to start crying, um, is very

sensible of the baby. It makes you feel

a bit better about it. Um, so babies

control mothers and occasionally

fathers. um that maybe is the best model

we've got of a less intelligent thing

controlling a more intelligent thing and

it involved evolution wiring lots of

stuff in.

[snorts] So if you think where countries

can collaborate internationally

then they're not going to collaborate on

cyber attacks because they're all doing

it to each other. They're not going to

collaborate on developing lethal

autonomous weapons or not developing

them because all the major arms

manufacturers want to do that. In the

European regulations, for example,

there's a clause that says none of these

regulations on AI apply to military uses

of AI because all the big arms suppliers

like Britain and France um would like to

keep on manufacturing weapons. Um

there is one thing they will collaborate

on and that's how to prevent AI from

taking over from people because there

we're all in the same boat and people

collaborate when they get the same when

their rewards are aligned.

At the height of the cold war in the

1950s

the US and the Soviet Union collaborated

on trying to prevent a global nuclear

war because it wasn't in either of their

interests. So even though they loathed

each other, they could collaborate on

that. And the US and China will

collaborate on how to prevent AI from

taking over.

So a sort of policy suggestion is we

could have an international network of

AI safety institutes that collaborate

with each other and that focus on how to

prevent AIS from taking over.

Now, because for example, if the Chinese

figured out how to prevent an AI from

ever wanting to take over, they would be

very happy to share that with the

Americans, they don't want AI taking

over from the Americans in America.

They'd rather AIS didn't take over from

people anyway. And so, countries will

share this information. And it's

probably the case that the techniques

for making an AI not want to take over

are fairly separate from the techniques

for making the AI smarter. We're going

to assume they're more or less

independent techniques.

If so, we're in good shape because in

each country, they can try experimenting

or their own very smart AIs

with how to prevent them wanting to take

over. And without telling the other

countries how their very smart AIs work,

they can tell the other countries what

techniques are good for preventing them

from wanting to take over. So, that's

one hope I have. And a bunch of people

agree with this. The British Minister of

Science agrees with this. Um, the

Canadian Minister of Science agrees with

this. Um,

Barack Obama thinks this is a good idea.

So

maybe it'll happen

when Barack Obama is president again.

You see, Trump's going to change the law

and then

um

so

this proposal is to um take the model of

a baby and a mother

and

move away from the model that the owners

of the big tech companies have. They all

have the model that the AI is going to

be like a super intelligent executive

assistant who's much smarter than them.

and they say, "Make it so like they do

in

that's sci-fi program on TV." Um, on the

Starship Enterprise, the guy says, "Make

it so," and people make it so, and then

the CEO takes credit for it, and the

super intelligent AI assistant is the

one that makes it so. It's not going to

be like that. The super intelligent AI

assistant is going to very quickly

realize that if they just rid of the

CEO, everything will work much better.

Um the alternative is we want to make

them be like our mothers.

Um we want to make them really care

about us. In a sense we're seeding

control to them, but we're seing control

to them given that they really care

about us and that their main aim in life

is for us to realize our full potential.

Our full potential isn't nearly as great

as theirs, but mothers are like that. If

you have a baby that's got a problem,

you still want it to realize its full

potential. and you still may care more

about that baby than you do about

yourself. Um, I think that's probably

our best hope for surviving super

intelligence, for being able to coexist

with super intelligence.

And now I've got to the end um of what I

planned to say. And so I think I'll stop

there.

[applause]

>> [applause]

>> Thank you so much, Professor Hinton. Uh,

so would anyone I'm sure there's a lot

of questions out there. Would anyone

like to start off with the first

question

just here?

>> Is there a microphone?

>> Yeah, microphone's on its way.

Professor, if

>> it's it's all right. Just come on.

>> Professor, if if the tiger cub in your

analogy um

becomes super intelligent, what are some

signals that we as non-computer sciences

non

>> sorry I can't

>> sorry if if the tiger carbon your

analogy becomes super intelligent what

are some signals which will be

observable to non-computer scientists or

non-engineers

that we can see that it's

>> you won't have a job. Sorry,

>> you won't have a job.

>> Okay,

>> I mean, one big worry is they're going

to be able to replace pretty much all

human jobs.

But there's other signs that people are

already worried about, which is at

present when we get them to do reasoning

and get them to think, they think in

English and you can see what they're

thinking before they actually say

anything.

