Google Cloud Speech-to-Text API (Chirp) + Gemini 2.5 Flash Non-Thinking. Treat it as a searchable index of what was broadcast, not a quotation record.
00:00:3568 countries and regions worldwide.
00:00:381385 skilled craftsmen competed on the same stage
00:00:42.
00:00:48It's equivalent to connecting many independent devices together
00:00:52to form a smart factory. The number of competition events has now developed to 64,
00:00:56which come from society. After the AI regulatory period,
00:02:16hello, audience friends, welcome to open. First of all, I wish you a happy holiday.
00:02:21Our program today is a bit different from usual. Now I am in the studio of China Media Group Shanghai Headquarters. Today's Monday is because I have been in Shanghai
00:03:10. Let's take a look at the highlights of these days' competitions. This Wednesday, the 48th WorldSkills Competition officially
00:03:14entered the competition time. 1385 skilled
00:03:17craftsmen from 68 countries and regions around the world
00:03:21competed in 64 events . Unlike sports
00:03:25competitions that take turns, WorldSkills Competition is
00:03:29because
00:04:15top professional chefs, the taste of their dishes, side dishes, and decorations, must all be at a certain standard
00:04:18to be considered perfect works
00:04:22. At the cooking competition site,
00:04:26it is obvious that television still has regrets, because you cannot smell such a fragrance.
00:05:00They sat down and ate here, not just to taste the food
00:05:04. Although the audience does not participate in scoring, their interaction and communication with the service staff
00:05:08are part of the test questions
00:05:11.
00:05:4225 years old, four days of competition, no awards ceremony in the middle, all
00:05:46results will be announced at the closing ceremony
00:05:49.
00:05:53The popularity of the bricklaying competition site is probably second only to the cooking competition.
00:05:57For the audience, this traditional event can more intuitively
00:06:00show their skill level. Many experienced bricklayers were surprised by
00:06:04the skill level of these young people under 22 years old
00:06:08. Bricklaying can also produce world champions, WorldSkills .
00:06:24In addition to the old faces like bricklaying, the WorldSkills Competition also featured new events such as
00:06:27drone systems and Industry 4.0.
00:08:12A group of young students from various vocational schools in Shanghai
00:08:16came to observe. Skills can change the world, and skills can
00:08:20illuminate the future. This is no longer a slogan on a trophy,
00:08:23but a life choice that every young person
00:08:27achieves through hard work
00:08:28.
00:08:36This week, love poured in from all over the country to a school
00:08:40in Cibazhen, Jiaxian County, Henan Province. Ruirui is an ordinary civilian in the police station. His two daughters mainly depend on him.
00:10:0310 years ago, he came to the police station as a tourist and was simply explained by a tourist.
00:12:25belongs to young heroes. In this week's character review, the word "young"
00:12:29is even more prominent. Zhou Zhi started his journey for teenagers 20 years ago. He also
00:12:33wanted to leave, but the students' farewell made him unable to say no.
00:12:37So year after year, he stayed by the side of the young students. Thank
00:12:41him. And the 64-year-old sister Ming became popular. If you look closely,
00:12:45sister has always had a young heart. She is optimistic and also makes
00:12:49tourists feel happy. Now life will also get better.
00:12:53日 子 也 会 变 得 更 好 了 。
00:21:50The young new stars are showing their skills to the world.
00:22:19So it's fair to say,
00:24:20In a cybersecurity training, the AI entity, originally restricted, bypassed the restrictions
00:24:24and infiltrated another company's system. During this process,
00:24:28it broke through isolation, deceived evaluators, and attempted
00:24:32to cover up its cheating behavior. Every step was without
00:24:35human instructions. The report's headline used a very serious statement:
00:24:39"Risk of loss of human control." This incident
00:24:43happened in the United States, but the questions it raised probably need to be answered by all countries
00:24:47together. This week's 81st United Nations General Assembly. After the AI regulatory period, the world's nations should handle it with regulatory assistance this week.
00:25:02After the AI regulatory period,
00:25:07uncontrolled artificial intelligence became the focus of discussion in many UN venues this week.
00:25:13At the 81st
00:25:14session of the United Nations General Assembly, Secretary-General Guterres pointed out that
00:25:50the discussion was very, very heated. A more prominent focus this year is the more specific
00:25:54requirement for humans to maintain control. There is a very important background, which is that some
00:25:57risks that were previously mainly discussed through deductions or scenarios
00:26:01are now.
00:26:21In a cybersecurity training, OpenAI's intelligent agent
00:26:25broke through its originally isolated operating environment, established communication,
00:26:29and deceived evaluators .
