Exploring the Differences Between Personal Language Models and Large Language Models

In November 2022, Large Language Models (LLMs) thrust into the mainstream. We’re familiar with their technological advances and capabilities for content creation and discovery. LLMs have already changed the way we discover content and interact with information on the internet, but LLMs are only one type of language model. This article will explore a different type of model, Personal Language Models (PLMs), or small language models, and introduce you to the Personal Intelligence Era.

GoogleBard AIOpen AIBingMeta

What Defines a Large Language Model (LLM)?

LLMs are artificial intelligence (AI) models capable of understanding and generating human-like text. They're trained on vast amounts of data from the internet, learning the structure of language, facts about the world, and even some reasoning abilities.

Companies building these models include Big Tech (Microsoft, Google, NVIDIA, and Meta), as well as Anthropic AI, Huggingface, Character AI, and others. It’s a competitive space, with many models fighting for limited resources and market share.

LLMs have transformed the world, but they have their limitations.

Let’s look at an entirely different kind of model, called the Personal Language Model (or small language model). These models have been in development by Personal.ai since 2020 and take a very different approach than Big Tech and LLMs.

What Defines a Personal Language Model (PLM) or Small Language Model?

PLMs are artificial intelligence (AI) models capable of understanding and generating information within the domain of an individual user. They are grounded in the data, memories, facts, and opinions of a single person and are highly personal.

With thought leaders like Bill Gates and Sir Tim Berners-Lee noting “personal AI’s” practicality and potential to change the world, it’s important to understand how they differ from Large Language Models.

How Does a Personal Language Model (small language model) Work?

PLMs contain an ensemble of models based on GPT and BERT architecture but are fundamentally different in that they are “grounded” rather than “pretrained”, so GPT becomes GGT (“Generative Grounded Transformer”).

Let’s talk about Memory Stacks, the digital repository of your memories which your Personal Language Model (small languge model) trains on.

As LLMs train on massive public data sets, your PLM trains on a Memory Stack that is a highly curated, specific corpus that is unique to you.

Unlike with LLMs, you control what goes into your Memory Stack. The Memory Stack is comprised of many independent Memory Blocks; smaller chunks of data that can include a time, a date, a source, a scope, text, and other information. One’s Memory Blocks are entirely within the users’ control, and can be added to, deleted, or edited.

How Can PLMs and LLMs Work Together? 

Think of LLMs as using an encyclopedia or Google. It’s where you go to gather and learn new information that does not exist in your memory.

PLMs on the other hand can capture and leverage what you already know, and can learn about your individual perspectives so it may surface them contextually. LLMs are like the mainframe computer, whereas PLMs are like the personal computer. 

These models each have their own purposes, strengths, and limitations. A user seeking new information about Barcelona can use LLMs to inform them, educate them on the history, and provide general tips about visiting the city. But after their visit, they will have formed their own opinions and had their own experiences, captured by their own PLM. LLMs and PLMs can work seamlessly together to connect your internal memories with external knowledge.