What You May Have Missed #9

Stability.ai AMA / Stable Diffusion 2 vs Midjourney v4 / More generative AI news / LLMs: From GPT-3 to Galactica / NVIDIA + Microsoft AI supercomputer / Seeing beyond the brain

Left: Sonic. Upper right: Mario. Lower right: Mijdourney’s imagination. Credit: Aleksandr (shared by Hidreley Diao)

Stability.ai Ask Me Anything session

Stability.ai held an AMA this week. They shared ideas, goals, and the company’s vision going forward. Here's a brief summary.

Stable Diffusion 2 (SD 2) is coming. It’ll be under a “more permissive” open-source license than its predecessor (MIT/Apache vs CreativeML/OpenRAIL-M). SD 2 will also have better language interpretation (more like DALL·E).

They’re also training language models (EleutherAI/CarperAI). In contrast to the general trend those won’t be super large (not 175B+ parameters), but more on SD’s range (~1B).

In Emad Mostaque's words, there's no need to go big to get impressive performance (Chinchilla scaling laws, reinforcement learning from human feedback, fine-tuning, etc.) And lower-size models are more versatile from a user point of view.

Mostaque made a couple of intriguing predictions. First, he said text-to-video models with a quality comparable to today’s SD will be here in 2-3 years. That’d be a revolution for the movie and video game industries.

Second, another world-changing-if-true prediction: in 1-2 years models “way above MJ v4 or DALLE 2 quality” will run on edge devices (e.g. smartphones). Everyone could have access to the best design tools we haven’t yet seen.

To afford all this (training and deployment costs) the company has recently increased its A100 GPU count from 4000 to 5408 (and many inference chips).