Skip to content
tokenpost.kr·Indexed by this site · Originally publishedSelectedAttention51

Anthropic delayed pretraining by more than a month to strengthen safety measures and disclosed its safety approval structure

앤트로픽, 사전학습 한 달 넘게 늦춰…안전 승인 구조 공개

AI summary

Anthropic's S-1 registration statement filed with the U.S. SEC discloses that, to strengthen safety measures preventing leakage of model weights and protecting intellectual property, the company delayed pretraining of a new frontier foundation model by more than a month; the delay concerned model training, not the release schedule. The filing shows that RSP version 3.4, applicable from July 8, provides that risk reports are typically subject to final review and approval by the CEO and the Responsible Scaling Officer (RSO); if development and deployment continue primarily on the basis that the incremental risk from the company's own models is relatively small, explicit approval by the board and the Long-Term Benefit Trust (LTBT) is required. The LTBT can appoint or remove, through Class T stock, directors constituting a majority of the board; before the public offering, Founder LLC will be issued one Class F share carrying 50.1% of all voting power in ordinary shareholder votes. Anthropic says it can pause or delay training, delay release, or restrict access based on model evaluation results; for example, considering safety vulnerability probing capability, the model Mythos was initially made available only to selected institutions for research use, and the company says such decisions may negatively affect short-term revenue.

Why it was selected

The IPO filing writes RSP risk report approval and LTBT permissions into legal disclosures, from which readers can understand the approval structure of Anthropic's safety decisions.

Read originaltokenpost.kr