WEBVTT

00:00:01.880 --> 00:00:05.317
There are many things that AI
cannot and should not automate.

00:00:05.400 --> 00:00:08.797
AI should be built for people,
not simply to turn the crank crank of the

00:00:08.880 --> 00:00:12.117
machines faster for their own sake.

00:00:12.200 --> 00:00:15.197
Some of the things people working to build
AI have said are so dystopian that

00:00:15.280 --> 00:00:18.597
they sound like the plot
of bad science fiction movies.

00:00:18.680 --> 00:00:22.517
Anything this power is also
powerful is also intimidating.

00:00:22.600 --> 00:00:24.117
This moment is complicated as we feel

00:00:24.200 --> 00:00:26.197
both tremendous potential and very

00:00:26.280 --> 00:00:29.917
understandable anxiety at the same time.

00:00:30.000 --> 00:00:32.917
These models can help people discover
new knowledge, new medicines, better

00:00:33.000 --> 00:00:36.277
health care, stronger education,
and build more productive economies. They

00:00:36.360 --> 00:00:40.277
can be new tools for people to pursue
dreams and ambition that today feel out

00:00:40.360 --> 00:00:43.517
of reach. The complexity
of life has increased so

00:00:43.600 --> 00:00:46.077
dramatically for the average person.

00:00:46.160 --> 00:00:49.637
AI can be a defensive tool against
this complexity as much as it can be this

00:00:49.720 --> 00:00:53.997
tool of unprecedented
discovery and innovation.

00:00:54.080 --> 00:00:56.397
On the other hand, as AI systems become

00:00:56.480 --> 00:00:59.037
more capable and more autonomous, they

00:00:59.120 --> 00:01:01.157
could move faster than our institutions,

00:01:01.240 --> 00:01:03.997
concentrate power in two two few hands,

00:01:04.080 --> 00:01:09.077
or make decisions that people
no longer understand or control.

00:01:09.160 --> 00:01:12.437
That concern becomes especially
important as we approach systems that

00:01:12.520 --> 00:01:16.357
can improve themselves and future
versions of themselves, often called

00:01:16.440 --> 00:01:20.477
recursive self-improvement.
As the process of building AI becomes

00:01:20.560 --> 00:01:22.997
more automated, the pace of AI progress

00:01:23.080 --> 00:01:24.797
could accelerate rapidly.

00:01:24.880 --> 00:01:28.037
This moment calls for extreme care.

00:01:28.120 --> 00:01:30.597
I largely agree with Professor Bengio.

00:01:30.680 --> 00:01:33.677
Uh although I'll say two instead of three.
I believe there are two ways that

00:01:33.760 --> 00:01:37.397
AI progress could go very
badly that we have to avoid.

00:01:37.480 --> 00:01:41.997
First, we could lose
control of the future to AI.

00:01:42.080 --> 00:01:44.877
The risk is that it moves so fast
that people can no longer follow what's

00:01:44.960 --> 00:01:49.397
happening or intervene when needed.
This would obviously be terrible.

00:01:49.480 --> 00:01:52.677
The industry must not accept too much
technological risk just because the

00:01:52.760 --> 00:01:56.237
benefits are too great and that they
feel too important to just down.

00:01:56.320 --> 00:02:00.317
Beating companies in a race is not
a reason to make rash decisions, nor do we

00:02:00.400 --> 00:02:02.917
believe we are locked in a race where
we are unable to do that. We have

00:02:03.000 --> 00:02:06.517
unilaterally slowed down in the past. We
will do so in the future. We do not want

00:02:06.600 --> 00:02:09.037
to build systems. We think it's
important that everyone avoids building

00:02:09.120 --> 00:02:12.797
systems where we cannot make
the alignment, monitorability, and safety

00:02:12.880 --> 00:02:16.237
guarantees that we must.
We need to understand what these systems

00:02:16.320 --> 00:02:20.077
are doing and have strong evidence that
they will do what people intend even as

00:02:20.160 --> 00:02:25.037
they get very, very smart. Actually,
especially as they get very, very smart.

00:02:25.120 --> 00:02:28.997
It doesn't matter whether people put the
risk of catastrophe at 10% or 1% or 12%

00:02:29.080 --> 00:02:33.237
or 0.1%. None of these levels are
remotely acceptable. And we should not

00:02:33.320 --> 00:02:36.557
train models that we cannot make
an extremely strong case they will be able

00:02:36.640 --> 00:02:41.597
to keep under human control. So,
that's one way things can go wrong.

