Visual Learn

How machine-learning systems work, explained with figures you can operate

Each explainer follows one idea from the problem it solves to the mechanism that solves it. The figures run in your browser: move a slider, step through a computation, compare two designs side by side. Where a figure uses a simplified or simulated model, its caption says so.

Start with JAX and PyTorch to follow a computation, quantization-aware training to examine rounding and memory, or reinforcement learning to understand learning from outcomes. Each tutorial defines its terms and links its claims to sources.

Systems

• Speculative decoding

  Step through draft verification, see how probability correction preserves the target model’s output, and explore when a smaller draft model saves time.

  6 sections, 5 figuresabout 15 minutesspeculative-decoding/

• JAX and PyTorch

  Follow one Python function from tracing and compilation to the accelerator, and compare how the two frameworks fuse, differentiate, batch, randomize, and shard it.

  10 sections, 12 figuresabout 25 minutesjax-vs-pytorch/

• Google Cloud accelerators

  Compare the GPUs and TPUs Google Cloud rents on memory, bandwidth, compute, interconnect and price, and see which of them limits a given model.

  7 sections, 7 figuresabout 15 minutesgcp-accelerators/

Training

• Reinforcement learning for LLM agents

  See how reward signals change a language model, from a single policy-gradient step through GRPO, reward models and KL penalties to reward hacking and tool-using agents.

  10 sections, 9 figuresabout 30 minutesrl-agentic-llms/

• Quantization-aware training

  Round weights onto a low-precision grid, follow the gradient through the rounding step, and train a small network both ways in your browser.

  9 sections, 8 figuresabout 20 minutesqat-explainer/

Models and papers

• Open-weight language models

  Explore the mechanisms behind Llama, DeepSeek, Mixtral, Kimi, GLM, gpt-oss and other open-weight families through worked examples and interactive figures.

  40 tutorials, 86 figures2023–2026open-weight-llms/

• K2 Horizon

  Read a released model family through its configs: dense and mixture-of-experts layers, routed value projections, diffusion drafting, how the largest model was trained, and what the reported benchmarks show.

  7 sections, 8 figuresabout 18 minutesk2-horizon/

• Register tokens

  Why large vision transformers write global information into a few background patches, and how a handful of extra tokens gives it somewhere else to go.

  6 sections, 8 figuresabout 12 minutesregister-tokens/

Data

• Intelligence versus cost

  Plot language-model benchmark scores against what each run costs, find the frontier of best trade-offs, and see how it filled in quarter by quarter.

  3 sections, 2 figuresabout 7 minutespareto-front/