Research overview
Four domains. One direction.
01Domain
Intelligence Systems
Language models are now general-purpose components, but the cost of running them, and the reliability of the systems built on top of them, limit where they can be used. We study both sides: how much of the computation spent on inference is actually necessary, and how software agents can reason about intent before acting on it.
Read the full domain brief →Open questions
- 01How much of the computation in a forward pass contributes to the output for a given input, and can that be determined cheaply at run time?
- 02Can inference-time optimizations be made model-agnostic, so that they apply to existing models without retraining?
- 03When a request is underspecified, how should an agent decide between asking, assuming, and testing a hypothesis?
- 04What does a verifiable plan look like for a coding agent, and how do you check that the work matches it?
Cost of computation
Infrastructure
Inference efficiency only matters if it survives contact with real hardware, drivers, and runtimes. Measurement methodology is shared with Computing Infrastructure.
Agents in tools
Simulation
AETHRA's agent bridge is a working case of an AI system editing a complex, stateful environment through reviewable transactions.
Ambiguity and planning
Applied
Reasoning techniques only prove themselves on real, messy tasks, where failures are concrete and observable. Applied systems are where that testing happens.
02Domain
Computing Infrastructure
Capability depends on access. Much of modern computation assumes a narrow set of hardware and software stacks, and the cost of leaving that path is paid in engineering time. We build tooling that makes alternative accelerators and reproducible workflows practical, and that reports honestly on what has and has not been measured.
Read the full domain brief →Open questions
- 01What does it take to prepare an open model for a non-default accelerator stack reproducibly, from metadata to a verified run?
- 02How should a tool report a measurement it could not take, so that missing data is never mistaken for a result?
- 03Which parts of an inference or rendering workflow are bound by hardware, and which by tooling and configuration?
Cost of computation
Intelligence
Inference-efficiency hypotheses need a measurement harness that controls for hardware, drivers, and run-to-run variance.
GPU pipelines
Simulation
Real-time rendering and model inference compete for the same accelerators and share the same profiling problems.
Reproducible measurement
Applied
Experimental applications are only informative when their runs can be repeated and compared.
03Domain
Simulation & Real-Time Systems
Real-time systems have a fixed budget: every frame must finish in a few milliseconds. That constraint makes them a demanding test bed for questions about efficient computation, and interactive environments are also where people and software agents increasingly build together.
Read the full domain brief →Open questions
- 01Which rendering techniques that rely on hardware ray tracing can be integrated into an existing forward renderer without regressing the baseline?
- 02How should an editor expose a complex scene to an automated agent so that its changes are previewed, validated, and reversible?
- 03How do you measure a renderer change honestly when the differences are within run-to-run variation?
Agents in tools
Intelligence
The engine's agent bridge applies reasoning-system ideas to a stateful, safety-sensitive environment.
GPU pipelines
Infrastructure
Rendering research depends on the same accelerator access and profiling discipline as inference work.
Developer experience
Applied
An engine is a product that other people build with, which keeps research questions tied to practical use.
04Domain
Applied Research Systems
Ideas from the other three domains are tested by putting them into working software and observing where they fail. Applied systems are instruments, not a separate product line: each one exists to answer a question raised elsewhere.
Open questions
- 01Do reasoning techniques that look promising in isolation still help on real, messy software tasks?
- 02Which failure modes appear only when research prototypes are used for sustained work?
Ambiguity and planning
Intelligence
Reasoning ideas from intelligence research are tested here against real tasks rather than benchmarks alone.
Reproducible measurement
Infrastructure
Applied experiments reuse the measurement and provenance practices developed for infrastructure.
Developer experience
Simulation
Engine tooling is evaluated by what it lets people actually build.