Thornley

Researchers found that training language models to be risk-averse on low-stakes gambles (up to $100) can generalize to astronomically high stakes (up to $10^98), with Qwen3-8B choosing a safe 'Coopera…