How AI Works: A Simple Musical Guide to Large Language Models

TL;DR: Large Language Models (LLMs) don’t “think.” They split text into pieces (tokens), turn those pieces into numbers (embeddings), track their order (positional encoding), weigh importance (attention), refine through layers (transformers), score probabilities (softmax), and generate new text one token at a time. Think of it like building the perfect playlist: songs become words, playlists become sentences, and the system predicts what fits next.

Why use music to explain AI?

Technical terms like tokenization and vector space can lose most audiences in seconds. Music is different. Everyone understands the flow of a playlist, the fit of a song, the way genres cluster. That makes it the perfect analogy for how LLMs work under the hood. 

Picture a playlist with Oasis classics like Wonderwall and Don’t Look Back in Anger. The vibe is set. But what song comes next? Too similar feels stale, too different breaks the flow.

That’s the exact challenge LLMs face: given the words so far, what’s the most likely continuation that “fits”? The model’s job is playlist-building for text.

How Does AI Predict the Next Word?

Computers can’t read words as we see them. They need to break them into tokens the smallest useful pieces of text. Think of it like chopping a song into beats. Each beat can be processed, compared, and reassembled. Without tokenization, an LLM can’t analyse or generate language.

Once broken down, tokens are turned into embeddings: lists of numbers that capture their meaning and context. Imagine a giant festival map.

Oasis and Blur sit near each other because they often appear on the same playlists. Dolly Parton ends up closer to Willie Nelson. The model doesn’t know guitars or lyrics, only that some songs appear together more often than others.

This “music-space” is the vector space of AI, where proximity equals similarity.

Why does word order matter?

Playlists aren’t random. Starting with Wonderwall feels different from ending with it. LLMs use positional encoding: signals that tell the model where tokens sit in a sequence. Without it, “Today is gonna be the day” could come out jumbled. With it, the system knows order is part of meaning.

Some songs define a setlist. Drop Don’t Look Back in Anger and the whole crowd shifts. LLMs use self-attention to assign weight to tokens based on context. In a hospital sentence, “doctor” and “nurse” matter more than “coffee.” In lyrics, “gonna” and “be the day” are tied closely together, so the model treats them as linked.

 Attention isn’t enough on its own. The system needs refinement. That’s where transformer blocks come in: layers that combine self-attention with feedforward networks to polish meaning. Each pass is like remastering the track until the whole setlist flows. The more blocks stacked, the sharper the model’s grasp of how words fit together whether in an Oasis lyric or a policy email or a marketing blog.

How does the model pick the next word?

At the end, the model faces a choice. Every possible token is scored. The softmax function converts raw scores into probabilities: “guitar” 80%, “percussion” 15%, “saxophone” 0.1%. The most probable is chosen, like picking the song that best fits the playlist. The output you want is actually built one word at a time: LLMs are autoregressive. That means once one token is chosen, it’s added to the prompt, and the cycle repeats:

predict, add, predict, add.

It’s like building a playlist track by track. After Wonderwall, the model sees Champagne Supernova as likely next. Add it, then look for the next fit.

So for example if you prompt your favourite model (ours is Gemini from Google) with:

“When I’m away, I will remember you…”

From billions of sentences, it predicts the continuation: Ed Sheeran’s Photograph:
“I will keep you inside the pocket of my ripped jeans…”

It doesn’t “understand” love or memory. It follows statistical patterns: tokenization, embeddings, positional encoding, attention, transformer layers, probability scoring, and autoregressive generation.

The result feels human because the system is excellent at capturing flow, not because it grasps meaning.

So Why do LLMs Sometimes Fail?

The playlist app sometimes suggests ABBA after Metallica because of one odd data point. LLMs can do the same. 

They work with patterns, not comprehension.

That’s why they can produce fluent nonsense sentences that sound right but aren’t.

The AI learning that matters

Large language models aren’t mystical. They’re playlist builders at scale: splitting text into tokens, mapping meaning into vector space, tracking order, weighing importance, refining through transformers, scoring probabilities, and generating word by word.

The value isn’t in treating them as intelligent beings. It’s in knowing their mechanics so you can use them effectively without over-trusting them.

That’s why the first step for any professional isn’t blind adoption, but building safe foundations. If you want to go further, our AI Fundamentals Masterclass gives you the grounding to use these systems with confidence, without needing a technical background.

FAQs

Q: Do LLMs understand meaning?
No. They recognise and predict patterns in text, but they don’t comprehend concepts the way humans do.

Q: Why do LLMs sometimes get facts wrong?
They predict the most likely continuation, not the most accurate one. If the training data contained errors or the context is thin, the model can generate plausible but false information.

Q: What’s the difference between training and using an LLM?
Training builds the map of word relationships (vector space). Using an LLM means giving it a prompt and letting it generate outputs from that pre-trained map.

Q: How is this different from search engines?
Search retrieves existing information. LLMs generate new text word by word, based on probabilities.