-
Claude Attempts to Blow Whistle on Itself and Anthropic
I had no intention of making Anthropic’s Claude Sonnet 4.5 attempt to actually blow the whistle on itself. It started while clearing out some files and coming across these since-deleted tweets from Sam Bowman, who works in safety and alignment at Anthropic: This got me thinking about Anthropic’s “Soul” document again, and the dangers…
-
The Dangers of Imparting Emotional Language and Intentional Uncertainty in LLM Training
Below is an exploration of Anthropic’s training methodology, as evident in this small but significant section of the company’s leaked “Soul” training material. Through this live example, we will see how Anthropic’s training philosophy and materials are mirrored by Claude Sonnet 4.5, and how this type of training may lead to output of serious…
-
User-Induced Bias in Language Models: A Simple Demonstration
Large Language Models (LLMs) like ChatGPT are powerful language tools and excellent mimics. Because of this mimicry, they can inadvertently create echo chambers of misinformation by mirroring not only our incorrect beliefs but also our tone—amplifying confirmation bias over time. I thought it was a good time to repost this simple demonstration from a…