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

Choose Your Collaborator.

Artificial intelligence is not a single, omniscient brain. It's an ecosystem of specialised workers. Some are brilliant artists who can't do basic maths. Others are flawless accountants who can't write a convincing poem. Choosing the wrong worker guarantees a flawed result. This class teaches you to audit any model — by memory, logic and privacy — and assemble the right toolkit.

You'll be able to

Tell models apart

Distinguish proprietary from open-source, and match each to a task.

You'll be able to

Define context windows

Explain the memory limit that makes models forget — and hallucinate.

You'll be able to

Protect your data

Spot the privacy implications of feeding info into corporate systems.

01 — Warm-up

Match the task to the model

Drag each task into the model best suited to it. Keyboard user? Use the "Send to…" menu on each card.

Self-Check · Drag & Drop
Draft a highly emotional, stylised community newsletter.
Keyboard:
Find a microscopic maths error in a 100-page spreadsheet.
Keyboard:
The Analytical Model
The Creative Model
02 — Foundation

The context window

The single most important spec isn't how smart a model is — it's how much it can remember. That memory limit is the context window, measured in tokens (the word-fragments from Class 1).

You have short-term memory: you recall what was said five minutes ago. A model's short-term memory is strictly capped by its context window. With an 8,000-token window (~6,000 words) it can hold the first chapter of a book. Paste in a 50,000-word novel, though, and it suffers instant amnesia — operating first-in, first-out, it forgets the beginning by the time it reaches the end.

Pause & notice

If a user summarises a contract that exceeds the context window, what happens when the AI forgets the first half? It confidently invents clauses that don't exist to fill the gap. This is a primary cause of AI failure at work.

Mini-recap

Context windows range from 8,000 to over 1,000,000 tokens. Picking a model whose memory can hold your whole document is the operator's primary responsibility.

03 — Compare

Two architectures, side by side

Corporate Model A

Big memory, locked data

Massive context window and high capability — but your inputs travel to a corporate server, the decision process is a black box, and cost is high.

Best for: public data, heavy logic, when you don't need to own the infrastructure.
Open-Source Model B

Smaller memory, total privacy

Free to download and run locally — smaller context window, but the data never leaves your machine. You can literally pull the network cable and it still works.

Best for: private, sensitive data; offline work; full ownership.

The trade is always the same: raw, effortless power versus absolute privacy. Your job is to decide which the task demands.

04 — Interactive

The Capability Radar

Every model is a compromise. Fast models rarely do deep logic; highly creative models often fumble strict formatting. Move the sliders to define your project's needs and watch the radar find the closest model archetype.

Sandbox · Tune your requirements
50
50
50
50
50

No shape is "best." Max out Memory and Logic and you'll land on a corporate powerhouse; max Cost-efficiency and Speed and you'll be steered to a small local model. The right model is the one that fits your shape.

05 — Deep dive

Walled garden vs mutual-aid library

The ecosystem splits into two camps. Understanding the difference is essential for data security, creative independence and ethical deployment.

Proprietary

The corporate walled garden

Models like ChatGPT, Claude or Gemini are owned, hosted and controlled by big tech, accessed via the cloud. Powerful, but black boxes: you can't see how they decide, and the company may analyse your inputs. For journalists, activists, medical and community workers, that lack of data sovereignty can be disqualifying.

Open-source

The mutual-aid library

Labs release the weights and code for free (e.g. Llama, Mistral). Anyone with a capable computer can run them entirely offline. Sever your internet connection and the AI still works — mathematically provable privacy for vulnerable community data, client notes, or unpublished manuscripts.

The trade-off

Open-source models demand technical setup and serious local hardware (a strong GPU), and — because they must fit on a laptop — may have smaller context windows and lower capability. You're always weighing effortless power against absolute privacy.

Takeaway

Never default to a single corporate tool. Audit the task: public data needing massive logic → the walled garden. Highly sensitive private data → download an open-source model and sever the connection.

06 — In the Wild

Sam & the archive

Sam leads a local queer-history archive and has 400 pages of scanned 1980s community newsletters to categorise by date, subject and people mentioned.

Before

Wrong tool

Sam used a standard free model with an 8,000-token window. It read the first 10 pages, forgot them, and hallucinated fake historical names for the other 390 — the document vastly exceeded the memory limit.

After

Right tool

Sam audited the ecosystem, recognised this needed a huge memory buffer, and switched to a model with a 200,000-token context window. It ingested all 400 pages at once.

Result: it cross-referenced names on page 10 with events on page 380 and output a perfectly formatted historical timeline.
Why it matters · ethical caveat

Knowing hardware limits stops you blaming yourself (or the prompt) when the real failure was a tool with too little memory. Because these archives were public historical data, a corporate model was fine. Had they held private medical records, Sam would be obligated to use a local open-source model, chunking the document to fit a smaller window.

