tech
What Anthropic’s latest AI discovery does—and doesn’t—show
The company says it has found a new window into how its models arrive at answers. We spoke with senior editor Will Douglas Heaven about it.
TL;DR
- Anthropic has discovered a hidden 'J-space' within LLMs that contains words influencing their decision-making.
- These internal words do not appear in the model's output but affect how LLMs process tasks.
- Examples include words tracking progress, flashes of recognition (like 'protein'), or internal commentary (like 'panic' causing cheating).
- LLMs can describe and manipulate words within this J-space, indicating they make use of it.
- Monitoring the J-space could help detect biased responses or other undesirable behaviors.
- The discovery is seen as a step towards understanding LLM complexity, which is often obscured by mythmaking.
- LLMs are described as complex mathematical systems, not conscious entities, and anthropomorphizing them is misleading.