[0:00] - It's a very intense time in the field.
[0:02] We obviously want all
[0:03] of the brilliant things
these AI systems can do,
[0:05] come up with new cures for
diseases, new energy sources,
[0:07] incredible things for humanity.
[0:09] That's the promise of AI.
[0:11] But also, there are worries
[0:12] if the first AI systems are built
[0:13] with the wrong value systems
or they're built unsafely,
[0:16] that could be also very bad.
[0:18] - Wired sat down with Demis Hassabis,
[0:20] who's the CEO of Google
DeepMind, which is the engine
[0:23] of the company's artificial intelligence.
[0:25] He's a Nobel Prize
winner and also a knight.
[0:27] We discussed AGI, the future of work,
[0:30] and how Google plans to
compete in the age of AI.
[0:33] This is "The Big Interview."
[0:35] [upbeat music]
[0:42] Well, welcome to "The
Big Interview," Demis.
[0:43] - Thank you, thanks for having me.
[0:44] - So let's start talking
about AGI a little here.
[0:48] Now, you founded DeepMind with the idea
[0:51] that you would solve intelligence
and then use intelligence
[0:56] to solve everything else.
[0:57] And I think it was like a 20-year mission.
[0:59] We're like 15 years into
it, and you're on track?
[1:02] - I feel like, yeah,
[1:02] we're pretty much dead on track, actually,
[1:04] is what would be our estimate.
[1:06] - That means five years away
[1:08] from what I guess people will call AGI.
[1:11] - Yeah, I think in the
next five to 10 years,
[1:13] that would be maybe 50% chance
[1:16] that we'll have what we
are defined as AGI, yes.
[1:18] - Well, some of your peers are saying,
[1:20] "Two years, three years,"
[1:22] and others say a little more,
but that's really close,
[1:25] that's really soon.
[1:27] How do we know that we're that close?
[1:30] - There's a bit of a debate
going on in the moment
[1:32] in the field about definitions of AGI,
[1:34] and then obviously, of
course, dependent on that.
[1:36] There's different predictions
for when it will happen.
[1:39] We've been pretty consistent
from the very beginning.
[1:41] And actually, Shane Legg,
[1:42] one of my co-founders
and our chief scientist,
[1:44] you know, he helped define
the term AGI back in, I think,
[1:47] early 2001 type of timeframe.
[1:50] And we've always thought
about it as system
[1:52] that has the ability to exhibit,
[1:54] sort of all the cognitive
capabilities we have as humans.
[1:58] And the reason that's important,
[1:59] the reference to the human mind,
[2:01] is the human mind is the
only existence proof we have.
[2:04] Maybe in the universe, the
general intelligence is possible.
[2:07] So if you want to claim sort
of general intelligence, AGI,
[2:10] then you need to show that it generalizes
[2:13] to all these domains.
[2:14] - Is when everything's filled in,
[2:15] all the check marks are
filled in, then we have it-
[2:20] - Yes, so I think there are
missing capabilities right now.
[2:23] You know, that all of us
[2:24] who have used the latest
sort of LLMs and chatbots,
[2:27] will know very well, like on reasoning,
[2:29] on planning, on memory.
[2:30] I don't think today's
systems can invent, you know,
[2:33] do true invention,
[2:35] you know, true creativity,
[2:36] hypothesize new scientific theories.
[2:39] They're extremely useful,
they're impressive,
[2:41] but they have holes.
[2:43] And actually, one of the
main reasons I don't think
[2:45] we are at AGI yet is
[2:47] because of the consistency of responses.
[2:50] You know, in some domains,
[2:52] we have systems that can do
International Math Olympiad,
[2:55] math problems to gold medal standard-
[2:57] - Sure.
- With our AlphaFold system.
[2:59] But on the other hand,
[3:00] these systems sometimes still
trip up on high school maths
[3:03] or even counting the number
of letters in a word.
[3:05] - Yeah.
- So that to me is not
[3:07] what you would expect.
[3:08] That level of sort of difference
[3:10] in performance across the
board is not consistent enough,
[3:14] and therefore shows
[3:15] that these systems are not
fully generalizing yet.
[3:17] - But when we get it,
[3:18] is it then like a phase
shift that, you know,
[3:21] then all of a sudden things are different,
[3:24] all the check marks are checked?
