THE CURIOSITY LAB
A little less scrolling.
A little more wondering.
Follow a connection. Question a claim. Build a tiny brain. There’s a whole world hiding behind a good question.
13 ways inTake your time
FIND YOUR WAY IN
What are you in the mood for?
13 experiences
ResearchOS
Good ideas need room.
Articles, notes, radio, and experiments. Arrange them in your own research desktop.
Research Constellation
Ideas have neighbors.
Move the map. Follow a connection. Find the question between two questions.
The Hallucination Room
Confidence is not evidence.
Six investigations. Three convincing claims. One leap too far.
Build a Tiny Brain
The system is the experiment.
Connect perception, memory, planning, tools, and safety. Then see what happens.
Signal Radio
Something is on the air.
Tune a quiet, imaginary radio. Discover a thought between frequencies.
Ask This Article
Keep the evidence close.
Ask a question. Inspect the passages. Follow the references.
The Hidden Frequency
A second look pays off.
Find five fragments tucked around the site. Unlock a little something extra.
Your Research Desk
Make an idea your own.
Collect stories, organize projects, write notes, and take a copy with you.
Lost Signal
Take the scenic route.
A tiny maze, three fragments, one portal. No timer. No rush.
The Unanswered Room
A question needs room.
An interactive fiction about evidence, permission, and the space between them.
Your Personal Observatory
Curiosity leaves a little light.
An opt-in, local constellation of the ideas and places you explore.
Research Time Machine
Good questions travel.
Revisit eight selected papers. Compare their findings, limits, and open questions.
The Quiet Library
A thought needs a margin.
Read beside another story. Keep exact-passage highlights and your own notes.
SMALL IDEAS, MADE VISIBLE
Stay a little longer.
01 / ATTENTION, MADE VISIBLE
What gets your attention?
Same sentence. Different focus. Watch a tiny attention calculation shift as you change the query.
“robot” receives the most weight: 32.8%. This query emphasizes the first dimension.
Look under the hood +
Each word has a hand-authored 3-dimensional key. The chosen query is also a 3-dimensional vector. We calculate dot products, divide by √3 and the temperature, then apply softmax. These labels describe our toy vectors; they are not learned semantic categories.
softmax(Q · Kᵀ / (√3 × temperature))| Token | Key vector | Weight |
|---|---|---|
| the | [0.1, 0.2, 0.1] | 6.51% |
| curious | [0.8, 0.2, 0.5] | 22.08% |
| robot | [1, 0.9, 0.2] | 32.81% |
| finds | [0.4, 0.7, 1] | 13.46% |
| a | [0.1, 0.1, 0.1] | 6.41% |
| path | [0.6, 0.8, 0.9] | 18.72% |
This is one unmasked attention distribution with invented vectors. It does not run an AI model or show a model’s internal reasoning. Real transformer layers also use learned projections, values, masks, and multiple heads.
SPACE TO THINK
Longer context. Bigger footprint.
Explore how tokens, attention heads, and concurrent sequences change raw KV-cache memory.
EXPLORE THE CALCULATION
How context changes the raw cache
Fixed example: 32 layers and head dimension 128. Change the retained tokens, KV heads, and sequence count.
Excludes weights, workspaces, allocator and quantization metadata. This is tensor arithmetic, not a serving-capacity benchmark.Curiosity is a useful habit.
These experiments illustrate ideas. For the evidence, assumptions, and questions behind them, spend some time in the journal.
Keep exploring ↗