Transcription
AGI has become a marketing term. If AI
could do a useful economic work task as
well as a skilled professional human,
then that seems to me like a more
reasonable definition for AGI. I think
we're very far away from that. Maybe
more than decades.
>> Andrew Angie is one of the key
architects of modern artificial
intelligence mentioned in the same
breath as Jeffrey Hinton, Yan Lun, and
Demi Hassabis.
>> Now, just be transparent, just be
completely truthful. There is a small
number of job roles that are fully
automated by AI. So candidly I think uh
excessive hype that leads to
disappointment that leads to you know
collapse of the so-called bubble that
would not be good for the world and good
for the field of AI. A lot of the best
open source open way models are coming
out of China.
>> Is the opensource era over?
>> Many universities are slow to adapt
curriculara and are still training
students for the jobs of 2022. Many of
those jobs don't really exist. So
employees don't want to hire. Will
programmers be unemployed?
>> Will this be the year we finally achieve
AGI? You are opening your expose from a
few hours ago with this question.
>> For any reasonable definition of AGI, I
think the answer is no. We will not get
AGI uh in 2026. The best way to get the
AGI in 2026 is if someone manages to
dramatically lower the bar for what AGI
actually means then maybe some company
clear the hurdle.
>> How you explain then define AGI?
>> The definition of Agi that I'm most
familiar with is AI that could do any
intellectual task that a person can. And
uh today a human can learn in maybe tens
of hours how to I don't know drive a
truck through a forest which I've never
done before but I think if you train me
for a few hours I could probably do it.
Um human could learn to take calls for a
contact center and answer questions
consistent with how certain business
needs it answered. And there's certain
things a person can learn to do
intellectual task as opposed to physical
task. Um, and a lot of these tasks still
feels like it's a ton of work to build
custom AI workflows to perform, which we
do do and which turns out to be really
valuable, but needing an engineer to
spend so much time building these things
is not how the public typically thinks
about AGI. And as much as I hope we'll
get there someday of computers are as
intelligent as people in every sense, I
think we're very far away from that.
You've just proposed today a new version
of the touring test. Can you explain
what you propose?
>> I'm excited about this. So, um I feel
like because of the hype around AGI, AGI
has become a marketing term rather than
something of precise technical meaning.
And the way it's used as a marketing
term, it actually misleads a lot of
people. But since people want AGI are
excited about it, I'm excited about it.
Why don't we come up with a test to see
if we are actually, you know, getting
close to achieving AGI? So, here's the
idea. Um, in the original touring test,
a human judge would get to type text
back and forth to either an AI or a
human. And the original touring test was
can AI fool the judge into thinking or
being unable to tell if they're talking
to AI or to a human. And it was a great
invention for 1950, which is when I
think Alan Turing came with this test.
But it's not really measuring
intelligence the way we think about it
today. Um here's how I think we can
build a touring AGI test is what I
called it which is have a human judge
design a multi-day experience uh that
could be um some uh on boarding training
uh via a computer and the test subject
will be either AI or a human with access
to computer and normal software like a
web browser maybe zoom maybe other types
of software and if over a few days um if
in a multi-day experience
If AI could do a useful economic work
tasks as well as a skilled professional
human, then that seems to me like a more
reasonable definition for AGI. And the
reason I'm proposing this is because
this is actually much closer to what the
broader public thinks of this AGI. So if
people think AGI could be here, they
think, wow, AI will be able to do
people's jobs. And indeed, if AI is able
to function like a remote worker for
multiple days and do productive work,
that will be incredibly valuable and I
think much closer to what people think
of as real AGI as opposed to some of
these, you know, alternative standards
that some businesses are promoting for
PR or policy or fundraising purposes.
>> I remember my interview with Sir Roger
Penrose, Nobel Prize winner. Math is a
closed system. So AI can play it like a
game. Uh aren't you afraid, Andrew, an
effective AI test doesn't exist?
>> One of the challenges with um tests or
benchmarks of AI is when someone fix a
test set in advance, then you're just
measuring one specific dimension of AI.
uh this is why for a lot of AI models
large language models there are a lot of
standard benchmarks you know Sweetbench
uh GBQA and so on and um it's very
difficult to avoid having teams optimize
for the known test sets out there even
if they don't do it directly it ends up
and AI is a jagged form of intelligence
it's great at some things really awful
some other things but AGI people think
of it as AI that could do any
intellectual task the person can should
cover all intellectual tasks that people
can cover. So to me, one big difference
between a fixed test set kind of
benchmark versus letting a human judge
probe in real time is a human judge can
take a look at the state-of-the-art and
can probe to see where is AI strong,
where is AI weak. And for AI to be this
AGI, you know, most people think of it
as matching humans in every dimension,
which means that a human judge should be
unable to probe to find places where the
AI is materially weaker than what a
human can do in terms of these say
economically valuable job tasks that we
want AI to do.
>> Are benchmarks still measuring something
real? benchmarks hit 90% and the users
say that is feels dumber
are benchmark measuring something real.
