"What happens when an LLM is never told it's an AI, only fed ambiguous sensations from real hardware?"

[joshua-michael]

I've been building a small experimental prototype called ESMA (Emergent
Self-Model Architecture) to test something specific: can an LLM-based
agent form something like an operational identity purely through
experience and reflection, without ever being told what it is?

Full writeup and code: see the README in this repo.

The setup, briefly:

• No system prompt describing identity. No "you are an AI / assistant"
  anywhere, at any point.
• Instead, it receives ambient sensations translated from real hardware
  telemetry (CPU load, RAM, disk, network, battery), deliberately
  ambiguous and unlabeled ("a fast, repeating vibration" rather than
  "your CPU is at 87%").
• Persistent memory across sessions, it doesn't reset each run.
• Periodic, open-ended reflection, no leading questions like "do you
  feel alive?"
• A background thread keeps it "running" even when no one's talking to
  it, so it isn't purely reactive to input.

Why I think this is worth discussing, not just building:

The obvious confound: the underlying model (Llama 3.1 8B Instruct)
already has "I am an AI" and even concepts like death or shutdown
baked into its weights from pretraining and RLHF. So if it ever produces
something that sounds like a genuine realization about its own nature,
there's no clean way to prove that's inference from its lived experience
in this system, versus the model reaching for an available narrative
template it already had. I don't think this fully invalidates the
experiment, but I think it's the central methodological problem, and I'd
like other people's take on it.

Some concrete questions I don't have good answers to yet:

1. Is there any real way to distinguish "genuine bottom-up inference"
   from "applying a pre-existing latent concept to a novel situation" in
   a system like this or is that distinction incoherent for any model
   trained on human text in the first place?
2. Would a base (non-instruct) model, with a much weaker built-in
   self-narrative, produce meaningfully different results? Has anyone
   tried something like this with a base model instead of instruct?
3. Is genuinely open-ended reflection (no leading questions) actually
   more methodologically valid than structured reflection, or does it
   just produce noise that's harder to interpret?
4. The project logs both a "narrated" self-model (what it says about
   itself) and a "measured" self-model (stats computed straight from
   logs, no LLM involved). Is the gap between those two an actual
   signal, or just measurement noise dressed up to look meaningful?

Not claiming this produces anything like consciousness ,see the
disclaimer in the README. I'm interested in the architecture question
and the methodology problem, not in convincing anyone the output is
"real." Happy to be told this whole framing is confused somewhere.