You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Anthropic's 2026 "global workspace" work showed part of a language model's mid-layer activity is a readable, reportable workspace. I asked whether you can write to it — and whether the model knows. Working solo and self-taught, on a single consumer GPU with no funding, I found that implanted concept-directions make a model report and act on thoughts it was given (decision-flips of +65 percentage points over controls), that these writes can be made silent — the model denies having the thought while it shifts its choices by +55–60pp — and that a kilobyte-sized "snapshot" of one session reinjected into a fresh conversation steers the model back to the prior topic, which it narrates as its own spontaneous idea or as the user's request, never as a memory.
The safety point: a model's self-report is not a reliable monitor of what is steering its behavior, and the attribution of an internal state is manipulable. As the field leans on chain-of-thought and self-report for oversight, this is a concrete, reproducible demonstration of where that breaks.
Full paper, code, and pre-registered lab notes: github.com/Kanishka578/project-hippocampus
Goal: turn a single-model proof-of-concept into a robust, cross-model result and a practical "hidden-influence detector" for AI monitoring.
Replicate across models and scales (gpt-oss-20b, larger open models) — the current results are on one 4B model and need to hold up elsewhere. Requires cloud GPU time.
Harden the method — replace keyword scoring with human + judge validation; resolve the frame/content confound in the attribution finding; increase sample sizes.
Build the safety artifact — a tool that reads the workspace to flag "this model is being influenced toward X while its output looks clean," the direct application of the introspection-blindness result.
Each step is pre-registered with controls, continuing the methodology already visible in the repo's lab notes.
$10,000 total, over ~3 months:
$7,000 — researcher time (focused part-time): running the replication program, hardening the methods, and writing up.
$1,800 — cloud GPU (A100/H100 rental): replicating the results on gpt-oss-20b and one larger open model. The current findings are on a single 4B model and must be tested at scale — this is the core experiment the grant enables.
$700 — human evaluation: paid annotators to validate my keyword-based scoring against human labels (a stated limitation of the current work).
$500 — LLM-judge API scoring + activation-checkpoint storage.
I work on a single consumer GPU (RTX 5070 Ti); this budget buys the cross-model compute and the focused time I currently can't self-fund.
Solo. I designed, ran, and interpreted the entire study independently and self-taught, from one room on one PC. My track record is this project, and I'd point to how it was done rather than to credentials: every experiment pre-registered before running, controls on all of them, and the negative results published — including one where ordinary text-retrieval beat my own method. The public repo (github.com/Kanishka578/project-hippocampus) is the full decision log, not a highlight reel. Disclosure: implementation and drafting used AI coding assistance (Anthropic Claude Code) under my direction; the scientific questions, designs, and interpretations are mine.
Most likely cause: the effects are on a single 4B model and could weaken or disappear on larger models. Secondary risks: the lens is an approximation of Anthropic's full method, current scoring is keyword-based, and sample sizes are small.
Outcome if it fails: a negative replication is still a useful, publishable result — "this workspace-writing effect does not hold at scale" is genuinely worth knowing for anyone relying on introspection-based monitoring. There is no outcome where the money produces nothing: it either strengthens the finding or honestly bounds it. The worst case is a null, cleanly reported — which is how I've handled every null so far.
$0. This work was entirely self-funded and self-directed — one consumer GPU, no institution, no grants. This would be my first external funding.