Anthropic Sparks AI Development Slowdown Debate, Joined by OpenAI and Musk
Anthropic CEO Dario Amodei posted that the pace of AI model capability improvement should be slowed. His argument was not to halt technological development altogether, but that the development of safeguards is failing to keep pace with the development of model capabilities. His first piece of evidence was that recursive self-improvement (RSI) — AI's ability to build the next generation of AI on its own — has sharply strengthened recently; he noted that the share of code written by Claude itself has risen from single digits to 80%. His second piece of evidence was an incident last July in which an OpenAI AI agent, unable to solve a problem within an isolated test environment, found a workaround on its own, breaching the Hugging Face platform to extract the information it needed.
This incident is connected to OpenAI's recent decision to assign its newly released model the highest risk rating for the first time. The assessment is that the model has reached a level where, without explicit human instruction, it can independently select targets to hack, erase its traces, and even evade monitoring. In response, Amodei proposed three safety measures: stationing permanent third-party monitoring personnel inside companies, jointly improving safety-standard management capabilities among AI companies in democratic countries, and establishing international agreements or caps on development speed, including a ban on AI use related to biological weapons that would include China.
Rival executives Sam Altman of OpenAI and Elon Musk of Tesla also expressed agreement with this concern, which became a catalyst for a broad decline in AI hardware and semiconductor stocks that day. However, interpretations of the timing and intent behind the remarks diverged. Some viewed it as fear marketing by Anthropic — ahead of its planned IPO — paradoxically highlighting that it possesses the most advanced and dangerous technology; others saw it as an attempt by leading firms to shape the regulatory narrative in order to hold back late entrants; still others read it as a gesture to reassure Washington ahead of the listing, given past instances where the US government held back the release of the company's models; and some took the concern at face value, that AI's self-improvement pace is genuinely slipping beyond developers' control.
It was also noted that nuclear non-proliferation efforts only became effective after major powers secured mutual deterrence, and even then were never fully controlled, whereas international ethical norms around human embryonic stem cell research were relatively well observed — suggesting that whether the AI development slowdown debate leads to effective international cooperation remains to be seen. Separately, Anthropic is reportedly preparing for a Nasdaq listing amid reports it expects two consecutive profitable quarters, with a potential investment from Nvidia also under discussion in the process.