Documentation / What is ML Learning Lab

What is ML Learning Lab

Edit: src/content/docs/introduction/what-is-ml-learning-lab.md

A focused learning platform for transformer systems and AI agent framework internals.

ML Learning Lab is a technical learning platform for people who want to understand how modern language-model systems and AI agent frameworks actually work.

The current platform has two first-class tracks:

  • Transformer Systems - from tokens and embeddings to attention, positional encoding, KV cache, quantization, decoding, and efficient LLM serving.
  • AI Agent Frameworks - OpenHands-style agent internals, including agents, controllers, runtimes, events, memory, tools, microagents, and conversation state.

The product combines long-form technical articles, structured lessons, roadmap navigation, and Anki-style flashcard review. The intended experience is close to a guided learning site like Brilliant for difficult concepts, but with a retention layer built into the platform.

Learning by Understanding

The platform is not trying to be a generic blog. It is organized around the idea that technical depth comes from three repeated activities:

  1. Reading a precise explanation.
  2. Reconstructing the mechanism from memory.
  3. Returning to the idea later until recall becomes reliable.

Transformer and agent systems have many moving parts. It is easy to read one article about attention, KV cache, or an agent event stream and feel that the concept is understood. The harder part is remembering how the pieces connect one week later.

ML Learning Lab treats recall as part of the product, not as a separate study habit.

What the Platform Offers

Articles and Lessons

Deep explanations rewritten into lessons, with diagrams, equations, implementation notes, and links to the original article routes while the platform is still being reshaped.

Roadmaps

Visual learning maps that break the content into stages, dependencies, and recommended order instead of a flat archive of posts.

Flashcard Challenges

Small review decks attached to lessons and roadmap nodes. Each card targets a specific concept, mechanism, equation, or implementation detail.

Daily Practice

A dashboard queue that selects review items across the key categories and turns progress into a recall score for each roadmap.

Why Transformer Systems

Most transformer tutorials stop at the architecture diagram. This platform treats transformers as systems: tokenization, embedding tables, QKV projection, attention masks, positional information, normalization, FFNs, KV cache layout, decoding strategy, quantization, and serving constraints all matter.

The Transformer Systems track is meant to answer questions such as:

  • Why does attention need a mask, and how does it change between training and decoding?
  • What exactly is stored in the KV cache, and why does it dominate memory?
  • How do MQA, GQA, MLA, FlashAttention, and quantization change the serving tradeoffs?
  • How do decoding strategies such as sampling, speculative decoding, Medusa, and lookahead decoding reduce latency or improve generation?

Why AI Agent Frameworks

Agent frameworks are often explained at the product level: prompt, tool, result. The AI Agent Frameworks track looks below that surface.

The initial content focuses on OpenHands-style internals:

  • agent loops and controller state machines
  • action, observation, and event streams
  • runtime boundaries and sandbox execution
  • tool calling and function-call flows
  • memory, microagents, and repository-specific context
  • session startup, services, and conversation lifecycle

The goal is to make agent frameworks feel inspectable. A learner should be able to explain how the framework moves from a user request to tool calls, runtime events, observations, and a final answer.

What This Is Not

ML Learning Lab is not currently a remote execution playground. The earlier playground idea has been reduced to lesson-attached practice and flashcard challenges.

Future versions may include simulations or notebooks for selected topics, but the core product is:

structured technical content plus spaced repetition, organized by roadmaps.

The End Goal

The long-term goal is a platform where a learner can open a roadmap, read the next lesson, answer targeted cards, and see whether they actually remember the concepts that matter.

For machine learning systems, this matters more than passive completion. A finished article is not the same thing as an internalized model of the system.