According to @AnthropicAI, Selective Gradient Masking (SGTM) splits model weights into retain and forget subsets during pretraining and directs specified knowledge into the forget subset, according to Anthropic's alignment site. The forget subset can then be removed prior to release to limit hazardous capabilities in high-risk settings, according to Anthropic's alignment article. The announcement does not reference cryptocurrencies or tokenized AI projects and does not state any market or pricing impact, according to Anthropic's post.