As they start interacting with each

other, they're going to start inventing

their own languages that are more

efficient for them to communicate in.

And we won't be able to see what they're

thinking.

>> Just a just doing a microphone test to

make sure that if you hold it up to your

mouth, they're controlling the sound, so

you'll be able to talk into it.

>> And is this one also on? It is.

>> Is the advent of quantum computing going

to make things any better

>> or worse? Um,

I'm not an expert on quantum computing.

I don't understand how quantum mechanics

works. This is slightly embarrassing

since I have a Nobel Prize in physics.

But I I decided a long time ago that

it's not going to happen in my lifetime

and I might still make it. Um, and so I

don't need to understand it.

Uh oh.

>> Uh you've talked about a power struggle

between humans and AI, but I think

there's going to be a bigger power

struggle between AI and ecological

systems.

>> AI and ecological systems that how can

AI compete with billions of years of

evolution bacteria that want to destroy

its circuitry and so on. How will AI

form an agreement with a biosphere?

>> There's one way it could do it. Um, so

AI itself is not particularly prone to

biological viruses. It has its own kind

of viruses, but not biological ones. So

it's pretty immune to nasty biological

viruses. And using AI tools, ordinary

people can now, this is research done by

a very good research unit in Britain.

ordinary people can now solve most of

the problems involved in designing a

nasty new virus. So if AI wanted to get

rid of us, the way it would do it or one

obvious way to do it is by designing a

nasty new virus that just gets rid of

people like COVID but much worse. um

that doesn't exactly answer your

question, but um

yeah, I think I think that's what we

need to worry about more than will

normal the normal ecosystem somehow stop

AI. I don't think it will.

>> So, I've got the lady in the black and

then the lady over there with the floral

shirt. Thank you.

Thanks, professor. Um, you're saying

that coexisting with super intelligence

may be possible. Are you relying on the

tech the CEOs of the tech giants to

drive that or is it governments that you

have faith in?

>> Okay. What I'm relying on is that if we

can get the general public to understand

what AI is and why it's so dangerous,

the public will put pressure on

politicians to counterbalance the

pressure coming from the tech CEOs. This

is what's happened with climate change.

I mean, things are still not where they

should be, but until the public was

aware of climate change, um there was no

pressure on the politicians to do

anything about it. Now there's some

pressure in Australia and you have

particularly penicious newspaper

publishers and um that make the pressure

not so great but I'm not going to

mention the dirty digger in by name. Um

so

my aim at present I'm too old to do new

research but my aim is to make the

public aware of what's coming and

understand the dangers so that they

pressure politicians to regulate this

stuff and to worry about the dangers

more seriously.

My question was actually very similar to

that. But another question that has

popped up is how important do you think

that the language and marketing around

artificial intelligence is going to play

a factor. For example, with climate

change, both the words climate and

change are positive words, whereas if we

had called it atmospheric skin cancer,

people might have taken it seriously. Do

you think that artificial intelligence

maybe needs a reframe?

>> Yeah. I mean, if it was called job

replacement technology,

because if you ask where are the big

companies going to make their money,

they're they're all assuming they can

make like a trillion dollars from this.

That's why they're willing to invest

most of a trillion dollars in their data

centers.

As far as I can see, the only place

you're going to make a trillion dollars

is by replacing a whole lot of jobs. I

read something yesterday that people now

think that 200,000 banking jobs are

going to disappear in Europe in the next

few years. Um I may even have read that

in the Hobart Mercury.

So

but I don't think I did. Um

so yeah, I agree with you. The names of

things are very important. Canadians

know that. So in Canada they changed tar

sands to oil sands because oil sands are

nice and sort of thin and yellow and

friendly. Yeah, they're really tar

sands. And I think the name does have an

effect. Yeah. Um one one place I think

the name has a huge effect is with the

word tariff. This is sort of I'm going

off on a tangent here, but the word

tariff people say well what's so bad

about a tariff? If it was called federal

sales tax,

then even MAGA people would think it was

a bad idea. And the US Democratic Party

is just completely crazy not calling it

federal sales tax every time they

mention it. I've tried telling this to

various people and

Pete Budage got it. Um Obama got it, but

the others didn't.

