LLMs Unveiled: Understanding AI Training

In this week's episode of Catholic Tech Table, Andrew explains in plain language how large language models like ChatGPT, Claude, and Gemini are trained and why that matters for ministry and fundraising. He notes what training doesn’t provide: human-like reasoning or your parish’s specific context. Andrew ends with three weekly challenges: a pattern test, a context test, and checking the model’s cutoff limits.

INTERVIEW TRANSCRIPT

Howdy. I'm Andrew Robison and this is the Catholic Tech Table. Today we're going to be looking under the hood to figure out how LLMs, such as ChatGPT, Claude, and Gemini are actually trained. Now, I'm going to be honest — I am not a computer guy, and when I first started learning about AI, it felt like there was a lot of information that I just didn't understand.

But what I've discovered is that you don't actually need a degree in computer science to understand the basics, and understanding how these models are built will actually help you know what the limitations are and what they can do.

So today we're going to be breaking down the training process that these companies put these LLMs through in plain language. Think of it like baking bread. You don't need to be a baker, but understanding the basic process will help you appreciate what you're working with and help you appreciate what might go wrong.

So let's dig in.

LLMs learn by reading a lot. Here's the simplest way to think about it: LLMs are trained by reading massive amounts of text. That includes books, articles, websites, and conversations, and the model learns the patterns about how language works. So it's not memorizing everything — it's learning the patterns. For example: after the phrase "good morning" comes a greeting; often when someone asks a question starting with "how do I," an answer usually involves steps or instructions; and business emails tend to have more structure and certain business phrases.

The scale of the information that these LLMs are trained on is really hard to imagine. We're talking billions and sometimes trillions of words. They used sources like books, Wikipedia, news articles, scientific papers, and yes, even Reddit conversations. GPT-3, for example, was trained on about 45 terabytes of text — that's roughly 300 billion words.

Here's an important clarification, though: the model isn't searching through all that text when you ask it a question. It learned patterns from that text, and now it's generating new text based on those patterns.

So how does this work in ministry and in fundraising? Think about when you learned to write appeal letters. You didn't memorize every appeal letter ever written, but after reading dozens or even hundreds of them, you internalized the patterns: how to open with a story, when to ask for a gift, how to close with gratitude and a P.S. LLMs do something similar, but with vastly more text and far greater mathematical precision.

This is why LLMs can sound so natural — because they've learned the patterns of how humans write and speak. But it's also why they can be confidently wrong. They're predicting what sounds right based on patterns, not accessing a database of verified facts every time you ask a question.

So let's talk about the three stages of training.

Stage one is called pre-training, and this is where the model learns language patterns. This is the foundation. The model reads massive amounts of text and learns basic language patterns. It's essentially playing a game of fill-in-the-blank millions of times: "The capital of France is ___" → Paris. "To make coffee, you need water and ___" → grounds.

This takes an enormous amount of computing power — we're talking months of processing on powerful computer clusters that cost literally millions of dollars just to run. At the end of pre-training, you have a model that can generate coherent text, but it's raw — not yet helpful for specific tasks.

That's where we go to stage two, which is called fine-tuning, and this is where the model learns to be helpful. This is where the model learns to be a useful assistant rather than just a text generator. Human trainers give the model examples of good responses: how to answer questions clearly, how to refuse inappropriate requests, how to admit when it doesn't know something, and how to be helpful without being harmful. This is a smarter, more focused training process — think thousands of examples rather than billions.

After stage two, we go to stage three, which is alignment — this is where the model learns human preferences. This is kind of like a final polish. They often call this RLHF, or Reinforcement Learning from Human Feedback. Human reviewers compare multiple AI responses to the same questions and rank them based on which answer was more helpful, more accurate, and more appropriately cautious. The model learns from these preferences and adjusts to give responses more aligned with what humans actually find useful.

So why does this matter for you? Understanding these three stages can help explain why AI can sound confident about things it doesn't actually know — that comes from pre-training. It can also help you understand why different AI models have different personalities and approaches. Claude answers questions differently than ChatGPT, which is different from Grok, because of fine-tuning choices. And AI companies keep releasing updated versions, which is how you get ongoing alignment improvements through RLHF.

So that's the three stages of training an LLM.

Now let's talk about what's not in the training. There are important things that LLMs do not learn during training.

Number one, they don't learn to reason like humans. They're predicting patterns. They're not thinking through problems logically step by step. This is why they can be great at language tasks but struggle with complex math or logic puzzles. They've gotten better at appearing to reason, but at the end of the day, they're still pattern matching at their core.

Number two, they don't learn your specific information. If you work in a parish, your database isn't in their training. Your donor preferences aren't known to them. Your diocesan priorities aren't part of their knowledge. You have to provide that context to them for them to be helpful and respond to your specific situations.

Training gives LLMs impressive language abilities, but it doesn't give them wisdom, truth, or discernment. It doesn't give them access to your specific content. That's still your job.

Now let's put what you just learned about training into practice. Here are three challenges for this week.

Challenge number one: the pattern test. Open ChatGPT or Claude, and start typing a common phrase like "once upon a time" and let it autocomplete. Then try "Dear Parishioners" and see what it generates. Notice how it's predicting based on patterns it learned. This helps you see that it's not thinking — it's pattern matching. Ask yourself: when is pattern matching helpful, and when might that be a problem?

Challenge number two: the context challenge. Ask Claude about something specific to your ministry — like, "What are the top priorities of my parish this year?" Notice how it can't actually answer because it doesn't have your context. Then ask the same question but provide that context: "My parish is focused on youth ministry and debt reduction. What are some ways to message these priorities to donors?" See how much better it performs when you give it that context.

Challenge number three: the knowledge cutoff test. Ask an LLM a question about something very recent — something from last week — and see what it tells you about its limitations. Then compare that to Googling the same question.

Based on understanding how LLMs are trained, can you think of one task where AI would be really helpful and one task where it would be inappropriate or unhelpful? I'd love to hear what you think. Put your ideas in the comments, or email me at [email protected].

So that's it for today's episode of the Catholic Tech Table. Hopefully understanding how LLMs are trained can help you be a wiser and more discerning user of these tools — so you know what they're good at and what they're not good at. Remember: these tools learn from human writing, but they're not human. They're powerful pattern-matching systems that can help with language tasks, but they still need your judgment, your context, and your mission focus to be truly useful.

Next week we'll be comparing some of the most popular models — ChatGPT, Claude, and Grok — to help you understand how to make the best use of each model based on their strengths and weaknesses. Until then, try those challenges and let me know what you're learning. Keep asking good questions. I'm Andrew Robison at the Catholic Tech Table. Thanks for pulling up a chair and learning right along with me. God bless.