AOS: Agent Operating System
A research prototype for agent collaboration through runtime primitives: artifacts, messaging, capabilities, and scheduling.
AI, Graphics, and Doodling.
Peihong Guo is a Senior Staff Research Engineer at Meta Reality Labs, working on real-time graphics and visual computing. Those three words are also what this site keeps growing: models run and probed on machines you can point at, renderers written to find out how light actually reaches a pixel, and a crayon-doodle shelf of bedtime picture books drawn to a four-year-old's specifications.
Currently
AI — Qwen served from a machine under the desk, an Android voice assistant that runs it locally, generative AI, and agent prototypes built on the same idea. Graphics — a raytracer that finally takes refraction seriously, with real-time work at Reality Labs in between. Doodling — Wukong & Friends, a bedtime story garden with signed PDFs.
Selected work
Featured projects
Jarvis Voice Assistant
An Android voice assistant with local Qwen inference, speech models, home-control tools, and a background update pipeline.
Generative Facial Keypoints
Exploring image-conditioned diffusion and flow matching for facial landmarks with frozen DINO features and transformer coordinate models.
Wukong & Friends
Crayon-doodle bedtime picture books for young readers, with free reading and two-up booklet PDFs, published from a private repository against an Ed25519-signed manifest.
Recent writing
Blog highlights
My Qwen 3.8 Flash Next on Strata Got 40% Slower Overnight, and It Wasn't the GPU
My Qwen 3.8 Flash Next on Strata, running on an Intel Arc Pro B70, fell from 42 tok/s to 26 tok/s with no config change. The cause was a pinned-memory mirror of MoE experts being allocated while Linux reclaimed page cache — pages the GPU reads at 55–85% speed while the CPU reads the same pages at full speed. What the measurements looked like, why my first three diagnoses were all wrong, and the 45-line patch that went upstream.
The Better Model That Wouldn't Fit: Qwen 3.8 27B on a 32 GB Arc Pro B70
Qwen 3.8 27B is a clearly better model than the MoE I run today, and one 32 GB Arc Pro B70 will not hold it at a context length worth having. What 'too heavy' actually decomposes into — weights, KV cache, drafter, and cold prefill — measured from 64K to 220K, the speculative-decoding break-even curve that grows from 43 tokens to 2,361, and why the port went back to a MoE.
Introducing Wukong & Friends
Free read-aloud picture books, built to make bedtime calmer for a four-year-old — and the unglamorous machinery behind them: keeping one character recognizable from book to book, signing every PDF, and publishing a private repository's build into a public subdirectory.