Researchers at Anthropic used a new approach to study a large language model called Claude, allowing them to see how it thinks and make decisions. They found that Claude has a universal "language of thought" and can plan ahead when writing, but sometimes makes up plausible-sounding arguments.