Archives

Speculative decoding with the roles reversed: the large model proposes a block of 16 tokens in one pass, a verifier reads it in one more. 2.5x exact (output equals greedy AR), 2.7x with a 3B verifier.

A photo wall of Peyto's first year, from April 2025 to July 2026, earliest to latest.

One SFT checkpoint, two modes: full-quality AR at τ=1.0, or 1.8 tokens per forward pass for 1.56–2.16× speedup, with an exact, free fallback to plain AR.

Same size, same memory ceiling, but the diffusion model needs ~10x the training to reach it, and eventually just can't.

A dog doesn't get reasons, only consequences. Treat it like a reinforcement-learning agent and almost every training problem collapses into one: what exactly are you rewarding, and what are you punishing?

Publication in Arxiv

Publication in Arxiv

A full-screen photo essay from a week in Japan, September 2024. 54 frames, in order. Tap to open the cinematic view.

Publication in NAACL 2025 (Main Conference)

Publication in Findings of ACL 2023