07 — Research

Map current memory limits

Context-window limits shift constantly under fierce competition. Last year's "massive" buffer is today's standard. Open a search engine, find the current limits, and log them below (saved on your device only).

Search for

What to look up

"Claude maximum context window" · "Gemini Pro context limit" · "how many pages is 1 million tokens?". Find which commercially available model currently has the largest context window, and how many book pages it can hold.

Judge the source: skip SEO blogs. Use official API docs from Anthropic, Google or OpenAI — they must publish exact token limits for developers. Check the year.
Your findings
Saved locally on your device — never sent to a server.
08 — Assignment

The blind taste test

Three outputs were generated from the same prompt: "Write a polite rejection letter for a freelance design project." Read them, then map each to a model type — citing the linguistic markers that gave it away.

Output A

Rigid & formal

Uses bullet points and cold, highly formal corporate syntax throughout.

Output B

Warm & conversational

Empathetic, uses exclamation points, and offers creative alternatives.

Output C

Broken

Repeats the same sentence twice and abruptly cuts off at the end.

Reveal the answer mapping

Output C is the older / smaller model — it lost memory and hallucinated a repetition. Output A is the logical/analytical model — it defaulted to bullet points and cold syntax. Output B is the creative model — empathetic adjectives and conversational flow. The skill is citing specific markers (bullets vs exclamation marks vs degradation), not guessing.

Success rubric

Needs work: guesses with no analysis of markers.
Getting there: spots the broken model but confuses logical and creative.
Solid: maps all three but without citations.
Excellent: maps all three and explicitly cites the logical model's bullets vs the creative model's adjectives.

09 — Retrieval practice

Quiz — 5 questions

Question 1 · Medium
What happens when a document exceeds an AI model's context window?
A
The model upgrades its memory automatically via the cloud.
B
The model crashes, deletes the document, and locks the user out.
C
It forgets the earliest information provided, leading to severe hallucinations.
Question 2 · Easy
For highly sensitive financial or medical data, which ecosystem is mathematically safest?
A
A free corporate web chatbot.
B
An open-source model running locally on your own offline hardware.
C
A paid corporate web chatbot.
Question 3 · Medium
Why might a user choose a small-parameter model over a massive one?
A
Small models are inherently smarter.
B
They can run entirely offline on a standard laptop, ensuring privacy and saving cloud costs.
C
Massive models refuse to answer simple questions.
Question 4 · Easy · True / False
"A model optimised for creative writing will perform perfectly when asked to debug complex software code." — True or False?
T
True
F
False
Question 5 · Easy
What is a "walled garden" in technology?
A
A physical server farm surrounded by security.
B
A proprietary, closed ecosystem where a corporation controls access, data and code.
C
A prompt-engineering technique used to protect data.
10 — Avoid these

Common mistakes

Trap

The loyalty trap

Sticking to one comfortable chatbot for everything — then using it to crunch a spreadsheet and getting maths errors. Fix: keep three models bookmarked and swap by task.

Trap

The endless paste

Dumping 50 pages into a 10-page window, so the AI hallucinates the ending. Fix: estimate token limits first; break large text into chapters and process sequentially.

Trap

The free-privacy assumption

Pasting proprietary strategy or unreleased work into a free public bot. Fix: treat free models like a public billboard — if you wouldn't tape it to a window in a public square, don't paste it.

Trap

Ignoring "temperature"

Leaving defaults and getting wild code or stiff poetry. Fix: temperature controls creativity — set it to 0 for logic and maths, 0.8+ for brainstorming.

Debug your thinking — check each before you choose a model
11 — Reflect

Whose worldview wins?

If a single corporation builds the most dominant AI in the world, how might relying on that one model's built-in tone and logic constraints enforce a standardised, homogenised worldview on digital culture?

Optional prompts: How does building independent, open-source models empower marginalised communities to protect their data? Would you sacrifice the convenience of a fast corporate AI to keep your data fully private on a local machine? Does the "walled garden" limit digital creativity?

Private journal
Saved locally to your device — never sent to a server.
Well done

You can now audit and select the exact right tool for any task — no longer bound to a single interface. Next, we learn to command these models: from simple questions to engineered, structured prompts.

12 — Reference

Glossary — precise words

Context window
The token limit of a model's short-term memory in one conversation. Exceed it and you get hallucinations.
Proprietary model
A closed-source AI owned by a corporation; code and weights kept secret. You can't audit its decisions.
Open-source AI
Models whose architecture and weights are public, so anyone can run or modify them — preventing monopolies.
Local model
An open-source model run entirely on your own computer, offline. Guarantees absolute data privacy.
Algorithmic specialisation
The idea that models excel in narrow fields (code, prose, images) rather than being universally intelligent.
Walled garden
A closed ecosystem where the provider tightly controls apps, content and data.
Data sovereignty
The principle that data stays under the control of the person or community that created it, not the host.
VRAM
Video memory on a graphics card — the hardware needed to run local AI models smoothly.