Phoenix Treats Refresh Drift as a Recoverable Runtime StatePhoenix breaks DDR5 TRR by detecting missed refreshes and repairing phase. A technical reading of the evidence, exploit, limits, and design lesson.7 min read·connectedresearchers.hashnode.devAug 9
Designing LLM Assistants for High-Stakes Operations: Evidence Before VerdictsWhat 3,090 real SOC queries teach us about permissions, state, observability, and human approval in operational AI systems.8 min read·connectedresearchers.hashnode.dev1Aug 8
Target-Region Reconstruction as a Training Contract for Robot PoliciesAn engineering reading of ReconVLA: where the auxiliary loss sits, what the ablation proves, and what a production evaluation still needs.6 min read·connectedresearchers.hashnode.devAug 7
PIPEMORPH: Treating Pipeline Parallel Training as a Runtime Control ProblemAn NSDI 2026 Outstanding Paper shows how dependency bubbles and GPU queue blocking amplify communication stragglers, then fixes each failure at a different system layer.7 min read·connectedresearchers.hashnode.devAug 6
Encode Once, Query Any Point: The Interface Behind D4RTOne global video representation, an independent five-index query, and task cost that follows requested output.8 min read·connectedresearchers.hashnode.devAug 5
