0:00
So, how do we build AI that [music]
0:03
doesn’t just get smarter, but gets safer
0:06
as it grows? What does this super
0:08
intelligence actually mean? I mean, I
0:10
know you’ve spent decades designing
0:12
architectures for safe super
0:15
intelligence, but you know, for the
0:16
everyday viewer, what does safe super
0:20
intelligence mean in practical
0:23
day-to-day terms?
0:25
Sure. Yeah. Uh so maybe it’s helpful to
0:27
have a little context just if we zoom
0:30
out a little bit and talk about sort of
0:32
the phases of AI. AI is very popular now
0:35
but everybody knows about it but um you
0:39
know 10 years ago it was mainly nerdy
0:40
researchers that were sort of you know
0:42
focused on this very briefly the field
0:44
started in 1956. That’s when the field
0:47
was named at a conference at Dartmouth.
0:49
And then for the first 30 years or so
0:51
until the mid1 1980s, it was all about
0:54
programming rules into a computer. You
0:56
would program the knowledge directly
0:58
into the computer. And that was the age
0:59
of symbolic AI. Then in the ’90s, and I
1:02
kind of like lucked out. I I did my
1:04
graduate work at Carnegie Melon right
1:06
when it was the transition from the
1:08
symbolic AI to the next phase, which was
1:11
machine learning. And the idea was it
1:14
was too difficult to program all of the
1:16
knowledge about the world into a
1:17
computer. just too much knowledge and
1:19
very difficult to get a practical
1:21
system. And so the idea was maybe the
1:23
computer could just learn the knowledge
1:25
on its own and that would be so much
1:27
easier. And so that was the idea behind
1:29
machine learning. And there were some uh
1:31
early pioneers Jeff Hinton uh Dr. Jeff
1:34
Hinton who won the Nobel Prize and
1:36
Turing award and is pretty well known as
1:38
the godfather of AI these days. He was
1:41
one of the early pioneers of the machine
1:43
learning approach that used neural
1:45
networks. And so the idea was you could
1:47
take a lot of data like a good chunk of
1:49
the information on the internet and if
1:51
you fed it into these data centers they
1:54
would just run algorithms and after
1:56
several weeks or maybe a month out would
1:58
come a GPT GPT5 or something that would
2:02
sort of just magically know all this
2:03
stuff and you didn’t have to program it
2:05
all in. And uh so that started in the
2:08
‘8s, but it actually took you know 20 30
2:11
years until computing power got fast
2:14
enough that it could actually learn
2:15
things quickly. So we had self-driving
2:18
cars at Carnegie Melon in 1986 but they
2:21
only went about 2 in an hour and that’s
2:24
because the computers were so slow.
2:26
Yeah. But then there’s been this huge
2:28
speed up. And so now those same
2:30
algorithms, the exact same car just with
2:32
the better processors could go like 200
2:35
mph. So it’s really been that speed up
2:38
of computing power that has enabled AI
2:40
to come into its own. And then kind of
2:42
the time when most people really started
2:45
paying attention to AI was Thanksgiving
2:47
3 years ago with the release of GPT,
2:50
ChatGpt. So November 2022
2:53
and that was kind of in my mind the
2:56
beginning of the third phase. So we went
2:58
from symbolic to machine learning to now
3:00
this age of generative AI and AI agents.
3:03
A lot of people are talking about AI
3:05
agents. And then from here I think
3:07
what’s going to happen is these agents
3:09
and these AI systems are going to become
3:11
smarter and smarter and at some point
3:13
they become smarter than humans in
3:16
pretty much every task and that’s what
3:18
super intelligence is. So super
3:20
intelligence hasn’t been invented at
3:22
least general super intelligence where
3:24
the AI can be smarter than you or I in
3:26
everything does not yet exist but there
3:28
is super intelligence in very specific
3:30
areas like playing chess or doing a
3:33
particular scientific problem folding
3:35
proteins or something already the AI
3:37
systems are better than the best humans
3:39
but to have that general super
3:40
intelligence that’s kind of the future
3:42
and those of us who’ve been working in
3:44
this field for a long time we’re super
3:46
excited about it on one hand and then on
3:48
the other hand there’s There’s always
3:49
this risk when you have this thing
3:50
that’s smarter than humans. What does
3:52
that mean for the humans? And so, how do
3:54
we design these systems more safely?
3:56
That’s kind of been my focus since about
3:58
2020\.
4:06
[music]