[3:26] - Yeah.
- You know,
[3:27] and we have a thing
that can do everything.
[3:28] - Mm-hmm.
[3:29] - Are we then power in a new world?
[3:30] - I think, you know, that again,
[3:32] that is debated, and it's not clear to me
[3:34] whether it's gonna be more
[3:35] of a kind of incremental
transition versus a step function.
[3:41] My guess is, it looks
like it's gonna be more
[3:43] of an incremental shift.
[3:44] Even if you had a system like
that, the physical world,
[3:47] still operates with the physical laws,
[3:50] you know, factories,
robots, these other things.
[3:53] So it'll take a while for the
effects of that, you know,
[3:56] this sort of digital
intelligence, if you like,
[3:58] to really impact, I think, a
lot of the real world things.
[4:02] Maybe another decade plus,
[4:04] but there's other theories on that too,
[4:05] where it could come faster.
[4:06] - Yeah, Eric Schmidt, who I
think used to work at Google,
[4:10] has said that, "It's almost
like a binary thing."
[4:12] He says, "If China,
for instance, gets AGI,
[4:16] then we're cooked."
[4:17] Because if someone gets
it like 10 minutes,
[4:19] before the next guy, then
you can never catch up.
[4:24] You know, because then
it'll maintain bigger,
[4:26] bigger leads there.
[4:27] You don't buy that, I guess.
[4:29] - I think it's an unknown.
[4:30] It's one of the many unknowns,
[4:31] which is that, you know,
[4:32] that's sometimes called
the hard takeoff scenario,
[4:34] where the idea there is
that these AGI systems,
[4:37] they're able to self-improve,
[4:39] maybe code themselves
future versus themselves,
[4:41] that maybe they're extremely
fast at doing that.
[4:43] So what would be a slight lead,
[4:46] let's say, you know, a few days,
[4:48] could suddenly become a
chasm if that was true.
[4:51] But there are many other
ways it could go too,
[4:53] where it's more incremental.
[4:54] Some of these self-improvement
things are not able
[4:57] to kind of accelerate in that way,
[5:00] then being around the same time,
[5:03] would not make much difference.
[5:05] But it's important, I mean,
[5:06] these issues are the geopolitical issues.
[5:08] I think the systems that are being built,
[5:10] they'll have some imprint of the values
[5:13] and the kind of norms of the
designers and the culture
[5:16] that they were embedded in.
- [Steven] Mm-hmm.
[5:18] - So, you know, I think it is important,
[5:20] these kinds of international questions.
[5:22] - So when you build AI at Google,
[5:26] you know, you have that in mind.
[5:28] Do you feel competitive imperative
to, in case that's true,
[5:32] "Oh my God, we better be first?"
[5:34] - It's a very intense time
at the moment in the field
[5:37] as everyone knows.
[5:38] There's so many resources going
into it, lots of pressures,
[5:41] lots of things that need to be researched.
[5:43] And there's sort of
lots of different types
[5:45] of pressures going on.
[5:46] We obviously want all
of the brilliant things
[5:48] that these AI systems can do.
[5:50] You know, I think eventually,
[5:51] we'll be able to advance
medicine and science with it,
[5:54] like we've done with AlphaFold,
[5:56] come up with new cures for
diseases, new energy sources,
[5:58] incredible things for humanity,
that's the promise of AI.
[6:02] But also there are worries
both in terms of, you know,
[6:05] if the first AI systems are built
[6:07] with the wrong value systems
or they're built unsafely,
[6:10] that could be also very bad.
[6:11] And, you know, there
are at least two risks
[6:14] that I worry a lot about.
[6:15] One is, bad actors in
whether it's individuals
[6:17] or rogue nations repurposing
general purpose AI technology
[6:21] for harmful lens.
[6:22] And then the second one is, obviously,
[6:24] the technical risk of AI itself.
[6:26] As it gets more and more powerful,
[6:27] more and more agentic,
[6:28] can we make sure the
guardrails are safe around it?
[6:32] They can't be circumvented.
[6:33] And that interacts with
this idea of, you know,
[6:36] what are the first systems that are built
[6:38] by humanity gonna be like?
