AI research
Inside ASI: models, experiments, and research tools
Connecting neural network ideas to inspectable implementations, experiments, and tools for working with the models.
By Curtis Tech Solutions3 min read
An AI research idea becomes easier to examine when the reasoning, implementation, and experiment are available together. Each provides a different kind of evidence: what the idea proposes, how the code expresses it, and what happened when it was tested.
ASI is Curtis Tech Solutions’ public research repository for experimental neural network architectures, learning methods, and memory systems. The work includes model implementations, research write-ups, recorded experiments, and tooling for training and interacting with the models. This article looks at how those pieces fit together.
Connect research notes to model code
The research directory develops ideas around cyclic graphs, sine-based activation functions, gradient behavior, and Double-Negative Reinforcement Learning, or 2NRL. The write-ups connect the reasoning and worked examples to specific parts of the implementation.
RadixCyclicNN is one of the model implementations. Its design explores a cyclic graph structure, learned activation behavior, and prediction through graph search. Related variants let the project investigate different structures and ways of retaining or changing memory.
This organization gives a reader a concrete route through the work: begin with a proposed mechanism, find the code responsible for it, and then look for an experiment that examines the claim. That path makes the research easier to inspect than an isolated demonstration.
Give experiments a place of their own
The experiments directory includes activation comparisons, learning-method trials, gradient-normalization work, and other focused investigations. Separating these runs from the main model keeps the question under investigation visible.
The recorded material includes mixed and unsuccessful outcomes as well as promising observations. These results help describe the conditions under which an idea was tried. A result from one task or configuration still needs that context when someone considers its wider significance.
The useful engineering habit is to preserve the question, setup, and observed result together. That makes the next experiment easier to plan and gives another reader something specific to examine or try to reproduce.
Separate the model from its tools
ModelKit contains much of the software around the models: teaching loops, command-line tools, an HTTP API, a React interface, model clients, and agent tools. Its dependency rule runs in one direction. The toolkit can depend on the model, while the core model remains independently usable.
Some ModelKit components can also work without the experimental model installed. That lets utilities such as client interfaces and tool connections be reused without pulling the model’s entire training and execution path into another application.
This boundary is valuable during research because the model and the surrounding workflow change for different reasons. A new interface or teaching tool can be developed without making it part of the model’s core responsibilities.
Make the work inspectable
CTS’s contribution spans both the experimental models and the practical tools used to work with them. Designs, implementation choices, and experiment records provide a way to follow the project beyond its headline ideas.
ASI is best explored through those concrete artifacts. The research directory explains a proposal, the model directories show its implementation, and the experiments document observations and limits. Together they support the continuing work of testing and refining the ideas.