WEBVTT

00:00:00.000 --> 00:00:03.517
I thank Mr. Altman for his briefing.

00:00:03.600 --> 00:00:07.480
I now give the floor to Mr. Dario Amodei.

00:00:08.360 --> 00:00:12.317
>> Good afternoon. I'd like to thank
France for convening this meeting and the

00:00:12.400 --> 00:00:15.157
members of the council
for inviting me to speak.

00:00:15.240 --> 00:00:19.197
AI is advancing very quickly.
Four years ago, it could barely write a

00:00:19.280 --> 00:00:23.517
line of code or solve a high school math
problem. Today, it writes most of the

00:00:23.600 --> 00:00:27.757
code at Anthropic and solves
world-famous unsolved math problems. The

00:00:27.840 --> 00:00:29.797
trajectory only needs to continue for a

00:00:29.880 --> 00:00:32.397
tiny period longer, one or two years,

00:00:32.480 --> 00:00:34.797
maybe less, to reach what I've called a

00:00:34.880 --> 00:00:37.477
country of geniuses in the data center.

00:00:37.560 --> 00:00:42.197
This rapid pace of progress will produce
incredible benefit for the world. Just

00:00:42.280 --> 00:00:45.757
today, Anthropic announced the Claude-led
discovery of a new molecular

00:00:45.840 --> 00:00:50.277
machine. It's a preliminary finding,
but it may constitute a new gene-editing

00:00:50.360 --> 00:00:54.597
mechanism that could have applications
in in gene therapy. Claude led the

00:00:54.680 --> 00:00:56.637
literature review, theorizing, and

00:00:56.720 --> 00:01:01.797
experimental design, and humans ran
and verified the experiments in the lab.

00:01:01.880 --> 00:01:06.077
We believe this type of application will
accelerate rapidly in the years ahead

00:01:06.160 --> 00:01:09.997
and that we can make profound progress
in the near future on diseases that have

00:01:10.080 --> 00:01:12.757
plagued humanity for millennia.

00:01:12.840 --> 00:01:14.957
But these remarkable capabilities also

00:01:15.040 --> 00:01:17.557
come with risks that must be managed.

00:01:17.640 --> 00:01:22.357
I see two main categories of risk.
The first is misuse of AI models. For

00:01:22.440 --> 00:01:26.557
example, misuse by bioterrorists
to create biological weapons.

00:01:26.640 --> 00:01:28.837
The second risk is loss of control.

00:01:28.920 --> 00:01:33.437
Model capabilities accelerating beyond
developers' ability to control them.

00:01:33.520 --> 00:01:37.997
These questions are global in scope and
concern all of humanity. They are grave

00:01:38.080 --> 00:01:41.917
enough that I believe they are important
questions for this council, not just for

00:01:42.000 --> 00:01:46.837
economic or ICT ministries.
If managed poorly, I even believe that

00:01:46.920 --> 00:01:49.917
AI could be a risk to humanity as a whole.

00:01:50.000 --> 00:01:53.597
At Anthropic, we believe we have
the best track record in the industry for

00:01:53.680 --> 00:01:57.557
preventing these risks,
both operationally and in our research.

00:01:57.640 --> 00:02:01.557
We test every model for dangerous
capabilities in areas like cybersecurity

00:02:01.640 --> 00:02:06.477
and biology. We put strong safeguards
on our most capable models, and we publish

00:02:06.560 --> 00:02:10.917
our work in these areas.
We have pioneered research in areas like

00:02:11.000 --> 00:02:14.597
mechanistic interpretability
and alignment science that improve our

00:02:14.680 --> 00:02:17.717
ability to steer and control our models.