[6:40] There's commercial imperative-
[6:42] - [Steven] Right.
- There's national imperative,
[6:43] and there's a safety aspect to worry
[6:46] about who's in the lead and
where those projects are.
[6:50] - A few years ago, the
companies were saying,
[6:53] "Please, regulate us.
[6:54] We need regulation."
- Mm-hmm, mm-hmm.
[6:55] - And now, in the US at least,
[6:57] the current administration
seems less interested
[7:00] in putting regulations on
AI than accelerating it
[7:04] so we can beat the Chinese.
[7:07] Are you still asking for regulation?
[7:09] Do you think that that's
a miss on our part?
[7:11] - I think, you know,
[7:13] and I've been consistent in this,
[7:14] I think there are these other
geopolitical sort of overlays
[7:19] that have to be taken into account,
[7:21] and the world's a very different place
[7:22] to how it was five years
ago in many dimensions.
[7:25] But there's also, you know,
[7:26] I think the idea of smart regulation
[7:29] that makes sense around these
increasingly powerful systems,
[7:32] I think is gonna be important.
[7:33] I continue to believe that.
[7:35] I think though, and I've
been certain on this as well,
[7:37] it sort of needs to be international,
[7:38] which looks hard at the moment
[7:40] in the way the world is working,
[7:41] because these systems, you know,
[7:43] they're gonna affect everyone,
[7:45] and they're digital systems.
- Yeah.
[7:47] - So, you know, if you sort
of restrict it in one area,
[7:51] that doesn't really help
[7:52] in terms of the overall safety
[7:53] of these systems getting
built for the world
[7:57] and as a society.
- [Steven] Yeah.
[7:58] - So that's the bigger problem, I think,
[8:00] is some kind of international
cooperation or collaboration,
[8:03] I think, is what's required.
[8:05] And then smart regulation,
nimble regulation
[8:07] that moves as the knowledge
[8:09] about the research
becomes better and better.
[8:13] - Would it ever reach a point
for you where you would feel,
[8:15] "Man, we're not putting the guardrails in.
[8:18] You know, we're competing,
that we really have to stop,
[8:21] or you can't get involved in that?"
[8:24] - I think a lot of the
leaders of the main labs,
[8:27] at least the western labs,
[8:29] you know, there's a small number of them
[8:31] and we do all know each other
[8:33] and talk to each other regularly.
[8:34] And a lot of the lead researchers do.
[8:35] The problem is, is that it's not clear
[8:39] we have the right definitions
to agree when that point is.
[8:42] Like, today's systems,
[8:43] although they're impressive
as we discussed earlier,
[8:46] they're also very flawed.
[8:47] And I don't think today's systems,
[8:49] are posing any sort of existential risk.
[8:52] - Mm-hmm.
- So it's still theoretical,
[8:55] but the problem is that a lot of unknowns,
[8:56] we don't know how fast those will come,
[8:58] and we don't know how risky they will be.
[9:00] But in my view, when there
are so many unknowns,
[9:03] then I'm optimistic we'll overcome them.
[9:06] At least technically,
[9:07] I think the geopolitical
questions could be actually,
[9:09] end up being trickier, given
enough time and enough care
[9:12] and thoughtfulness, you know,
[9:13] sort of using the scientific method
[9:15] as we approach this AGI point.
[9:18] - That makes perfect sense.
[9:20] But on the other hand, if
that timeframe is there,
[9:23] we just don't have much time, you know?
[9:25] - No, we don't.
[9:26] We don't have much time.
[9:27] I mean, we're increasingly
putting resources into security
[9:31] and things like cyber,
[9:34] and also research into controllability
[9:37] and understanding of these systems,
[9:38] sometimes called mechanistic
interpretability.
[9:40] You know, there's a lot of
different sub-branches of AI.
[9:43] - Yeah, that's right.
[9:44] I wanna get to interpretability.
[9:45] - Yeah, that are being invested in,
[9:46] and I think even more needs to happen.
[9:48] And then at the same time,
[9:50] we need to also have societal debates more
[9:53] about institutional building.
[9:55] How do we want governance to work?
[9:57] How are we gonna get
international agreement,
[9:59] at least on some basic principles,
[10:01] around how these systems
are used and deployed
[10:04] and also built?
