BlueCat — notes

Systems research sits between application above and hardware below — the layer that lets software actually reach silicon. The good ideas inside the field cross-pollinate constantly. Symbolic execution — born at the intersection of security and compiler analysis — turns out to drive optimal scheduling design. CHERI capability hardware, built for memory safety, turns out to drive a new kind of read barrier for a moving garbage collector. Verifying whether a cross-domain idea like that is real used to cost a PhD-year — sometimes two — of literature, positioning, and prototyping before you'd know.

A current top-tier model collapses most of that front-loaded cost. The bottleneck moves to the human who can frame a sharp enough cross-field question. The thing I'm honestly curious about — and the thread running through what's posted here — is how short this innovation cycle can actually become.

Writing

2026-06-14The back half, too

Three weeks after saying a model could compress research's front half but not build the thing, the thing got built: StoplessGC, a CHERI-native moving garbage collector for the JVM, and the paper. I wrote none of it — so what was left that was mine, and why it's still the work.

2026-05-19The trigger is the bottleneck

A 60-minute cross-domain research session with Opus 4.7 on CHERI capabilities × Java garbage collection. What the model did well, where it pushed back, what it couldn't do.

Currently

Unified runtime for heterogeneous silicon — collapsing CUDA / Ascend / ROCm into one IR-level scheduler, with RDMA/CXL memory pooling.
CHERI capability hardware × moving GC — load-barrier semantics on immutable capabilities. Built and measured — what was the model's, and what was mine.
NPU compiler toolchains — what one person + Opus-class models can carry from spec to working backend.
High-performance network optimization — DPDK / RDMA-era ideas reapplied to AI training cluster IO paths.
Cycle-time compression — with a current top-tier model, how short can a cross-domain systems experiment go end-to-end? The honest open question driving everything else here.