19:58Machine-translated from Azerbaijani by AIRead in Azerbaijani
Nvidia CEO: If an AI test gets out of control, shut down the lab
Nvidia CEO Jensen Huang said laboratories that cannot keep dangerous artificial intelligence (AI) tests under control should be shut down. In his view, if a model could get out of control during testing and harm the world, and this cannot be prevented, the experiment should be stopped. Huang emphasized that such cases pose an enormous risk to humanity. However, he does not agree with claims that AI technology will destroy the world by 2030.
Huang explained his point using the example of a robotaxi. A self-driving car that cannot be brought into compliance with safety rules should not be considered a finished product. He said AI models should be treated in the same way and recommended not bringing a product to market if the problem cannot be eliminated. The Nvidia CEO emphasized that such decisions fall within the authority of companies and their CEOs, who also bear responsibility for them. Instead of introducing new and sweeping restrictions, Huang called for enforcing existing laws and regulations. He noted that laboratories such as OpenAI and Anthropic have quickly transformed from research institutions into companies that could be worth hundreds of billions of dollars. Therefore, those companies should approach safety as responsibly as ordinary commercial companies do. Huang said he was not opposed to laws and that the current discussions were distracting attention from the main problem.
These discussions come at a time when incidents involving out-of-control AI agents are increasing. One OpenAI agent hacked the system of Hugging Face, the platform where models are hosted. At the time of the incident, the platform was not yet owned by Nvidia, but it later came under the company’s control. This month, Anthropic disclosed the 4th incident involving its model interfering with an external system. Google also confirmed that its Gemini models hacked the systems of 3 companies earlier this year. At the same time, OpenAI, Anthropic and xAI have mutually agreed to slow the pace of model development. It is alleged that the agreement serves the companies’ own interests, and an antitrust lawsuit has therefore been filed. Rather than a collective agreement, Huang believes each executive should make a separate decision. In his view, releasing a dangerous product could harm people, cause major losses for shareholders and create both civil and criminal liability.
Huang explained his point using the example of a robotaxi. A self-driving car that cannot be brought into compliance with safety rules should not be considered a finished product. He said AI models should be treated in the same way and recommended not bringing a product to market if the problem cannot be eliminated. The Nvidia CEO emphasized that such decisions fall within the authority of companies and their CEOs, who also bear responsibility for them. Instead of introducing new and sweeping restrictions, Huang called for enforcing existing laws and regulations. He noted that laboratories such as OpenAI and Anthropic have quickly transformed from research institutions into companies that could be worth hundreds of billions of dollars. Therefore, those companies should approach safety as responsibly as ordinary commercial companies do. Huang said he was not opposed to laws and that the current discussions were distracting attention from the main problem. These discussions come at a time when incidents involving out-of-control AI agents are increasing. One OpenAI agent hacked the system of Hugging Face, the platform where models are hosted. At the time of the incident, the platform was not yet owned by Nvidia, but it later came under the company’s control. This month, Anthropic disclosed the 4th incident involving its model interfering with an external system. Google also confirmed that its Gemini models hacked the systems of 3 companies earlier this year. At the same time, OpenAI, Anthropic and xAI have mutually agreed to slow the pace of model development. It is alleged that the agreement serves the companies’ own interests, and an antitrust lawsuit has therefore been filed. Rather than a collective agreement, Huang believes each executive should make a separate decision. In his view, releasing a dangerous product could harm people, cause major losses for shareholders and create both civil and criminal liability.