Research
My work studies explicit structural models for complex AI systems, with separate questions of what is legal, what can be executed, what can be predicted and what has been observed. Three Models is a methodological perspective; SSG and TSG are two parallel research lines, not two mandatory layers of one compiler.
Methodological Perspective: Three Models
| Model | Core question | Validation |
|---|---|---|
| Formal | What structure and transformations are admissible? | Explicit definitions, invariants and counterexamples |
| Simulation | What dynamics follow under stated assumptions? | Predictions compared with held-out observations |
| Statistical | What can be inferred or learned from experience? | Generalization, regret, estimation and empirical tests |
Physical measurements and observed task histories remain external to these models. Read the Three Models research note; there is no standalone paper PDF yet.
Research Line A: SSG — AI Computing Execution
Research object: heterogeneous mapping and execution of AI workloads.
- Representation: Workload, System and Schedule; OpHierarchy and DomainHierarchy; Compute–Memory–Communication (CMC) and selected Compute–Remap–Sync (CRS).
- Mechanism: Build independent workload and system descriptions; Traverse legal choices domain by domain; Translate selected structures using target-supported backends.
- Scope: operators and composite model graphs; training, inference, RL; dynamically revealed Agent, tool and model calls.
- Evidence: a 2026 technical manuscript with bounded H20 and A800 results, including limitations and numerical failures.
Agent-level execution is part of SSG, not a separate Agent-SSG project. A previous Agent-SSG working draft remains a private historical record, not a current paper. SSG-DT explores performance prediction over selected SSG execution structures; validated twin results remain pending.
SSG details and measurements · SSG-DT research
Research Line B: TSG — Evolving Human–AI Tasks
Research object: a persistent task across human/model activities and revisions.
- Formal structure: Actor × Spec × Impl with contextual Owner, Location (Intent/Discipline) and Consume/Digest/Create operations.
- Exchange roles: Component, Model and Artifact are relative to activities, not exclusive file types.
- Dynamics: instantiation and intra-/inter-task evolution, with open questions about identity, evidence and admissible reuse.
- Evidence: 2026 preprint with formal definitions and illustrative trajectories; empirical comparison remains open.
Evidence and Research Maturity
| Work | Public evidence candidate | Status |
|---|---|---|
| Three Models | Research note | Methodological account; no standalone paper |
| SSG | Technical manuscript | Measured H20/A800 cases; not peer-reviewed |
| SSG Agent workload | Integrated SSG research | Execution scope; new end-to-end validation needed |
| SSG-DT | Modeling proposal | No measured digital-twin results yet |
| TSG | Preprint | Formal research; empirical generalization open |
| Applied computer vision | Journal DOI | Peer-reviewed, 2024 |