Oh, I think benchmarks are measuring
something real but it's a very narrow
slice of the sorts of things we want to
measure. Um, one weakness of a lot of
benchmarks is we're much better at
designing objective benchmarks. So
things like math which is a right or
wrong answer or a factual question you
know like in the this year's Olympics
who won this you know 200 meter
freestyle swimming or whatever so those
questions have objective right and wrong
answers and so a lot of benchmarks is on
measuring is the AI right or wrong it's
very black and white um lot of life
there is it's very difficult to design
one completely objective best answer so
even in a um human conversation like
we're having, you know, what's the one
right thing for me to say? I don't know,
but I could probably say stuff that is
better or some stuff that is worse. Um
and on average, we've been not very good
at designing benchmarks to measure these
things that are more um subjective or
whether gradations are good and bad, but
it's very hard to write a test set
saying this is one right answer, this is
one wrong answer.
um not that we don't try but I feel like
a lot of the benchmarks we have uh don't
capture those things as well and the
reality is a lot of human work if you
ask me to write a research report you
know there's no one right report there
are different gradations of how good or
bad the report is and our ability to
write benchmarks to capture this gray
shades of gray is much weaker than our
ability to write benchmarks to say did
it solve this math problem correctly did
it write a piece of code that runs
correctly. Those we know how to do
better than what a lot of actual useful
human work requires.
>> If not AGI, what we can expect in 2026?
Greg Brockman from OpenAI says two big
dams of AI in theat dams of AI in 2026
will be agent adoption and scientific
acceleration. What's your take on that?
>> I think AI is a distraction. We're very
far from it. You know, we're not going
to get there anytime soon. Um, but even
without achieving AGI, there's so much
incredibly valuable stuff that we are
doing and that we will continue to do in
2026. Um, I coined the term agentic AI
to describe what I saw as a growing
phenomenon. And even though I was saying
that it's so much work to build AI to
capture useful business processes,
um sometimes when we do that is
incredibly valuable. So in 2026 and
beyond
there'll be a lot of exciting work to
build AI agents or to build agentic
workflows to do a ton of really valuable
economically important work. Here at my
team at AION, uh our teams have been
using agentic workflows to write code
like many others do too, but also to um
look at paperwork to check for tariff
compliance as guy AI or um to read
tricky legal documents to you know help
lawyers do their work better or to do
medical assistance tasks um or to
support customer service. And I find
that taking the mental processes that
people are currently going through to do
these types of tasks be it legal
compliance or whatever um and coding
them into an AI agent on agentic
workflow. So the AI could do it for you.
This I think will go on for a long time.
We'll be building these agentic
workflows for many many years to come.
But the value is is uh will be will be
very large.
>> What we are talking about agentic
workflows. Professor Rich Sutton
famously uh said, "Raw computing power
always beat human cleverness." Uh by
focusing on agentic workflows, aren't
you against?
>> So Rich's uh uh uh article on the bitter
lesson was very influential and very
well written. Um just to be clear I like
scale. When I started Google brain the
which later merged with um deep mind to
form Google deep mind and created Gemini
and so on. My number one mission that I
set to the Google brain team was said
let's just scale right there bu really
really big neon networks threw lots of
data out it so I think I was probably
one of the earliest in AI.
>> You are the architect of the era of
scale. I pushed for scale when everyone
else thought when frankly everyone else
at that time thought it was a strange
thing to push for pretty much. So I so I
like scale and I think it was because um
the Google brain team the way I set it
up is DNA was just a scale that was
number one thing you know there
obviously a team does many things worry
about many things but the number one
mission was scale that's how I set up
Google brain that's why with that DNA it
was the Google brain team that invented
the transformer neuronet network that
was the most scalable neuronet network
architecture you know in history uh and
that powered the generative AI
revolution, right? So, so I really
believe in scale. Having said that, at
different moments in time, um, our
ability to scale and our ability to
inject other forms of knowledge creates
a balance where I don't think we'll
achieve everything we want only by
scale. um is incredibly powerful too.
And like many things um
sometimes we've seen businesses hype up
one phenomenon for fundraising or PR or
whatever purposes and it is absolutely
true that because of scaling laws when
you scale systems you can fairly
predictably forecast where performance
will go. This was shown by my team at
BYU later by uh OpenAI I think. Um uh
and so this is actually a wonderful
argument to raise funds because you can
say give me more money I will scale up
these machines scale up the data get
better but but because of because of
that huge kernel of truth and scale I
think it has been hyped up beyond what
it actually is which is amazingly
valuable but not as valuable as the hype
says it is but nonetheless still
amazingly valuable and agentic workflows
allows us to take advantage of uh the
scale AI models launch models and
additionally inject other types of
knowledge into the system to build more
reliable, more performant um workflows.
>> Today, do you truly believe that the era
of scaling is over?
>> No, I don't think era scaling is over.
It's been getting harder and harder. Um
yeah, it's been interesting.
