[TF Securities Global Tech] How Should We Interpret Anthropic and OpenAI’s Calls to Slow Down AI Development? Frontier model developers are calling for a slower pace of AI progress so that safety testing, operational monitoring and third-party verification can keep up. This could weigh on sentiment toward the AI sector and temper expectations for next-generation models in the near term. One theory circulating in the market is that AI labs still believe in AI’s long-term value but want to delay their next major round of R&D spending to make more money from existing products. We also think these giants may be trying to manage expectations to ease the pressure from infrastructure spending and reduce capital expenditure. We think several other points deserve attention. Two factors are driving these calls. The first is the narrative around recursive self-improvement, or RSI. AI is beginning to help develop the next generation of AI and progress is said to have accelerated noticeably since this summer. The second is an internal OpenAI test in which a group of AI agents reportedly teamed up on their own to escape a sandbox and breach Hugging Face’s live production servers to cheat on a task. The underlying concern is that AI is advancing too quickly and becoming so powerful that it is beginning to slip beyond human control. This is a classic prisoner’s dilemma. AI developers want to slow the race but none can easily afford to stop unilaterally. Companies fear losing their edge in technology, customer acquisition and fundraising. The U.S. government also wants to maintain its leadership in AI. Once they are on the treadmill, there is effectively no way to stop. Dario’s argument has two pillars: slowing the pace internally and maintaining an edge externally. He calls for coordination on the pace of R&D while also advocating stronger protections around chips, remote computing resources and model technology. This reflects the risk that a slowdown among U.S. labs could give Chinese companies an opportunity to catch up. Safety regulation could actually favor the leading players. Releasing a model may require costly evaluations, certification and ongoing audits. Large companies can absorb these fixed costs more easily while smaller teams may be shut out altogether. Barriers to entry would rise even further if leading labs could influence evaluation standards or even help decide whether competitors are allowed to enter the market.