What is Neural Automata?
Neural Automata is a local research environment for two separate experiments: a cellular memory model and a causal byte-language model. Native PyTorch performs training and inference on the selected CPU or CUDA device, while a browser visualises snapshots of real model state. It is not a combined online-learning chatbot or a general-purpose language-model product.
The separation is intentional. Writing an association into a model’s transient memory is not the same operation as changing durable model weights. The lab makes memory replay, memory training, language generation and language training distinct modes so the behaviour of each can be inspected.
Four modes with different questions
Memory replay loads baseline weights, freezes them and writes fresh associations into private state. Memory training learns a new set of shared weights and records loss and snapshots. Language generation uses the supplied byte-model checkpoint, while language training learns from the included toy text. New experiment outputs are written separately from the supplied baselines.
The bundled language checkpoint uses a short context and original template text. Its outputs should be interpreted as evidence about this constrained experiment, not evidence that it matches a general-purpose assistant. Browser recordings demonstrate execution; they are not comparative performance benchmarks.
Native compute, browser visualisation
The Node server binds to loopback and exposes fixed experiment modes rather than executing uploaded browser scripts. Python and PyTorch run on the local machine. The browser’s WebGL2 renderer displays snapshots; this is not Python running in WebAssembly.
Selecting several GPUs starts independent seeded experiments. It does not pool their memory or shard a single model. That makes comparisons easier to observe while keeping the meaning of multi-GPU execution explicit. CPU is an intentional execution option rather than a hidden fallback.
Reproduce before generalising
Use the research guide, recorded results and experiment commands to establish the original baseline. Confirm the selected device and inspect the saved logs and checkpoints. Small workloads need not saturate the GPU, and functional CUDA acceptance should not be reported as a speedup measurement.
The research project is separate from ZIPP. ZIPP retains its language VM and generic browser GPU examples; Neural Automata runs full native PyTorch experiments without requiring a ZIPP checkout. Review the migration record for provenance and licence scope.
Common questions
Is this a self-learning chatbot?
No. The current lab contains separate memory and language experiments, not a unified conversational system.
Does selecting two GPUs combine their VRAM?
No. It starts separate runs with different seeds. A single model is not split across the selected devices.
References & source
Read the project’s own documentation for implementation details, current releases and supported boundaries.
This overview describes the documented project scope. Check the linked source, licences and release notes for the exact version you plan to use.
