Same brain, different schooling.

Same brain, three schools.

This page talks to Ollama running on your machine. The demo is a three-slide deck — same prompt, three different ways of delivering it. You drive the story.

1 · pull the models

ollama pull llama3:text
ollama pull llama3:instruct

2 · let the browser reach ollama (CORS)

On macOS, set the env var, then quit & relaunch the Ollama app:

launchctl setenv OLLAMA_ORIGINS "*"

Or run from a terminal: OLLAMA_ORIGINS="*" ollama serve

3 · open the file

Just open demo.html. If your browser blocks file:// requests, serve the folder instead: python3 -m http.server and visit localhost:8000.

the three slides

Slide I — BASE (raw). The base model has read the internet but never been taught to "answer." Hit run and watch it ramble or pose more questions instead of replying.

Slide II — BASE + scaffold. Same model, same prompt — but wrapped in a Q&A scaffold (a fake conversation the model continues). Suddenly it answers cleanly. Same brain, different priming.

Slide III — INSTRUCT. The instruct-tuned variant of the same base. It's been trained on millions of Slide-II-style examples, so it answers cleanly without us doing the scaffolding. The scaffolding is baked in.

The punchline: from the audience's seat, Slide II and Slide III look the same. Instruct-tuning is, basically, automating Slide II.

flow for live demos

• Click a tab — or press 1, 2, 3 — to jump to that slide.
• Use ← / → to step between slides.
• Press ↵ enter in the prompt to run the active slide.
• Click a slide's ▸ run button to re-roll just that one.
• Click ↻ reset all to clear and start over with a new prompt.

more tips

• Try "The " on Slide I — base just keeps writing.
• Bump temperature past 1.0 for more dramatic rambling.
• Swap in mistral:7b-text-v0.2 for an even more unhinged base.
• Project in a darkened room — the slides are tuned for that.

got it