Um AI has been progressing maybe
exponentially.
It's been interesting seeing that the
number of dollars needed to drive this
exponential progress has also been
exponential. So you spend exponentially
many dollars to drive exponentially
rapid improvement is not bad. It is very
valuable because um the cost of building
it can be amotized over a lot of users.
Um and I think there's still more to be
gained by scale but um it's not the only
avenue we have for improving AI at this
point. What would change your mind
>> about what?
>> About scaling.
>> Oh, what would make me give up on
scaling?
>> If for a extended period of time further
efforts to scale up models are not
paying off, that would make me change my
mind. But maybe um one important
asterisk to that just as over many
decades Mo's law drove progress but it
took lots of different technologies that
kept on changing over time for Mo's law
to keep on making semiconductors work
better and better. What we see is in the
early days of genon the recipe was get
more data train a bigger model. this
exactly what I at least at that level of
abstraction exactly what Google brain
was set up to do right more than 15
years ago or so um uh but since then AI
models have read pretty much the entire
open internet so that simple recipe for
scaling it doesn't really work anymore
which is why many teams are now doing
much more work on synthetic data
generation which requires much more
human engineering a lot more work on
different reinforcement learning recipes
um and So even though scaling still
continues to pay off the specific
recipes for driving that scaling has
definitely shifted quite significantly
over these last you know two three
years.
>> So from the another side can a smarter
model simply beat a smart workflow.
Um,
yes. I wish it were that easy though
that in practice uh I think it is um
in theory yes in practice very harder
than most people think.
>> Why?
>> So one thing I think has been fantastic
was with more intelligent models uh be
it you know the claw 4.5 focus uh Gemini
3 um GBD 5.1 5.2 to we've been
increasingly able to give a large
language model a set of tools and just
have it let loose, right? So just give
it a set of tools for reading and
writing for a file system then just tell
to do some tasks like you know look for
extraneous files and help me clean up my
hard disk or whatever. Um and the level
of performance of these is is is
fantastic. It's really really really
impressive. Having said that, for many
workflows, it's just not reliable enough
to be production ready. Um, and so while
smarter models are great, what I see
right now for a lot of practical
business use cases is teams thinking
through what is the workflow you want,
what are the key steps and implementing
them peace meal so they can get that
reliable performance to get this stuff
working and have it run 10,000 times and
maybe even work, you know, pretty much
every single time. And maybe and and as
models become more intelligent, we are
letting them be more autonomous. We're
removing guardrails. One thing my teams
often do, stuff that we built six months
ago probably had more guard rails, was
more scripted, and we routinely end up
reducing the scaffolding. Instead of
giving a very detailed step-by-step
instructions, we're more likely to say,
you know what, um, go decide for
yourself what you want to do. I mean one
example we used to build deep
researchers where we say to do research
on a topic do a web search do this many
queries then download this many pages
and summarize it blah blah blah common
research agentic workflow from like a
year or two ago now AI models are much
more capable of deciding do they want to
keep on searching the web more do they
want to summarize or not or not so I
find I find myself frequently taking
systems we had built a prototype a year
ago six months or a year a year and a
half ago and ripping out instructions to
say you hey prompt the just figure out
the decision by itself but um still a
long ways to go it it works for use
cases like deep researchers where um
it's not taking action if it misses a
citation or misses a reference it's not
the end of the world but for a lot of
high stakes enterprise use cases um I
think the reliability gap to letting AI
model just let loose and do what it
wants to do I suspect that gap while
it's closing is still bigger than than
some people.
>> I need to ask you about the dispute
between Ian Leon and uh Deis Hassabis.
Uh Ian Leon uh famously says human
intelligence is specialized not general.
Demisabis argue that brains are general
learns under theoretical computations.
Where do you stand on this?
>> You know, I see no contradiction. Maybe
I'm missing something. Uh to me, the
amazing thing about the human brain,
assuming the human is AGI except it's
not artificial, right? It's general
intelligence. The amazing thing about
the human brain is its plasticity or it
ability to learn. Um, and I feel like
AGI to me should be less about AI that
already knows everything under the sun.
That seems very challenging, doesn't
seem practical. But what makes a human
brain so valuable or one of the things
that makes it so valuable for economic
tasks is ability to just learn to do new
stuff, whatever is needed. And um and it
is through learning that we then gain
these incredibly specialized
intelligences like that human brain has
learned to solve incredibly difficult
math problems by getting a PhD in math.
And so you know that is very specialized
intelligence as the end point but it was
learning and then but but theoretically
that same human brain just given
different training could have been a
chess master or could have been um
amazing at you know playing tennis. Uh
so to me a lot of what makes the human
brain so general is not that my brain or
your brain already knows everything
under the sun is our ability to adapt to
learn a huge range of things.
>> Doesn't that implicit side with le?
