INDEPENDENT LEARNING RESEARCH
What changes
when a model
has experience?
Two research directions. A shared question about language, memory, and the difference between predicting an action and learning how to act.
Explore the research ↘Conceptual illustration Separate environments · no live connection
How do we tell the difference between remembering an answer and learning from experience?
01 / TWO WORLDS
Change the environment.
Examine the difference.
Adam and Eve are names for research systems. They are not claims about personhood, consciousness, or human development.
Language meets a bounded world.
Adam’s project records describe a small language model trained from random initialization, followed by a bounded environment with observations, actions, and a supervised learning workflow.
Early reports describe context-sensitive representations alongside repetitive action policies. That gap motivates the research; the reported numbers still need independent verification.
02 / FOLLOW AN EXPERIENCE
From an observation
to a defensible claim.
Select a stage to explore the documented workflow. This is an explanation, not a live run or an experimental result.
An environment becomes an input.
Adam’s documented environment offers observations and actions. Eve’s Loom proposal replaces familiar places with token streams. In both cases, the structure of the environment shapes what can be learned.
Check that the recorded input matches what the model actually received.
03 / THE EVIDENCE LENS
“Better” needs
a precise meaning.
A DIFFERENT CLAIM. A DIFFERENT COMPARISON.
Does language become more predictable?
Measure language prediction under a fixed evaluation. A lower loss alone does not establish effective action, broader knowledge, or understanding.
What is documented?
The Adam README and harness describe limited-data training and host-supervised verification. The August 25 Eve documents describe a matched pretraining protocol and a proposed stream-based world. Those documents establish what was specified; they do not verify every current implementation or result.
What remains to verify?
A claim-to-run appendix must connect numerical results to checkpoints, evaluation code, immutable records, and controls. Current Eve world implementation, comparative outcomes, and live interaction are not established by the sources reviewed for this page.
Where do the explanations come from?
Source basis: Adam project README, HARNESS.md, RESULTS-policy.md; Eve pretraining freeze and The Loom v0.3 design draft dated August 25, 2026. This page is a communication draft. The source documents were inspected locally; the underlying runs were not independently reproduced for this page.