1 · Before the conference (on good wifi — the pull is gigabytes)
# macOS
brew install ollama # or download from ollama.com
ollama pull gemma3:4b # primary — fast + smart enough on any M-series (~3.3GB)
ollama pull gemma3:1b # emergency spare for old/loaner laptops (~0.8GB)
ollama run gemma3:4b "Say: offline fallback ready."
Why Gemma 3 4B: fits in laptop RAM, ~20–40 tok/s on an M-series (fast enough to feel live on a projector), 128K context, and a nice on-stage line given the Google history. After the pull it needs zero network.
2 · Day of — warm it up before you walk on
ollama run gemma3:4b # first response loads the model; do it backstage
# bump terminal font: Cmd+ {plus} until readable from the back row
3 · Which workshop prompts work offline — and the two that don't
| Stage / prompt | On local Gemma | Adjustment |
| 01 · Ideation (generate, score, attack, obituary) | WORKS | Ask for 10 angles, not 20 — small models pad past 10. |
| 02 · HTML animation + logo concepts | WORKS | One file HTML: fine. Logo: Gemma describes concepts (no image gen) — sketch on paper. |
| 03 · Positioning 50 → 1 | WORKS | Ask for 20 lines, top 3 — quality drops past 20 on 4B. |
| 04 · Khosla seed deck + skeptic pass | WORKS | Split into 2 turns: deck first, then "now the skeptic pass." One mega-prompt overloads 4B. |
| 05 · GTM buyers / channel / pricing | WORKS | As-is. |
| 05 · "Pull public signals on [prospect]" | NEEDS VARIANT | No web. Use: "Here are their last 3 posts (pasted below). Find the sharpest signal and draft the 60-word reach-out." |
| 06 · "Simulate 20 users on my product at [URL]" | NEEDS VARIANT | No fetching. Use: "Here is my landing page copy / HTML (pasted). Simulate 10 [ICP] users on it…" 10 personas, not 20. |
4 · Night-before smoke test (5 minutes, one per stage)
- S1: "Generate 10 startup angles for AI testing tools. Score 1–10. Kill 8."
- S2: "Build a single HTML file that animates a bug scanner finding 3 bugs." → save, open in browser
- S3: "20 positioning lines for an escaped-bug finder. Pick the best."
- S4: "Build a simple Khosla-style seed deck outline for it." then "Skeptic pass: 3 weakest slides."
- S5: "10 most likely first buyers + one channel + 3 price points."
- S6: paste any landing copy → "Simulate 10 on-call engineers using this. Top 3 bugs."
If any check feels weak on 4B, the demo line for the room is honest and on-message: "this is a 3GB model on my laptop with no internet — imagine the frontier model". The loop is the lesson, not the eloquence.
5 · Escalation ladder (announce at the open)
- Plan A — venue wifi → cloud AI (Claude / ChatGPT / Bolt)
- Plan B — phone hotspots, one per pod → cloud AI on phones
- Plan C — presenter laptop + local Gemma on the projector, room follows along
- Plan D — printed sessions: DogVan + Apology Cookies, pods run the loop on paper