00:02:17.800 --> 00:02:22.277
We also have unparalleled transparency
practices, releasing risk reports that

00:02:22.360 --> 00:02:24.317
run to hundreds of pages. But, we

00:02:24.400 --> 00:02:26.917
understand that as AI models become more

00:02:27.000 --> 00:02:29.557
powerful, even more stringent standards

00:02:29.640 --> 00:02:31.917
of safety will be required. And we are

00:02:32.000 --> 00:02:35.237
always aiming to improve our safety and

00:02:35.320 --> 00:02:39.877
alignment practices, our abilities
our ability to prevent misuse, and we are

00:02:39.960 --> 00:02:42.117
always striving to do better. We will

00:02:42.200 --> 00:02:44.997
slow down as much as necessary in order

00:02:45.080 --> 00:02:50.357
to make sure that every successive AI
technology that we release is actually

00:02:50.440 --> 00:02:53.957
safe. But, regardless
of what we do, managing

00:02:54.040 --> 00:02:57.757
these risks is ultimately bigger
than any one company. And it has an

00:02:57.840 --> 00:03:00.677
industry-wide and global scale.

00:03:00.760 --> 00:03:02.757
On September 12th, I called for a

00:03:02.840 --> 00:03:04.637
collaborative effort to pace the

00:03:04.720 --> 00:03:07.717
frontier, not to stop AI progress, but

00:03:07.800 --> 00:03:09.917
to work together to make sure the

00:03:10.000 --> 00:03:12.157
industry and the world can proceed at a

00:03:12.240 --> 00:03:17.157
rate where we can be sure that safety
keeps up with capability for everyone in

00:03:17.240 --> 00:03:21.397
the industry. This call
had three parts. First, we

00:03:21.480 --> 00:03:23.797
committed to embed external evaluators

00:03:23.880 --> 00:03:25.837
inside Anthropic with employee-like

00:03:25.920 --> 00:03:28.917
access, similar to a food inspector. And

00:03:29.000 --> 00:03:33.877
And we recommended that other companies
across the world do the same. Some have

00:03:33.960 --> 00:03:37.717
already agreed uh uh
to adopt this measure.

00:03:37.800 --> 00:03:40.317
Second, we called for cooperation across

00:03:40.400 --> 00:03:44.877
the industry to set standards
and modulate the pace of progress. And

00:03:44.960 --> 00:03:49.637
finally, we called for global
cooperation across the world between

00:03:49.720 --> 00:03:54.557
governments to set international
standards. It is this last area that

00:03:54.640 --> 00:03:59.197
most squarely falls within this
council's remit. I want to offer three

00:03:59.280 --> 00:04:03.557
ideas in this regard. First,
we should begin with narrow

00:04:03.640 --> 00:04:05.317
agreements that every member can

00:04:05.400 --> 00:04:08.877
support, such as a ban on using AI to

00:04:08.960 --> 00:04:11.637
make biological weapons or permitting

00:04:11.720 --> 00:04:14.757
your AI technologies to be used to make

00:04:14.840 --> 00:04:17.797
biological weapons. Second, we should

00:04:17.880 --> 00:04:22.477
build evaluation and verification
systems that keep pace with AI

00:04:22.560 --> 00:04:26.477
development so that states can
have visibility into frontier model

00:04:26.560 --> 00:04:30.677
capability and can verify
each other's commitments.

00:04:30.760 --> 00:04:33.117
Third, we should establish common global

00:04:33.200 --> 00:04:36.157
standards for testing AI models for loss

00:04:36.240 --> 00:04:39.517
of control risks and misuse risks, and a

00:04:39.600 --> 00:04:42.637
notification system for AI incidents

00:04:42.720 --> 00:04:45.677
that are significant to global security.

00:04:45.760 --> 00:04:50.037
I believe that this is the most
important global security issue facing

00:04:50.120 --> 00:04:54.917
the world today. For all the differences
of the people in this room, we must put

00:04:55.000 --> 00:04:56.997
aside those differences in order to

00:04:57.080 --> 00:05:00.037
confront this global opportunity and

00:05:00.120 --> 00:05:04.957
global threat that is being presented
to us at the same time. No leader, no

00:05:05.040 --> 00:05:08.037
company, and no nation
can manage this alone.

00:05:08.120 --> 00:05:09.917
We commit to working with governments in

00:05:10.000 --> 00:05:12.637
this room on this urgent work.

00:05:12.720 --> 00:05:15.760
Thank you for your attention.

00:05:16.200 --> 00:05:20.037
>> I thank Mr. Amodei for his briefing.
I now give the floor to Mr. Klemann de

00:05:20.120 --> 00:05:22.520
Lange.

00:05:23.600 --> 00:05:27.080
>> Mr. President, members