An LLM Is (Not Really) a Black Box Full of Sudoku and Tic Tac Toe Games (mikenotthepope.com)

Playing with Ollama and the qwen2.5-coder:7b model, the author moved from playful prompts to a useful realization: an LLM isn’t a mystical black box but a multigigabyte file of numeric grids (matrices) — the model’s weights. Text is tokenized into numbers, converted to embeddings, then repeatedly multiplied through those matrices across many layers. Each pass produces scores for the next token (logits → probabilities), the chosen token is fed back in, and the cycle repeats until a reply is produced. In short: words in → linear algebra operations → words out — a “pinball machine” of matrix math rather than hidden game logic. This matters for the AI/ML community because tools like Ollama let developers run models locally, demystifying deployment and enabling experimentation, privacy-conscious uses, and better model selection (e.g., text-only vs multimodal). Understanding that models are collections of weights clarifies why model size, architecture, and training data matter, why some models can’t generate images, and how performance relates to linear algebra primitives (matrix multiplies, embeddings, autoregressive decoding). For engineers and learners, grasping this pipeline — tokenization, matrices, layer-wise transforms, and autoregressive sampling — is enough to bridge web-facing services (ChatGPT.com) and the underlying model being queried.