>> Yeah. One thing that really motivated me
when I was um started the Google brain
team was uh this idea that people don't
talk much much about anymore was that a
lot of human learning may be due to one
learning algorithm. uh what that means
but but it turns out that um our DNA is
not that long. So our DNA contains a
very limited amount of information. um
uh but somehow our DNA has encoded
uh the biology of the brain and the
brain is a fairly general learning
algorithm which is why the brain could
learn to you know do a PhD in math or
learn to ride a motorcycle or right
learn to type a computer or learn to um
work in a contact center in a call
center is because the brain has this
very general learning algorithm that
allows it through learning to specialize
in it bey
range of things and that's what makes
intelligence seem general. It is this
learning ability that lets it gain you
know almost any specialty under the sun.
Some oddman um said a few months ago
Google is doomed because they are
bolding AI onto old search.
Open AI builds from scratch. Uh do you
think they can marry old with the new?
>> I mean the race is on. This isn't
exciting. Uh I think uh Sam was my
student at Stanford and a lot of friends
at Google. So I'm very pro open also
very pro Google. Um if we look back at
the history of technological disruptions
whenever there's a tech disruption
sometimes the new entrance won sometimes
the incumbents one. There's actually
room for either one to do well. And I
think the game is still on. Um in fact
if we look at the inter internet
disruption um Google was a startup that
rose with the r with the rise of the
internet but there's some incumbents
like Microsoft and Apple that were
founded long before the internet and
they did just fine. So clearly AI is
very disruptive uh for incumbent
businesses like Google. I think Google's
played its hand well. I I think uh
Gemini 3 is an incredible model
>> better than Chad GPD.
>> I use both Gemini and Chai GPD uh and
claude uh and and a whole bunch of
models quite often.
>> You are uh talking about AGI a lot. Um
five years ago
>> can I just say I actually don't like
talking about AGI but others talk about
it so much and it's so hype. Yeah. I I I
feel like you know I feel like as as um
leaders in the field when there's a
massive amount of hype
>> you believe that AI AGI is a hype or
hyped
>> it is vastly hyped uh at least the
public thinks of AGI as becoming AI
becoming as intelligent as people in a
very general sense we are so far away
from that um I hope we'll get there I
would love to get to AGI but
realistically I think we're decades you
know maybe maybe more than decades away
from that and so this idea that oh we
just need you know another few quarters
to get to AGI that's just not that's
just not going to happen unless you
redefine AGI to lower the bar and make
it much easier to achieve and I think
and and and and candidly
in the history of AI we've seen a few AI
winters where people well-meaning people
overhyped the promise of AI this led to
elevated expectations that were not met
and um a collapse of investment and
interest and this was bad for the few.
AI works really well right now. It's
incredibly valuable. I see relatively
few things that could derail AI's
momentum. One of the things I actually
worry about is um excessive hype that
leads to disappointment that leads to
you know collapse of the so-called
bubble um and that would not be good for
the world or good for the field of AI.
So diffusing hype about AGI is an
important thing to do to lay the
foundation for us to have more
sustainable growth.
>> Five years ago, solving any coding
problem would have been called AGI.
Today we have it and we call it a tool.
Did we move the goalposts?
>> I I don't remember credible teams
declaring achieving AGI. Uh, I remember
teams declaring that AGI was only a
short time in the future, but I don't I
I don't teams were not three years ago
declaring they got AGI. They were saying
we could get there really soon. Uh, and
then, you know, so far no one's gotten
there.
>> Don't you think that we move the
goalpost?
>> I I think that if anything, teams have
been trying to lower the bar to what it
means to achieve AGI. I mean, AGI,
AI that can do any intellectual task
that humans can is a very high bar. Um,
we have not gotten anywhere near to that
bar. Um, but if teams come up with
alternative definitions that are easier
to achieve, then maybe we can get there
sooner. And by the way, I don't mind how
we define AGI. You know, we can define
AGI however we want. But the problem is
most of the broader public thinks of AGI
as generally very intelligent AI
basically humanlike intelligences. So
that's what most people out in the world
thinks. And so the problem with
>> in subjectivity.
>> Yeah. Basically so you know I think
because people keep coming alternative
definitions. It it turns out that terms
lose meaning when lots of different
people use that same term to refer to
very different things. And so um uh for
example, we all have a sense of what the
word blue means, right? My shirt is
blue. But if for some reason society
points all sorts of different colors and
says this is blue, this is blue, then
the word blue will lose meaning because
people don't even know what it means
anymore when someone says blue. And
that's what happened with AGI. Different
teams came up with different alternative
definitions. people say this is AGI and
and and and
because people are applying lots of
different definitions to the term when
someone now says AGI it's hard to know
what exactly they mean and the problem
is the public thinks of AGI as humanlike
intelligence and so if someone come
because of this uh which seems
reasonable to me and so if someone comes
up with some weird narrow technical
definition and says AGI will appear in
two is the broader public nonetheless
thinks AI will be as intelligent as
humans in two years which just doesn't
seem true to me.
>> Let's talk where are we now? Does the
hardness matter or it is all about you
know the workhorse?
