[Misc] Eric Schmidt on AI

Download MP3
taking the former Google CEO seriously on AI.
Watch: https://www.youtube.com/watch?v=UQkeRxeh34I


eric schmidt is a business leader and

software engineer that served as

google's chief executive officer from

2001 to 2011.

under his leadership google grew from an

early silicon valley startup to arguably

the most important technology company on

the planet

schmidt is currently co-founder of

schmidt futures and sits on the board of

many public and private institutions

he is still involved with technology

consults with the us department of


also talks about ai in his latest book

the age of ai and our human future

written alongside former u.s secretary

of state henry kissinger and computer

scientist daniel huddenlocker

schmidt was a guest at the milken global

conference and here he anticipates some

of the ai innovations that we will

certainly see in five years

he also predicts what we might see in 20


here are the details

recently in the last couple of years

there have been

extraordinary gains so for example a

team at google and at the baker lab


figured out a way to actually understand

if you take dna what proteins are

generated and what their structure is

that's an extraordinary achievement in

my opinion worth a nobel prize

there are drugs being designed now that

could not be possibly

designed by humans

in any way because of their complexity

there's evidence that ai can be used in

biology ai is mapped to biology the way

math is to physics in other words

biology is so complicated that ai will

be used to interpret biology and predict

its outcome

over and over again ai will arrive in

your life

another example is the hottest area in

my industry right now are large language

models uh recently a set of startups

have been funded between 100 and a

billion 100 million and a billion


they have no current product or revenue

plans um

the the belief of the power of this

technology these large language models

are interesting because you suck all the

information in

like you read all the web which

computers can do but we can't and then

they discover things they appear to

discover a structure of language and an

example of recent google product last

week can actually translate from one

computer language to another and we

didn't give it any examples of one to

translate to the other it discovered a

structure and it can predict it

these are the beginning of general


the the current um excitement stems from

a technology called transformers that

was invented three or four years ago and

what transformers do is it can predict

the next word after a set of words so if

you give it a sentence it can predict

what the word will be and it's done

using a complicated mathematical

technique it turns out predicting the

next word is mathematically the same

thing as predicting the next sound the

next video the next image

all of that and so you have a

unification a multi multi-modal

unification of video text and speech so

these systems sound and look like

they're intelligent

a good example is gpt3 which came out

last year

which kicked the current revolution off

you asked it

do you think like a human and it says no

i do not

because i am a large language model and

you are a i think a human who has been

taught to think in this way

now is that

it thinking about you or is it pattern

matching we can't tell and the truth is

and i'm as part of philanthropic work

i'm funding projects to try to

understand this we don't actually

understand why this works we don't

mathematically understand why it works

and we also don't understand its failure


so you wouldn't want to use this as a

replacement for something that's live

critical because we can't say when it

fails when does it just crash

the current large language models for

example have trouble with the notion of

gravity so if you say to them i moved it

i moved the object from here to here and

then i put it up here and i put it down

there and so forth now everyone just

followed what i did the large language

model gets confused because it doesn't

understand gravity so the computer

scientists say we're going to now add



so with concepts and then with planning

maybe you get to the point where it

looks like a human-like intelligence

which has all sorts of issues

if i were 24 today this is exactly what

i'd be working on this is where the

hardest and most challenging computer

science systems problems are with the

greatest payoff

now remember that the system can predict


and if you can predict a pattern you can

also generate an artifact there's a

duality in these systems where they can

generate things

so part of the issues that we face now

is that these systems can generate

speech i'll give you an example

within five years

the following will be true

you'll be able to take a system

um take one of these language models

which would be infinitely expensive to

make but you didn't pay for it

it shows up in your doorstep and it fine

tunes the technical term is literally

fine-tuning it you fine-tune it to you

who are you what do you care about it

sort of watches you and learns from you

it learns your voice

right all of a sudden it can generate

videos with you in it

now you could think of this as a second

ai right

now the interesting thing is imagine

five years from now i install this thing

and i use it for a few years and

eventually we all die unfortunately well

it lives on


as a pretty good impersonation of me

and what happens when i'm dead and it's

still learning is that me

is that an artifact of me or is it just

a stupid artifact of history that you'll

keep in a box and some future will say

oh eric was so stupid back then but it's

entertaining to watch him right because

he didn't keep learning we don't know we

have no way of discussing these things

this stuff is incredibly powerful it

will be the basis of enormous gains in

human health

language translation communication

summary and education

all the things that milken represents

will be affected in an almost always

positive way having said that there's

terrifying consequences as well so the

first question has to do with jobs

does this fundamentally mean there are

more jobs or less jobs i spent my whole

life people saying computers will

