L4 Classroom Demos

Start: bash start_demo.command
Keep the launcher running during class. Ctrl+C stops services owned by this launcher.

1. L3 Feedback: http://127.0.0.1:7894/l3_feedback.html
2. What Does an LLM Know? http://127.0.0.1:7894/
3. What Does an LLM Remember? http://127.0.0.1:7894/memory.html
4. Remember Bob: http://127.0.0.1:7894/context.html
5. Who Won the World Cup? http://127.0.0.1:7894/web.html
6. Instructions Shape the Reply: http://127.0.0.1:7894/instructions.html

Demo 1 needs no model or internet connection.
Demos 2–6 use the same Qwen3.5-9B Instruct Q4_K_M model from L3.
Existing complete downloads are reused; there is no automatic model download.
Shared file: ~/Library/Application Support/AILT1001/l3_demo/models/qwen_3_5_9b.gguf
For a manually stored model, set AILT_MODEL_DIR to its containing folder.
If an older incompatible L3 demo service is running, stop its terminal with Ctrl+C before starting L4.
A compatible running service is reused; only one model process is needed.

Fresh Mac: Apple Silicon, Homebrew, Python 3.11+, llama.cpp.
Run bash install_demo.command once; download Instruct through Demo 2 only if absent.
Allow about 6 GB disk and roughly 7–9 GB runtime memory (estimate).
Node.js and Firefox remain Deckroom prerequisites. No extra Python or Node packages.

Demos 2–6 use Send and Start Again. Start Again clears the conversation or result and restores the example prompt. Enter sends; Shift+Enter adds a new line.

Demo 4 includes earlier messages. Introduce Bob, then ask Who am I? Click What Did the LLM See? to inspect the conversation context. This is conversation history, not saved memory.

Demo 5 performs a fresh lookup for each question. The application requires an initial search; Qwen generates the query. Search uses Wikipedia introductions and Google News RSS headlines. Internet is required; no API key. Inspect the sources because retrieval does not guarantee correctness.

Demo 6 adds editable instructions, initially Always answer in German, to the same web-search workflow. Inspect the highlighted instruction in What Did the LLM See? Weights remain fixed.

The advanced inspector at /context_explorer.html is not a classroom tab. Its optional MEMORY.md lives in ~/Library/Application Support/AILT1001/l4_demo/ and is sent only when enabled there.

Demo 7: Create, Publish, and Email a Quiz
Use the same Apple Silicon Mac, Python and llama.cpp setup as Demos 2–6.
No extra account, setup code or login is needed. The package includes a restricted
demo key for the M3 gateway. Never replace it with a personal/provider credential.
Internet and the M3 gateway are required; this does not use Dirk's presenting Mac.
Open Demo 7, enter a consenting recipient privately and confirm consent.
The [recipient] placeholder becomes their actual address in the model prompt;
projected logs keep it masked. Run the goal, inspect the tool results, and open
the live quiz. Try a revision to preserve the URL. Emails contain just the link.
Qwen can call play_sound() to play Glass on the presenting Mac when done.
If email is unconfigured or fails, the log reports it; no success is simulated.
Limits: 12 recipient sessions/day across all copies, four emails/recipient/day,
40 emails/day total, eight publishes/session. Contact Dirk if a limit is reached.
The agent uses an 8,192-token context on the same model, so allow extra RAM.
