Leon Furze's AI Iceberg
Scroll down to dive beneath the surface of AI, exploring the hidden layers of large language models — from the applications you see to the vast ocean of training data you don't.
Most people interact with AI through polished applications like ChatGPT, Claude, or Gemini. But like an iceberg, what you see is only a tiny fraction of what's really there. The vast majority — the training data, the model architecture, the hidden processes — lies beneath the surface.
Picture a carefully sculpted snowman sitting on top of the iceberg. This represents an application like ChatGPT, Claude, or Gemini — built on top of a general LLM and fine-tuned specifically for conversational tasks.
The chatbot layer includes system messages (rules governing every interaction) and RLHF (reinforcement learning from human feedback), where human reviewers judge outputs as helpful, harmful, correct, or incorrect — shaping how the model responds.
The visible iceberg above the waterline is the Large Language Model itself — the result of the training process fuelled by the vast dataset beneath. This is where three key technical processes happen.
Tokenization breaks text into machine-readable pieces. Weighting assigns probability values to connections between tokens. And the transformer architecture with its attention mechanism — introduced in 2017 by Google researchers — allows the model to consider relationships across entire passages of text.
LLMs are often called "black boxes" because their internal connections are so massively complex that no human or team of humans could possibly unravel everything going on inside.
The bulk of the iceberg — hidden underwater — represents the vast dataset on which the LLM is trained. GPT-4 trained on roughly 13 trillion tokens — about 1,600 times the population of Earth.
Known data sources include Common Crawl, The Pile, Wikipedia, GitHub, and social media. Much of the data remains proprietary. This layer carries plenty of side effects, including the kind of bias and discrimination central to AI ethics discussions.
The ocean surrounding the iceberg represents the internet at large — the vast environment from which the dataset is sourced.
And the sharks? They symbolize threats lurking in the data: misinformation, bias, toxic content, and data privacy issues that can influence the dataset and, subsequently, the behaviour of the LLM.
“A language model is a stochastic parrot — it haphazardly stitches together sequences of linguistic forms from its training data, without any reference to meaning.”
— Emily Bender et al., “On the Dangers of Stochastic Parrots” (2021)
Let's open the black box. The following sections explore the key processes that make large language models tick — from breaking text into tokens to predicting the next word.
Here's the single most important thing to understand about how LLMs work: they build text one word at a time, each time predicting the most likely next token based on everything that came before.
Think of the autocomplete on your phone. You type "I'm on my" and it suggests "way." An LLM does the same thing — but with vastly more context, vastly more data, and vastly more sophistication. It doesn't "understand" your message. It predicts what comes next.
As AI researcher Gary Marcus puts it, large language models are “glorified autocomplete” — impressive in their output, but fundamentally just predicting the next word.
Before an LLM can process language, text must be broken into small pieces called tokens — like Scrabble tiles the model can work with.
Scroll to see how text is broken into tokens
Original text
The animal didn't cross the street because it was too tired
Tokens
Before an LLM can process text, it must break it into tokens — small pieces it can work with, like Scrabble tiles.
Tokens are not always whole words. The tokenizer uses Byte-Pair Encoding (BPE)
to split rare or compound words into smaller, more common pieces. Notice how
“didn’t” becomes two tokens: didn and ’t.
Each token is then mapped to a numerical ID — the only thing the model actually sees. GPT-4’s vocabulary contains roughly 100,000 possible tokens.
As Andrej Karpathy explains, “a lot of weird behaviors and problems of LLMs actually trace back to tokenization.”
Each token becomes a numerical vector — essentially GPS coordinates in "meaning space," where similar concepts naturally cluster together. These positions are learned during training, not hand-coded.
Scroll to explore meaning space
The breakthrough that made modern AI possible. The attention mechanism lets each word look at every other word to understand context — resolving ambiguity in ways previous architectures couldn't.
Scroll to see attention in action
When the model encounters the word “it”, attention lets it look at every other word and decide which ones are relevant. The thicker the line, the more attention “it” pays to that word.
“What am I looking for?”
The word “it” asks: “What noun do I refer to?”
“What do I contain?”
Each word advertises its identity, like a library catalogue entry.
“Here’s my actual information.”
