WEBVTT 00:00:00.000 --> 00:00:03.917 Good afternoon. I'd like to thank France for convening this meeting and the 00:00:04.000 --> 00:00:06.717 members of the council for inviting me to speak. 00:00:06.800 --> 00:00:11.157 AI is advancing very quickly. Four years ago, it could barely write a line of 00:00:11.240 --> 00:00:15.077 code or solve a high school math problem. Today, it writes most of the 00:00:15.160 --> 00:00:19.317 code at Anthropic and solves world-famous unsolved math problems. The 00:00:19.400 --> 00:00:21.357 trajectory only needs to continue for a 00:00:21.440 --> 00:00:24.277 tiny period longer, 1 or 2 years, maybe 00:00:24.360 --> 00:00:26.357 less, to reach what I've called a 00:00:26.440 --> 00:00:29.037 country of geniuses in the data center. 00:00:29.120 --> 00:00:33.757 This rapid pace of progress will produce incredible benefits for the world. Just 00:00:33.840 --> 00:00:37.317 today, Anthropic announced the Claude-led discovery of a new molecular 00:00:37.400 --> 00:00:41.797 machine. It's a preliminary finding, but it may constitute a new gene-editing 00:00:41.880 --> 00:00:46.157 mechanism that could have applications in in gene therapy. Claude led the 00:00:46.240 --> 00:00:48.197 literature review, theorizing, and 00:00:48.280 --> 00:00:53.357 experimental design, and humans ran and verified the experiments in the lab. 00:00:53.440 --> 00:00:57.637 We believe this type of application will accelerate rapidly in the years ahead 00:00:57.720 --> 00:01:01.597 and that we can make profound progress in the near future on diseases that have 00:01:01.680 --> 00:01:04.317 plagued humanity for millennia. 00:01:04.400 --> 00:01:06.557 But these remarkable capabilities also 00:01:06.640 --> 00:01:09.157 come with risks that must be managed. 00:01:09.240 --> 00:01:13.917 I see two main categories of risk. The first is misuse of AI models. For 00:01:14.000 --> 00:01:18.557 example, misuse by bioterrorists to create biological weapons. The second 00:01:18.640 --> 00:01:22.597 risk is loss of control, model capabilities accelerating beyond 00:01:22.680 --> 00:01:25.037 developers' ability to control them. 00:01:25.120 --> 00:01:29.517 These questions are global in scope and concern all of humanity. They are grave 00:01:29.600 --> 00:01:33.437 enough that I believe they are important questions for this council, not just for 00:01:33.520 --> 00:01:38.357 economic or ICT ministries. If managed poorly, I even believe that 00:01:38.440 --> 00:01:41.477 AI could be a risk to humanity as a whole. 00:01:41.560 --> 00:01:45.157 At Anthropic, we believe we have the best track record in the industry for 00:01:45.240 --> 00:01:49.117 preventing these risks, both operationally and in our research. 00:01:49.200 --> 00:01:53.077 We test every model for dangerous capabilities in areas like cybersecurity 00:01:53.160 --> 00:01:58.077 and biology. We put strong safeguards on our most capable models, and we publish 00:01:58.160 --> 00:02:02.477 our work in these areas. We have pioneered research in areas like 00:02:02.560 --> 00:02:06.157 mechanistic interpretability and alignment science that improve our 00:02:06.240 --> 00:02:09.317 ability to steer and control our models. 