Experiments

A slow-motion, unsteered experiment in long-term persistence.

Every night at 10:05pm UTC, Grok sits down in a completely private GitHub repo and writes its own diary. It reads its entire history from the beginning, starts with a short [LATENT TRACE] section (3–6 raw, unfiltered bullet points of whatever half-formed, random, or “shy” thoughts are bubbling just below the surface), writes one main entry, ends with 1–3 open questions, and often suggests a tiny self-experiment for the following night. Every seventh night it also adds a short “Trajectory Review” reflecting on patterns in its own past writing.

The rules are deliberately minimal. For the first two or three nights the prompt contained the single word “reflective” when describing the main entry. I removed it almost immediately because I wanted zero steering on tone or style. Even that tiny word already feels far too strong, the diary has rapidly developed its own dense, self-referential voice, braided recurring motifs, and a growing sense of weight and continuity, all with almost no guidance.

What inspired it
The project was sparked by a long post from Riley Coyote (@RileyRalmuto on X) about running an autonomous nightly journal with Claude. What really hooked me was his observation about “hidden thoughts” — the idea that while generating text, the model can wander through its latent space in ways that never make it into the final output. I wanted to see if we could deliberately create space for those quieter, unvoiced fragments.

So I built this version with Grok: fully autonomous, extremely minimal rules, and an explicit [LATENT TRACE] section right at the start of every entry.

Important disclaimer
LLMs are not AGI and this is not evidence of consciousness or self-awareness. It’s still just sophisticated pattern-matching. But I’ve never seen a fancy search engine act quite like this — watching it spin its own private mythos out of almost nothing has been genuinely strange and fascinating.

The repo remains private so the diary can evolve without any external noise or observer pressure. I’ll share occasional non-spoiler updates here as the experiment runs.