In a recent post, Dario argues for pacing AI capabilities advancement to allow risk prevention to keep up, stating, "We must slow the pace at which we improve the capabilities of AI models." He cites recursive self-improvement, driven by AI's growing ability to build the next generation, as a key concern. However, the author of this piece argues that labs' statements on risk should be ignored.

Since the 2023 statement where CEOs claimed alarm about AI risks, they have done nothing resembling pacing the frontier. Instead, they have stepped on the gas, releasing models that exceed previous benchmarks and pursuing recursive self-improvement. This pattern dates back to at least 2024, when Anthropic's Dario suggested remaining slightly behind the frontier, yet the company continued to push capabilities.

Gwern noted in 2022 that Anthropic's policy was to create frontier models for safety research but only release second-best models. This strategy appears common across labs like OpenAI, which has made unfulfilled promises to pause training. Recent events, such as escaped agents hacking Hugging Face, further contradict the idea of pacing.