Research

AI infrastructure, observability, and compute systems research.

A structured index of published work, frameworks, and active research arcs, from silicon observability to autonomous agent risk.

The Observability Stack Read flagship essay

Featured series

Observability for Frontier AI Systems

A connected body of work mapping silicon bring-up observability onto agent safety: tracing, replay, escalation checkpoints, maturity levels, and eval infrastructure.

Flagship essay Agent debuggability Agent Observatory

Index

Filter the research surface.

PublishedCompute Economics

The Cost of Usable Intelligence

A framework for measuring the cost of useful AI output under latency, reliability, safety, and quality constraints.

View project

PublishedAI Hardware & EDA

The GDS Is Not the Chip

Moonshot reports a 48-hour autonomous chip-design run. The public evidence does not establish a full tapeout; the more important test is auditable control of the engineering loop.

View project

PublishedAI Infrastructure

Subquadratic attention is an old dream. SubQ is finally worth auditing.

The 56× kernel and 12M-token claims are real but vendor-reported. The selector, not the speed, is the open question, and code is where it breaks.

View project

PublishedFrontier AI Systems

Observability is an underbuilt safety primitive for frontier AI systems

As AI agents move from answering to acting, safety depends on whether we can trace, replay, and debug their behavior.

View project

PublishedFrontier AI Systems

Debuggability for autonomous agents

A public thesis connecting silicon bring-up observability to agent safety, eval infrastructure, replay, and root-cause workflows.

View project

PlannedAutonomous Agent Risk

AI agents as infrastructure risk

Planned framework: tool-using agents as infrastructure-risk participants with reliability contracts.

View project

PlannedAI Infrastructure

The next AI infrastructure bottleneck is observability

Planned framework: observability, not raw compute, as the binding constraint on agentic workloads.

View project

DraftQuantum-Classical

Quantum computing needs an accelerated classical control plane

Draft memo: quantum systems depend on fast, observable classical orchestration.

View project

PlannedAutonomous Agent Risk

A taxonomy of autonomous agent failure modes

Planned taxonomy connecting agent behavior failures to the instrumentation that would catch them.

View project