30 March 2025

An Introduction to Capabilities and Limitations of Large Language Models and Generative AI Technology

Daniel G. Shapiro

Many approach the rise of machine learning tools like large language models (LLM) and multimodal generative artificial intelligence (GAI) with great expectations. These systems have multiple capabilities and diverse applications. For example, they can answer questions, analyze sentiment, generate images from text, and follow instructions. Common applications include content creation, translation, code generation, cybersecurity, candidate screening, storytelling, and virtual assistants. LLM and GAI capabilities are growing at an enormous rate, with major new systems and applications announced each week.

This storm of development inspired IDA researcher Dr. Daniel Shapiro to conduct an assessment of what these tools can achieve in principle, with the goal of tying readers’ expectations for their capabilities and limitations to a core understanding of the technology. This report explains why generative AI and large language models are able to demonstrate “a level of intelligence that has never been seen before in a computing system”, and are at the same time, “impaired models of cognition”. Given this understanding, they draw implications for future development.

LLM is a statistical model of a large training corpus that creates new content from prompts by generalizing past examples. Training an LLM on every sentence in every book ever written results in a statistical model of what comes next given what was seen before. Because that model reflects the knowledge expressed in training texts, LLM prompts can mine it to produce new output.

No comments: