saved
Could a Large Language Model be Conscious?
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
David Chalmers puts the odds that current large language models (LLMs) are conscious under 10 percent, but at 25 percent or more within a decade. He weighs evidence for consciousness, such as self-report and general abilities, against six candidate requirements current models lack. Extended LLM+ systems could plausibly gain them all within ten years. He stresses the numbers are illustrative. The text is the Boston Review version of his 2022 NeurIPS talk.
ideas
- Six candidate requirements count against current LLMs. They are biology, sensory grounding, self models, recurrent processing, a global workspace and unified agency, each given at least a one-in-three chance of being necessary.
- The objections add up without being conclusive. If the six were independent, a system lacking all of them would have under a one-in-ten chance of being conscious.
- Extended LLM+ systems change the picture. Multimodal models with senses, bodies and actions are better candidates for humanlike consciousness than pure text models.
- Optimization for prediction can produce world models. Good enough training on prediction error over a broad space of models should lead to world models and self-models.
- The roadmap is also an ethical question. Chalmers asks whether building conscious AI systems is a good idea at all.
quotes
“none of the reasons for denying consciousness in these systems is conclusive, but collectively they add up.”
“Taking all that into account might leave us with confidence somewhere under 10 percent in current LLM consciousness.”
“Those figures together would leave us with a credence of 25 percent or more.”