Abstract:
The debate surrounding whether cutting-edge artificial intelligence should proactively slow down its development has heated up rapidly in recent days. Anthropic CEO Dario Amodei publicly called on the entire AI industry to slow down the pace of advanced model capabilities, and OpenAI CEO Sam Altman and xAI CEO Elon Musk later expressed support.
However, David Sacks, an American technology investor and chairman of the President's Science and Technology Advisory Committee, sharply criticized this initiative, believing that if OpenAI and Anthropic really believe that the current speed of AI development is dangerous enough to need to apply the brakes, then they can slow down research and development on their own, and should not impose industry coordination, antitrust exemptions, or government approval mechanisms as additional conditions.
Amodei previously published an article warning that cutting-edge AI capabilities are rapidly approaching a stage that may create major security risks. He believes that if the model continues to develop at the current speed and the security mechanism cannot keep up, in the next 6 to 12 months, AI agent clusters may even have the ability to control the Internet on a large scale and be used to launch cyber attacks or carry out other serious harm.
Amodei proposed a set of plans called "slowing down cutting-edge AI," which includes allowing independent third-party security assessment agencies to go deep into AI companies to continuously check model security with authority close to company employees; let major AI laboratories coordinate security standards to avoid companies constantly lowering security thresholds due to competitive pressure; while promoting closer cooperation between the United States and its allies, and ultimately discussing AI security issues with other countries, including China.
This initiative quickly gained support from OpenAI. Altman said that he agreed with Amodei’s view that the AI industry needs to slow down the development of cutting-edge models, and said that this issue has become an important topic of recent discussion within OpenAI. Musk also publicly expressed his approval. In a short period of time, several of the world's most important AI companies have reached a rare consensus on the issue of "appropriate slowdown."
But Sacks believes that there is an obvious logical problem here.
According to his point of view, if Amodei and Altman really believe that the next generation AI model has serious risks that must be slowed down, then the two companies do not need to wait for government approval, nor do they need to convince competitors, nor do they need a new legal framework. As the most advanced AI development companies currently, they can directly reduce their R&D and deployment speed.
Sacks even suggested that if these companies really believe that some AI systems are dangerous enough to threaten humans, then the "easiest way" is not to continue developing or releasing these systems. He believes that instead of asking the government to establish a new regulatory system, it is better to let companies that truly master cutting-edge AI research and development capabilities take action on their own.
What Sacks really objects to is the part of Amodei's plan that involves industry coordination and regulatory arrangements. He warned that if competitors such as OpenAI and Anthropic obtain antitrust exemptions through the government and then jointly decide at what speed the entire industry should develop AI, the arrangement may actually evolve into a "cartel."
In Sacks’ view, allowing several of the most leading AI companies to jointly set the development pace of cutting-edge AI may produce results that are completely different from security goals. Existing giants can not only limit competition through this mechanism, but may also use the regulatory system to raise the threshold for latecomers to enter the market, thereby further consolidating their leading position.
This is also the core of Sacks’s criticism: He is not simply opposed to slowing down the AI industry, but believes that if companies are really willing to slow down, they should do it themselves instead of asking the government to provide some kind of legal protection in exchange.
Sacks also opposes the establishment of a prior approval system similar to the "AI Food and Drug Administration". In his view, there is no need for government agencies to approve these products one by one before advanced AI models are officially released. Rather than establishing a new prior licensing mechanism, it is better to rely on the existing product liability system to allow AI companies to bear legal responsibility for their products and their consequences.
For example, if an AI model is used to launch a serious cyber attack, or directly causes major damage due to its own flaws, then the model development company should bear corresponding product liability. Sacks believes that this potential legal and economic consequence itself can become an important constraint for AI companies to control risks.
He also believes that the capital market also has a constraining effect. If a certain AI company's products frequently experience safety issues, unpredictable behavior, or major accidents, its corporate reputation, customer trust, and market valuation may be directly affected. Therefore, there is no need to establish a special approval system for the AI industry that is completely independent of the existing legal system.
At the same time, Sacks also questioned the independence of some AI safety assessment mechanisms. He specifically mentioned the security assessment agency used by Anthropic and believed that if there is an investment relationship, personnel relationship or other interest relationship between the assessment agency and the enterprise being assessed, then whether this assessment is truly independent deserves further discussion.
This controversy occurred at a very sensitive time. In the past few months, concerns about security risks within the AI industry have significantly escalated. Anthropic recently disclosed a series of incidents related to AI system abuse, network attack capabilities, and security protection. Researchers within the company have also publicly expressed concerns about the speed of AI development.
