LLMs versus LGMs
Though Large Language Models (LLMs) and Large Graphical Models (LGMs) both fall into the Generative AI category, they are structurally very different and serve unique purposes. LLMs primarily work on text, using sophisticated algorithms and neural networks to process and generate language. They are trained on publicly available, internet-scale text data, allowing them to learn patterns, context, sentiment, and relationships within the data to answer questions and create new text and images.
LLMs analyze text documents in a linear fashion, typically left to right, top to bottom, whereas LGMs are designed to find and evaluate relationships across multiple dimensions. This makes LGMs highly effective for tabular and time series data, which is by nature multivariate and multi-dimensional (aka, non-linear). Simply stated, an LGM is a generative AI model that produces a graph to represent the conditional dependencies between a set of random variables. The goal is to create connections between the variables of interest in order to find relationships, create synthetic data, and predict how the relationships may change in the future based on various conditions.
Differentiators between LLMs and LGMs:
Data Sources
One major differentiator of a Large Language Model is that it primarily leverages publicly-available data to train its models, such as Wikipedia entries, articles, glossaries, and other internet-based resources. This lends itself well to use cases like creating marketing content, summarizing research, creating new images, and automating functions such as customer support or training via chatbots. In contrast, the Large Graphical Model is designed for enterprise-specific or proprietary data sources, making it highly effective for business use cases that depend on an organization’s historical data. For example, forecasting the demand for a product may require looking at past orders, marketing promotions, discounts, sales channels, and product availability, and projecting that into the future – sometimes with different assumptions, such as a new supplier.
Computational differences
The drive to improve the power and accuracy of LLMs has led to a massive increase in the number of parameters that the models must evaluate. More parameters enable greater pattern recognition and calculation of weights and biases, leading to more accurate outputs. As such, LLMs have gone from evaluating millions of parameters to evaluating billions or even trillions of parameters in the space of a few years. And all of this comes at the growing cost of network resources, compute cycles, and energy consumption, making it cost prohibitive for all but the deep pocketed enterprises to enable LLMs at enterprise scale.
LGMs, on the other hand, are more computationally efficient and require significantly less data to generate precise results. By combining the spatial and temporal structure of data in an interactive graph, LGMs capture relationships across many dimensions, making them highly efficient for modeling tabular and time series data. In contrast to LLMs, the LGM model was found to deliver on a commodity laptop compared to modern infrastructure using 68parallel machines. This makes LGMs accessible to businesses of all sizes and budgets.
Privacy and security
Privacy and security are a top concern for organizations interacting with commercial LLMs over the internet. A 2023 survey by Fishbowl found that 43% of the 11,700 professionals who responded say they are using AI tools for work-related tasks, raising concerns that proprietary data may be finding its way onto the internet. Since LGMs are designed to work with enterprise data, they are governed by an organization’s own privacy and security policies and do not pose the same concerns for data leakage or hallucinations that LLMs do.
Learn more here.
