INITIALIZING BUNKROS IDENTITY LAB
LOC UNDERGROUND
SYS --:--:--
Learning / AI Models

Speak to the Machine.

Artificial intelligence does not read English. It does not read Spanish, French, or Japanese. It reads mathematics. Before you can command a machine, you have to see how it perceives the world. This class dismantles the myth of the "thinking machine" and rebuilds it as what it really is: a fast, limited, statistical calculator — and shows you exactly how to work with one.

You'll be able to

Explain prediction

Tell the difference between human reading and AI next-word prediction.

You'll be able to

Define a token

Say what a "token" is, and why slang shatters into many of them.

You'll be able to

Split training vs inference

Know what the AI "knows" versus what it merely calculates.

01 — Warm-up

Quick self-check

Before the mechanics, get a feel for the scale of artificial memory. How many words did the latest AI model have to "read" before it could converse fluently? Drag the fader, then check your guess.

Self-Check · Estimate
How much text does a modern AI train on?
1 Million1 BillionSeveral Trillion
Your guess: One Million Words
02 — Foundation

How AI actually reads

It is tempting to imagine a tiny, hyper-intelligent being inside the server, reading your question, thinking, and typing back. That is not what happens.

Imagine a vast library of billions of books, run by a head librarian who speaks no human language at all. This librarian is a master statistician. Handed the phrase "The sky is," they don't look out a window — they calculate probability. Having read trillions of sentences, they know "blue" follows "The sky is" about 92% of the time, while "falling" follows it 4% of the time. So they answer "blue." The machine does not know what a sky is. It only knows the mathematical probability of words appearing near one another.

Pause & notice

If the machine only knows the probability of words from what was written in the past, what happens when it's asked to predict the future — or to describe a marginalised culture historically excluded from published books?

Mini-recap

AI models are not thinking entities. They are probability engines guessing the most likely next piece of data based on historical patterns.

03 — Visualise

Raw data → patterns

Picture a mechanical blender. Raw data goes in at the top — newspaper clippings, wikis, handwritten zine pages. The algorithm extracts patterns in the middle, discarding the actual words and keeping only how often they appear together. Probability weights emerge from the spout at the bottom.

In

Raw data goes in

Books, articles, websites, zines — trillions of words of human text poured into the machine.

Process

Patterns extracted

The algorithm throws away the words themselves and keeps only the statistics of how they co-occur.

Out

Probability weights emerge

What's left is a giant map of likelihoods — the model's entire "knowledge."

Diagram, in words: raw human text is fed into an algorithm; the algorithm breaks it down, discards the words, and saves only the mathematical patterns of how often words appear together. That saved pattern is called a probability weight.

04 — Interactive

The Token Transformer

Computers can't process letters — they process numbers. To read your prompt, the AI slices your words into chunks called tokens and gives each chunk an ID number. Type a complex word or a piece of slang below and watch the machine chop it apart. Hover a token to see its ID.

Sandbox · Chop text into tokens
Notice

A single word fractured into multiple numbers. Those are tokens — exactly what the AI sees. Standard English words are often one token; unfamiliar slang gets shattered into many, because the model has to fall back on tiny letter-chunks it recognises.

05 — Deep dive

Training vs inference

To command AI and protect your data, you must separate two phases of a model's life: training and inference. Blurring them causes most beginner misunderstandings.

Phase 1

Training — building the archive

Like a community archivist compiling a history of local queer nightlife from thousands of flyers and zines, the model aggregates billions of documents over months, on thousands of chips, adjusting its probability map. Then the library is built — and frozen.

Phase 2

Inference — consulting the archive

When you ask a question, the model is not learning from the live internet. It references the probability map it built months ago to infer the next word. It calculates; it does not learn.

The privacy trap

A dangerous myth: that typing a private document into a prompt instantly adds it to a global database. During standard inference the model doesn't permanently learn new facts — it holds your conversation in short-term memory until you close the tab. Caveat: the company that owns the platform may log your chats to use in a future training run. That's why opt-out toggles and privacy policies matter for anyone protecting sensitive community data.

Takeaway

Training is studying for the exam. Inference is taking the exam. When you prompt an AI, you're talking to the inference phase: it isn't learning, it's calculating.

06 — In the Wild

Elias sorts the inbox

Elias is an independent zine creator running a busy online store, drowning in hundreds of emails a week. He wants AI to categorise them by urgency.

Before

The emotional prompt

Elias asked: "Can you please be a good assistant and figure out my emails for me?" The AI hallucinated and scrambled the categories — the request was too vague and emotional.

After

The structured prompt

Remembering AI is a statistical engine, Elias fed it predictable structure:
Rule 1: if the email contains "broken" or "missing" → label URGENT.
Rule 2: if it contains "collab" or "art" → label LOW PRIORITY.
Then he pasted the raw emails below the rules.

Result: the tokenizer recognised the rigid pattern and sorted 100 emails accurately in three seconds.
Why it matters · ethical caveat

Treating AI as a probability engine, not a human assistant, stops you sending vague emotional requests that confuse tokenization. Structure is the only language the machine respects. And Elias first strips any sensitive data (card numbers, addresses) before pasting, since corporate platforms may log the interaction.

