Perhaps it’s just me; but, I always like to be specific when I’m looking at technology and working out how it’s going to be (or potentially not be) useful. It’s all very well to get on board with the latest fad; but if I don’t understand the fundamentals of the technology at hand, I find it difficult to correctly embrace it. I could, of course, approach the topic by just doing whatever everyone else is doing, or whatever the sellers of the technology pitch the use cases as, but is that really the best approach for me, or the work that I’m carrying out?

Today’s topic is a burning one. ‘AI’.

I put the quote marks around the letters, because in technology, we should be immediately asking what part of AI am I referring to? A simple linear regression? Machine Learning? A deep learning neural network? A transformer model approach? Narrow AI? Or Large Language Model? The options are near endless (and start to actually approach endless when we consider ensemble approaches, and all the options you could have in a Random Forest; But people and influencers today, are usually meaning an LLM model that is available in the market place (and for the sake of this post, I am considering open source to still be ‘the market’).

So what is a market-available LLM actually useful for in the Enterprise space? That depends who you’re speaking to. To start with the people selling such services, you might look towards Microsoft’s CEO, Satya Nadella, who in April 2025, said that ‘as much as 30% of Microsoft code is written by AI’ (source: https://www.cnbc.com/amp/2025/04/29/satya-nadella-says-as-much-as-30percent-of-microsoft-code-is-written-by-ai.html). This sounds truely marvellous as we know that code at such a company like Microsoft is heavily relied upon and needs to be correct and hallucination-free. This speaks to the Co-Pilot LLM that Microsoft retails to its Enterprise clients, and shows them that they are trusting their own model. Fantastic. Of course, fast forward literally four months later, and we get Microsoft putting Co-Pilot into Excel, but also telling users to not use it in ‘any task requiring accuracy or reproducibility’. (Source: https://www.pcgamer.com/software/ai/microsoft-launches-copilot-ai-function-in-excel-but-warns-not-to-use-it-in-any-task-requiring-accuracy-or-reproducibility/).

I’m sorry, pardon me? If Excel is known for NOT being something, it would be it’s tendency to go off the rails from time to time, and just decide to put different formulas in spreadsheets. So, what gives? Why would Microsoft issue such a warning, whilst simultaneously brag about the same system writing 30% of their code?

That’s not for me to say. Truely, it’s not (court cases exist, and I’d like to remain a viewer of them, instead of someone partaking in one).

Do I use LLMs? Yes, often. Do I write code to build ML and NN models? Also yes, less often than I would like. Do I hate the AI hype? That depends. If someone is aware of the benefits and the things to avoid, and understands how the underlying systems work, then no. Hype can be a perfectly reasonable reaction to such an amazing technology.

But they don’t think. They don’t know. They aren’t understanding.

Over a series of posts (I’ve no idea how many), I intend to explain the way that LLMs are trained, how they function, where their brilliant use cases are, and some things to be aware of. This will be a somewhat technical blog, but after a number of conversations where I’ve explained these things to non-technical people, I know that I can share this in a way that can empower people to understand LLMs.

As Arthur C Clark put it: ‘Any sufficiently advanced technology is indistinguishable from magic’.

I want to have as many people as possible become the magicians, instead of the awe-struck audience members. Because the more magicians we have in the world, the better the choices can be with the magic.