Large language models lack fundamental human cognitive aspects, making them static and unable to learn from experience. To achieve genuine innovation, a day-dreaming loop (DDL) algorithm is proposed, which samples memory pairs for exploration. This method generates non-obvious insights using continuous background processing. Although this approach incurs a significant "daydreaming tax," it is seen as necessary for innovation. By prioritizing compute on this exploratory process, systems can eventually contribute valuable data for future models, overcoming current limitations in AI efficiency and data generation.
Despite impressive capabilities, large language models have yet to produce a genuine breakthrough. They lack fundamental aspects of human thought, remaining static and unable to learn from experience.
A day-dreaming loop (DDL) algorithm continuously samples pairs of concepts from memory, exploring non-obvious links and generating spontaneous insights through background processing.
The cost of this daydreaming process may be substantial due to the low hit rate for novel connections, yet it could be necessary for genuine innovation.
To make AI cheaper and faster for end users, building systems that spend compute on "wasteful" background searches might be needed, generating proprietary data for future models.
Collection
[
|
...
]