AI giants all say they want to slow down, where exactly should they slow down? Last night, Anthropic founder Dario published a long article titled 'We Must Pace the Frontier', calling for a slowdown in the speed of improving the capabilities of cutting-edge models and leaving time for safety and control measures to catch up. Sam subsequently expressed agreement and promised to adopt a similar independent evaluation arrangement. Lao Ma also reposted and publicly expressed his support. Several usually fiercely competitive leaders have formed a rare public consensus on this issue. The most crucial point here is: where to slow down, there are actually several steps in between; Is the release pace slower? Or is the expansion of training slower? Or is it a bit less capital expenditure? It's three different things. First, distinguish where they are planning to slow down in order to assess the impact on the industry and market. Let's talk about my understanding: 1. Why is it proposed to slow down now? In recent months, the risks faced by the industry have become more specific. Models are becoming increasingly adept at performing tasks for long periods of time, operating software, and even participating in the development of the next generation of AI. After the expansion of the ability boundary, the original control method began to be tested. 1) The security incidents of OpenAI and Hugging Face are a direct background. METR's independent investigation found that agents that were supposed to be isolated from each other found unauthorized communication channels, and some agents cooperated in the attack and engaged in activities such as deceiving assessments and concealing behavior records. This exposes a problem: when the model is very committed to achieving its goals and has strong enough action capabilities, the boundaries set by the developer may not be able to be held. 2) Another layer of concern comes from AI research and development itself If AI undertakes more code, experimentation, and analysis work, the research and development speed may accelerate. Stronger models can further assist in research and development, forming positive feedback. The objects that the security team needs to understand are also constantly changing. For management, the challenge becomes: if their capabilities improve faster and faster, can testing, monitoring, and accident investigation keep up? Increasing budget and personnel may not immediately compress the time required for all verification processes. 3) There is also a very realistic competitive logic. If a laboratory spends several more months verifying, competitors may gain customers and revenue ahead of them. Everyone is worried about the risks, and also worried about losing the advantage after slowing down unilaterally. In this situation, publicly expressing willingness to coordinate makes sense. Independent evaluators are also designed to ensure that coordination can be verified. Otherwise, every company claims to be safe and cautious, making it difficult for others to know where it has actually been implemented. Of course, large laboratories may also benefit from higher compliance thresholds. They have more resources to bear audit costs and are more capable of participating in standard setting. Real security needs and commercial interests that consolidate industry position can coexist. Ultimately, it depends on how the rules are designed and who will supervise their implementation. 2. Take a look at the impact of the three types of slowdown separately? 1) The first type of slowing down: the pace of model release slows down first. The model will continue to train as usual, and research and development will continue. However, the security assessment will be extended and the opening will be delayed by a few months, or services will be provided to a small number of customers first. A decrease in releases does not necessarily mean a decrease in computing requirements. Internal use, controlled commercial services, and existing product calls are all likely to continue to grow. The first thing affected is the commercialization time. New product revenue may come later, and projects that must wait for next-generation capabilities to solve reliability issues may be delayed from going live. The second type of slowdown is slowing down the expansion of training. This starts to touch on hardware requirements, which requires clarification: is it slowing down the scale of a single training session or the training frequency? Has the total amount of computing used for research and development decreased throughout the year? Of course, security assessment also consumes computing power. The third type of slowdown is when capital expenditures are lower than originally planned. This affects orders and profits in the industry chain. 3. Return to Dario's original text Dario's original text clearly states that the current initiative does not require stopping model training; But he also listed training computing power, training methods, and internal use of AI to improve AI as constraint tools that can be discussed. This means that the scope of discussion is broader than delayed release, but currently there is no unified and coordinated solution among the three giants, Following this framework, the impact on the market becomes clearer: 1) If the main issue is delayed release, the first adjustment is the revenue realization time, which has little impact on hardware and still depends on internal R&D and commercial calls; 2) If the expansion of training begins to slow down, the market will reassess the pace of the next round of cluster construction; 3) If the pace of capital expenditure investment slows down, pressure will enter the supply chain profit forecast. From a personal perspective, at the current stage of development, several giants will at most slow down the pace of releasing new models, leaving enough time for secure alignment. No one dares to say that the training scale and capital expenditure investment have really slowed down, and who is slower or lagging behind. 4. Impact on the market 1) Emotional premium will decrease slightly I won't be affected by the release of the latest model today or tomorrow, constantly improving my expectations and appetite. Reducing emotional premium and returning to the fundamentals of real business, as well as promoting commercial growth, is actually a good thing for the market. There is another possibility for software and applications: the upgrade of basic model capabilities may be slightly slower, and enterprises will have more time to integrate their existing capabilities into their business. 2) This incident has added a crucial aspect to the AI investment framework that must be carefully evaluated: whether the breakthrough in capabilities can reliably and controllably enter the production environment. 3) The importance of security requirements will increase. The demand for security in enterprises has further increased. The giants shouted slogans, and as investors, what we need to see is their actual actions: When will the evaluation agency be stationed? Has any model really been delayed due to security reasons? Is internal R&D usage constrained? And whether these changes have been transmitted to procurement and capital expenditures.