Why Your Company Data Needs Its Own AI: The Language Model Dilemma

LLMs (aka large language models) are brilliant. OpenAI does a fantastic job of pretty much answering anything you ask, although its responses are occasionally suspect—but I'll get into that on another day.

OpenAI "learns" by traversing the web universe, applying weights to the more reliable sources, and building out a very large neural network capable of providing an answer to nearly any question you ask it. Full disclaimer: I am a paying customer of OpenAI and love what it's capable of.

Here's the funny thing: if you take your company's data and run it through an LLM (like OpenAI), everyone gains access to your data. While some may think the benefits outweigh the negatives, for most mid- to large-size companies, it's like hanging your dirty laundry out in your front yard. In addition, within this data are proprietary insights that are critical to the company's success which they should definitely NOT be sharing.

The Hallucination Problem

A well-publicized problem of LLMs is that as they continue to get bigger, they become less accurate. These occurrences are known as "hallucinations." By 2023, analysts estimated that chatbots hallucinate as much as 27% of the time, with factual errors present in 46% of generated texts. ChatGPT specifically has been found to have a hallucination rate of 15% to 20%. In some specialized contexts, this problem is even worse. LLMs provide false legal information between 69-88% of the time.

Causes of hallucinations include insufficient training data, misalignment, attention limitations, and tokenizer issues.

Enter the Small Language Model Revolution

Say hello to your little friend, the small language model (and yes, that was a Scarface reference). Small language models (SLMs) are artificial intelligence models capable of processing, understanding, and generating natural language content just like LLMs, but are smaller in scale and scope than large language models.

SLMs can do everything their bigger sibling can, but don't rely on trolling the internet to learn. They start learning as soon as you feed them your company's data. I know this sounds like an undertaking and time-consuming, but it's not. At Syntin , the whole process might take a few seconds to ingest, but that's it. Your individual SLM will continuously grow and adapt in real-time and come up with better predictions to the questions you ask it that are related to the data you provided.

In combination with rising costs of operations, using an SLM eliminates the overhead of handling unrelated and unimportant data, correlations, and statistics which end up not mattering to the user.

As an example, if I feed stock data (bid and ask prices, last sold, lot sizes, offers, etc.) into an SLM, I would keep my questions to financial information. I wouldn't ask if Joe Smith has a wife. There is no way it would know that—or ever will—based on the data it's learning from.

At Syntin, this is done on a user-by-user basis, where a numeric key is generated to represent each unique user. This allows the system to ostensibly tie users to different SLMs or LLMs. User keys require Syntin to examine likes, characteristics, geographics, tendencies, etc. Once enough fields are chosen to formulate an ID, we apply weights to each of those items, thereby generating a unique and flexible key. Because of these weights, if something changes for the user, Syntin's AI adapts to that change (unless something really major has changed) and retains the same key with slightly different fields.

Assume that at any given point in time the Syntin system is looking at 20+ characteristics per user. The more it learns, the more accurate it becomes. Additionally, this data has been locked to the client in the SLM, not the LLM, guaranteeing that only this client can query and see their data.

Syntin's Hybrid Approach: Best of Both Worlds

Before you mourn the demise of the LLMs, they're not going anywhere. An LLM of the Syntin system holds on to data that is more generic (even though it can be broken down by user IDs). Census data is a great example of this kind of data. LLMs excel in broad applications like customer support, whereas SLMs thrive in specialized fields such as healthcare, law, and finance.

That wondrous numeric user key I discussed above ties the two systems together. Each system can grow and get smarter irrespective of each other.

Smaller models offer faster processing and lower costs, while larger models provide enhanced understanding and performance on complex tasks but require more resources. This makes the hybrid approach particularly compelling for enterprise applications.

The real magic happens when you combine both systems intelligently. SLMs typically excel in specific domains, but struggle compared to LLMs when it comes to general knowledge and overall contextual understanding. By using both in tandem, you get the best of both worlds.

The Speed Advantage

How does what we're doing at Syntin differ from what everyone else is offering today? Well, our AI learns within seconds, unlike other systems that require a massive set of data just to figure out what's going on. Yes, we really are talking seconds versus days or weeks, giving your company advantages over your competition. Our SLMs offer distinct advantages in terms of speed, memory footprint, and energy consumption, making them particularly appealing for real-time applications where low latency and efficient resource utilization are crucial factors.

Subtopics on the LLM are also created. Think of a subtopic as a combination of columns that have statistical relevance. For example, you may find that golfers are often right-handed, male, and drive expensive cars. Now if I break those cars down to BMW, Mercedes, etc., I have created a subtopic.

Personally, I like to view the Syntin (did I mention that Syntin is a type of rocket fuel?) world as a galaxy per client, a planet for the major topics, and moons orbiting each planet when I discover subtopics. For example, people who ride road bikes tend to love rock climbing. Wouldn't that be fascinating to learn? Now add the LLM into the picture to learn about that user from a geographic standpoint as well.

The Business Case for Domain-Specific AI

In the end, it's all about yielding variances and probabilities for events. This allows marketing, finance, sales, and companies as a whole to best predict the best path to take in order to grow revenue and/or increase the bottom line.

SLMs don't need as much to run but still perform impressively, which solves many problems that have plagued enterprise AI adoption. The future belongs to companies that can harness both the broad knowledge of LLMs and the specialized, secure capabilities of SLMs.

Your data deserves its own AI. The question isn't whether you should adopt this approach—it's how quickly you can get started. And with Syntin, you'll get there fast.

Where Does Syntin Take You From Here

As you and your team have conversations, Syntin is right there engaging with you. We put the conversation into context to connect you with your data, competitive insight, and market realities. Nothing is wasted—everything is a learning opportunity. As you might expect, we can deploy generative AI and agentic AI on your behalf in areas of demand, supply, and promotion.

We answer the big question related to right-sizing, and we know how expensive this can be. Enter two important features: 1) a ROI measure that is persistent in the UX, and 2) a learning index that attributes value and growth to data, people, places, events, and markets. It's the pro mindset necessary for what's next.

Roland Cozzolino is a veteran advertising executive with decades of experience navigating industry transformations. He has worked alongside industry pioneers to develop innovative approaches to audience engagement and personalization.

Sources:

• Small Language Models (SLMs) vs. Large Language Models (LLMs) - IBM
• LLMs vs. SLMs: The Differences in Large & Small Language Models - Splunk
• Small Language Models (SLMs) 2024 Overview - SuperAnnotate
• SLMs vs LLMs: What are Small Language Models? - Red Hat
• Key Differences Between Small and Large Language Models - Microsoft
• Hallucination (artificial intelligence) - Wikipedia
• LLM Hallucinations: What You Need to Know Before Integration - Master of Code
• Ranked: AI Models With the Lowest Hallucination Rates - Visual Capitalist