{"repo":"Jeba-Jebarsan/insideai","free":true,"listed":false,"github":"https://github.com/Jeba-Jebarsan/insideai","clone":"git clone https://github.com/Jeba-Jebarsan/insideai.git","description":"Watch a real AI model think - live 3D visualization of a transformer's actual forward pass: tokens, embeddings, attention, neurons, sampling","language":"TypeScript","stars":20,"topics":["education","interpretability","llm","nextjs","pytorch","threejs","transformer","visualization"],"license":"MIT","category":"ai-agents","readme_excerpt":"InsideAI Watch a real AI model think. InsideAI runs a genuine language model on your machine and turns every step of its computation into a live, interactive 3D visualization — tokenization, embeddings, attention, neurons firing, probabilities, sampling, and the answer streaming out. No videos, no papers, no mock data: what you see is the model, running. Is there a real LLM behind this? Yes. By default InsideAI runs Qwen2.5-0.5B-Instruct (494M parameters, 24 layers) locally in the Python backend — it actually answers your prompt. Every visual is driven by tensors extracted from its live forward pass: - attention arcs and heatmaps are the model's real softmax(QKᵀ/√d) matrices - the neuron diagram lights up with actual GELU/SiLU activations, wired by the model's true weight matrices - probability bars are the exact distribution the sampler draws from - the layer stack glows by how much each layer really changed the token's meaning (‖Δresidual‖) Where data must be reduced to fit a screen (4864 neurons → 128 groups, 896 dimensions → 3 PCA axes), the reduction is documented and labeled in the UI — never invented. Details: docs/ARCHITECTURE.md. What you'll see --- --- Neurons — the classic textbook network, live: green input → blue hidden neurons → red output. Wire brightness = real connection strength × the signal flowing through it, one pulse per generated token. Embeddings — every token becomes a point in space (PCA view of the real vectors); similar meanings sit close together.","default_branch":null,"files":null,"tree":[],"storefront":"/r/Jeba-Jebarsan","claimed":false,"request_supported":{"post":"https://gitbuyer.com/r/Jeba-Jebarsan/insideai/request-supported","requests":0},"note":"indexed from public GitHub; nothing is for sale on this page. Clone it from GitHub. Paid listings live at /search."}