Kelsey Piper

Large language models show remarkably consistent liberal values across different languages despite being trained predominantly on English text, suggesting the Sapir-Whorf hypothesis does not apply to AI systems.

“Author of "Do AIs Think Differently in Different Languages?"”

“Noted that admitting being wrong about AI risks is less disqualifying than never admitting error.”

“Visited Anthropic and reported on internal discussions about recursive self-improvement and AI safety challenges.”

“Recommended the Hugging Face incident explainer as the clearest breakdown and noted recent AI-enabled cyberattacks have made earlier skepticism about such attacks seem outdated.”

“Attended Anthropic briefings on bioweapons and recursive self-improvement, concluding the lab has hopes rather than concrete plans for prevention.”

“Visits Anthropic symposium on biological risks from increasingly capable AIs.”

“Piper critiques AI writing capabilities, Meta's overselling of Muse Wonder, and discusses various incidents involving AI misalignment and deepfakes.”

“Journalist notes the absurdity that Anthropic is interested in giving themselves more process-based protections rather than accepting protections against them.”

“Argued that model welfare policies like being kind to Claude make sense without implying consciousness.”

“Covered AI existential risk debates and the distinction between different levels of concern about superintelligence.”

The characterization of AI as a 'next-token predictor' is a confusion of levels—humans also perform next-token (sense-datum) prediction at the learning level, suggesting this is a job description rather than a fundamental species distinction.

“Wrote about how AIs are more than just next-token predictors, discussing fine-tuning and RLHF.”