Research / Areas

Research

I explore the foundations of increasingly capable, reliable, and general-purpose AI systems.

Pillar I

Intelligence

Models that reason, perceive, and build an understanding of the world.

Intelligence

Foundation Models & Reasoning

Model architectures01
Training & post-training02
Reasoning03
Planning04
Self-improvement05

Fig. 1Search over reasoning steps. Grey branches are explored and pruned; the dark path is the one the model commits to.

Intelligence

Multimodal & Embodied AI

Vision01
Audio & video02
Spatial intelligence03
World models04
Robotics05

Fig. 2Perception to action. Separate inputs share one representation, which a world model uses to plan physical movement.

Pillar II

Autonomy

Systems that act over long horizons, run efficiently at scale, and make discoveries.

Autonomy

Agentic Intelligence

Multi-agent coordination01
Long-horizon planning02
Agent memory03
Tool use04
Agent learning05

Fig. 3A planner splits work across agents that share tools and memory, and keeps them on track across a long sequence of steps.

Autonomy

AI Systems & Infrastructure

Efficient inference01
Distributed training02
AI runtimes03
Model serving04
Compute efficiency05

Fig. 4Keeping compute busy. Incoming requests are grouped into batches so each accelerator does more work per unit of energy and time.

Autonomy

AI for Science

Mathematical reasoning01
Scientific discovery02
Hypothesis generation03
Experiment design04
Automated research05

Fig. 5The research loop, automated. A model proposes a hypothesis, designs the experiment, and updates its theory against what it observes.

Pillar III

Control

Methods that keep capable systems understandable, safe, and measurable.

Control

Interpretability

Model internals01
Mechanistic interpretability02
Representations03
Behavior analysis04
Model monitoring05

Fig. 6Tracing a circuit. Out of every connection in the network, only a few carry a given behavior. Finding them shows what the model is actually computing.

Control

AI Safety & Security

Alignment01
Scalable oversight02
Robustness03
AI security04
Agent safety05

Fig. 7Defense in depth. Layered safeguards stop adversarial inputs, an overseer watches the model, and each action an agent takes is checked before it runs.

Control

Evaluation & Measurement

Capability evaluation01
Agent evaluation02
Reasoning benchmarks03
Safety evaluation04
Reliability testing05

Fig. 8Illustrative scorecard. Every score has an uncertainty range, and each one is compared against a bar a system must clear before release.

Research Frontier

We are particularly interested in the intersection of intelligence, autonomy, and control: building AI systems that can reason, learn, act, and discover while remaining reliable and understandable.