·
Possible-State Greedy Search for Game Strategy
A technical reflection on a moving-number game strategy, where a 60% greedy baseline was strengthened with explicit (N, P) state tracking.
Blog / Notes
Research notes and engineering reflections on reliable AI systems, agents, retrieval, robotics, multimodal interfaces, and applied machine learning.
·
·
DeepSeek V4 and Curved Weight Space: Full Rank Is Not the Whole Geometry
A mathematical note on why full-rank parameter matrices can still live near a much lower-dimensional curved manifold.
·
Ferrari: Federated Feature Unlearning via Feature Sensitivity
A technical reading note on the NeurIPS 2024 Ferrari paper, which studies feature-level unlearning in federated learning through sensitivity minimization.
·
Why Harness Engineering Matters for Reliable AI Agents
A short note on why reliable agents need more than prompt engineering.
·
Retrieval Is Not Grounding: What Makes RAG Reliable?
Retrieval gives a model context, but grounding requires evidence selection, answer constraints, and failure-aware evaluation.
·
Agent Evaluation Beyond Task Success
Why agent evaluation should track traces, recovery behavior, tool choices, and controllability instead of only final answers.
·
Medical AI Needs Explainable Failure, Not Only Accuracy
A reflection on why medical AI systems should report uncertainty, failure modes, and evidence rather than only predictions.
·
Multimodal Interfaces for Human-Centered Robotics
Notes on combining gesture, voice, state machines, and feedback loops for robotics systems that feel controllable.
·
Reward Design for Grounded AI Agents
Why reliable agents need reward signals for evidence use, refusal, tool discipline, and recovery rather than only final-task success.