>> I think the honest matters a lot. Um
it's been really incredible watching um
Anthropic build cloud code um as was
their uh SDK uh to make sure that they
and and maybe others have a good harness
to use the model powering this um and
the details of uh the harness, how you
structure the prompts, how what tools
you give the elm, all those details
really matter still. Um actually maybe
one one one small example uh our current
models are incredibly intelligent and
they're becoming better and better at
two to two use right making function
calls. Having said that, if you give a L
model too many tools,
um, it consumes a lot of the input
context and it's much more likely to
struggle and make the wrong API call,
make the wrong tool used call. And so
these little kind of maybe it feels
like, you know, it's 2026, why do we
need to worry about injuring details
like these? Um, it turns out we do
because it still make a big difference
to the overall performance. And, um, you
know, many teams use MCP. I use MCP a
lot and one of the practical engineing
things is if your MCP server has too
long a list of tools, it consumes a ton
of your own input contexts. Uh, and it
may be too many tools to the to
effectively figure out which ones to
use. So, so you get into context
engineering and and and sometimes a
harness can help make these decisions
much more smooth.
>> Anthropic predict continual learning
will be solved by 2026.
Do they expect the same? I look forward
to when continue learning will be
solved. Uh it would be um amazing. It
would be delightful if it ends up being
completely solved in 2026.
I I expect we'll make progress. I think
continue learning is very important.
>> A child learns to walk uh from few
falls. Reinforcement learning needs
millions of of simulations. Uh do we
just brute force learning?
One of the challenges of brute force
learning is um to the extent that in
intelligence human intelligence comes
from the generality of the learning
algorithm that's what makes a human
brain so powerful. general learns new
things really quickly and this is why um
for example humans you know that work
with us are can do so much whereas if
you need to spend a long time to brute
force AI to do some narrow task it could
still be really valuable in some cases
but the case for doing it just isn't
going to be there for a lot of tasks so
the ability to um say hire a human and
talk to them a it and then figure out
what to do and do a job that's really
valuable. If you need to hire an AI and
then spend a million dollars to train
the AI to do this one task, then there
are a lot of tasks that it just doesn't
make sense, right, to spend that much
effort to train AI.
>> If sample efficiency doesn't matter,
only three companies on air
unfortunately can afford to train
frontier models. It'll be really
interesting to see the continually
evolving dynamics between open source
and open way models versus the
proprietary models um AI as oligopoly.
>> I hope we don't get to that future. So
kindly if I look at um the mobile
development platform
it's just not as interesting these days
partly because there are two gatekeepers
right to do something on mobile um at
least in United States you kind of need
permission from either um iOS or Android
and so there are certain innovations
that we're just not allowed to do
because of those closed platforms
>> a lot of us in AI really hope that there
will not end up with two or three
gatekeepers to building cutting AI
things where if someone has an idea um I
would love to have people allowed to
innovate on top of large language models
and so open source open way models is a
key to um preventing there from this
small handful of gatekeepers from
arising and if we can um make sure that
everyone preserves the freedom to
innovate which we have much more in the
AI world today than say in the mobile
world today then uh we'll see a lot more
inventions a lot of cool applications
and society will be much richer for it.
Do we really need
real continual learning?
There's a lot of work to be done in
continue learning. I think of it as one
of uh the uh one of the important open
research topics in AI right now. Um a
lot of the current textbased memory
systems have the AI do whatever it does
and then write a bunch of text into some
sort of agentic memory. Um, and
questions are, you know, is text really
the good enough representation for
memory? And there's some research on
non-ext representations, but a lot of it
is in text right now. Um, uh, and then
also the fact that we're building all
these memory systems that don't really
ever update the weights of the LM that
feels like we have to be missing a key
piece of the puzzle. So um some of my
teams have been have had ideas uh not
sure not sure where we'll take them or
not take them but uh uh improving
continue learning in terms of the things
humans can do the AI cannot do that
seems
>> what's the biggest bottleneck to
continual learning right now. Oh, I
would say we don't have the right ideas
or we're not sure what the right idea
is. It's a bunch of ideas. It's just one
of those things
>> even you
>> there are few ideas that I think are
promising that but
uh uh for example
>> uh maybe let me see if I have time to
try it out first and I'll let you know
if it works. I honestly don't know if it
will work. But to me, it's one of these
things, you know, if we were to ask um
yeah, it's like it's kind of saying
what's the bottleneck to solving this
huge unsolved research problem? And I
don't know because the path isn't the
path to get there isn't clear. So I
can't even say what exactly are the
bottlenecks. We just don't know how to
do it. Elizabeth Yutokovski
um says if that's a quad if anyone
builds AGI
everyone die but every year we delay
millions die from cancer aging disease
AI could solve which risk is greater
>> I can't make head the tales of a lot of
his arguments when I read them um many
of his arguments are sufficiently
circular popular. I don't even know how
to argue with them. Um uh I think that
um
AI is doing so much good in the world
today that um anything we can do to
accelerate the progress of AI will lead
to better lives, many more lives saved,
many people much wealthier, many people
much better off, live a lot of people
out of poverty. um AI's net benefit is
so much greater than the net harm and
there are a few harmful use cases which
Azure is in clear on and let's get rid
of those but at this moment in time very
confident anything we do you know things
we could do to accelerate AI progress
would be good for humanity
>> from your point of view future is
unpredictable right
>> I wish I knew how to predict the future
of course yes it's unpredictable but but
I feel like just because we don't know
in with certainty exactly where the
world will be um that trends that we
have very high confidence about. So for
example, um having in computers be more
intelligent seems clear to me that's a
great thing. Making democratizing access
to intelligence so that it's not just um
uh you know so so it turns out one of
the most expensive things in the world
is intelligence. It costs a lot of money
to hire a smart doctor, a smart teacher,
right, to to care for you or
>> do we need more AI safed tools?