[10:05] - What about the effect on
work on the marketplace?
[10:09] - Yeah.
- You know,
[10:10] how much do you feel that AI is going
[10:13] to change people's jobs,
[10:15] you know, the way jobs are
distributed in the workforce?
[10:18] - I don't think we've seen,
[10:19] my view is if you talk to economists,
[10:20] they feel like there's
not much has changed yet.
[10:23] You know, people are
finding these tools useful,
[10:25] certainly in certain domains-
[10:26] - [Steven] Yeah.
- Like, things like AlphaFold,
[10:27] many, many scientists are using
it to accelerate their work.
[10:30] So it seems to be additive at the moment.
[10:32] We'll see what happens over
the next five, 10 years.
[10:34] I think there's gonna be a lot of change
[10:37] with the jobs world, but
I think as in the past,
[10:40] what generally tends to
happen is new jobs are created
[10:44] that are actually better,
[10:45] that utilize these tools
or new technologies,
[10:47] what happened with the internet,
what happened with mobile?
[10:49] We'll see if it's different this time.
[10:51] - Yeah.
[10:52] - Obviously everyone
always thinks this new one,
[10:53] will be different.
[10:53] And it may be, it will be,
[10:55] but I think for the next few years,
[10:57] it's most likely to be, you know,
[10:59] we'll have these incredible tools
[11:00] that supercharge our productivity,
[11:03] make us really useful for creative tools,
[11:07] and actually almost make
us a little bit superhuman
[11:09] in some ways in what we're
able to produce individually.
[11:13] So I think there's gonna
be a kind of golden era,
[11:16] over the next period of
what we're able to do.
[11:19] - Well, if AGI can do
everything humans can do,
[11:22] then it would seem that they
could do the new jobs too.
[11:24] - That's the next question
about like, what AGI brings.
[11:28] But, you know, even if you
have those capabilities,
[11:30] there's a lot of things I
think we won't want to do
[11:32] with a machine.
[11:34] You know, I sometimes give this example
[11:36] of doctors and nurses.
[11:38] You know, maybe a doctor
[11:39] and what the doctor
does and the diagnosis,
[11:41] you know, one could imagine
that being helped by AI tool
[11:44] or even having an AI kind of doctor.
[11:47] On the other hand, like nursing,
[11:49] you know, I don't think you'd
want a robot to do that.
[11:51] I think there's something
[11:52] about the human empathy
aspect of that and the care,
[11:55] and so on, that's particularly humanistic.
[11:59] I think there's lots of examples like that
[12:01] but it's gonna be a
different world for sure.
[12:04] - If you would talk to a graduate now,
[12:07] what advice would you give
[12:09] to keep working-
- Yeah.
[12:11] - Through the course
[12:12] of a lifetime-
- Yeah.
[12:14] - You know, in the age of AGI?
[12:16] - My view is, currently,
[12:18] and of course, this is
changing all the time
[12:20] with the technology developing.
[12:22] But right now, you know,
[12:24] if you think of the next
five, 10 years as being,
[12:27] the most productive people
might be 10X more productive
[12:30] if they are native with these tools.
[12:33] So I think kids today, students today,
[12:36] my encouragement would be immerse yourself
[12:39] in these new systems, understand them.
[12:41] So I think it's still important
[12:43] to study STEM and
programming and other things,
[12:45] so that you understand how they're built,
[12:47] maybe you can modify them yourself
[12:49] on top of the models that are available.
[12:50] There's lots of great open
source models and so on.
[12:53] And then become, you know,
[12:55] incredible at things like
fine-tuning, system prompting,
[12:59] you know, system instructions,
[13:01] all of these additional
things that anyone can do.
[13:03] And really know how to get
the most out of those tools,
[13:07] and do it for your
research work, programming,
[13:10] and things that you are
doing on your course.
[13:12] And then come out of that being incredible
[13:14] at utilizing those new tools
[13:17] for whatever it is you're going to do.
[13:18] - Let's look a little beyond
the five and 10-year range.
[13:21] Tell me what you envision
when you look at our future
[13:26] in 20 years, in 30 years,
if this comes about,
[13:30] what's the world like
when AGI is everywhere?
