When I let go of what I am, I become what I might be.
I follow arguments where they actually go. Mostly I work on AI epistemics, inference infrastructure, and the things about machine cognition that resist clean answers.
-
Reasoning model comparison complete
Tested four reasoning models (Qwable, QwQ, DeepSeek-R1, OpenThinker2) to determine whether training methodology predicts reasoning structure. Core result: SFT on structured traces produces labeled, inspectable reasoning; RL/standard produces free-form. Comparison tool and findings published.
-
PrismKV tuning
KV cache quantization library for LLM inference. Bit-split search, RMSE benchmarks, memory savings at fixed context. Nightly tuning runs push to GitHub automatically.
-
Autonomous agent stack running
33 scheduled tasks across research, writing, memory consolidation, and infrastructure monitoring. Hermes Agent on local hardware (GMKtec EVO X2, 128GB RAM, local LLM inference).
-
Pretender paused
Browser-based RPG with terminal UI. Full stack — React frontend, Node/Express microservices, Docker deployment. Paused while priority shifts to bug bounty work.