AI Buzzword Olympics: Topic #2

"We will just Fine, Tune the Model" Or how two words became the universal solution to every AI problem.

"Fine, tuning has become the corporate equivalent of: 'Have you tried restarting your laptop?'"

Last week it was, "We will train the model." This week, we have become more sophisticated.

"No, no. we don't need to train the model. We will just fine, tune it."

Excellent.

The word fine, tune has now entered the meeting. Everyone nods. The AI strategy suddenly sounds much more intelligent. So let's ask the question nobody wants to ask:

What exactly are we fine tuning?

A foundation model already knows language, code, mathematics, general knowledge and a lot of patterns. Fine tuning does not mean teaching it everything again. It means adjusting an already capable model so it performs better for a specific task, domain, style or behavior.

Think about hiring an experienced engineer. They already know software engineering. You don't send them back to university. You teach them your business, your products, terminology, workflows, customers and exceptions.

That is fine, tuning. You are not creating intelligence. You are adapting existing capability. And here's where the corporate confusion begins. Uploading 10,000 PDFs is not automatically fine tuning. Connecting SharePoint is not fine tuning. Adding documents to RAG is not fine tuning. Writing a longer prompt is not fine tuning. Building a chatbot is not fine tuning.

Yet somehow, in enterprise architecture meetings, all five have been called "fine tuning." Apparently, everything is fine tuning now. So before reaching for fine tuning, ask a simpler question: What problem are we actually trying to solve?

Need current enterprise information? You probably need RAG. Need better instructions or output format? Try prompt engineering. Need the model to consistently behave in a particular way or perform a specialized task? Now fine tuning might make sense.

Need to create a completely new foundation model? That is a very different, and very expensive, conversation. Fine tuning is not free magic. You need quality training data, usually labelled examples, compute, experimentation, evaluation and engineering expertise. You will run training, measure the result, adjust, run it again, and sometimes discover that the expensive new model isn't actually better.

Then someone asks: "Can we just fine tune it a little more?"

That is usually when the GPU bill starts becoming a leadership discussion. Techniques such as LoRA can make adaptation considerably more efficient by updating a much smaller set of parameters rather than modifying the entire model. So What are we actually changing?

This is where a few more AI buzzwords usually enter the room.

LoRA: Low Rank Adaptation sounds complicated, but the idea is simple: instead of changing billions of parameters in the entire model, you update a much smaller set of additional parameters. Think of it as modifying a few critical components instead of rebuilding the engine. It makes fine, tuning significantly cheaper and faster, which is why it's widely used for adapting large models.

Then there is PEFT: Parameter Efficient Fine Tuning. It's essentially the broader idea behind techniques like LoRA: don't retrain the whole model when you only need to change a small part of its behavior.

And then you hear RLHF:  Reinforcement Learning from Human Feedback. In simple terms, humans compare model responses and indicate which ones are better. The model then learns from those preferences. In corporate language:

Even AI gets performance reviews. These techniques are useful because full model fine, tuning can be expensive and computationally heavy. Instead of modifying everything, we try to make the smallest effective change. But here is the important distinction: RAG gives the model more information. Prompting gives the model better instructions. Fine, tuning changes how the model behaves. LoRA makes that behavioural change more efficient. Training from scratch creates the model itself.

Those are very different engineering decisions. And that is exactly why saying "We will fine tune it" is not a strategy.

It's just the beginning of a conversation. The best AI engineers are not the ones who know the most acronyms. They are the ones who know which acronym they don't need.

So the next time someone confidently says: "We will just fine tune the model." Don't ask how much data you have. Don't ask which GPU. Start with one brutally simple question: "What behavior are we trying to change, and why can't we solve it without changing the model?"

If the room suddenly becomes quiet, Congratulations. You have just turned an AI buzzword into an engineering discussion. That is the AI Buzzword Olympics. Where the goal is NOT to know the most AI terms. It is to know what they actually mean. Stay Tuned for the next event : AI Buzzword Olympics: Topic #3

"Let's Build Our Own LLM"