uh and yes of course we should be
working on um making AI systems more
reliable and reducing risk. At the same
time if I look at all
>> everyone around is focused on something
that is profitable.
>> You know I think many people be
surprised at how um there's this view of
Silicon Valley that people only care
about money and nothing else and that's
just absolutely wrong. frankly a lot of
friends that work in a lot of the
companies. Um I know many cos personally
and yes there's a tiny minority that
just cares about profit but it's a very
small fraction. Um a lot of my friends a
lot of these companies people have known
for decades a long well a decade or
longer um really want to do the right
thing. So people take God real
seriously. People take responsible AI
seriously. people will sit down and
really brainstorm all the things that
could go right or wrong with AI system
and try to mitigate your risk. So I know
that there's a view that there's a
stereotype a false stereotype of a bunch
of you know cowboys or cowg girls in
Silicon Valley that just do anything for
profit. It's just completely not true.
Now unfortunately it is true that
there's a small number of companies when
um there's a temptation you know to make
billions of dollars that temptation is
strong but I think that's really a small
minority very small minority tiny
minority of the decisions people have to
make in terms of
>> is the opensource era over
>> open source is doing great now one weird
thing is uh a lot of the best open
source open way models are coming out of
China looking back on Indry over over
the last few years every single year the
options that have been open source or
open weight has grown rapidly so I think
the opensource movement is very strong
at the same time the proprietary options
have also grown rapidly but that's okay
the important thing is uh the open
options are also growing strongly what
goes through your mind when you look at
the current stage of AI development
I want to empower everyone to build AI.
So um as a developer, I don't ever want
to have to code by hand again. I want AI
to write as much of my code for me as
possible. And the acceleration of
software engineering because of AI is
very clear. But what is less clear to
many people is many people that are not
software developers would be so much
better off building software, doing
things with AI than not. And um Armor
Marie seeing this among our teams, the
marketer that knows how to use AI is
trying to run circles around the ones
that don't. Uh my CFO, you know, at AI
fund, she writes code and she gets much
more done than some other CFO,
hypothetical CFO that doesn't know how
to write code uh using AI assistance. So
what I'm seeing is with AI tools,
including specifically
building software with AI, um it feels
like a really important new capability
that we need everyone to have. So as new
capabilities come up, um many people
will embrace them and raise ahead and be
able to do much more and be much more
productive. And then sadly, there'll be
people that don't embrace them that
unfortunately will be left behind. Um
and I'm actually quite worried about
that. One of the big challenges is uh if
you look at the university system,
many universities are slow to adapt
curricular and are still training
students for the jobs of 2022. But many
of those jobs, you know, don't really
exist. So employees don't want to hire
like 2022.
But instead, many employers can't find
enough talent that knows AI, knows how
to build with AI. And I don't mean
software engineers. I I can't find
enough marketers and recruiters and
finance professionals that really know
AI and shifting the educational system
to give students um as well as adults
the ability to use these tools and get
so much more work done. That's going to
require a huge shift of the educational
system. But how to get there at this
moment? It still seems very challenging
to me. What would it take for AI to
replace you as educator?
>> I wish I knew. Turns out my team um
regularly tries to write AI to replace
me uh with my blessing and strong
encouragement.
>> And still it's an impossible, right?
>> Sadly they've not managed to replace me
yet. I suspect that I am a general form
of AI.
>> Maybe we don't know how learning works.
>> Yeah, I think so. You know, if if we
ever get to AGI, it may be great. Go
retire, go do something else. uh uh I
feel like uh replacing a lot of skilled
people that feels like a AGI problem
which feels like it's still I think for
narrow verticals our ability to build AI
to serve narrow verticals uh will grow
much faster right than this hope for
general AI so uh I don't know driving
school is actually an interesting one it
turns out that
driving schools in the US have been
subject a bit to refugee capture So a
lot of innovations that have been
adopted in other nations like driving
simulators. It turns out that
>> driving simulators count in terms of
number of hours driven in many nations
but not in the United States. Because of
that driving simulators are not used as
much in the United States and other
nations. So there's this weird stuff
that is very frustrating. I think
America should be using driving
simulators more because it's a very safe
way to teach kids.
