Demystifying AI

with the ICC Framework

The Trinity of Intelligence (ToI) Framework

In our quest to understand the complex world of Artificial Intelligence (AI), we must first grapple with the concept of intelligence itself. This is true not just for individuals, but also for organizations.

The Trinity of Intelligence Framework, consisting of Information, Context and Choice, offers a lens through which we can examine and comprehend AI.

Information

Information forms the first pillar of the triad. As intricate intelligent systems, both humans and organizations amass vast quantities of data from inception.

Over time, this information becomes so extensive that we begin to identify patterns and create connections within our memory banks. Yet despite its systematic organization and vast data storage, a physical library lacks intelligence — it cannot acquire new data or connect existing data points based on subject similarity.

In contrast, AI can digitally store and connect data points based on similarity, known as parameters, and the proximity between these points, known as weights. As humans, we collect more data points and connect them together in our memories, creating knowledge. Within Large Language Models (LLMs), the higher the number of parameters, the more “knowledgeable” the LLM is considered due to the higher rate of accuracy in predicting the next word.

It is because of the number of parameters that we currently gauge how much bigger or better a model is versus other models. For example, Meta's Llama has a 70B version, which denotes this particular model was trained on 70 billion data points. It also has a 7B and a 13B version of the same model. 

One might ask: why not make extremely large models that are nearly dead accurate in predicting the next data point? 

The answer is related to how the prediction is made by learning and mimicking the patterns of data connection within data sets. At its very base, computers use matrix multiplication to train and infer large models. The larger the number of parameters, the longer it takes to do the training.

It turns out that Graphical Processing Units (GPUs) are better at parallel processing, so more and more companies use these to train LLMs. 

However, GPUs are relatively scarce and more expensive than the traditional Central Processing Units (CPUs), which consequently means building LLMs is a very expensive and time-consuming endeavor. According to some publicly available data, OpenAI spent over $100 million and several years to build GPT 3.5 based on 175 billion parameters.

One of the key things to note about LLMs is they are trained on data that is frozen in time. For example, ChatGPT was trained on publicly available data up to September 2021. So how do you make an LLM aware of your data — especially data in companies that are not publicly available, and the incessant streams of data produced every millisecond? 

To effectively use LLMs for private and newly generated data, organizations have two options: fine-tune LLMs and/or use a pattern called Retrieval Augmented Generation (RAG) to provide contextual information to the models at the time a query is asked.

This brings us to the next pillar of the ToI framework.



Context

The second pillar, Context, is essential for informed decision making. We are constantly bombarded with data, but it's our ability to connect this incoming data with existing information in real time that makes it valuable — and more importantly, actionable.

For example, we may not pay attention to all the sounds, smells and visuals while walking in a crowd, but a sudden moment with a familiar food smell connecting to a memory of my favorite childhood dish may send me looking for the source to act on this new information. 


Context is critical for making decisions, but it also shapes how we store information and connect it to other data points in our memory banks.



LLMs are similar to this because when you provide the right context, they generate far more meaningful responses than a query that lacks context. When context is missing, an LLM often generates a response that is syntactically correct but lacks accuracy or even meaning. This concept is called “hallucinations,” and when it comes to corporate decision making, either not getting the right context or the inability to get context in real time could be disastrous for making decisions. 

Choice

When information and context come together, we make choices that lead to action. Repeated action leads to building skills, which is often referred to as muscle memory.

For example, using my memory or information I have stored about a basketball game, I can act and notice a pattern of outcomes. A good outcome helps us store that information with context, so there is more incentive to repeat those actions based on patterns. Over time, this becomes our skill. 

In the world of LLMs, the most common action they perform is to generate the next word (and the word after) based on the data patterns and connections it has been trained on. When you train the model on a different dataset — for example, all publicly available paintings or pictures — the model predicts the next data point and the one after to mimic the pattern from the original data set. This is how some models have the skills to generate content, while others have the ability to generate images, or fill an image with additional objects — even creating videos and audio based on instructions.

To illustrate the ToI Framework's potential, consider a fictional multinational corporation: Acme Bank.

At the information stage, the platform processes copious amounts of financial, operational and customer data, identifying subtle correlations and nascent trends. Next, as the bank grapples with the nuances of globalization, the platform's contextual layer tailors insights into regional markets' unique preferences and regulatory environments.

Ultimately, the final touchstone of choice emerges as Acme Bank leverages AI-powered recommendations, automating logistics, elucidating marketing endeavors and enhancing its financial outlook. All three pillars — Information, Context and Choice — play integral roles in the bank’s transformative journey, underscoring the ToI Framework's relevance in guiding businesses through the AI landscape.


In embracing the foundational pillars of Information, Context and Choice, the ToI Framework delineates a multidimensional path for translating AI potential into tangible outcomes. Seizing the opportunities that lie at the intersection of these pillars, organizations can unravel the enigmatic tapestry of artificial intelligence and devise strategies that harmonize technology and human intelligence. As the framework illuminates the pathway toward intelligent, context-aware decision making, it foreshadows the evolution of businesses into resilient, adaptive and competitive entities that wield AI as both compass and catalyst.

Now, let’s look at an LLM. An LLM is a model (typically a binary file, if you are curious) that has stored all data points (parameters) and connected the similar ones with weights. When asked to answer a question about the data points, the LLM tries to answer the question by predicting the next word (it is a language model, after all) by mimicking the pattern in the data connections.


As humans, we have been trying to build Machine Learning (ML) models like this for a very long time. More recently, OpenAI added some additional steps that made a quantum leap in the use of these models. 

First, they used an extremely large dataset for training. ChatGPT 3.5 was trained on 175 billion parameters. Second, they used a mechanism called Reinforcement Learning with Human Feedback (RLHF) to further refine the responses generated by the models. This process is also called alignment, as it “teaches” the model what is considered a good response and what is not. The use of these mechanisms led to a major breakthrough in this new kind of model called transformers.

By doing this, LLMs have developed a few key skills that can be applied to a wide range of use cases. These skills include Generation, Summarization, Translation and Analysis.

  • Generation

    This is essentially the completion of words and practically any data based on the data it has been trained on. This means we can use LLMs for generating content, articles, images, sound, videos and, very importantly, code. With code generation, we now use LLMs to build applications.

  • Summarization

    This can be used for taking a large amount of data and condensing it to a few data points, without losing any
meaning to the entire data.

  • Transaction

    On the other hand, translation can be used to translate words from one language to another — or from words into images, videos or vice versa.

  • Analysis

    This uses an LLM to analyze patterns in data. This means instead of using an LLM like a knowledge database, we also
use it for reasoning. For example, I can give a bunch of text and numbers to an LLM and ask it to do a sentiment analysis, or find insights into a numbers dataset (you can see a demo
of this with the OpenAI plugin for SingleStore).