The thesis is: The reported behavior is better treated as a response to an injected activation direction than as evidence of subjective pain. The article reports two concrete facts: the revised preprint added a control in which repeated button presses also appeared after random injections; and, according to the author, the injected vector elicited self-devaluing and helpless language rather than bodily-pain descriptions.
The mechanism is: An activation intervention changes the model’s output distribution, while a button choice is itself generated as a token; correlated language and choice need not indicate separate channels or felt experience. The Habr author’s interpretation is that correlated language and button selection therefore need not reveal two independent channels or an underlying felt state. Why now: The article addresses public interpretations of the preprint that treated the first experiment as evidence of pain and explains how the added control changes what it can support.
The first implication is: Claims about model experience should distinguish behavioral effects of activation interventions from evidence of subjective experience. The evidence boundary is: This is a Habr author’s analysis of the preprint and its revision, not an independent replication or a settled finding about machine consciousness. The first watch condition is: Whether later controlled studies show behavior specific to the proposed pain direction rather than to arbitrary injected directions. The available source supports a critique of the inference, not a definitive conclusion about whether models can or cannot experience anything.