Pack the brain
Drag content blocks — system prompt, chat history, retrieved docs, the question — into a fixed context window sized to a real model. Overflow the window and the lowest-priority blocks get evicted; even for what stays, an attention heat-strip shows the middle of a stuffed window going cold. Then watch a buried answer get the wrong response. Runs entirely in your browser (0 uploads, works offline).
How to use it
- Pick a context window size and choose which blocks to include (checkbox) and their order (↑/↓).
- Watch the fill bar and the attention heat-strip — brighter means more attention, dim means the model barely reads it.
- Read the verdict: whether the answer-bearing doc gets enough attention to be used — or was evicted or lost in the middle.
What this clears up (the fundamentals)
- A bigger window isn't a bigger memory — attention doesn't spread evenly. Models attend most to the start and end of the context and least to the middle ("lost in the middle").
- Context rot is real — as a window fills, attention degrades across the board. A fact that would be used in a lean prompt gets ignored in a stuffed one.
- Eviction is a priority decision — when content overflows, something gets dropped. If you don't decide what, the system drops it for you — sometimes the answer.
- Placement matters — the same retrieved doc helps at an edge and gets ignored in the middle. Where you put a fact is part of the prompt.
Where it's used
An AI-engineering literacy tool — the fastest way to feel why "just paste everything into the giant context window" produces worse answers, not better. It's a miatz build-lab concept playable; learn to build it yourself.
FAQ
What is "lost in the middle"?
An observed behaviour where language models retrieve information reliably from the beginning and end of their context but poorly from the middle. Bury the relevant fact in the middle of a long prompt and the model often ignores it.
What is context rot?
The general degradation of attention and answer quality as a context window fills up. More tokens dilute the model's focus, so relevant content gets less effective attention than it would in a shorter prompt.
How is attention estimated here?
With a U-shaped position curve (high at the edges, low in the middle) multiplied by a fill penalty that grows as the window approaches full. It's a teaching approximation of the real, measured effect.
Is anything uploaded?
No. The packing and attention math run entirely in your browser — nothing is transmitted or stored. Turn off your Wi-Fi and it still works.
Limits
A teaching model with fixed blocks and an approximated attention curve — real models vary in how strongly they show these effects, and good retrieval/re-ranking mitigates them. Use it to build the "lean, ordered context" instinct, then measure on your own prompts.
Related
Part of the miatz playgrounds. Explore the rest from the playground catalog.
Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.