Each note lives somewhere on the spectrum from seedling (a half-formed thought) to budding (worked out but still malleable) to evergreen (stable, revisited often). Click any note to read it, then follow the backlinks at the bottom to wander.
Garden
A slow-growing collection of interlinked notes. Some are budding, some are taking root, some are in full bloom. Nothing here is final.
- What even is a world state?
We talk about world models as if 'the state' were given. The representation is the hard part.
- The bottleneck is rarely the model
Across four very different projects, the thing that capped quality was almost never the network. It was a systems or representation cost hiding underneath.
- The test distribution is the only one that counts
A strong validation number can be a trap. My clearest version of this lesson came from a 0.58 that became a 0.35.
- Cheap structure beats heavy machinery
The simplest mechanism that clears the bar keeps winning over the impressive one that doesn't. A pattern across a startup, a challenge entry, and a rendering project.
- Read the data before you add machinery
Supervision and structure are often already in the dataset, hiding in plain sight. The second vehicle in a sentence was free labels.
- Build on the representation that already travels
PedroVerse stylized 3D assets by editing only their UV maps, because UV maps render everywhere. Choosing the substrate that is already universal buys portability for free.
- When discovering the action space is the game
Text adventures let you guess at an open verb set; modern games hand you a closed one. The same tension shows up in agent action spaces.
- What would it actually take to build Royal Road?
Through a researcher's lens: the properties that make Royal Road great, how close today's LLM-driven games get, and the gap that remains.
- The best stories in Royal Road were never written
What makes the world feel infinite isn't more content, it's emergence: outcomes no designer anticipated, born from player-world synergy.
- Royal Road, before I had the words
Legendary Moonlight Sculptor is one of my favorite novels, and the things that hooked me as a kid turned out to be my research.
- A world in a grain of sand
Blake's image, read through world models: the whole inferred from the detail, both as a design principle and as evidence the model worked.
- A shared channel is not a shared world
Generative agents coordinate through language and memory, but they never share a world. That ceiling is the whole point.
- Shared state is a tiny world model
A multiplayer game restated my research agenda in miniature: two agents can only coordinate when they agree on one authoritative model of the world.
- The loss decides what the model ignores
Two projects where the objective, not the architecture, dictated what the model paid attention to and what it left blank.
- World models as shared substrate
Why a common representation of the world is upstream of believable multi-agent behaviour.
- Why now
A personal note on the path from audio processing at Huawei to wanting to spend six years on world models.
- Reachy pilot study: what 38% means
Our autonomous HRI pilot fooled participants 62% of the time, but the way it failed mattered more than the number.
- Grounding signals beyond language
If we want agents to share more than what can be said, the grounding signal has to be richer than text.
- Every conditioning choice is a hidden assumption
Lessons from extending IMLE to text-conditional generation: failures almost always traced to latent mismatch.