Research from Anthropic on the ability of large language models to introspect