Picking up where we left off
In the previous two articles we covered two things:
- AI isn't magic. It's mathematics, plus data, plus computing power (article 1).
- A model learns through optimization, and is only as good as its data (article 2).
Today we step into the most visible part of this revolution: LLMs.
Everyone is talking about them. ChatGPT, Claude, Gemini, Llama. Few people know what they actually are, what makes one different from another, or how to pick the right one for their case. Let's make it clear.
What an LLM is
LLM stands for Large Language Model.
At its core, an LLM does one and only one thing: it takes a piece of text and predicts the next word.
More precisely, the next token (which can be a word, part of a word, or a punctuation mark). That's it.
Give the model: "The capital of Greece is..." and it will predict: "Athens."
Looks simple. It is simple. The strange part is that from this one ability, all the behavior we see emerges. Conversations, summarization, translation, code, explanations, poetry.
How? Because in order to correctly predict the next word at a random point in a random sentence, the model needs to have internal representations of:
- grammar
- semantics
- the world
- logic
- tone
All of these are learned implicitly. Nobody programmed them explicitly. They emerged because without them, the prediction is bad.
Why we call them "large"
The L in LLM stands for Large, and it's literal. We're talking about models with billions or trillions of parameters (the "knobs" from the previous article).
To picture the scale:
- GPT-3 (2020): 175 billion parameters.
- GPT-4 (2023): not public, estimated over 1 trillion.
- Llama 3.1 from Meta (open source): 8 to 405 billion.
- Claude and Gemini don't publish numbers for their latest versions.
The trend is clear. The bigger they get, the better they generalize. Up to a point. In the last several months researchers have been talking about a plateau: quality no longer scales proportionally with size. New training methods (like reasoning training) are becoming more important than pure scale.
How they actually work
Every modern LLM is built on the same architecture: the Transformer, born with the paper «Attention Is All You Need» (2017) from Google researchers.
I won't go into detail. The only thing you need to remember:
- The model sees all of its input at once, not word by word.
- For each point in the text, it computes how much attention to pay to every other point. That's "attention."
- From there it produces the next prediction.
Every LLM you use today is a variation on this idea.
The major players, one paragraph each
ChatGPT (OpenAI): the one that went mainstream first. Very good at general tasks, best output for content creation in English. Depends entirely on API calls to external servers, currently in the US.
Claude (Anthropic): focused on reasoning quality and safety. Many researchers consider it the most accurate model for complex tasks. Handles Greek well, better than ChatGPT for specialized text.
Gemini (Google): strong on multimodal tasks (text + image + video). Natively integrated with Google Workspace. Good choice if your company already lives in the Google ecosystem.
Llama (Meta): open source. Meaning you can run it on your own servers. Crucial for companies with sensitive data. A little behind the top closed models in quality, but wins the increasingly important control dimension.
There is no single "best." There's the right one for your case.
How to choose
Practical criteria when deciding which model to base a solution on:
-
Sensitive data: if you process personal data or critical information, an open model like Llama on your own servers is the safest option. Otherwise check whether the provider offers zero data retention.
-
Greek language: Claude currently gives the most natural Greek output. ChatGPT is still good. Gemini is mediocre. Test with your own texts before deciding.
-
Latency: if you're building a real-time chatbot, the smaller model usually wins. A 2-second delay is acceptable, 8 seconds is not. Every provider has smaller, faster variants (Haiku, mini, flash).
-
Cost: the gap between a small and a frontier model is 10x to 100x. For 80% of tasks, the small model is enough. Don't route all traffic through the most expensive one.
-
Context window: how much text the model can "hold in mind" per call. Today ranges from 8,000 to over 1,000,000 tokens. If you want to feed it a 200-page PDF, you need a large context window.
What an LLM is NOT
A few misconceptions that exist:
-
Not a search engine. The LLM doesn't "search" the internet. It generates text based on what it learned during training. When you see ChatGPT searching online, that happens through an external tool (web search), not from the model itself.
-
Doesn't know what happened today. Every LLM has a knowledge cutoff, a date after which it knows nothing. Usually 6 to 12 months behind the current date. If you need fresh data, you need retrieval from external sources.
-
Not deterministic. Two runs of the same prompt can produce two different answers. That's by design. For decisions that require consistency, you need to tune the temperature parameter or wrap the call in structured output validation.
-
Doesn't remember past conversations. Every API call sends the entire history. The "memory" in chat applications (ChatGPT, Claude.ai) lives outside the model, in external storage.
What this means for your business
Two thoughts I want you to leave with:
First: model choice is not a one-time decision. In a serious AI deployment, we switch models per task. Small model for classification, mid for summarization, large for complex reasoning. That keeps cost low and quality high.
Second: don't lock yourself into one provider. The field changes every month. The "best model" today is not the same in three months. Build your infrastructure so you can switch providers in hours, not weeks.
What's coming in the next article
In the next article we'll look at where AI fits in a business. Concrete examples, no hype. Where it works today, where it doesn't, and how we tell the two cases apart.
As always, if you have questions along the way, email me →. I read every email and reply personally.
Thank you for following along.