LLMs Explained: How GPT, Gemini & Llama Actually Work

Large language models feel like magic until you understand the four pieces under the hood: tokens, the transformer block, pretraining, and alignment. Once those four click, the differences between GPT, Gemini, and Llama become a vocabulary you already know.

Tokens, not words

An LLM does not see words. It sees tokens — sub-word chunks produced by a tokenizer. “Innovations” might be one token; “ChennaiMetro” might be three. Token counts drive context windows and pricing, which is why one of the first skills any AI engineer learns is counting tokens before sending a request.

The transformer in one paragraph

A transformer block does two things: self-attention (each token looks at every other token to decide what matters) and a feed-forward network (it transforms that information). Stack 40 of those, train on trillions of tokens, and you have a base model that predicts the next token well.

What makes GPT, Gemini, Llama different

• GPT (OpenAI) — closed weights, strongest reasoning at premium tiers, mature tool-use ecosystem.
• Gemini (Google) — strong multimodal (image, video, audio), tight Google Workspace integration.
• Llama (Meta) — open weights, you can fine-tune and self-host, smaller variants run on a single GPU.

Why this matters at work

Picking between hosted GPT and self-hosted Llama is a real business decision involving privacy, cost per token, and latency. Understanding the three families lets you contribute to that conversation as a junior engineer.