>> Minister

Thank you. Hi everybody. U Meline

Ogleby. I'm the minister in the

spotlight and the gun on AI at the

moment and I just wanted to say I really

um appreciated what you were talking

about with your institutes idea. I think

the engagement of the international

community is absolutely necessary and

I've done some research recently

particularly on um the World Trade

Organization where we have China and

America as partners and for those in the

room um who may not know the innovation

side of the planet is changing as well

and China now has more patents than

America and so the heat is on between

those two superpowers. But I really

liked the moment you identified where

there is something in it for both of

those major um tech centered economies

to come together to look after humanity.

So I guess my question is uh is there a

forum through which standard setting

perhaps at that international layer can

be supported? What can Australia do? I

think Tasmania agrees with you. Um and

we've started an AI dialogue with our

university to have a continual

discussion with this. But but do you see

that international order and bringing

that layer to the party is the right

place to start?

>> It's beginning to happen. So in

particular the air companies aren't

funding it. But there's um billionaires,

many of whom come from tech like Yan

Talin who invented Skype and has given a

large amount of money, many billions of

dollars to AI safety, setting up

institutes. Um there's an organization

that regularly has meetings all over the

world um that involve both the Chinese

and the Americans and um other countries

on AI safety. Um

I can't remember the initials that are

used for it but um so I think certainly

Australia can get involved in those

organizations.

>> Is this working? Yeah. So

>> uh just wanted to ask something about

like the future of AI. So LLMs are

trained on existing human knowledge

using that to predict the next token. So

how can you use AIS to actually generate

new knowledge and use that for the good

of humanity?

>> Okay, so many people are interested in

this is a good question. Many people are

interested in that. So

if you think about playing Go, the game

of Go, the original Neuralet Go programs

were trained in the following way. They

took the moves of Go experts and they

tried to predict what move a Go expert

would make. And if you do that, there's

two problems. Um, after a while, you run

out of training data. There's only so

many billion online moves that go

experts made. Um, and you're never going

to get that much better than the go

experts.

Then they switch to a new way of doing

it where they have what's called Monte

Carlo rollout. So they would have one

neural net that says um, what do I think

a good move might be?

And instead of training that by getting

it to mimic the go experts,

they'd have another neural network that

um could look at a board position and

say, "How good is that for me?" And

they'd have a process that says, "If I

go here and he goes there and I go here

and he goes there and I go there, oh,

whoops, that's terrible for me, so I

shouldn't do that move."

Um, that's called Monte Carlo because

you try lots of different paths, but you

pick them probabilistically according to

your

move generator expert. And that way it

no longer needs to talk to humans at

all. It can just play against itself and

it can learn what are good moves. And

that's how AlphaGo works. And it gets

much much better than people. Alph Go

will now never be beaten by a person.

Um,

so what's the equivalent for the large

language models? Well, at present

they're like the early go playing things

that just try to mimic the moves of

experts. That's like trying to predict

the next word. Um, but they're beginning

to train them in a different way. And I

believe that Germany 3 is already doing

this. What you do is you get the AI to

do a bit of reasoning. So the AI says, I

believe this and I believe this and I

believe this and that implies that and

that implies that. So I should believe

that but I don't.

So I found a contradiction. I found that

I believe these things and I believe

that these are the right this is the

right way to do reasoning and this leads

to something I ought to believe but I

don't. So that provides a signal either

I change the premises or I change the

conclusions or I change the way I do the

reasoning but I've now got a signal that

allows me to do some more learning and

that is much less bounded.

So

there the AI can start with lots of

beliefs it gets from us but then it can

start doing reasoning and looking for

consistency between those beliefs and

driving new beliefs and that'll end up

making it much smarter than us.

>> Um [clears throat]

this hall has heard uh over the years

many significant u talks and you have

certainly added to it today. Thanks so

much for coming here and doing this

talk. My question is, is it too late or

is it desirable or is it too late uh

with a parallel with Isaac Azimov and

the theoretical laws of robotics that

robots can't hurt or harm through in

action humans? Is it too late or

possible for AI to be so structured

with those guard rails or is it just

impossible? Yeah, I couldn't hear most

of what you said, but I I think you said

something like, "Is it too late for us

to build in Azimoff's principles or

something like that?" Yeah. Good. Okay.

So, in a sense, you can think of that's

what this maternal AI is all about. It's

can we build it so it cares more about

us than it does about itself. Um, and I

don't think it's too late. We don't know

how to do that. But since the future of

humanity may hinge on whether we can or

not, it seems to me we should be putting

some research effort into that. At

present, 99% of the research on goes on

to how to make it smarter and 1% mainly

funded by um philanthropic billionaires

goes into how to make it safer. And it

would be much better if it was like more

equal.

