Agents & Autonomous Systems
Reasoning, planning, memory and real-world agency under explicit safety constraints.
Work at the Frontier
Original inquiry across intelligent systems, mathematical foundations and human understanding.
Foundations of Intelligence
Mathematics, computation and the nature of minds — what learning systems can and cannot do.
Human–AI Collaboration
Augmentation, alignment and the future of work between people and machines.
Current major project
The Reflective Agent
A persistent multimodal agent designed to observe, reason, remember and act — evaluated against simpler baselines, with memory provenance, inspection and deletion built in from the first version.
Active research
Research Laboratories
Frontier Mirror Lab
A privacy-aware, multimodal agent embodied in a physical mirror — from V0 voice prototype to V4 commercial prototype.
Language Model Lab
Tokenizers, Transformers, training loops, kernels and post-training — built from scratch, measured honestly.
Agent Systems Lab
Tool use, retrieval, memory and evaluation harnesses, with red-team reports for every system that touches real accounts.
AGI Research Lab
Falsifiable hypotheses about general intelligence, preregistered protocols and defensible negative results.
How Research Progresses
- 01Question
- 02Literature map
- 03Falsifiable hypothesis
- 04Preregistered protocol
- 05Experiment
- 06Discrepancy analysis
- 07Reproduction & review
- 08Paper or defensible negative result
Reproducible
Every experiment records hypothesis, protocol, code and data version, configuration, seed, result, analysis, limitations and next action.
Reviewed
Results are reproduced and externally reviewed before they become claims. Negative results are kept and defended.
Logged AI use
AI assistance is allowed in research but logged; every claim and derivation must remain independently defensible.