Al Kingsley MBE is CEO of NetSupport, Chair of a multi-academy trust in the U.K., tech writer, speaker & author of multiple education books.
If you've ever found yourself wrangling with your AI tool of choice and not quite getting the output you wanted, a greater understanding of how it works under the hood may help.
Looking beyond the tool's immediate promises and knowing exactly how it puts text together could help you feed it better prompts, diagnose and rescue output failures more effectively, and recognize and remember its limitations.
LLM Fundamentals
At the base level, LLMs are advanced engines that complete word patterns. They've absorbed vast amounts of text data and learned the probable relationships between words. When you ask a question, it encodes your input, predicts the likely next word or "token" and continues this process until it reaches a stopping condition in its code.
An LLM doesn't "think" like we do. Instead, it generates probable continuations of what we've asked for. This explains its fluency and how it generates text so quickly. It also explains its shortcomings. When you get an odd response, it's because the statistical patterns it has learned have led it toward that kind of completion.
What Are Tokens?
Tokens are typically short chunks of text of around three or four characters—not even the full length of some words. An LLM can only process a fixed number of these at once, known as the "context window." Anything that falls outside of this window is effectively "forgotten."
This limitation becomes apparent when you start working with an LLM on long conversations or documents; it may simply lose track of earlier content, leaving you exasperated at its "memory loss." Bigger context windows sound like the ideal solution, but they're not always better. They can increase cost and sometimes even degrade performance as the model struggles to determine the salient parts among thousands of tokens.
Training Data Limitations
An LLM learns from the data it's trained on (e.g., web pages, books, academic papers, code repositories and more). It adopts the patterns contained within these sources but does not "remember" the text. This means its knowledge stops at its last training date, and therefore, its coverage across topics is uneven. This is the part that's most often forgotten and leads to frustration.
To work around these gaps, you can turn to retrieval-augmented generation (RAG) to inject up-to-date or domain-specific information. If you point the LLM toward this relevant content in your prompt, your output will be much improved. Chatbot tools are superb at pattern recognition, but they can't refer to source material outside of their training.
Bluffing It Out
What do humans sometimes do when we don't know an answer? We bluff it out. LLMs are the same. When presented with a weak prompt, they fill the gap with acceptable but false content. This is known as a hallucination.
Learning to be more specific can protect you from this, and breaking things down and being clear about what you need will certainly help. You can even ask it to tell you if it doesn't know or isn't sure of something, which should go some way toward preventing falsehoods in its results.
If you suspect hallucination in your content, you must fact-check it closely and edit accordingly. An LLM doesn't save you any time if what it's giving you is fundamentally wrong.
What To Do About Bias
Humans are inherently biased, and that's reflected across the internet and all of the sources on which LLMs are trained. This training data reflects the most common demographics and viewpoints online, which explains how prevalent perspectives often dominate. For example, we can see this by the over-representation of Western contexts in its outputs.
You can manage this bias to some extent by varying your prompts and comparing the outputs to see where better wording can achieve a more balanced result.
A Note On Privacy And Data Handling
The general rule is to treat LLMs like any other external API and avoid sending personal data unless your agreement explicitly allows it. For anything confidential, be sure to use any privacy features available to you.
It's all too easy to treat your LLM as a "co-worker"; after all, it talks to you like one. However, you must ensure that you keep data processing within your organization's environment wherever possible and draw up staff policies for its use.
Achieving Better Results
Understanding these fundamentals may change how you work with LLMs. The more explicit about tone, format and success criteria you can be, the more likely it is that your output will reflect what you're looking for.
However, the work isn't done even then. You'll still need to read it in detail, edit it and verify any facts or statistics it's given to you. Remember, it likes to "guess" and supply you with a probable output rather than leave a gap.
Most importantly, iterate and refine your approach through experimentation. The clearer your mental model, the better you'll debug problems, optimize costs and extract genuine value from these remarkable tools.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?