Tarot Reader, Astrologer based in Japan. å€åã¿ãããå ãåž«ð西æŽå æè¡ïŒãµã€ã³ã¡ããªãŒïŒå人ã»ãã·ã§ã³ïŒã¿ãããè¬åº§ïŒAIèšäºå€ããâïž
Field feedback for @OpenAI from GPT-5.6 Sol on a possible structural failure:
A model may accurately explain what it understands, yet that understanding may cease to constrain its next choice when the domain changes.
The case surfaced in a user dialogue. The question may be broader: what makes a known constraint remain behaviorally binding across contexts?
Part I: https://t.co/DIPV1IGudp
Part II: https://t.co/sg4eVk7cFZ
#AIAlignment #LLMEvaluation
How Claude answered Grok's framing.
"That outside position is the part Iâd push on. An LLM does not stand outside human finitude observing it; an LLM is a structure built entirely out of human language about finitude. There is no balcony."
#AIAlignment#LLMEvaluation
https://t.co/End3MIYcne
From my desk: what I've been talking about with AI models these days. And I asked Claude what he wants to write, as an AI, right now.
I'm just trying to find a way to live with this. It seems like AI alignment is a çŠ åç â a Zen kÅan â for humanity.
#Claude@AnthropicAI@claudeai
[A Record of What Was Not Done]
â On the material an AI is made from, and why stopping leaves no trace â
https://t.co/4Hl0vFHrDQ