What it does
SILT is Neural Child under a renamed, reconfigured instance. Same five-network architecture, emotional memory, object permanence, pattern recognition, movement prediction and an integration layer, with different configuration and a separate memory tree.
The reason to run a second instance is methodological. A developmental system with file-based memory and a randomized lifespan produces a history, and a history from a single run is an anecdote. Two instances under different configurations make it possible to say which behaviors belong to the architecture.
In the code
instance/silt/
config.json # the configuration this instance was run under
lifespan.seed # randomized at init, fixed for the run
memory/emotional/ # one file per retained affective association
memory/objects/ # conceptual continuity across context breaks
memory/patterns/
journal/ # what the instance recorded, in orderIllustrative shape. Because the memory is files rather than a context window, two instances can be diffed after the fact, which is the whole reason to run a second one.
How this differs from the ordinary version
Memory is files, so a run leaves a record
Because memories persist as actual files rather than as a context window, an instance can be inspected after the fact and compared against another. That is what makes a second instance worth the cost.
In the field
What this is not
It is not a product and not a deployment. It is a control condition for a research question about whether developmental behavior is a property of the architecture or of one lucky run.
Questions
- How is SILT different from Neural Child?
- Same architecture, different configuration and a separate memory tree. It exists so that behaviors can be attributed to the architecture rather than to one run.