What You Need to Know About Large Language Models (LLMs)
In the past year, we’ve seen a pattern emerging. As you’re typing in a box, the application tries to finish your sentence. You may have interacted with a chatbot on a website. Your browser’s search bar seems to “understand” what you meant, not just what you typed. These moments are all powered by the same kind of system: a large language model (LLM). In this post, I’ll explain in plain language what LLMs are, how they work, and what you actually need to know about them – no math, no jargon, just a clear guide for non-specialists. This post is a companion to my post on Supervised Learning and the post on Understanding Machine Learning.
What an LLM Is
An LLM is a system that predicts the next word. LLMs are trained using lots of text. You can think of it as being similar to someone who has read the entire contents of the New York Public Library and is very good at finishing sentences.
How to Train an LLM
The simple explanation of training an LLM is feeding it a lot of data and having it guess the next word in a particular piece of text. You can think of training as putting the model through trillions of tiny next‑word quizzes. It reads a snippet of text, guesses the next word, sees the real answer, and tweaks itself to do a little better next time. The LLM learns patterns and probabilities, helping it choose the next word.
Over time, this constant cycle of guessing and correcting teaches it not just grammar, but patterns of reasoning and style across many topics. Specific uses like support, coding, and drafting call for some fine-tuning and testing. After the model becomes proficient at the task, the engineers “freeze” the model, telling it that it is no longer in training mode. They make the LLM available behind an interface like a chatbot, where it uses what it has already learned to respond to your prompts in real time.
What LLMs are Good At
LLMs do very well at drafting (reports, emails, and blog posts – but not this one, I write it myself), brainstorming (helping me come up with ideas for this blog), and rephrasing (when a sentence just feels awkward). I have used LLMs for translation when a contractor on our property spoke only elementary English, and I needed him to understand a complex task. I asked an LLM to translate my English instructions into the contractor’s language, and then I took that translation to another LLM and asked it to translate it back to English to verify the translation. It did a great job at it, and we got exactly what we wanted.
LLMs are good at these tasks, true, but there are benefits beyond simple availability. LLMs are very fast at providing a response, because, despite Perplexity’s (my preferred LLM) message of (Thinking…), it’s not really “thinking.” It’s simply composing the next most likely word along the string.
Persistence
You can also count on an LLM to be persistent, both in a session and over time. Within a session, an LLM will remember what you told it earlier. It becomes sort of a conversation. If you ask a question that is relevant to something earlier in the conversation, the LLM will often reference that point. Additionally, if you ask a question and receive an answer in January, and then ask the same question in June. That assumes the conditions are the same, you’ll very likely receive pretty much the same answer.
Your LLM will also be available anytime your connection to it is available. In other words, you can run an LLM on your computer, but it takes a lot of resources. However, as long as your internet connection is stable, your LLM will be available to you online. Additionally, Perplexity will cite the sources it uses to produce the responses, so that I can verify the information or dig deeper on my own. More LLMs are offering that now, which can help guard against some of the drawbacks.
Where They Fail or Mislead
LLMs aren’t perfect, and you shouldn’t consider an LLM to be an authority on — well, anything. That’s why the sources are so important. The second L in LLM is LANGUAGE, not LOGIC, and that’s important to remember. Because the LLM is predicting the next word, despite how many times the LLM gets it right, it’s never going to be 100% accurate. As a result, the LLM can produce a piece of information and present it to you, and tell you that it is perfectly confident in the accuracy of the information – even if that information is completely wrong. We call that a hallucination, and it’s gotten some people into trouble.
LLMs can also contain bias and blind spots from training data. If you tell a piece of technology that everything with four legs is a dog, that technology will identify every four-legged object as a dog, including tables and chairs. If you tell that same piece of technology that the only good vehicle is a New Holland tractor, it may guide you away from buying a car in favor of a New Holland tractor. Obviously, these are simple examples, but if the training contained enough similar biases and blind spots, the LLM would produce responses that reflect those limitations.
One of the most significant aspects that users should consider when using an LLM is that the LLM has no “experience.” It lacks understanding, and it lacks lived context, and context is often helpful for true comprehension. The best it can do is explain what users have found. Sometimes, it’s helpful for a real person to provide some perspective.
Everyday Implications
There are tasks that can be handled quickly by an LLM. For example, I can take a transcript or notes from a meeting, and an LLM can summarize those notes and identify action items and the assignee. I’ve used LLMs to help me troubleshoot problems and figure out PowerShell commandlets to perform common tasks. Sometimes it takes a few tries to get the command right, because, as I’ve emphasized, it’s Language, not Logic.
LLMs can also be helpful for education and learning. However, we need to remember that thing about hallucinations. Unless you can verify the information, relying solely on an LLM for education can be risky. But for an outline of topics to study, it can be very useful.
At a very high level, data owners are still arguing with LLM engineers about copyright questions. Ethical and legal questions will pop up as well from now till eternity. We’ve never had to answer these questions with regard to a non-human that is producing new material from someone else’s work without direct intervention.
How to Use Them Wisely
LLMs can be safe to use as long as you exercise some basic cautions and common sense. Treat all of the LLM’s outputs as drafts, not final answers. This is less important for casual curiosity than for academic or professional work. However, using an LLM’s answers in an online debate may leave you with egg on your face. Always fact-check details, especially numbers, names, and legal or medical information. I saw a post on social media where a guy said a LLM finally agreed with him that 2+2=5 after he said his wife told him that was the correct answer.
Always be careful what personal or sensitive data you feed the LLM. Remember that the processing of the data occurs in the cloud, on servers you don’t own. If you need specific information, provide it as a hypothetical case rather than using your own information. The LLM lacks strong logic and judgment capabilities. That means that it will not be able to decide that your hypothetically-provided information is actually you.
Your Turn
Do you use LLMs for work or personal use? I’d love to hear your use cases. What other questions have I inspired that you’d like cleared up?
My photography shops are https://www.oakwoodfineartphotography.com/ and https://oakwoodfineart.etsy.com, my merch shop is https://www.zazzle.com/store/south_fried_shop.
Check out my New and Featured page – the latest photos and merch I’ve added to my shops! https://oakwoodexperience.com/new-and-featured/
Curious about safeguarding your digital life without getting lost in the technical weeds? Check out ‘Your Data, Your Devices, and You’—a straightforward guide to understanding and protecting your online presence. Perfect for those who love tech but not the jargon. Available now on Amazon:
https://www.amazon.com/Your-Data-Devices-Easy-Follow-ebook/dp/B0D5287NR3
