IT HootClub — AI Community Newsletter

Most AI newsletters teach you how to use AI. This one is going to teach you how to build it — or at least, how to make it yours.

Over the next several weeks, we're going to take an open-source language model, feed it data that doesn't exist anywhere in its training, and produce something the big three — Gemini, Claude, and GPT — simply cannot replicate: a model that genuinely understands the Milwaukee tech ecosystem. Local colleges and their AI programs. Regional companies. Events, initiatives, and the people building things here. Ask any of the major models a specific question about Milwaukee's tech scene right now and you'll get a generic, often inaccurate answer. That's the gap we're attempting to close.

I want to be upfront about something: I don't know what the outcome of this will be. I've run a fine-tuning process before — once, on a small dataset, without fully understanding what I was doing. This series is me learning in public, documenting every step, every decision, and every result — including the ones that don't go as planned. If the fine-tuned model turns out to be worse than what you can already get from ChatGPT, we'll say so. That's the deal.

When it's done, we'll put our model in the ring with the big three and ask them all the same questions. Side by side. Unedited. You'll see exactly where a small, focused, locally-trained model wins — and where it doesn't. There are real questions here worth exploring honestly: Can a smaller model trained on specific data actually outperform a much larger general model on a narrow domain? And if the big three have already crawled the same public internet data we're training on — does fine-tuning even give us an advantage? We're going to find out together.

But here's the larger point — and it matters beyond this newsletter. The outcome of this project isn't just about Milwaukee or one model. It's a demonstration of something the industry is still coming to terms with: a well-executed, domain-specific fine-tuned model can be more effective for a business than a general-purpose model from one of the big three. Not because it's smarter — it isn't. But because intelligence applied to the right problem, with the right data, beats raw capability applied broadly every time.

That's not a guarantee. It's a hypothesis. And proving or disproving it requires something the big three can't sell you: genuine, targeted work. Not prompting. Not clicking through settings in a vendor's fine-tuning portal. Real decisions about data, architecture, and evaluation — made by people who understand what they're building and why. That's the skill this series is about. There is real room to stand out in this industry precisely because most organizations are still treating AI as something you subscribe to rather than something you build.

This isn't a theoretical exercise. The model will run on real hardware — the GPU lab right here at WCTC — and every step of the process will be documented in this newsletter as we go. The tools, the decisions, the things that didn't work, and what we learned from them. If you're a student, a practitioner, or someone advising organizations on AI adoption, this series is built for you.

Before we get into the project itself, we need to settle one foundational question — because everything else builds on it: what does an AI model actually know, and how does it know it?

The honest answer is that it doesn't "know" anything the way you do. A language model is a massive compression of patterns extracted from text. During training, it reads an enormous amount of human-written content and learns, statistically, what tends to follow what. It doesn't store facts in a filing cabinet. It encodes relationships between words, concepts, and ideas as billions of numerical weights — and when you ask it something, it reconstructs a plausible response from those weights.

That's why it can be wrong with total confidence. It's not lying — it's pattern-matching to the most statistically probable answer, even when the correct answer isn't the probable one.

Here's the analogy that makes this click: imagine someone who read every book in a library, then the library burned down. They can tell you a lot. They can reason, summarize, explain, and synthesize. But if you ask them about a document that was never in the library — or what changed last Tuesday — they'll try to answer anyway. That confident attempt at an answer they don't actually have is what the field calls a hallucination. It's a structural limitation, not a bug someone forgot to fix.

This is the core problem we're working on. The big three don't have Milwaukee in their library — not really. Their training data is so vast that local, regional content gets diluted to near-zero signal. A sentence about the Milwaukee tech scene that appeared somewhere on the internet gets averaged across billions of other tokens. It's there, statistically. It just has almost no influence on what the model actually says.

Fine-tuning is one way to change that. The three techniques that make this possible — and that we'll be working through over the coming weeks:

• Fine-tuning — Taking an existing model and retraining it on new data so it internalizes specific knowledge. This is the core technique for our project. It changes what the model structurally knows, not just what it can look up.
• Embeddings — The mechanism that converts text into numbers representing meaning. "Similar" ideas end up numerically close to each other. Understanding this is essential for understanding why some training data works better than others.
• RAG (Retrieval-Augmented Generation) — Giving a model access to external documents at query time, rather than baking everything into the weights. We'll cover this as the natural next layer after fine-tuning — and explain when you'd use one versus the other.

Next week: the model choice. What the numbers actually mean — 7B, 13B, 70B parameters — and the question nobody seems to ask directly: if bigger models are more capable, why aren't we just fine-tuning one of those?

Building Intelligence runs weekly. All editions archived at hardais.com/newsletter.

How do I train an AI model?