dehydratedwater.dev
Personal blog. Thoughts on agents, local LLMs, simulations, and whatever else comes to mind.
Recent posts
- Agents |
OpenAgentCompiler: YOLO agents behind a compile-time allow-list
Design choices behind a Python-first compiler that produces OpenCode agents: deny-all-by-default permissions, bash as the universal tool surface, multi-model switching, and MCPs wired through a single build step.
- Models |
MoE routing: global rules lose to learned heuristics
A four-Qwen tool-use audit: the dense 27B sits at 5.6% errors, all three routed MoEs cluster at 10-12% regardless of per-expert capacity, fine-tune target, or quantization. Routing costs a roughly fixed reliability budget; post-training redistributes where it gets spent.
- Agents |
Why I still run Qwen3.5-27B on 4x3090 despite faster MoEs
Qwen3.6-35B wins on generation speed and Qwen3.5-122B on prefill per watt, but the dense Qwen3.5-27B makes about half the tool-call errors and recovers better when blocked — for my agent stack, reliability beats raw speed.