Notes from the lab
Writing
On visual representations, controllable generation, and the evidence needed to understand what an agent can actually do.
Evaluate an agent by the state it leaves behind
Agent quality depends on task completion, constraints, and recovery. Inspect the trajectory and verify the outcome.
From perception to action: the architecture of grounded AI
Controllable video, editable 3D content, predictive world models, and software agents—connected through state, action, and outcome verification.
Start with one workflow and a measurable outcome
A practical way to scope an AI pilot: define the task, establish a baseline, and count the cost of failure and review.
The model and the system matter
Reliable AI needs strong learned representations and sound system design. Our updated view on models, memory, and evaluation.
Data quality is a model-design question
Representation, supervision, and data quality need to be studied together. A practical approach to diagnosing multimodal failures.