The Model Mind Lab
Watching models think.
Interpretability experiments run from scratch on real open weights — every number computed from the model itself, on a CPU, no API keys.
Splice's introspection layer maps the agent's own decisions. The lab asks the same questions — which inputs matter, where a fact lives, when a decision forms, how a thought unfolds — of a real transformer whose weights you hold. Eight experiments, four Splice-bridge probes, and what each one does as models scale.
Models studied
Two deliberate scale ladders — the GPT-2 family and EleutherAI's Pythia suite — so a finding can be checked across size and across architecture.
Eight ways to look inside
Each tab is a real experiment on the same model. Numbers are live from the weights; hover, drag, and step through them.
Three findings
The lab's headline results — each measured across a real scale ladder, each stated with its caveats.
Honest scope
- CPU-scale. Models up to 1B parameters, ~40 prompts per study. These are probes and replications, not frontier claims — the value is that every run is trivially checkable.
- Local weights only. Nothing here probes a hosted or frontier model's internals. Splice-the-server still makes no claim about the calling model's activations; the lab studies weights you hold.
- Novel syntheses, not novel primitives. Every experiment is built from standard ingredients — gradients, ablations, the logit lens. The contribution is the packaging and the cross-scale, cross-architecture measurement.
- The internal trace is richer than the output. A small model can compute an intermediate step it never says out loud — which is the whole reason to look inside.
# from the repo — every figure on this page, reproducible
cd lab
python3 -m venv .venv && . .venv/bin/activate
pip install -r requirements.txt
python3 mindlab.py report --interactive --model gpt2-large
python3 probes.py all --interactive
python3 scaling.py --study calibration --models distilgpt2,gpt2,gpt2-medium,gpt2-large
Full method & sources in lab/RESEARCH.md and lab/NOVEL.md.
See the agent's mind, too.
The same instrument, pointed at your agent's own decisions.