Run real code
Python in your browser
Pyodide runs NumPy, matplotlib, and your code right in the page. No setup, no Colab tab dance.
Run real code
Pyodide runs NumPy, matplotlib, and your code right in the page. No setup, no Colab tab dance.
Your key, your model
Anthropic, OpenAI, Google, OpenRouter, Ollama, DeepSeek — anything. We never spend on AI, you do.
Made for beginners
Drag arrows, slide softmax temperature, step through matmul — feel the math before you see it.
The vector + gradient + probability foundation every LLM idea rides on.
Write training loops you understand down to autograd. Debug exploding gradients with confidence.
From word vectors to attention — and why attention won.
Implement multi-head attention and a full decoder-only Transformer from a blank file.
Llama 4, Qwen3, DeepSeek V3/R1 — what's actually inside the frontier.
Estimate compute-optimal trade-offs; use FSDP and parallelism that doesn't OOM at 3am.
Take a base model to instruction-following with the right algorithm for your budget.
From ad-hoc prompting to a real vocabulary — CoT, self-consistency, budget forcing, reasoning models.
Build a production-grade RAG that beats naive cosine-similarity on your own eval.
Multi-tool, stateful, durable agents. Expose your tools to the world via MCP.
Trustworthy evals before you trust any output. Goodhart's law in practice.
Serve at >1000 tok/s/GPU. Quantize down to consumer hardware without losing your benchmarks.
Red-team your own model. Reproduce a basic SAE. Understand current limits of mech interp.
Read 1 paper/week with full understanding. Reproduce one paper per quarter. Ship a preprint.