07 — Research

Map the frontier

Models are measured in parameters — the mathematical connections formed during training. New ones drop weekly, so any number printed here would be stale fast. Open a search tab and investigate, then log what you find below (saved on your device only).

Search for

What to look up

"latest open-source LLM parameter count" · "Llama parameter sizes" · "GPT estimated parameters". Find one small model built to run on a phone, and one massive model built for a data centre.

Judge the source: check the date. If it's over six months old, the numbers are likely obsolete. Trust direct announcements from engineering teams (Meta, OpenAI, Anthropic, Google) over secondary blogs.
Your findings
Saved locally on your device — never sent to a server.
08 — Assignment

Breaking the tokenizer

Algorithms struggle with words that were rare in their training data. By watching how a model breaks down words, you can spot its cultural blind spots.

Steps

Do this

1. Write a two-sentence paragraph using highly specific community slang, queer vernacular, or niche hobby terms.
2. Open an external tokenizer (e.g. the free OpenAI Tokenizer page) or use the sandbox above.
3. Paste a formal English sentence; note the token count.
4. Paste your slang paragraph; compare the count.
5. Write 2–3 sentences comparing the word-to-token ratio of standard English vs your slang.

Example: "'The weather is nice' was 4 words and 4 tokens. My community's drag slang — a 5-word sentence — fractured into 18 tokens, because the AI had to chop the unrecognised words into two-letter chunks."
Success rubric — tap to expand

Needs work: uses standard English for both tests, getting identical ratios and missing the point.
Getting there: uses moderate slang, notes a slight increase, but doesn't connect it to training data.
Solid: uses complex slang and correctly notes unknown lexicon is chopped into many pieces.
Excellent: notes the fragmentation and argues it proves the training data skewed toward mainstream, formalised text rather than diverse community dialects.

09 — Retrieval practice

Quiz — 5 questions

Question 1 · Medium
Which best describes the "inference" phase of an AI model?
A
The model is actively reading the entire internet to learn new facts.
B
The model is slicing words into individual letters to save hard-drive space.
C
The model is using its pre-calculated, frozen probability map to generate a response.
Question 2 · Easy
Faced with a brand-new slang word it has never seen, what does the tokenizer do?
A
Delete the word from the prompt to avoid errors.
B
Slice the word into smaller, recognisable mathematical chunks of letters.
C
Reject the prompt and ask the user to speak formally.
Question 3 · Hard
Why do LLMs excel at poetry but struggle with complex mathematics?
A
They are programmed with emotional intelligence rather than logic.
B
They predict the most statistically probable next word — great for language, poor for strict calculation.
C
The developers forgot to include math textbooks in the training data.
Question 4 · Medium · True / False
"Typing a private document into a corporate AI instantly posts it to a public website for all users to read." — True or False?
T
True
F
False
Question 5 · Easy
What happens during the "training" phase?
A
The user types a prompt and waits for an answer.
B
The AI is fed billions of documents to build its internal probability map.
C
Developers manually type out the answers to every possible question.
10 — Avoid these

Common mistakes

Trap

The sentience fallacy

The AI says "I feel" / "I think", tricking your empathy. Fix: remind yourself it's an advanced autocomplete calculator — no intent, no feelings, no memory of you beyond this chat.

Trap

The truth fallacy

Treating AI like a search bar and trusting its confident, well-formatted answers. Fix: it predicts plausible-sounding words; it doesn't verify facts. Always check critical facts elsewhere.

Trap

The one-word-one-token myth

Counting words instead of tokens, then overflowing the model's memory. Fix: rule of thumb — 100 tokens ≈ 75 English words; far more for slang or technical text.

Trap

The emotionally vague prompt

"Make this sound better." "Better" is subjective. Fix: give structure — "Rewrite this: cut 20% of the words, use active voice, professional and urgent tone."

Debug your thinking — check each before you prompt
11 — Reflect

Whose voices are missing?

If an AI can only predict language from the documents it was fed in training, whose voices, histories and cultural contexts might be entirely missing from its internal library?

Optional prompts: How does over-representation of Western, English-language data shape the "personality" of global AI? If you built a dataset for your own community, which zines, archives or voices would you insist on including? Does the phrase "artificial intelligence" grant the technology unearned authority?

Private journal
Encrypted to your device via local storage — never sent to a server.
Nice work

You've shattered the illusion of the "thinking machine" — you can see the numbers behind the curtain. Next, we use this to pick the exact right model for your projects.

12 — Reference

Glossary — precise words

Large Language Model (LLM)
An AI trained on vast text to predict the next word. The master statistician in the library.
Token
The unit of data an AI processes — usually a fragment of a word. The syllables of machine language.
Parameter
A mathematical connection in the network — the "synapses." More = more capable, but slower and pricier.
Training data
The raw text fed in during creation. Biased or incomplete data → biased, incomplete output.
Training phase
The massive period where the AI reads data and builds its probability map. Studying for the exam.
Inference phase
The operational phase where the frozen model generates responses. Taking the exam.
Hallucination
When the AI confidently states false info because its probability map guessed wrong. Proof it knows patterns, not facts.
Probabilistic prediction
Guessing the most likely next variable — the sole underlying mechanic of text generation.