Two years of post-training vanilla transformers to generate audio, without a diffusion head
kastnerkyle · x · 2026-11-11
yaboilyrical published a long-form article, "I have no diffusion head, and I must scream," distilling 2 years of hands-on work on post-training multimodal language models. The piece covers how post-training a vanilla transformer to generate audio actually works — without a diffusion head — with first-hand engineering lessons and pitfalls, making it a rare practical deep-dive for anyone working on multimodal or audio generation.
More from Multimodal
- The 'Boeing Bench': AI models race to build a Boeing 747, Fable 5.2 beats Gemini 4 Carbon — victormustar · 2026-11-11
- Same prompt face-off: FABLE 5.2 vs Gemini 4 Argon draw an illuminated-manuscript magpie — teortaxesTex · 2026-11-11
- Can you tell which is real? AI-edited conference photo goes viral for its realism — threepointone · 2026-11-11
- Turning household sounds into instruments with Gemini 4 (FFT + onset detection) — alexanderchen · 2026-11-11
- Midjourney dual-sref combo yields whimsical storybook fantasy art — Kyrannio · 2026-11-11
- Blind judge scores Fable 5.2 at 8.4 vs Opus 5.5's 6.4 in same-prompt video showdown — nateliason · 2026-11-11