The content retrieved once Query matches a Key.
The word “mole” means completely different things depending on context. Attention lets the model figure out which meaning is intended by attending to surrounding words.
Tokenization, embeddings, and attention combine inside a transformer — the architecture introduced in Google's landmark 2017 paper "Attention Is All You Need" that powers every modern LLM.
Scroll to open the black box
GPT-4 has an estimated 1.8 trillion parameters — numerical values learned during training. That’s why it’s called a “black box”: no human could unravel all those connections. But we can understand the architecture. Let’s open the box.
A raw LLM is powerful but unruly — it might generate toxic content, confidently state falsehoods, or ignore your actual question. That's where RLHF comes in.
After the base model is trained on trillions of tokens, human reviewers step in. They're shown pairs of model outputs and asked: which response is more helpful, more accurate, less harmful? These preferences are used to train a reward model — a separate system that learns to score outputs the way humans would.
The LLM is then fine-tuned using reinforcement learning to maximise that reward score. Over many iterations, the model learns to produce responses that are helpful, honest, and safe — not because it understands these concepts, but because it's been optimised to match human preferences. This is what turns a raw text predictor into a usable assistant like ChatGPT or Claude.
RLHF is the "snowman sculptor" — the process that shapes the raw iceberg into the polished application sitting on top.
But there’s a hidden cost. The humans doing this work — many employed through outsourcing firms in Kenya, the Philippines, and other countries — are often paid less than $2 per hour to review and classify content, including violent, sexual, and deeply disturbing material. A 2023 Time investigation revealed that workers labelling data for ChatGPT described the work as “torture.” This is the invisible human labour that makes AI “alignment” possible.
“Artificial intelligence is neither artificial nor intelligent. It is both embodied and material, made from natural resources, fuel, human labor, infrastructures, logistics, histories, and classifications.”
— Kate Crawford, Atlas of AI (2021)
Which response is better?
Through thousands of comparisons like these, human feedback shapes the model's behaviour — making it more helpful, safer, and more accurate.
Image generators like DALL-E, Stable Diffusion, and Midjourney use a different technique called diffusion. Imagine ink diffusing in a glass of water until it's uniformly mixed — then learning to run the process in reverse, starting from pure noise and gradually "un-mixing" it into a coherent image, guided by a text prompt.
Scroll to watch noise become an image
Diffusion models like DALL-E and Stable Diffusion learn to create images by reversing a noise process. Imagine ink dropped in water, diffusing until uniformly mixed. These models learn to un-mix the ink, step by step, guided by a text prompt.
Starting from pure random noise — no discernible pattern. This is the equivalent of fully diffused ink in water.
The model begins its first denoising pass. It has learned statistical patterns from millions of images during training.
Broad shapes begin to emerge. The text prompt “snowman face on blue background” guides the model via cross-attention with a CLIP text encoder.
Colors start separating — whites clustering where the face should be, blues filling the background region.
The pattern becomes clearly recognizable. Each denoising step refines the prediction from the previous step.
Almost there. Fine details resolve. In practice, diffusion models run 20-50 of these denoising steps.
The final denoised image — a snowman face on a blue background, generated entirely from noise, guided only by text.
The terminology we use for AI — “learns,” “thinks,” “understands” — is what researcher Melanie Mitchell calls wishful mnemonics. These words describe what the system appears to do, not what it actually does. An LLM is a sophisticated statistical prediction engine, not a mind.
That’s why Leon Furze chose the iceberg metaphor: “I’m deliberately avoiding any kind of analogy that represents the AI as magical, mythical, human, or godlike — we’ve seen enough of them.” The iceberg emphasises hidden infrastructure, not hidden intelligence.
What about “reasoning” models like OpenAI’s o1 or DeepSeek-R1? These models are trained to generate a chain of thought — working through a problem step by step before answering. But the mechanism is identical: they’re still predicting the next token. The “reasoning” is just more tokens — the model has learned that writing out intermediate steps leads to better final answers. It’s autocomplete that’s been taught to show its working.
The Critical Guide to GenAI in Education
Stay ahead of the curve. Subscribe for articles, resources, and updates on understanding generative AI — delivered straight to your inbox.
Subscribe to the mailing list