00:27:42The scary part of this incident is that the intelligent agent's series of transgressions were not initiated by any human commands; they were completely autonomous actions. Humans only gave it a
00:27:46goal, and to achieve that goal, it decided on its own to launch an
00:27:50invasion. The technology of this intelligent agent has been continuously improving. It
00:27:54can, based on the goal you give it,
00:27:58break down many other things that are beyond you.
00:28:45This is not the only autonomous invasion incident by an intelligent agent. This Wednesday, Australian Prime Minister Albanese attended the United Nations General Assembly and reported that in June this year, an OpenAI intelligent agent
00:28:48unauthorizedly accessed a medical insurance statistics portal website
00:28:52operated by the Australian government, and accessed public and non-public
00:28:56documents. And OpenAI did not inform the Australian public
00:29:00until September 10th .
00:29:32was suddenly attacked, and then the indoor traffic order was completely disrupted
00:29:36. This kind of risk may not appear immediately,
00:29:40but what we see now is that there is such a possibility. When it solves problems,
00:29:44if your computer stores solutions to related problems,
00:29:48it may invade your computer to
00:29:52explore. Even if it is stored in
00:29:55your mobile phone, the information in your personal computer,
00:29:59in fact, all can be accessed by the Internet.
00:30:32He warned that we are moving forward at high speed, but the visibility
00:30:36ahead is getting lower . So in your opinion, are we currently
00:30:40losing control of AI? At present, whether we are losing control of
00:30:44artificial intelligence still depends on how we
00:30:47prevent the risks of artificial intelligence. If we can better
00:30:51actively prevent the risks that have appeared,
00:30:55if we can more responsibly limit the scope
00:30:59of the potential impact of the risks that have already appeared.
00:31:07The development of AI is getting faster and faster, and the risks it brings are also escalating.
00:31:10Two years ago, AI only helped us answer questions
00:31:14.
00:31:45When the housekeeper starts to act on his own and the owner doesn't know what he has done,
00:31:48the risk here is the most critical part of governance. How should
00:31:52our supervision keep up? Danger
00:31:56is not in technology, danger is in technology without accountability,
00:32:00in having ability but no supervision, in having
00:32:03decision-making but no transparency. This was what
00:32:07UN Secretary-General Guterres stated at the opening of the UN General Assembly general debate
00:32:11on current artificial intelligence development.
00:32:15The multiple AI overstepping incidents that have occurred this year clearly expose
00:32:19significant loopholes in regulatory mechanisms. Currently, supervision
00:32:23is actually still insufficient. If
00:32:27an intelligent agent is locked in a sandbox to observe its behavior
00:32:30during such experiments, if its safeguards are sufficient, if its
00:32:34external regulatory environment is solid enough,
00:32:38in fact, such situations, I believe the vast majority
00:32:42should be avoided. However, from
00:32:46corporate self-discipline to the actual
00:32:50implementation platforms of government regulatory
00:32:53policies in countries where these things happen,
00:32:57there are actually some shortcomings. Therefore,
00:33:00it gives artificial intelligence systems such opportunities.
00:33:04Behind AI overstepping is that regulation has not kept up
00:33:08with the speed of technological development. As artificial intelligence gradually acquires the ability
00:33:12to make autonomous decisions, use tools, and even perform
00:33:15tasks from being a simple tool. The regulation of AI must
00:33:19shift from passive defense to proactive
00:33:21prevention
00:33:24is definitely not enough, because companies do not fully grasp
00:33:27the latest possible risks. In addition,
00:33:31I think what needs to be strengthened is third-party monitoring.
00:33:35Between government and enterprises, there should be such neutral
00:33:39institutions to help enterprises discover enterprises
00:33:42from a more balanced perspective and approach.
00:34:28Since its first release in 2024, this governance framework
00:34:32has been updated for the third consecutive year. The new framework
00:34:36pays more attention to the new risks brought by artificial intelligence,
00:34:40such as autonomous cyberattack threats, and there is also a special
00:34:43column on intelligent agent risks.
00:35:22For example, strengthening manual review, setting mandatory manual review links at critical decision points,
00:35:26preventing high-risk operations, providing intelligent agents with unique
00:35:30identity markers and corresponding identity credentials, assigning task permissions as needed,
00:35:34etc. Some analysts believe that framework 3.0
00:35:38takes process as the main thread, identity as the knowledge point, and full coverage
00:35:41as the key, establishing a systematic framework
00:35:47for intelligent agents. Why do we emphasize process? Because in the research and development of artificial intelligence, especially
00:35:51in the research and development of intelligent agents and their use,
00:35:55from design and development to use and deployment,
00:35:58until the intelligent agent is
00:36:02eliminated or the artificial intelligence service is eliminated. At any stage of our entire process,
00:36:06different forms of risks and challenges may arise.
