Archive
Published essays, notes, and implementation write-ups.
The previous post rejected MTP-1/2 and projected ~2× from NVFP4. We measured +152% aggregate at c=10 and discovered that published benchmark numbers are only honest when paired with the workload that produced them.
An agent that can't come back, is an agent with no agency, a function that has reached its end.
Benchmark results for Qwen/Qwen3.6-35B-A3B-FP8 (NVIDIA DGX Spark, GB10) serving via vLLM.
Why Lecta moved from markdown to runtime MDX, and what it took to make live, AI-generated slides fail clearly.
A blunt look at why companies that call themselves cautious are often just afraid to move in the AI era.