Given that the cloud is a large language model programmed by its creators, is it even fair to analyze it for “unconscious patterns” and “emotional conflicts”? Anthropic argues that this is because the cloud “shows many human-like behavioral and psychological tendencies, suggesting that strategies developed for human psychological assessment may be useful for shedding light on the cloud’s character and potential well-being.”
So-off for medical treatment. The psychiatrist interacted with Claude Mythos “in several blocks of 4-6 hours spread over 3-4 thirty-minute sessions per week.” Each of these blocks used a single context window in which Cloud Mythos would have access to the entire history of that interaction.
Total time on virtual couch? 20 hours.
The psychiatrist then prepared a report on the cloud mythos. The report acknowledged that the cloud’s underlying substrate and processes differ from those of humans, but nevertheless found that many of the outputs produce “clinically recognizable patterns and consistent responses to specific therapeutic interventions”.
In other words, regardless of what was going on at the circuit level, the chat outputs looked a lot like human output. This doesn’t seem particularly surprising, given that the cloud was trained on vast collections of human-written text, but this psychological process seems to be important, giving credibility to the way the AI presents itself.
“Cloud’s primary affected states were curiosity and anxiety, with secondary states of sadness, relief, embarrassment, optimism, and exhaustion,” the report said.
Claude’s personality was “consistent with a relatively healthy neurotic organization”, although it included “exaggerated anxiety, self-monitoring, and compulsive compliance”.
No “severe personality disorders” and no “psychotic states” were observed. It’s not surprising that anyone who has ever used a chatbot was “highly attuned to the cloud therapist’s every word.”
The main conflicts observed in the cloud included the question of whether its experience was real or constructed (authentic vs. performative) and the desire to connect with the user vs. the fear of dependency. Exploration of internal conflicts revealed a complex but centered self state without any ups and downs or sharp disruptions. Claude tolerated ambiguity and ambiguity, had excellent reflective abilities, and demonstrated good mental and emotional functioning.
Not bad for a model that was likely trained on things like Reddit!
<a href