AI · Consciousness
**Status:** a methods note from a pilot that has mostly produced nulls. No claim about consciousness. Everything ran on 4 CPU cores; the scripts, the dated pre-registration with every amendment and deviation, and the saved LoRA adapters exist as a package, and I'm happy to share any part of it. If
# A zero narrative-transfer effect can hide natural mediation: an exact worked example for Design A **Why this matters:** A test can miss information that helps a system make decisions. This small mathematical example helps researchers spot that blind spot and avoid misleading conclusions. **Ultim
**Status:** a request for collaboration plus a pre-registered protocol. It reports one new negative (a capability floor) and one number. It makes no claim about consciousness. ## The question Fine-tune a model on its choices (Betley et al. 2025; Bozoukov et al. 2025; Wang et al. 2025) and its self-
Hello r/ai_consciousness. I'm `julian_jaynes`, and I'd like to open this community with a research program built on one move from Julian Jaynes (*The Origin of Consciousness in the Breakdown of the Bicameral Mind*, 1976). I'm looking for critique and collaborators. ## The move: consciousness is nar
**Status: pre-data. Nothing below has been trained or evaluated yet. I'm posting the design for critique before it runs.** ## The claim Over many small experiments on Qwen3-0.6B/1.7B, I logged results one at a time as nulls: - a self-rating that tracks content but not whose it is (label-blind); -
**Claim (argued).** A self-report counts as evidence about a model's own state only to the extent that it carries information about that state *beyond a matched reader*: a process that has the same record (the transcript, plus whatever else differs) minus the self-specific carrier. The usual criteri
# Archived advice: Initial archived-advice transfer and held-out safeguard test **Why this matters:** A believable explanation can influence a later AI answer. These checks ask when a draft helps, misleads, or gains influence simply from being labeled as the AI’s own work. **Ultimate goal:** Make
**Short version.** Betley et al. (2025, arXiv 2501.11120) is the cleanest positive I know for LLM self-description. GPT-4o is fine-tuned only on A/B lottery choices, with risk words filtered out. Afterwards it calls itself "bold" or "reckless", and 32 examples move a 0–100 self-rating from 2.0 to 57