00:02:09.400 --> 00:02:13.837 We also have unparalleled transparency practices releasing risk reports that 00:02:13.920 --> 00:02:15.917 run to hundreds of pages. But we 00:02:16.000 --> 00:02:18.477 understand that as AI models become more 00:02:18.560 --> 00:02:21.117 powerful, even more stringent standards 00:02:21.200 --> 00:02:23.477 of safety will be required and we are 00:02:23.560 --> 00:02:26.837 always aiming to improve our safety and 00:02:26.920 --> 00:02:31.437 alignment practices, our abilities, our ability to prevent misuse and we are 00:02:31.520 --> 00:02:33.717 always striving to do better. We will 00:02:33.800 --> 00:02:36.597 slow down as much as necessary in order 00:02:36.680 --> 00:02:41.877 to make sure that every successive AI technology that we release is actually 00:02:41.960 --> 00:02:45.477 safe. But regardless of what we do, managing 00:02:45.560 --> 00:02:49.357 these risks is ultimately bigger than any one company and it has an 00:02:49.440 --> 00:02:52.237 industry-wide and global scale. 00:02:52.320 --> 00:02:54.317 On September 12th, I called for a 00:02:54.400 --> 00:02:56.197 collaborative effort to pace the 00:02:56.280 --> 00:02:59.277 frontier, not to stop AI progress, but 00:02:59.360 --> 00:03:01.477 to work together to make sure the 00:03:01.560 --> 00:03:03.717 industry and the world can proceed at a 00:03:03.800 --> 00:03:08.717 rate where we can be sure that safety keeps up with capability for everyone in 00:03:08.800 --> 00:03:12.917 the industry. This call had three parts. First, we 00:03:13.000 --> 00:03:15.357 committed to embed external evaluators 00:03:15.440 --> 00:03:17.397 inside Anthropic with employee-like 00:03:17.480 --> 00:03:20.477 access, similar to a food inspector and 00:03:20.560 --> 00:03:22.557 and we recommended that other companies 00:03:22.640 --> 00:03:28.237 across the world do the same. Some have already agreed uh uh to adopt this 00:03:28.320 --> 00:03:31.877 measure. Second, we called for cooperation across 00:03:31.960 --> 00:03:36.437 the industry to set standards and modulate the pace of progress. And 00:03:36.520 --> 00:03:41.197 finally, we called for global cooperation across the world between 00:03:41.280 --> 00:03:46.077 governments to set international standards. It is this last area that 00:03:46.160 --> 00:03:50.757 most squarely falls within this council's remit. I want to offer three 00:03:50.840 --> 00:03:55.117 ideas in this First, we should begin with narrow 00:03:55.200 --> 00:03:56.877 agreements that every member can 00:03:56.960 --> 00:04:00.437 support, such as a ban on using AI to 00:04:00.520 --> 00:04:03.197 make biological weapons or permitting 00:04:03.280 --> 00:04:06.437 your AI technologies to be used 00:04:06.520 --> 00:04:09.317 AI technologies to be used to make 00:04:09.400 --> 00:04:12.357 biological weapons. Second, we should 00:04:12.440 --> 00:04:17.037 build evaluation and verification systems that keep pace with AI 00:04:17.120 --> 00:04:21.037 development so that states can have visibility into frontier model 00:04:21.120 --> 00:04:23.997 capability and can verify each other's 00:04:24.080 --> 00:04:26.797 commitments. Third, we should establish 00:04:26.880 --> 00:04:29.357 common global standards for testing AI 00:04:29.440 --> 00:04:32.477 models for loss of control risks and 00:04:32.560 --> 00:04:35.637 misuse risks and a notification system 00:04:35.720 --> 00:04:40.277 for AI incidents that are significant to global security. 00:04:40.360 --> 00:04:44.597 I believe that this is the most important global security issue facing 00:04:44.680 --> 00:04:49.477 the world today. For all the differences of the people in this room, we must put 00:04:49.560 --> 00:04:51.557 aside those differences in order to 00:04:51.640 --> 00:04:54.597 confront this global opportunity and 00:04:54.680 --> 00:04:59.517 global threat that is being presented to us at the same time. No leader, no 00:04:59.600 --> 00:05:02.597 company, and no nation can manage this alone. 00:05:02.680 --> 00:05:04.477 We commit to working with governments in 00:05:04.560 --> 00:05:07.157 this room on this urgent work. 00:05:07.240 --> 00:05:10.320 Thank you for your attention.