00:02:41.680 --> 00:02:44.757
In the other direction, these systems
could concentrate too much power in too

00:02:44.840 --> 00:02:48.677
few hands. No one person
or company or country

00:02:48.760 --> 00:02:52.837
should be able to use the most powerful
AI models to impose their worldview on

00:02:52.920 --> 00:02:56.237
everyone else. A company
or country that believes only

00:02:56.320 --> 00:02:59.797
it can be trusted with this technology
can use that belief to justify almost

00:02:59.880 --> 00:03:02.917
anything else.
We have to reject that logic even when

00:03:03.000 --> 00:03:05.637
it's convenient for any one actor.

00:03:05.720 --> 00:03:09.317
I believe there is a pragmatic
path through both of these.

00:03:09.400 --> 00:03:12.997
At OpenAI, we approach these
questions with a clear goal.

00:03:13.080 --> 00:03:16.757
AI should give people more power over
their own lives, more freedom to live

00:03:16.840 --> 00:03:20.677
lives the way they want, more ability
to solve problems, big or small, and to

00:03:20.760 --> 00:03:25.117
improve their communities. We don't want
to fall into the trap of blind optimism.

00:03:25.200 --> 00:03:28.157
We also don't want to fall into
the trap of doomerism, either.

00:03:28.240 --> 00:03:31.597
Getting pulled into either extreme will
not help us successfully figure out the

00:03:31.680 --> 00:03:34.757
challenges in front of us,
especially at a time when the world and

00:03:34.840 --> 00:03:39.757
its people face such enormous challenges
that this technology can help solve.

00:03:39.840 --> 00:03:43.357
So, we try to walk the middle
path between these two poles.

00:03:43.440 --> 00:03:45.597
For us, that means three principles.

00:03:45.680 --> 00:03:50.677
First, as AI becomes more capable,
people must remain at the center of AI

00:03:50.760 --> 00:03:53.877
decision-making. Alignment
is not an abstract research

00:03:53.960 --> 00:03:57.797
question. It is the work of ensuring
that these systems reliably remain under

00:03:57.880 --> 00:04:01.197
human control, reflect
human values, and help people

00:04:01.280 --> 00:04:03.597
guide the next stages of development.

00:04:03.680 --> 00:04:08.117
We are not trying to and must not
automate human judgment or human values.

00:04:08.200 --> 00:04:12.317
Second, the benefits of scientific
progress and economic growth must be by

00:04:12.400 --> 00:04:15.117
people and for people.
AI should help researchers make

00:04:15.200 --> 00:04:18.557
discoveries, doctors treat patients,
teachers teach students, entrepreneurs

00:04:18.640 --> 00:04:21.797
start companies, communities solve
problems that felt out of reach, and

00:04:21.880 --> 00:04:26.877
much more. The continuous story of human
progress is not the tools replace human

00:04:26.960 --> 00:04:30.597
agency, it is that the right tools make
people more capable, and I believe this

00:04:30.680 --> 00:04:33.797
is going to go much further
than we believe possible.

00:04:33.880 --> 00:04:37.557
Third, this technology must
empower people individually.

00:04:37.640 --> 00:04:40.837
A great future will come from people
realizing their potential and building

00:04:40.920 --> 00:04:45.197
value for themselves, their communities,
their countries, and the broader world.

00:04:45.280 --> 00:04:48.237
We might be with the builders, one
of the builders of this technology, but we

00:04:48.320 --> 00:04:49.997
are not the heroes of this story.

00:04:50.080 --> 00:04:51.597
Our role is to enable people all over

00:04:51.680 --> 00:04:57.477
the world, and to trust in the magic
of humanity's skill and faculties.

00:04:57.560 --> 00:05:01.717
We now have models capable of discovering
knowledge that we have never had before.

00:05:01.800 --> 00:05:04.277
Consider the progress we've seen in math.

00:05:04.360 --> 00:05:07.957
Three summers ago, our models
were okay at grade school math.

00:05:08.040 --> 00:05:11.837
Two summers ago, they were pretty
good at high school math competitions.

00:05:11.920 --> 00:05:15.437
Last summer, we reached gold level
a gold level medal in the most prestigious

00:05:15.520 --> 00:05:17.597
international math competition.

00:05:17.680 --> 00:05:20.797
And just a few weeks ago this summer,
one of our models solved one of the

00:05:20.880 --> 00:05:24.477
millennium prize problems,
the Navier-Stokes equations.

00:05:24.560 --> 00:05:27.717
These equations are used for aircraft
design, weather forecasting, and the

00:05:27.800 --> 00:05:30.597
study of blood flow.
We are not solving math problems for

00:05:30.680 --> 00:05:34.517
their own sake, but to empower people to
discover new knowledge, and to use it to

00:05:34.600 --> 00:05:38.317
improve health care, raise standards
of livings st- of living, and expand

00:05:38.400 --> 00:05:40.277
everyone's potential around the world.

