Could AI face a liquidity crisis in 2027?
Big Tech AI CapEx may be pressing against a practical cash ceiling, while private AI capital is not a clean substitute for data-center financing.
Read note
Matrix Team
We build world models, video systems, agents, and robotic intelligence that learn from action, remember across time, and run efficiently enough to stay in the loop with real environments.
Vision
A world model is most useful when it is interactive: it receives actions, updates memory, predicts consequences, and returns feedback quickly enough for agents, robots, and humans to keep acting.
We aim for models that do more than generate plausible video. They should preserve structure, expose controllable state, support long-horizon reasoning, and become a substrate for perception, planning, and embodied learning.
Represent actions, constraints, rewards, controls, and environment responses as first-class signals.
Carry geometry, identity, temporal context, and learned experience across long rollouts.
Make inference and simulation fast enough for continuous feedback and scalable deployment.
Video models give us a visual substrate. World models add state and consequence. Agents close the loop by choosing actions. Robotics makes the loop physical.
Matrix Team connects these layers through research on controllable generation, interactive environments, memory, understanding, and efficient computation. The long-term target is an AI system that can perceive a situation, simulate alternatives, act, and learn from the outcome.
We model the loop between agents, controls, environments, and visual change so intelligence can be grounded in what a system can do.
+Persistent memory keeps identity, geometry, intent, and causal traces available across long interactions instead of resetting every frame.
+Interactive intelligence needs fast inference, stable rollout, and careful compute allocation so models can remain responsive while scaling.
+We connect video, robotics, games, and physical environments so learned models can transfer across scenes, tasks, and control modes.
+Research systems
Selected papers, datasets, and systems ordered by public release date, with lightweight previews for fast browsing.
No project index found yet. Run node scripts/generate-project-index.mjs to generate it.
Blog
Compact positions on AI systems, research taste, markets, and the infrastructure behind intelligence.
Big Tech AI CapEx may be pressing against a practical cash ceiling, while private AI capital is not a clean substitute for data-center financing.
Read noteWhy notes?
Short positions on bottlenecks, incentives, infrastructure, and strange signals around interactive intelligence.
View all notesThread to note
Working conversations become public only when the claim, evidence, and counterpoint are sharp enough to be useful.
Position essaysA compact group of researchers building interactive world models and agent systems. Interests are shown instead of hierarchy. View full list. Listed alphabetically.
About
Matrix Team is an independent research interest group bringing together young researchers across world models, video systems, agents, robotics, and interactive simulation.
We stay small, practical, and research-driven: build systems, test ideas in public projects, and collaborate with people who care about long-horizon memory, embodied interaction, and efficient computation.
Interested in our work, open research questions, or possible collaborations around interactive intelligence?
Reach out anytime at fengruili.frl@gmail.com.