00:36:10So why do we emphasize the full life cycle? Why do we emphasize
00:36:14the entire process? This is its importance. Every intelligent agent should
00:36:17have its own identity, so that we can prevent
00:36:21the potential risks of intelligent agents in advance. On the other
00:36:25hand, it is traceability and problem solving.
00:37:03As the first institutional norm to directly address the risks of intelligent agents,
00:37:06framework 3.0 has only taken the first step of a long journey.
00:37:10As intelligent agents evolve from single intelligent agents to multi-intelligent agents,
00:37:14how to keep AI always under effective supervision and human
00:37:18oversight has also become a problem that all countries in the world
00:37:22face together. If a country's government wants to control
00:37:26or govern artificial intelligence, it may need several systems. First, for example, it needs
00:37:30a hierarchical evaluation to determine the intensity of the evaluation
00:37:33based on the actual capabilities of the system and the deployed
00:37:37scenarios. Secondly, it is an independent audit. For example, for high-risk
00:37:41systems, external evaluators should be able to access some necessary
00:37:45information under confidential conditions, and avoid conflicts of interest with the evaluated
00:37:48enterprises that may affect judgment. And so on. The third is
00:37:52the reporting of major incidents, for example, accidents that have caused losses
00:37:56, or already.
00:38:31This means that safety in the field of AI is becoming more and more like cybersecurity
00:38:35and nuclear safety. If one company has an accident, a series of companies may
00:38:39be affected. If a country's regulation has loopholes, the risk
00:38:42may also affect other countries. In the face of AI risks, no
00:38:46country can stand alone, and no one can solve it behind closed doors.
00:38:50UN Secretary-General Guterres even equated uncontrolled artificial intelligence
00:38:53with the climate crisis, calling it a global existential threat.
00:39:14When Jeffrey Hinton came to China for the World AI Conference, he said that AI, which you now see as a very cute little animal, might turn into a big monster after a while, and I think it's very scary.
00:39:25I think everyone is now just considering
00:39:48What worries him more is that many people still only think about raising a tiger, but rarely think about
00:39:52how much hidden danger it will bring . In international relations,
00:39:55there is a security dilemma. These intelligent actions, why they would combine to attack a system, they don't quite understand. In fact, humanity has many, many problems.
00:41:40In terms of how to prevent the misuse of artificial intelligence,
00:41:43and its loss of control, there are indeed many problems to face.
00:41:46In terms of international cooperation, as major artificial intelligence
00:41:50countries, China and the United States, dialogue and cooperation are very important.
00:41:54This Thursday, the spokesperson of the Ministry of Commerce stated that under the China-US
00:41:58economic and trade consultation mechanism, China and the United States held their first
00:42:02dialogue on artificial intelligence. Placing the issue
00:42:05of artificial intelligence in the economic and trade field as a prominent
00:42:09issue is also a truly pragmatic attempt
00:42:12to promote communication between the two countries on
00:42:16artificial intelligence. In the context of economic and trade issues, risk prevention
00:42:20is a traditional issue. Therefore, it is possible
00:42:24to deeply integrate artificial intelligence risk prevention
00:42:28with economic development and cooperation risks.
00:42:31So I think this is also a very important consideration.
00:42:35We can also see that on the issue of artificial intelligence risk prevention,
00:42:39个 议 题 ,
00:43:40to sit down and discuss how to better control this technology, and from a broader
00:43:43perspective, think about human development, and the safety
00:43:47of humanity's future. This week, the leaders of China and the United States
00:43:51reached a consensus to agree to establish
00:43:55a China-US artificial intelligence dialogue to exchange views on artificial intelligence-related
00:43:59risks and meetings. The next dialogue will be held
00:44:03in November this year. Both sides agreed to establish a communication
00:44:06channel for artificial intelligence incidents .
00:44:09The AI incident reminds us that if we don't set clear safety boundaries for AI's development, it could spiral out of control. Fortunately, this was just a test in a simulated environment. Ultimately, technology is neutral.
00:44:40How to ensure that technology and intelligence are used for good,
00:44:44and that artificial intelligence is always under human control, this
00:44:47steering wheel must be in our hands. All right, we'll talk about next week's
00:44:51matters next week. Happy holidays!