| CARVIEW |
William Chan (陳俊樂)
I am Cofounder and CTO of Ideogram. I work on research, engineering, and product.
I was previously a student at George Brown College studying the Culinary and Baking Arts (never finished due to COVID). I also graduated from Carnegie Mellon University (PhD) and University of Waterloo (BASc) both in Computer Engineering. I was also previously a Staff Research Scientist at Google Brain Toronto where I had the fortune to have Geoff Hinton as my manager, and before that I was an intern at AMD, NVIDIA, Intel, Google, Amazon and TD Securities.
My research contributions span autoregressive, and non-autoregressive (diffusion, discrete) models across a variety of modalities, including: speech-to-text (SpeechStew, SpecAugment), image-to-image (SR3, Palette), text-to-text (Insertion Transformer, Imputer), text-to-speech (WaveGrad), text-to-image (Imagen, Imagen Editor), text-to-video (Imagen Video), text-to-3D (3DiM), and noise-to-music (noise2music).
I love to code. In grad school, I worked mostly in C++/CUDA/Python (no PyTorch/TF/JAX!) and built my own distributed GPU training infra. These days, I love JAX (❤️ shard_map), low-level optimizations like staring at profiler traces and explicitly controlling DMA transfers (CUDA/Pallas kernels), large scale data processing (Beam, Kubernetes, data understanding, filtering and deduplication at scale), product backend infra (Bigtable, PubSub, Redis, ❤️ Spanner), and frontend code like HTML/React/TypeScript (data/ML infra, Ideogram Canvas).
While I have no formal training, I am deeply passionate about computer architecture and semiconductors, especially as it relates to machine learning accelerators. I am an amateur home renovator - I gutted and rebuilt my own house (1886 Victorian townhouse), including underpinning and structural wall changes with my neighours attached. I am an avid public equity investor, with >10 year track record averaging ∼20% IRR (most of my returns come from $MSFT, $TSM). In my spare time, I like to squat and eat croissants.