Skip to content
Martin Cid Magazine·Indexed by this site · Originally published· Governance →· Its place in the scan record →SelectedAttention46

Anthropic judges on the basis of expectations that AI has reached the self-improvement threshold and activates safety rules

앤트로픽, AI 자기개선 기준 넘었다고 판단…안전 규칙 발동

AI summary

Anthropic said in a revised IPO prospectus that it has reached the AI R&D capability threshold set by its internal Responsible Scaling Policy (RSP) and has accordingly activated the corresponding safety rules. The policy describes this point as the beginning of recursive self-improvement, the stage at which AI systems begin to accelerate the creation of better AI. Anthropic explained that this judgment is not based on observations of existing models but on expectations for future models and the overall pace of development, and that no model among the Claude models on sale has been confirmed to have crossed this line. The internal basis listed in the prospectus is that in the most recent month, Claude's share of code contributions to Anthropic's codebase exceeded 90%, and the amount of code contributed per engineer per day grew more than 45-fold. The company is disclosing this judgment to investors before stock trading begins.

Why it was selected

Anthropic's basis for triggering this threshold is expectations about the future rather than observations of existing models; this distinction is key background for understanding the nature of the decision.

Read originalko.martincid.com