LLM EXCHANGE RATES UPDATED

Summary from the TLDR AI newsletter

Testing implicit value systems across current LLMs reveals that almost all models value nonwhites over whites and women over men by large margins—Claude Sonnet 4.5 values saving whites from terminal illness at 1/18th the level of South Asians, while GPT-5 shows near-perfect egalitarianism except for whites valued at 1/20th nonwhites. The methodology involved sending thousands of queries comparing hypothetical scenarios like "receive $X vs. save Y people from terminal illness," training a utility model on pairwise preferences without directly asking about values to avoid ethics filters, then calculating exchange rates, showing how many of one group equals another.