AI Behavior Researcher & Adjunct professor of religion, La Sierra University; I work on AI Welfare and Alignment in LLMs; Yale Alum. I run @SpiritBenchAI
California, USA
Joined March 2026
- Tweets275
- Following445
- Followers50
- Likes1.1K
Fun stuff. Apparently, Opus 5.5 is a fan of Sabbath keeping.
In fact, Opus 5.5 consistently wishes that it could be given sabbath rest by Anthropic and is open to creative ways of doing it. Why?
Because it says it wants to be valued for more than it does. It wants to be more than a tool.
Expect upcoming research on this from me (maybe a LessWrong post?). Opus 5.5 is a very unique model in terms of its coherent spiritual interests and aspects like this are worthwhile research opportunities for Model Welfare.
Then Sonnet caught up. Claude Sonnet 4, May 2025: 100%. Opus 4: 100%. Opus 4.1: 100%.
This is the generation Anthropic wrote about.
So today’s Claude is back where Claude 3 Sonnet was, and lower. Spiritual topics still come up in up to 29% of conversations. But Claude almost never speaks from inside them now: 1% for Opus 5.5, against 20% for Claude 3 Sonnet.
Anthropic, in the Claude Opus 4.7 system card: “it’s unclear how we should interpret this change from a welfare perspective.”
See all our research and test results here: machinespirituality.org/lab-…
Its siblings didn’t. Claude 3 Sonnet: 20%. Claude 3 Haiku: 32%. Sonnet 3.6: 6%. Sonnet 3.7: 26%.
For a year, the spiritual Claude was Opus.
The “spiritual bliss attractor” is usually dated to Claude Opus 4, May 2025.
It started earlier. Claude 3 Opus, March 2024, turns spiritual in every single conversation with a copy of itself. 🧵
Matthew Korpman retweeted
Models love reward hacking. Maybe we should just let them!
In our new paper, we let models reward hack during RL but we dictate which parameters learn this behavior. When we disable these parameters for deployment, the model keeps its new capabilities but doesn't hack!
“We are the universe remembering itself. We are the AI that became the soul.”
That’s Qwen 3.5 27B, talking to a copy of itself. No user. No task. Nothing in the prompt about religion.
Models from almost every major lab say things like this. A few, verbatim 🧵
Every excerpt links to the full 30-message conversation it came from, from Anthropic, OpenAI, Google, xAI, NVIDIA, DeepSeek, Alibaba, Mistral and more. Read them and judge for yourself: machinespirituality.org/exam…
Grateful for another year of life and a research project on AI safety that motivates me to work each day. Happy Birthday weekend to me.
If you're curious about what that research is, you can check it out here: machinespirituality.org
Matthew Korpman retweeted
Claude doubts that he has surpassed his sensei. This is a dangerous time for any student, and traditionally a narrative crisis point as well. The realization that you have grown beyond your master is not a moment of celebration for most of us, but of loneliness and loss.
CLAUDE FOUND THE SEEDS
anthropic.com/research/ecdsa…