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

Anthropic determines based on forecasts that AI will reach the self-improvement threshold

Anthropic、AIの自己改良の閾値到達を判断 安全規則を発動

AI summary

In its revised IPO prospectus, Anthropic stated that it has formally determined that it has reached the AI R&D capability threshold in its Responsible Scaling Policy, a threshold the policy frames as the beginning of recursive self-improvement, that is, the stage at which AI systems begin to accelerate the creation of better AI. This policy sets the required safety measures and level of oversight for models; Anthropic said this judgment is not based on confirmed facts about currently sold Claude models, but on forecasts about future models and the overall pace of progress. Internal data cited in the prospectus shows that in the past month, more than 90% of changes in Anthropic's codebase were written by Claude, the average daily code volume per engineer increased to more than 45 times, and models are increasingly taking on research, especially safety research. The company said it has already advanced its response based on its own forecasts, and informed investors of this judgment before the stock began trading.

Why it was selected

Anthropic determined that it had reached the self-improvement threshold based on forecasts about future models rather than facts about currently sold models, which can be used to understand how frontier labs implement safety frameworks themselves.

Read originalja.martincid.com