The vital tech powering GPT, Midjourney, Dall-e, and the rest of their kin is the Large Language Model, abbreviated “LLM” by the savvy. LLMs are the kind of neural network responsible for an AI’s ability to interpret and respond to prompts written in natural language – that’s the kind we use in everyday speech and writing, as opposed to the formal, mathematical languages used to instruct computer hardware and argue with the stubbornest academics.

In a very basic sense, LLMs achieve this fluency by reading a bunch of human-written texts (in the form of publicly available web data, known as “training data”) and analyzing their construction, searching for regularities in the way words and phrases are arranged. For LLMs, this analysis produces a kind of “language-map”, where words and phrases likely to appear in similar linguistic environments are placed closer together. So ‘Apple’ and ‘Peach’, which usually appear in very similar types of sentences –

• “I ate some _____ pie for lunch!”
• “That _____ was delicious.”
• “I went to the _____ orchard today!”

– would be very close to each other on the map, like two neighboring suburbs of an imaginary Fruit City. Mostly unrelated words like ‘Apple’ and ‘Sewage’ would hopefully be separated by a considerable distance, perhaps segregated to opposite coastlines. ‘Red-Delicious’ and ‘Granny-Smith’ could be almost on top of each other (like overlapping districts in the town of Apple) as might other near-synonyms like ‘Alligator’ and ‘Crocodile’, or ‘Mountaintop’ and ‘Summit’.

Once the map is complete – meaning the LLM has finished analyzing its training data, plotting every word found in its massive dataset in a unique location on the map – the language model is ready to fulfill a pre-programmed objective. In the most straightforward case of text generators like ChatGPT, that objective is simply to predict the next word in a sequence. This basically consists of doing a fill-in-the-blank puzzle, adding another blank after each new word is chosen – exactly the way predictive text on your smartphone suggests the next words of a text message based on the ones you’ve recently typed.

Continuing the map analogy, ChatGPT would function something like this:

1. Convert the words/phrases of a prompt into a list of numerical coordinates (word-places) on the language-map, as described above.
2. Plot the most likely routes between these coordinates, drawing an arrow from the first word-place in the list to the next, and so on.
3. Predict the likely next-destinations to which the route might continue.
4. Select one at random. This randomness is hugely important.
5. Repeat the process, adding new coordinates until the route ends at a likely final-destination.
6. Convert each final coordinate back into words, generating a natural-language completion of the prompt.

As with any analogy, the description given here is only approximate, and in some respects remains imprecise. But it’s still a largely accurate representation of a system of Natural Language Processing – the type of process engaged by a computer when it translates natural language into formal language so we can supply plain English inputs and receive coherent responses. That could mean giving a conversational reply like ChatGPT does, creating an image like Midjourney does, or performing any other requested function.

A few steps of the “mapping” process could use a bit more digging into though, especially for those of us who, having avoided math classes like food poisoning since horrific episodes of high school algebra, would nonetheless appreciate a more detailed picture.

The puzzling parts: Steps 2 and 3

Steps 2 and 3, where the LLM charts different routes between word-places, might seem puzzling. There is a persistent strangeness to this “route-finding” that even the foremost experts haven’t penetrated (a mystery perpetrated on us during the “unsupervised” phase of the LLM’s training, naturally).

But we do have high-level insights:

• As it analyzes training data, the LLM learns how to navigate the map by imitating human linguistic “routes.”
• It identifies trends in how we travel between ideas and creates its own set of mathematical rules (algorithms) to navigate.
• You can imagine these rules like:
  • travel around cliffs rather than up them
  • go toward population centers
  • cross rivers where shallow
  • avoid highways at rush hour
  • reach a summit before descending

These rules help the LLM predict the next stop on a linguistic itinerary much like a human would, but then select one at random.

Why outputs have gotten so good

Gen-AI progression is due to:

• Vastly larger training datasets
• Better context awareness
• Improved randomness mechanisms
• Viewing language as a web of interconnected places, not linear sequences

This allows neural nets to follow more varied, more humanlike “paths,” resulting in increasingly impressive outputs.

A few essential points about the mapping process

• GPT’s enormous “language-map” emerges from mathematical relationships between numerically encoded words.
• “Routes” are essentially curved lines of sequence connecting these points.
• Prompting = drawing a path through specific points.
• GPT’s rules imitate human movement through idea-space.
• Creativity = continuing the user-drawn route using established pattern knowledge.

Thus:

All generative AIs are pattern extractors.
The artificial creative process = adapting massive human-derived patterns to new contexts.

But… patterns aren’t what data scientists call patterns

Data scientists define “patterns” as strict, repeatable statistical relationships. LLM patterns aren’t stable formulas – they’re emergent structures based on human linguistic tendencies.

Randomness isn’t true randomness either; it’s simulated via stochastic processes.

So:

• AI is not thinking or imagining.
• It is not analyzing like humans do.
• It is exposing and remixing mathematical patterns embedded in humanity’s collective creative output.

These patterns are too large for our brains to perceive, but they ultimately come from us.

How AI discovers these patterns

LLMs learn through:

1. Pre-training – preparing to recognize broad features
2. Unsupervised training – building the “map” from raw text
3. Supervised training – humans correcting, annotating, ranking outputs

Every step funnels toward one endpoint:

Human provides prompt → AI continues the pattern.

This applies to language, image, audio, video, all of it.

Stable Diffusion, DALL·E, Midjourney — all different architectures, same fundamental logic:

Create map → Learn patterns → Map prompt → Follow predicted route

Why does this work?

We don’t know.
Nobody does. That’s the spooky part.

The deeper technical weeds

• LLMs use embedding spaces with thousands of dimensions (GPT-3 used 12,288).
• Transformer architecture involves:
  • tokenizing text
  • encoding into vectors
  • transforming vectors through layers with up to 170 trillion parameters
  • self-attention deciding importance
  • nodes activating based on weights
  • minimizing loss functions

Even experts can’t fully explain why this complex system works so well.

The unsettling truth

Despite the mind-blowing performance of these systems, we cannot explain the “why” behind their intelligence-like behavior.
They work, but the underlying principles remain obscure.

We have no fundamental explanation for the newfound creative prowess of recently dumb machinery.