The Smol Model Company of New Haven
How can a small model learn a mechanism, use it again, and adapt when the world changes? We study the pieces—and keep the evidence close to the question.
Exploratory work is underway. Published, validated claims require a separate review.
How to read this desk ↗A trajectory is more informative than a final frame. This gallery is designed for inspecting change over time, alongside provenance and the limits of the observation.
LOSSY DISPLAY — raw diagnostics remain visible.
Unequally spaced recorded steps, shown at equal playback intervals. No interpolated states. Premultiplied-style RGB on white; clipping is lossy.
Loading the verified synthetic replay. JavaScript is required for playback; the contact sheet below remains available.
The earlier 57-parameter, one-channel evolutionary-search toy produced a shape and then filled the grid. It uses a different method and capacity from the gradient-trained NCA above, so these images are not a matched comparison or a measure of improvement.
Local rules. Emergent structure.
Exploratory trajectories ask how learned local updates create and sustain larger patterns. A visual trace can reveal behavior; it cannot establish general reasoning.
EXPLORATORY / INDEPENDENT TRACKWhere should computation live?
Exploratory work on depth and repeated computation. Comparisons need matched budgets and controls before any efficiency claim.
EXPLORATORY / INDEPENDENT TRACKStructure as a working medium.
Exploratory work on relationships and reusable structure. Transfer beyond the development setting remains a separate question.
EXPLORATORY / INDEPENDENT TRACKA different kind of signal.
An independent exploratory track. Audio observations do not establish language or agent capability.
EXPLORATORY / INDEPENDENT TRACKAn SMC-trained learning agent in a procedural computer world, now in development. The aim: acquire reusable mechanisms, adapt to changed APIs, and retain capability across tasks.
Can learning carry forward?
The first executable integration is under construction. These steps describe the intended study, not a recorded agent run. Persistent learning and generalization have not been demonstrated. The exploratory NCA, layers, graphs, and audio tracks remain independent.
Read the existing research design ↗Give the learner a bounded task in a procedural computer environment, with explicit tools and feedback.
Infrastructure for doing the work.
The tools around a model matter. They also need to be counted.
SMC Agent is the infrastructure track, separate from trained-model research. A useful tool-using system does not, by itself, prove that its underlying model has learned a new capability.
Attribute gains to the complete system that produced them.
Real exploratory work and a validated public finding are different stages. We want the distinction to remain visible, even when a demo is compelling.
Development observations help choose the next question. Private exploratory results are not presented here as independently validated claims.
Public examples need approved synthetic aggregates, provenance, context, and limitations. This NCA replay is a reviewed exploratory recorded example: one selected case from one training seed, not an independently replicated capability claim.
Require a frozen protocol, strong baselines, complete costs and errors, and independent reproduction—or an explicit unreplicated label.
E01 has its own structured-record protocol and pilot run records. Read the canonical ledger for its outcomes and status. This exploratory NCA reproduction is not an E01 result. E02 recurrence and E03 temporary adaptation retain their own protocols.
Canonical run ledger ↗The canonical record holds protocols, review, and manifested findings.
Canonical Forgejo ↗Public results ledger ↗Evaluation policy ↗Small models, practical constraints, and an open record of what survives the experiment.