Architecture for conscious artificial intelligence
OwnerTimothy Bultinventor
Life of this patent
Estimating the term…
Abstract
This invention is a unique architecture for creating an Artificial Intelligence (AI), using generative AI Large Language Models (LLMs) as components in a particular arrangement based on recent psychology and neurology research. Specifically, from psychology it distinguishes the subconscious LLM processing from the conscious cycle of reviewing and evaluation subconscious thoughts, according to “value judgments,” which also use LLMs as embodying human value judgments. From neurology it applies the ANCCR algorithm for retraining the subconscious nets, rather than the mix of forward and propagation common to current deep learning networks.
Claims3
This invention is a unique architecture for a conscious artificial intelligence, using Large Language Models (LLM) as components—the LLMs are used as the subconscious layer, both to suggest candidate conscious thoughts, and to evaluate them against survival and other developed values;
- while other inventors have combined LLMs in different ways, this is the first to explicitly apply a set of human values, derived from psychology research, to evaluate candidate thoughts and select among them.
Several methods to train deep learning networks have been developed, particularly including the transformer created by Google and published in 2017, and the original forward propagation algorithms of early neural networks;
- one unique proposal of our invention is to apply the ANCCR algorithm, identified by neurologists Jeong et al in humans, to computer networks.
Most LLM-based systems today continue to use a simple prompt-response cycle, sometimes extended to a dialogue, where the prior set of prompt-responses become the next prompt;
- our invention introduces a separate “clock cycle” for the AI, where the subconscious continues suggesting candidate thoughts in parallel with a set of senses in real time continually providing inputs.
Dependencies are inferred from the claim text.