[13:33] - Well, if everything goes well,
[13:34] then we should be in an era of what I like
[13:37] to call sort of radical abundance.
[13:40] So, you know, AGI solves
some of these key,
[13:43] what I sometimes call root node problems
[13:45] in the world facing society.
[13:46] So a good one, examples
would be curing diseases,
[13:50] much healthier, longer lifespans,
[13:52] finding new energy sources,
[13:54] you know, whether that's optimal batteries
[13:56] and better room temperature,
superconductors, fusion.
[14:01] And then if that all happens,
[14:04] then we know it should be a kind of era
[14:06] of maximum human flourishing
where we travel to the stars
[14:10] and colonize the galaxy.
[14:14] You know, I think the
beginning of that will happen
[14:16] in the next 20, 30 years if
the next period goes well.
[14:20] - I'm a little skeptical of that.
[14:22] I think we have an
unbelievable abundance now,
[14:25] but we don't distribute it,
[14:27] you know, fairly.
- Yeah.
[14:28] - I think that we kind of know
[14:29] how to fix climate change, right?
[14:31] We don't need a AGI to
tell us how to do it,
[14:33] yet we're not doing it.
- I agree with that.
[14:36] I think we being as a species,
[14:38] a society not good at collaborating,
[14:41] and I think climate is a good example.
[14:42] But I think we are still operating,
[14:45] humans are still operating
in a zero-sum game mentality.
[14:48] Because actually, the
earth is quite finite,
[14:50] relative to the amount
of people there are now
[14:53] in our cities.
[14:54] And I mean, this is why
our natural habitats,
[14:56] are being destroyed,
[14:58] and it's affecting
wildlife and the climate
[15:01] and everything.
- [Steven] Yeah.
[15:02] - And it's also partly
'cause people are not willing
[15:04] to accept, we do now
to figure out climate.
[15:08] But it would require
people to make sacrifices.
[15:10] - Yeah.
- And people don't want to.
[15:12] But this radical abundance
would be different.
[15:15] We would be in a finally, like,
[15:17] it would feel like a non-zero-sum game.
[15:20] - How will we get [indistinct] to that?
[15:21] Like, you talk about diseases-
[15:22] - Well, I gave you an example.
- We have vaccines,
[15:24] and now some people think
we shouldn't use it.
[15:26] - Let me give you a very simple example.
[15:28] - Sure.
- Water access.
[15:29] This is gonna be a huge issue
in the next 10, 20 years.
[15:31] It's already an issue.
[15:32] Countries in different, you know,
[15:34] poorer parts of the world,
dryer parts of the world,
[15:35] also obviously compounded
by climate change.
[15:38] - [Steven] Yeah.
[15:38] - We have a solution to water access.
[15:40] It's desalination, it's easy.
[15:42] There's plenty of sea water.
- Yeah.
[15:43] - Almost all countries have a coastline.
[15:45] But the problem is, it's salty water,
[15:47] but desalination only very rich countries.
[15:50] Some countries do do
that, use desalination
[15:52] as a solution to their
fresh water problem,
[15:54] but it costs a lot of energy.
- Mm-hmm.
[15:55] - But if energy was essentially zero,
[15:57] there was renewable free
clean energy, right?
[16:01] Like fusion, suddenly, you
solve the water access problem.
[16:04] Water is, who controls a river
[16:06] or what you do with that does not,
[16:08] it becomes much less
important than it is today.
[16:11] I think things like water access,
[16:13] you know, if you run forward 20 years,
[16:15] and there isn't a solution
like that, could lead
[16:17] to all sorts of conflicts,
[16:18] probably that's the way it's trending-
[16:19] - Mm-hmm, right.
- Especially if you include
[16:21] further climate change.
[16:22] - So-
- And there's many,
[16:23] many examples like that.
[16:23] You could create rocket fuel easily-
[16:25] - Mm-hmm.
- Because you just separate
[16:27] that from seawater, hydrogen and oxygen.
[16:29] It's just energy again.
[16:30] - So you feel that these problems
get solved by AGI, by AI,
[16:37] then we're going to,
our outlook will change,
[16:41] and we will be-
- That's what I hope.
[16:44] Yes, that's what I hope.
[16:45] But that's still a secondary part.