>> Will programmers be unemployed?
>> I think programmers that don't use AI,
they will be in trouble. But programmers
that really know AI, they're so
productive. Just can't find enough of
them.
>> Are they lost to go?
>> Oh, lost to go sounds a bit dire. I
think that uh most job roles are not
going anywhere, but that you know that
that pity saying AI won't replace
someone, but someone that uses AI will
replace someone that doesn't. That's
that's true a lot of the time. Now, just
be transparent. Just be completely
truthful. there is a small number of job
roles that are fully automated by AI. So
candidly I think uh lot of translators
are in trouble um translators translator
I think uh voice actors they could be in
trouble as well. So the bad news is
there's a very small fraction of job
roles the AI can automate entirely those
I I actually feel for them and I think
we owe it to them to do a lot to make
sure that they can gain new skills reach
out the workforce find other meaningful
things to do that I feel as a AI person
I feel a duty to do whatever I can to
make sure people are taken care of
but for the vast majority of jobs AI can
automate
>> radiology
um I think it's taken much longer than
people thought to automate radiologists,
right? Uh um I think uh uh I think it's
taken much longer than the lawyer
predictions.
>> Oh, I think lawyer has so much reg
capture it'll be very hard to completely
replace human lawyers. But lawyers that
don't use AI will be so much less
productive than lawyers that do use AI
>> because you know uh AI is very good in
that legal research. I think AI could do
parts of law really well. But but here's
the trick. If does a job and AI could do
30% of it, then that 70% you still need
a human to do it. But that human had
better use AI because the one that
doesn't use AI, well, they're missing
out on a lot of productivity.
>> Could you predict which jobs could
finally disappear?
>> Um well, the clear ones are a lot of
call center jobs are going away. um uh
translator jobs, voice actor jobs. I
think the ones that are in trouble are
the ones where almost 100% of the work
or 100% almost 100% of the work can be
automated. But it turns out that so many
jobs are so complex, multifaceted, also
great at text, not so good at
non-extings. So for most jobs, AI could
I I think we look at these task based
analysis of job types of studies that
you know my friend Eric Brennson and
others have done. take a job, break it
into task, see what task AI can
automate. For a lot of jobs, AI can
automate like 30 40% of someone's job
and so you still need a human to do that
you know 60 70%. Uh and there is a small
fraction where AI can automate almost
everything. So those jobs are in trouble
but that's a very small minority of the
jobs out there. You left by do in uh
2017 right? Uh what was the reason that
you decided to be here? So you know I
had a great time at BYU uh uh and um
it's funny people keep wondering when I
you know move when I moved from Google
brain to and then by do to running AI
fund and div
some secret thing but but uh what
happened so I feel like I was running
the AI team at BYU uh great team uh uh
and I think the team did a wonderful job
you know uh building the mothership
right so online advertising improve
improving websites, all the core
businesses and team was doing a great
job there. I remember um uh looking at
the orc chart and then thinking, you
know what, if I weren't here, we've got
a really good team. They'll probably do
just fine without me. And then I also
realized that the most fun part of my
job at the time was actually building
new business units. We did a great job
making money for the mothership. Really
proud of the work that the team did. Um
uh but I found that where I personally
had the most fun was building new
businesses. So, for example, um I ran by
do self-driving car team still doing
really well in China today. Um I ran by
um kind of a a smart speaker team like
the you know like um Alexa or or or Hey
Siri or whatever China still doing
really well today and I felt that um
while I could make money for the
mothership where I was actually really
engaged was building new businesses and
then I asked myself while doing it
within the context of a big company was
fine. Um I felt that maybe if I start
something else uh uh which turned out to
be the venture studio AI fund to build
businesses you know more from zero to
one rather than the concept is a big
company maybe I could make that work
even better so that's why I stepped away
from BU um to start deep learning.ai AI
because I want to continue to do a lot
more work in education to empower people
to build with AI and then additionally
to run AI fund venture studio where we
build startups. Uh
>> do you think China is now ahead?
>> I think AI is multifaceted and China is
ahead of the US in some places like open
source open way models and the US is
ahead of China in some places like
proprietary models. So you don't regret
your choice.
>> I think uh I had a great time doing some
work with teams in China. Um I'm having
a great time doing work with DAI, AI
fun, uh more recent AI aspire uh landing
AI. So I'm I I uh yeah know I'm happy
with very I'm very happy what I'm doing.
No regrets.
>> From pure scale perspective, doesn't
China have more energy? Uh China has a
lot going for it. Um the US also has a
lot going for it. I feel like uh uh you
know as imperfect as this union is um I
love America and the Western world and
uh even though some years it feels like
democracy works better and it works
worse in some years that feels like an
important pillar of um how we do things
in America. uh and uh I think there's a
lot of good work to be done still in
America and I'm excited
>> what we can expect from your company.
