LLM Daydreaming

10 months ago

LLMs lack fundamental aspects of human thought, such as continual learning and a 'default mode' for background processing.
Proposal of a 'day-dreaming loop' (DDL) for LLMs to simulate spontaneous human insight by continuously sampling and linking concepts.
The DDL involves a generator model exploring links between concepts and a critic model filtering for valuable ideas, creating a feedback loop.
The 'daydreaming tax' refers to the substantial compute cost of this process, which may be necessary for innovation.
Strategic implication: expensive, daydreaming AIs could generate proprietary training data for efficient models, bypassing the 'data wall'.
Human researchers benefit from continual thinking and background processing, which LLMs currently lack.
The default mode network in humans is associated with spontaneous thoughts and creativity, a feature absent in LLMs.
The DDL is inspired by wake-sleep algorithms and the default mode network, proposing a method for LLMs to achieve similar creativity.
Potential obstacles include the high cost of the DDL and the challenge of optimizing the process for useful insights.
Implications suggest that only power-users or researchers may be willing to pay the 'daydreaming tax' for novel insights.

Hasty Briefsbeta