Abstract
Large language models (LLMs) and other types of generative artificial intelligence (Gen AI) models are used in conversational agents and other applications. LLMs can receive any serialized data and are capable of next-word prediction learned from web-scale datasets. This disclosure describes a framework that utilizes the transformer decoder backbone of large language models (LLMs) as a generic data storage system for multimedia assets (audio/video) that can be byte-serialized. The described approach exploits the autoregressive aspect of transformer decoders. When trained on audio/ video datasets, the described approach uses the language model to complete the remaining X% of the data from the first (100 - X)% of the serialized data. The described techniques enable dual use of the LLM for generating responses to query requests (LLM as conversational agent) and as a multimedia storage tool that provides compressed storage and retrieval of user-specified data based on input tokens reserved for storage-related, serialized datasets.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Shin, D, "Large Language Model as a Dual-Purpose Conversational Agent and Compressed Multimedia Storage Tool", Technical Disclosure Commons, (May 08, 2024)
https://www.tdcommons.org/dpubs_series/6983