replace humans humans won't have

anything to do so far that narrative has

been false notice that there's a huge

surplus of jobs and not people to fill

it certainly in the united states the

second one has to do with national

security something i've worked on for

almost a decade now

and in our in the kissinger book we talk

a lot about this

what happens when the decision time that

you have in a conflict is less than

human i'll give you an example

in the future there's a war and the war

is north korea attacks the us the attack

u.s attacks back and china says not a

good time for a war stops the war in in

north korea

and the entire war took five


now how are we going to organize around

that war how are you going to build a

system that can make all of those

decisions what are the rules

around automatic decision making i just

uh i on a team for working for the

congress published a recommendation that

this was a disaster

and that you have to have human in the

loop but you have but when time gets so

short another example if there's a

in the horrific scenario of an icbm

coming with a nuclear weapon in the

1960s they had about a half an hour or

45 minutes something like that for the

icbm to get to the u.s well during that

time they have time to figure out where

the president is wake the president up

have the president say why are you

waking me up and they say there's a

missile coming and you have to authorize

it and and then the president goes is

this a dream and they say no

okay and then eventually he or she goes

okay launch the retaliatory thing

right it's organized around human

decision time in the military it's

called an ooda loop for those of you

that are that are veterans

i'll give you another example


the question of social media

social media was originally organized as


feeds right linear uh we saw what our

friends were doing in order

and the systems the literal the software


social media networks

amped them up by amplifying content they

thought you would want

well here's a formula for you if you

were an evil social media ceo which none

of you are you would try to maximize

revenue to maximize revenue you would

maximize engagement to maximize

engagement you would maximize outrage

because you make more money

now how we're going to deal with this

and none of us are opposed to free

speech all of us believe in free speech

blah blah blah we have to sort these

issues there's issue after issue

so the strategic problems that i'm

describing are going to be solved by


the algorithms will be changed by



microsoft has published a very powerful

product called codex which finishes your

software so you start typing

right in a normal programming language

like python and it can actually finish


something like a third of the code

that's coming in was written by a


now what happens when it's half what

happens when it's 60 what happens when

it's 70 percent

this is the beginning of this

extraordinary revolution

i am sure in the next five years you're

going to see the in the integration of

conversation and multi-modality because

that's what everyone's working on

and what that means is that you'll have

a digital assistant that will

mark what you should do in the morning

make suggestions give you summaries have

a sense of humor say this guy's an idiot

ignore him on social media uh and by the

way generate your own tick tock videos

you don't have to go out there in the

wild and actually get photographed doing

it it'll just generate it for you and

you can submit it all of that's going to


then we get to speculation

today the systems and it's very


do not have the ability to set their own

objective function

humans very very smart people say we

want to solve this problem or this

problem or this problem

there are many people including myself

who believe that there's going to be a


and this is the

the median of the predictions is 20

years so i'll say 20 years which

technically means may 2042 so we're

clear you have it on record


these systems will be able to set their

own objective functions

now you sit there and you go oh that

means the computer can start to think

like a human

that's not actually what i mean

what i mean is that it will be a kind of

intelligence that's not human but feels

like it's kind of human maybe

it can reason it can think it can make

predictions it can make choices

but not based on a biological human


now how we're going to treat these


right so let's say i have my opponent

right so he's sitting here and i don't

like him and i don't like what he's

doing and i'm fighting him in some way

i know he's human i know he has

biological limits i know he has to sleep

every day i know he has concerns about

his own mortality he has to eat you know

the things that we all share as our

shared humanity

i know the limitations of his evilness

and his brilliance

because we have so many examples

so now we have the computer here

which is now doing its own thing

how do i

characterize its risk

what is it capable of doing

where does it find its limits

what how does it decide what to do we

don't even have a language to discuss

how we would regulate these things

so i'll give you a further prediction

that these computers when they emerge

and they're going to be a huge fight to

get these things because these things

are so powerful

but once they emerge they're going to be

like plutonium

i visited as part of my military work

where we keep the plutonium and it's

inside of a fence with guards and then

there's another set of guards with even

bigger guns and then there's a

a building with even more guns in it

because it's so dangerous

and there'll be a small number of these

because the computation required to

doing what i'm describing will exceed

the vast majority of most of our

capabilities but let's say there's one

or two in china and maybe one in north

korea and maybe one in two or three in

america and one in israel and so forth

how we're going to sort all that out

nobody knows we don't even have a

language to talk about the emergence of

a non-human

human intelligence

and that's incredibly interesting and

incredibly powerful maybe can solve the

gravity problem that einstein couldn't


but also you can imagine the

consequences negatively


[Misc] Eric Schmidt on AI
Broadcast by