Interactive Study Lab

Top AI engineers &
researchers in AI labs
build intuition.

We dissect signals, probe edge cases, and extract intuition that compounds. No hype. Just signal.

Real ML Source CodeArchitecture DiagramsTerminal LogsClear Explanations

We at FBA First Break AI & FBA Labs will be doing this on the weekend.

// tracks

Start with one lab track

Both tracks are self-contained. Start with speedrun if you want to understand how training improves a model. Start with inference if you want to understand how a model runs.

Phase 1 · Training

speedrun

Start with a baseline GPT-2. Watch six researchers' decisions — RoPE, Muon optimizer, logit capping, 32K context — each annotated in the training code and plotted on the loss curve. This is how models actually get better.

  • 6 iterations
  • 3.28 target loss
  • Python / nanoGPT
Start speedrun →Source: 6 runfiles on GitHub
Phase 2 · Inference

Qwen3 Pure C

The AI doesn't plan sentences. It picks one word, then the next — based on every word before it. Six stages trace how: loading the weight file, running attention, choosing the next token. ~1,200 lines of C. No black boxes.

  • 6 stages
  • ~40 blocks
  • C / GGUF
Start Qwen3 walkthrough →Source: run_viz.c on GitHub
fba-lab — bootready
FBA Lab study mode — pick a track to begin$ fba load-tracks: speedrun, qwen-csimulation ready — pick a lab track ✓
// preview

One block. Four synced views.

Pick a function on the left. The source code highlights in the center, the architecture diagram pulses on the right, the terminal logs what happened, and the explanation tells you why it matters.

Open the live lab →

Block-synced everything

One click updates the code highlight, diagram node, terminal log, and explanation — always the same block.

Shorts + block tour

Auto-advance presentation mode with a record-ready 9:16 layout for cohort content.

Teacher authoring

Annotate diagrams, merge blocks, build slide decks, export shareable JSON.

Repositoriesfba-labQwen3-RunLocallyddp-training