00:05:40.360 --> 00:05:43.157
There remains so much to be discovered.

00:05:43.240 --> 00:05:47.357
But it's very important that companies
not substitute for the democratic process.

00:05:47.440 --> 00:05:50.997
If AI is to be democratic, the most
important decisions cannot be made by

00:05:51.080 --> 00:05:55.757
labs in San Francisco alone. They must
be shaped through democratic processes

00:05:55.840 --> 00:05:59.237
and by governments accountable
to the people that they serve.

00:05:59.320 --> 00:06:03.157
At the international level,
this will require cooperation.

00:06:03.240 --> 00:06:06.277
In our history, there have been times
where countries who compete and don't

00:06:06.360 --> 00:06:09.917
always like each other very much still
come together for shared interests and

00:06:10.000 --> 00:06:13.357
the collective good in the face
of a powerful new technology.

00:06:13.440 --> 00:06:16.917
We believe this must
be one of those times.

00:06:17.000 --> 00:06:20.237
We need a mechanism for complementary
national and international frontier AI

00:06:20.320 --> 00:06:23.677
standards. Standards
for measuring capabilities,

00:06:23.760 --> 00:06:27.717
assessing risks, determining whether
safeguards are sufficient, and

00:06:27.800 --> 00:06:32.357
preserving meaningful human oversight
as systems become more autonomous.

00:06:32.440 --> 00:06:35.397
We need common standards so
countries can compare evidence,

00:06:35.480 --> 00:06:39.157
verify compliance, and have a shared
language and understanding about what is

00:06:39.240 --> 00:06:43.237
happening. We need to accurate
and speedy incident reporting,

00:06:43.320 --> 00:06:46.917
classification and reporting protocols
so the world can learn from failures

00:06:47.000 --> 00:06:49.677
before they become catastrophes.

00:06:49.760 --> 00:06:52.757
And we need secure channels among
governments, critical infrastructure

00:06:52.840 --> 00:06:56.397
operators, and technical experts
to share emerging vulnerabilities and new

00:06:56.480 --> 00:06:59.837
threats. Obviously,
these standards should not

00:06:59.920 --> 00:07:03.677
lock in incumbents or favor
one business model over another.

00:07:03.760 --> 00:07:07.157
They must support open and closed model
developers, new entrants and established

00:07:07.240 --> 00:07:11.157
labs, national authority,
and international cooperation.

00:07:11.240 --> 00:07:14.077
Each government should decide how to
incorporate standards into its own legal

00:07:14.160 --> 00:07:18.237
system, but we will all be better off if
we can agree on what good evidence, good

00:07:18.320 --> 00:07:22.717
safeguards, and good oversight
look like on the global stage.

00:07:22.800 --> 00:07:26.557
We do not have all the answers,
and we're very open to any better ideas.

00:07:26.640 --> 00:07:29.397
We're ready to work with you all and try
our hardest to figure out the path

00:07:29.480 --> 00:07:33.397
forward. We are at a crossroads.

00:07:33.480 --> 00:07:37.717
One path leads to a world where power
is concentrated, decisions are opaque, and

00:07:37.800 --> 00:07:40.557
people feel that the most important
technology of their lifetime is

00:07:40.640 --> 00:07:44.917
something being done to them.
The other path leads to a world where AI

00:07:45.000 --> 00:07:49.317
works for people, because people,
through democratic institutions, help

00:07:49.400 --> 00:07:53.157
decide how it should develop and guide it.
Where scientific progress improves

00:07:53.240 --> 00:07:56.717
real lives, where more people can
participate in the future, and where

00:07:56.800 --> 00:08:00.117
this technology continues to give
individuals and communities greater

00:08:00.200 --> 00:08:02.357
power to live life as they choose.

00:08:02.440 --> 00:08:05.837
This is a future of realized potential,
excitement, and wonder, and it is the

00:08:05.920 --> 00:08:07.277
future we believe is worth building.

00:08:07.360 --> 00:08:10.160
Thank you very much.

00:08:10.880 --> 00:08:13.517
>> I thank Mr. Altman for his briefing.

00:08:13.600 --> 00:08:17.440
I now give the floor to Mr. Dario Amodei.

00:08:18.360 --> 00:08:22.317
>> Good afternoon. I'd like to thank
France for convening this meeting, and the

00:08:22.400 --> 00:08:26.640
members of the council
for inviting me to speak.