According to @AnthropicAI, when controlling for general capabilities, models trained with SGTM perform worse on the undesired forget subset than models trained with data filtering, highlighting a reported performance gap between these unlearning approaches on targeted knowledge removal tasks, source: https://twitter.com/AnthropicAI/status/1998479611945202053. For trading context, the verified takeaway is the relative underperformance of SGTM versus data filtering on the forget subset under equal capability control, with no specific assets or tickers mentioned in the source, source: https://twitter.com/AnthropicAI/status/1998479611945202053.