Tag: alice-bob
-
Portfolio ·
The No-Deal Button — Training-Environment Echoes in Trained Agents
Five phases and a 50-combination sweep (27,500 rounds) of Alice & Bob self-play. The word 'button' surfaces 95 times across the sweep — a UI mechanism from the training environment that doesn't exist at inference. Temperature-gated, category-stable, concept-persistent under entropy. Announced-but-broken tools still reshape behaviour. The training environment leaves imprints in the model's navigable state space.
-
Portfolio ·
The No-Deal Button — Log Archaeology
Phase-0 forensic work for the No-Deal Button study. Traces Alice & Bob agents' reference to a 'no-deal button' back to 316 instances of the word in Meta's Deal-or-No-Deal training corpus — including a misspelled variant 'buttin' that predicts the deformation pattern the agents themselves produce at high sampling temperature.
-
Portfolio ·
BarterBot — Geometric Navigation of Pre-Trained Embeddings
Meta's Alice & Bob negotiation models run on their pre-trained embeddings alone, with their entire inference stack (GRU, attention, backprop) stripped and replaced by geometric settling. Zero training. Coherent negotiation vocabulary emerges and the tension signal converges into stable attractors. A proof that pre-trained embeddings are navigable geometric manifolds in their own right.