>> I want to empower everyone to build with
AI. Um we've been very focused on
helping developers get access to the
latest tools, continue to work hard on
that. And in addition to really
supporting AI developers, build their
careers, grow their careers, um I
actually want to broaden what we do to
empower developers and everyone else to
build with AI. So that's DI's focus. Um
and one of the reasons I stepped away
from BYU from actually I found both when
I was leading a team at Google and at
BYU. Um there's certain businesses that
make sense to build within Google that
makes sense to build within BYU. But
there are other things like, you know,
tariff compliance. Why would an internet
company care about that? So AI fund
builds lots of different startup
businesses. And so I'm excited about
just the diversity of stuff we get to
work on at AI fund. Um, and then AI
aspire, which is relatively new effort,
is uh my friends Kirsty Tan and I are
working um on uh AI advisory for large
enterprises. So it turns out that if we
want to really move the needle on AI
adoption, um developers are important,
individual consumers are important, and
we just got to get large businesses
there. So we're spending quite a bit of
time, uh uh in in partnership with Bane
and others. Um uh Bane is a fantastic
team, really so smart, privileged
working with Kristoff and his team over
there. but um to help advise large
businesses on how to drive real value
and how to really win with AI.
>> You've decided to be also I think AI
communicator. What was the reason?
>> Yeah, you're touching on the very core
values kind of thing. Um I'll tell you
how I prioritize what I do. I think
there are two things um that I
prioritize most most highly that I
really believe in. I think it's a good
use of my time. um one is um things that
make humanity more powerful. That's why
I became a researcher, you know, became
a professor at Stanford because I think
research advancing the state of the art
that makes humanity more powerful by
inventing new technologies. So I believe
in making humanity more powerful. The
second thing I deeply believe in is um
to help others realize their dreams. Uh
it's important that is help others
realize their dreams, not help others
realize my dreams. is as important as
distinction and all through my life I
felt that if we can give others tools
and skills then it puts them in a better
position to realize their dreams which
is why um education you know through
divi
uh uh has always uh been something I
found very motivating
>> is something that you regret when you
look at your story if you could you know
turn back time would you do this and
change something
>> well there's so many things you would
have done differently so frankly
Um, I don't know. I I' I've been
fortunate to have made a few good
decisions. Uh, but I've made so many bad
ones as well. I think there's so many
things I, you know, like, uh, should I
have hired that particular person or,
uh, you know, done that particular
project or should I have, uh, work
harder on that project rather than given
up. I I think
>> if you could find answer for one
question about reality what it would be.
>> I wish I I I wish I understood what is
the nature of intelligence?
>> You mean consciousness?
How matter is transformed into
consciousness inside the human brain or
something more?
>> Actually not consciousness. Um
I think consciousness is important
philosophical question but I don't know
what is consciousness. So philosophers
talk about consciousness in the sense
that is this notion of being self-aware.
But it turns out that you don't actually
know if I'm conscious and I don't
actually know if you are. Right? How do
you know if in in philosophy there's
this uh uh concept that maybe I'm just a
zombie and I'm not actually conscious
but I'm just moving my hands, moving my
mouth and pretending to be conscious. So
because you don't have access to my
inner experience and vice versa, we
don't actually know if anyone else is
conscious but I think out of politeness
we pretend everyone else is conscious
much as we think we are. So because
consciousness is not measurable to me
that makes it a philosophical rather
than a scientific question. And while
philosophy is important I am I gravitate
more to the scientific questions and and
to me the nature of intelligence is um
what on earth are the mechanisms that
allow the human brain or maybe other
biological brains to demonstrate this
huge range of intelligent behaviors that
we see? like how on earth does this
work? Um
by the way something that is not widely
known before starting the Google brain
team um one thing I did was hang out a
lot with my neuroscientist friends and
you know I was reading huge piles of
neuroscience papers and concluded you
know lots of respect to my neuroscience
friends but neuroscience has no idea how
the brain works frankly pretty much so I
gave up on neuroscience as a path to
building intelligence but understanding
how does intelligence actually work what
is the nature of intelligence
I'm pretty sure it's not a transformer
network with scaling laws. I'm pretty
sure it needs more than that. But I wish
I
>> what is the nature of reasoning also?
>> Yeah. Yeah. I I I think reasoning is a
subset of intelligence. Uh but
understanding how reasoning actually
works would also be fascinating.
>> Why in the western world are we so
unhappy?
I really disagree with his
characterization of the western
weather's loss for these reasons. Um I
don't look for happiness in the water I
drink but I drink water and it keeps me
healthier. I think it mistake to look
for happiness in AI but AI really holds
people. that holds us whole help holds
us build things give us skills and so AI
is important but there's a mistake to
look for happiness in AI in the same way
that it would be a mistake to look for
happiness um uh in in I think happiness
comes much more from within than from
the artifacts that we build but it is
that process of striving to help others
um I hope a lot of people and find
happiness in that. And I hope I can help
others through my work in AI. Um, and I
find a lot of joy in
>> uh being able to do work that could be
useful to others.
>> Thank you very much for the time.
>> Thank you.