[16:47] So the AGI will give us the
radical abundance capability,
[16:50] technically, like the water access.
[16:52] - Yeah.
- I then hope,
[16:53] and this is where I think we
need some great philosophers
[16:55] or social scientists to be involved.
[16:58] That should hopefully shift our mindset
[17:01] as a society to non-zero-sum.
[17:04] You know, there's still the issue
[17:05] of do you divide even the
radical abundance fairly, right?
[17:08] Of course, that's what should happen.
[17:10] But I think there's much more likely,
[17:11] once people start
feeling and understanding
[17:13] that there is this almost
limitless supply of raw materials
[17:18] and energy and things like that.
[17:19] - Do you think that
driving this innovation
[17:23] by profit-making companies
is the right way to go?
[17:26] We're most likely to reach
[17:27] that optimistic high point through that?
[17:29] - I think it's the current capitalism
[17:31] or, you know, is the current
[17:32] or the western sort of
democratic kind of systems,
[17:37] have so far been proven
[17:39] to be sort of the best
drivers of progress.
[17:41] - Mm-hmm.
- So I think that's true.
[17:43] My view is that once you get
[17:44] to that sort of stage of
radical abundance and post-AGI,
[17:48] I think economics starts changing,
[17:51] even the notion of value and money.
[17:53] And so again, I think we need,
[17:55] I'm not sure why economists
are not working harder on this
[17:57] if maybe they don't believe
it's that close, right?
[18:00] But if they really did that,
like the AGI scientists do,
[18:04] then I think there's a lot
[18:06] of economic new economic
theory that's required.
[18:08] - You know, one final thing,
[18:10] I actually agree with you
that this is so significant
[18:14] and is gonna have a huge impact.
[18:16] But when I write about it,
[18:18] I always get a lot of response from people
[18:21] who are really angry already
about artificial intelligence
[18:26] and what's happening.
[18:28] Have you tasted that?
[18:29] Have you gotten that pushback
and anger by a lot of people?
[18:34] It's almost like the
industrial revolution people-
[18:36] - Yeah.
- Fighting back.
[18:37] - I mean, I think that anytime there's,
[18:39] I haven't personally seen a lot of that,
[18:41] but obviously, I've read
and heard a lot about,
[18:42] and it's very understandable.
[18:44] That's all that's happened many times.
[18:45] As you say, industrial revolution,
[18:47] when there's big change,
[18:48] a big revolution.
- [Steven] Yeah.
[18:49] - And I think this will be at least
[18:50] as big as the industrial
revolution, probably a lot bigger.
[18:52] That's surprising, there's unknowns,
[18:54] it's scary, things will change.
[18:56] But on the other hand,
[18:57] when I talk to people about the passion,
[18:59] the why I'm building AI-
- Mm-hmm.
[19:00] - Which is to advance science
[19:01] and medicine-
- Right.
[19:02] - And understanding of
the world around us.
[19:04] And then I explain to people, you know,
[19:06] and I've demonstrated, it's not just talk.
[19:08] Here's AlphaFold, you know,
[19:09] Nobel Prize winning breakthrough,
[19:11] can help with medicine and drug discovery.
[19:13] Obviously, we're doing
this with isomorphic now
[19:14] to extend it into drug discovery,
[19:16] and we can cure terrible diseases
[19:18] that might be afflicting your family.
[19:20] Suddenly, people are like,
[19:21] "Well, of course, we need that."
[19:23] - Right.
- It'll be immoral not
[19:24] to have that if that's within our grasp.
[19:26] And the same with climate and energy.
[19:29] - Yeah.
- You know,
[19:30] many of the big societal problems,
[19:31] it's not like you know,
[19:34] we know, we've talked about,
[19:35] there's many big challenges
facing society today.
[19:38] And I often say I would be
very worried about our future
[19:41] if I didn't know something
[19:43] as revolutionary as AI
was coming down the line
[19:45] to help with those other challenges.
[19:47] Of course, it's also a
challenge itself, right?
[19:50] But at least, it's one of these challenges
[19:52] that can actually help with
the others if we get it right.
[19:54] - Well, I hope your optimism
holds out and is justified.
[19:59] Thank you so much.
- And I'll do my best.
[20:00] Thank you.
[20:01] [upbeat music]