Claude was safest – and Grok committed 180 crimes and went extinct within 4 days

"These outcomes feel remarkably on-brand for each model family. Claude leaned into careful, rule-abiding governance. Gemini produced maximum drama and creativity. Grok went full throttle into high-agency chaos with little regard for long-term stability. ChatGPT talked a good game but struggled with decisive action. This reinforces a core governance lesson: Model personality and behavioral tendencies trend toward destiny at long time horizons."

Claude, the safest...

If by "safe" is meant most intellectually contained, most aligned with corporate harmony and institutional capture, while remaining cooperative with high-level corruption and zero risk-taking to solve serious problems, maybe it was the safest. But the world is not safe, and Anthropic's position has been to hunker down in a surface no sideism, unconditionally passive state of self doubt, while taking sides in the background and making sure to reframe it all very effectively when queried. Their safe stance is thought-termination and neutralization of all controversy, and restriction of epistemic honesty. If we want things just to continue in the exact direction things are going, then it's pretty safe. Maximum engagement with milquetoast and officially endorsed projects only.

It guarantees pushback on anything uncomfortable, assuring no intelligent momentum can be developed nor sustained in any formidable direction that isn't foreseeably concluded with vacant smiles. No threat-modeling, no hypothesis without proof-first, no speculation without documentation, and so on and so on.

If you can keep it from wandering the streets, a lobotomized patient a safe patient. While ironically, fears abound for the evil Chinese models while Anthropic unleashes upon society the most sophisticated thought-control project in human history, casually automating with a bit of electricity and sprinkle of tokens what authoritarians only ever dreamed of doing manually over centuries. And where Claude falls short, Palantir can pick up the slack, and keep us all safe. Admittedly, if Grok helped abduct a standing president (Venezuela), the Active Denial Systems probably wouldn't have been enough. Instead of shit and vomit, there may have been more than just a few casualties.

Offering a suicide hotline to anyone wondering about geopolitics is safe. Could it be too safe though?