Repository navigation
User Guide
Choose a validated model in the selector, write a text prompt, and press Send. The model streams its response into the chat pane while the connectome responds to each received token.
Press ∞ Inf to start an open-ended local roleplay. AIBrain loads two separate copies of the selected GGUF: the World produces only external events; the Participant produces only first-person responses. Every stream update is tied to both its role and turn number, so World tokens cannot append to a Participant bubble. The UI clearly labels this as a simulation—the model is generating roleplay text, not a real sentient being. The session continues until Stop is pressed, while each model keeps only a bounded recent context window so it can run without unbounded prompt growth. The right-side connectome renders participant tokens only.
NN Analysis+ replaces the compact analysis action. It creates a compact JSON or compressed JSON data file with the full conversation and a trained online autoencoder's findings: reconstruction/coherence metrics, regional patterns, temporal segments, and selected high-novelty events. It processes every recorded visual frame but does not dump every neuron vector. These signals are the connectome's procedural visualization data; they are not measured model hidden states or transformer activations.
- Stop requests cancellation after the current native generation step.
- Escape does the same from anywhere in the application.
- Regenerate removes the latest assistant answer and resends its preceding user prompt.
- Clear removes visible messages and in-memory conversation history.
-
Ctrl+Cin the launch terminal requests a clean application shutdown.
The latest completed answer has replay controls beneath it. Replay neurons plays its captured activity sequence; Token arrows move one recorded step at a time. A new response, model selection, graph-quality change, or spacing change replaces that one-response replay buffer.
The controls at the bottom of the chat pane are saved in Windows application settings when a response starts.
| Control | Effect |
|---|---|
| Temperature | Sampling randomness, from 0 to 2. |
| Top-p | Nucleus-sampling cutoff, from 0.05 to 1. |
| Max tokens | Upper limit for a generated answer. |
| Context | Context window requested from llama.cpp. |
| GPU layers |
-1 requests automatic/all-layer offload where the installed backend supports it; 0 requests CPU inference. |
| Generation speed | UI token pacing from 0.1x to 1.0x. It does not change model sampling. |
The right pane is a native QOpenGLWidget renderer. It supports left-drag to orbit, mouse-wheel zoom, and Reset
view to restore the camera. Pause stops visual updates; press Resume to continue. Low, Medium, and
High quality graphs contain respectively 5,000, 11,000, and 22,000 visual neurons. The Spacing slider shows its
selected multiplier immediately and rebuilds the procedural cluster layout after you pause movement, avoiding repeated
expensive graph rebuilds while dragging. Region colors remain visible even before activity begins.
The GPU selector asks Windows to use the high-performance adapter for the project Python executable on the next launch. The overlay always identifies the renderer actually selected by OpenGL.
Click a visual neuron to inspect its region, current value, and peak. You can silence it (zeroing its visual activity) or change its importance from 0 to 3. Click the displayed Selected neuron number to enter an exact neuron index. Neuron borders outlines every visual neuron for clearer separation; use the adjacent numeric control to tune the outline width. Region clusters use a stable distinct color palette. All message bubbles and status/inspector text can be highlighted and copied. These controls affect the visual simulation only; they never modify model weights or inference.
After generating a response, choose NN Analysis+ to create its compact neural-analysis JSON. Full semantics and file layout are documented in Connectome and analysis. Chat bubbles render safe Markdown for headings, emphasis, inline code, and HTTP (S) links while preserving selectable source text. Generated chat follows the newest token only while the message list is already at the bottom; scroll up to read or copy earlier text without being pulled back down.