Research
Sharp notes on models, agents, and product craft. New pieces show up below when they're published.
Physics AI
Where generative models meet the physical world
Physical AI describes autonomous systems—cameras, robots, vehicles, spatial compute—that don't only generate text or pixels, but perceive, reason, and act under real-world physics constraints. Modern generative models excel at language and images yet have a limited native grasp of 3D space, mass, contact, and sensor noise; physical AI closes that gap by grounding policies in geometry, simulation, and multimodal sensing.
Teams typically combine large multimodal or world models with physics-based simulation, synthetic data, and reinforcement or imitation learning so behaviors are trained safely before deployment. That stack shows up everywhere from manipulation and mobile robots to autonomy, digital twins, and large-scale spatial understanding—complementary to the software- and model-focused research notes below.
Terminology and framing inspired by NVIDIA's overview of physical and generative physical AI — see Generative Physical AI (NVIDIA glossary).
Research that ships with the stack
Model choices, evals, and agent UX—plus chains, proofs, and infra notes when they matter. Written for builders, not buzzwords.
No filler. No deck theatre.
If it doesn't help you ship, it doesn't belong in this feed.
New notes publish here first
Nothing live yet—when posts are ready in the CMS, they'll stack in the list above this block.