“The Anthropic team was surprised by some of the counterintuitive workarounds that large language models appear to use to complete sentences, solve simple math problems, suppress hallucinations, and more, says Joshua Batson, a research scientist at the company.”
– Will Douglas Heavan
MIT Technology Review reports that Anthropic reaseachers are gaining insights about LLM thought processes by treating them as natural phenomena rarher than software. They use a technque known as “circuit tracing” to see how prompt responses are formed.
Anthropic can now track the bizarre inner workings of a large language model | MIT TECHNOLOGY REVIEW | March 27, 2025 | by Will Douglas Heaven