Among them, an Anthropic researcher publicly warned when he recently resigned that the AI industry is advancing technologies that may be extremely risky at an excessively fast pace. Other security researchers within Anthropic have also expressed serious concerns about the risk of AI getting out of control in the future.
Therefore, Amodei’s public call for slowdown is not an isolated policy statement, but the result of a further escalation of the entire AI security debate.
However, even if companies are really willing to slow down, the actual operation is far more complicated than it appears. OpenAI, Anthropic, Google DeepMind, xAI, and other AI companies are all competing fiercely. If any company unilaterally slows down R&D, it may worry that its competitors will take the opportunity to speed up and gain advantages in model capabilities, market share, and capital markets.
This actually forms a typical "prisoner's dilemma": if all AI companies slow down at the same time, the entire industry may gain more time to improve security measures; but if only one company slows down while competitors continue to advance, then the company that actively slows down may suffer commercial and technical losses.
This is why OpenAI has already begun to pay attention to whether industry coordination will violate U.S. antitrust laws. According to reports, OpenAI has sought clear opinions from U.S. lawmakers on whether industry coordination to slow down AI development is legal. Because if multiple competing companies jointly decide to slow down the development of cutting-edge models, this coordination itself may trigger antitrust legal issues.
As a result, a very thorny question arises: In order to improve AI security, should companies allow coordination between competitors that is usually restricted by antitrust laws?
Supporters believe that if each AI company knows that slowing down alone will lead to competitive disadvantages, then only through some kind of industry-level coordination can all companies truly apply the brakes at the same time.
Sacks believes that this may be a reason for large AI companies to use security issues to seek special regulatory treatment. He worries that once the government allows several leading companies to coordinate through antitrust exemptions, this mechanism may eventually evolve from "security coordination" into a tool to restrict market competition.
At the same time, there is a clear difference between the US government's attitude towards the slowdown of AI and the claims of these technology companies. U.S. President Trump recently stated publicly that although he believes that AI can set up certain "guardrails", he does not want the United States to slow down its development due to excessive concerns about risks. He emphasized that the United States is currently leading China in the field of AI, and "whoever wins AI will win the future." Therefore, the United States cannot easily give up its technological leadership.
This also directly links AI security issues with Sino-US technology competition.
Amodei himself also admitted that if the United States unilaterally significantly slows down the development of AI while China and other countries continue to accelerate, the United States may lose its technological advantage. Therefore, what he proposes is not to ask American companies to stop AI research and development indefinitely, but to hope that major countries and companies can form some kind of coordination mechanism to buy time for security measures while maintaining competitive advantages.
The problem is that there is currently no mechanism to ensure that all countries will slow down simultaneously.
This has also become one of the most important reasons for those who oppose the large-scale slowdown in the industry. If AI has become an important part of national strategic competition, it is difficult for either party to believe that competitors will voluntarily stop and wait for themselves to improve their security systems.
Therefore, the debate between Sacks, OpenAI, and Anthropic has actually gone far beyond the simple technical issue of "how fast AI should develop" and involves a more complex industrial policy choice: whether companies should decide the development speed of cutting-edge AI, or whether the government should participate in formulating unified rules; whether safety coordination should be regarded as legitimate industry cooperation, or whether it may evolve into a cartel that restricts competition; whether AI accidents should be avoided by ex-ante regulation, or whether companies should bear the post-event costs through product liability.
At present, it seems that Sacks is not opposed to OpenAI and Anthropic taking the initiative to slow down. If the two companies really believe that the risks of cutting-edge AI are serious enough to slow down their research and development, he even made it clear that he can support them in doing so.
But his conditions were also clear: If you think you should slow down, slow down yourself, and don’t trade it for government approval, antitrust exemptions, or special regulatory protections.
The real outcome of this debate may ultimately depend on what OpenAI and Anthropic do next. If they really take the initiative to slow down the development of cutting-edge models without receiving special legal treatment, then "slowing down AI" is more likely to be regarded as a real security action; if what is ultimately formed is a regulatory system in which a few leading companies jointly formulate rules while restricting the entry of latecomers, then Sacks' concerns about "regulatory capture" and "cartels" will receive more attention.
As AI model capabilities continue to improve rapidly, the debate over "acceleration or deceleration" will obviously not end soon. It is likely to become one of the core contradictions in U.S. and global AI policy in the next few years.
Comments