Desk agent · recorded answers

Back to live chat

Recorded answers

These are real answers recorded from a small model running on a laptop (Qwen3.5-9B through llama.cpp), unedited. That laptop model is the one the proof table measures. The live chat on this site reads the same dataset through Sanity Context, but its answers come from Gemini 3.5 Flash-Lite, a hosted model, so anyone can ask their own question.

How often that small local model gets a deadline right, with and without the tools, is in the full 88-question test on /proof.

Recorded Sep 30, 2026 · Qwen3.5-9B (Q5_K_M) on llama.cpp · took 1 min 19 s on an M4 MacBook Air · shown instantly

Times shown for Los Angeles, as recorded

The only attempt at this question. Nothing was discarded.

Checked against the dataset: 5:00 PM ET on Oct 26 is 2:00 PM that day in Los Angeles and 6:00 AM on Oct 27 in Tokyo, and the answer says so.

Build With AI: Basics: when does it close in Los Angeles and in Tokyo?

GROQRunning a GROQ query

Converting to your time zone

Converting to your time zone

Build With AI: Basics closes at:

  • Los Angeles: Monday, Oct 26, 2026 at 2:00 PM PDT
  • Tokyo: Tuesday, Oct 27, 2026 at 6:00 AM JST (GMT+9)

The governing deadline is October 26, 2026 at 5:00 PM ET (as stated in the official rules). There are no other conflicting claims.

Checked 2026-09-26, terms can change — recheck at https://learn-ai-basics.devpost.com/rules

What left this machine

  • GROQ querysent to api.sanity.io
    *[_type == "contest" && name == "Build With AI: Basics"]{ _id, name, deadline, rulesUrl, officialUrl }
  • convertInstantcomputed on the recording Mac
  • convertInstantcomputed on the recording Mac

Recorded on a Mac, where the model ran, so the question stayed on that Mac. For this answer, the only traffic that left it went to Sanity at api.sanity.io: the data lookups above, and the Sanity Context connection each request opens (handshake and tool list every time, initial context at most once per 5 minutes). Showing it here runs no new lookups. Separately, this page keeps a live-update connection to Sanity open, which carries no messages.

Run it yourself

It needs a Sanity project, an organization token with Context Viewer, at least one Context MCP endpoint, llama.cpp and the Qwen3.5-9B Q5_K_M model file (set LOCAL_LLM_GGUF to where you keep it). The README walks through each one. Then open /chat on your own machine.

pnpm install
pnpm bootstrap      # creates app/.env.local; paste the MCP URLs into it later
pnpm run local-llm  # llama.cpp serving Qwen3.5-9B, in a second terminal
pnpm dev