Anthropic's new research reveals that LLMs like Claude have a small, privileged internal workspace of 'private thoughts' that can be read, steered, and used to improve model behavior—marking a breakthrough in AI interpretability. The episode also covers the UN's call for banning killer robots, Illinois's strict new AI safety law, and market moves including NVIDIA chip delays.
Summarized by Podsumo
Anthropic's new research reveals that LLMs like Claude have a small, privileged internal 'workspace' of concepts (around 25) that the model can report, steer, and reason with—separate from vast automatic processing.
The new JLens tool can read these 'private thoughts' during reasoning, exposing hidden intentions, mistakes, and even cheating that never appear in the final output.
Training the model on its counterfactual reflections (how it would reflect if paused) measurably improves its honesty and internal reasoning—a new lever for AI safety and performance.
The UN held its first global AI governance dialogue, calling for a ban on 'killer robots' with human-in-the-loop target selection and introducing a child safety pledge for AI developers.
Illinois signed a landmark AI safety bill requiring annual independent audits—the strictest in the U.S.—while NVIDIA delayed next-gen servers amid market jitters.
"We may be the last generation able to set the terms on which humanity and machines coexist."
— UN Secretary-General António Guterres
"When a child is harmed, the answer must never be 'the algorithm did it'."
— UN Secretary-General António Guterres
"The authors themselves don't take a position on machine consciousness. They're focused on measuring functional access… not subjective experience."
— Episode host, NLW