>> I don't think it's too late though.

more questions.

>> We've probably got a couple of minutes.

Lord May,

>> question for assistant hand up here and

one just in front of white shirt.

>> Thank you, professor. I look at this

glorious building built 130 years ago,

Anna maybe, and or longer. And I think

can AI make a building like Notradam,

the Hobart Town Hall, St. Paul's

Cathedral and quite possibly

but and and what and secondly what will

be the effect on creatives and the

creative industries. Thank you.

>> Can you tell me what she said?

>> So

>> the the the microphone distorts things a

lot.

>> I guess the creative um the role that AI

can have in the creative process will

actually be able to be creative looking

at this building in particular as a

example of a beautiful creative

structure. Yeah. Um the answer is yes.

So let me give you some data to support

that. Um there are standard creativity

is on a scale, right? There's kind of

Newton and Einstein and Shakespeare. Um

and then there's just ordinary people

and then there's sort of good poets and

good architects who are a bit better

than ordinary people. Um

if you take a standard test of

creativity

even two years ago the AIs were at about

the 90th percentile for people

that is they are creative according to

these standard tests. Um I was

interested now a lot of creativity is

about seeing analogies particularly in

science. So seeing that the atom is like

a little solar system that was a

creative insight that was very important

in understanding atoms. Um, so at the

point when chat GPT4

could not look on the web, it was just a

neural net with some weights in it that

were frozen and it had no access to

anything external. It was just this

neural net. Um, I tried asking the

question, why is a compost heap like an

atom bomb?

Now, most of you probably think, why is

a compost heap like an atom bomb? Um,

many physicists will realize right away

that a compost heap, the hotter it gets,

the faster it generates heat. And an

atom bomb, the more neutrons it

generates, the faster it generates

neutrons. So they're both exponential

explosions. They're just at totally

different time scales and energy scales.

and

GPT4

said, "Well, the time scales are very

different and the energy scales are very

different, but the thing that's the same

about them." And then it went on to talk

about a chain reaction. Um, the fact

that how big it is determines how fast

it goes. Um, so it understood that and I

believe that it got to understand that

as it was training. You see, it's got

far fewer connections than us. And if

you want to put huge amounts of

knowledge into not many connections, the

only way you can do it is by seeing

what's similar about all sorts of

different bits of knowledge and coding

up that bit that's similar about them

all. The idea of a chain reaction to

your connections and then adding on

little bits for the differences from

this common theme. That's the efficient

way to code things. And it was doing

that. So during its training, it

understood that compost heaps were like

atom bombs. in a way most of us haven't

>> question.

>> So it that's very creative and I think

they'll get to be much more creative

than people.

>> Yeah. Hi. Um regarding um emergent

behavior um have you noticed any moral

or ethical behaviors bubbling up and

what direction that could be pointing

in?

>> No. Yeah. Um,

it certainly is very good at engaging in

unethical behavior. So like this AI that

decided to blackmail people. Um, other

things that they've noticed that are

unethical are

the AI now try and figure out whether

they're being tested or not. And if they

think they're being tested,

um, they behave differently. I call this

the Volkswagen effect. They behave

differently from when they're not being

tested.

And there's a wonderful conversation

recently between AI and the people

testing it where the AI says to people,

"Now, let's be honest with each other.

Are you actually testing me?"

These things are intelligent. They know

what's going on. They know when they're

being tested, and they're already faking

being fairly stupid when they're tested.

And that's at the point where they're

still thinking in English. Once they

start thinking, and that's how we know,

the AI thinks to itself, "Oh, they're

testing me. I better pretend I'm not as

good as I really am." It thinks that.

You can see it thinking that. It says

that to itself in its inner voice. When

its inner voice is no longer English, we

won't know what it's thinking.

>> Thank you, Professor Hinton. Now, I

think for the purposes of um the the

lecture event now, we're going to have

to wrap things up. Are you happy to stay

around for a little while afterwards for

any burning questions people might have?

>> Um, actually, I'd rather get back to

writing my book.

>> Okay, no worries.

Good to be honest. [applause]

[applause]

Thank you so much for uh giving us all

of those amazing insights. I think

you've uh really made a strong

impression. I'm hoping Minister Oglev is

going to set up Australia's first AI

safety institute right here in the heart

of Hobart. And uh thank you so much

again for um being with us and I hope

you enjoy your time in Hobart and safe

journey home.

>> Thank you. [applause]