'Neural twins' that talk in neuralese compared to mixture-of-transformers architectures
torchcompiled · x · 2026-10-30
Reacting to the 'neural twins' announcement — models trained for each other that communicate entirely in neuralese — torchcompiled notes it's a promising way to combine multiple models, but architecturally closer to mixture-of-transformers, the common pattern in multimodal diffusion models jointly attending across modalities. The exchange offers a useful architectural framing for this new model-to-model communication paradigm.
More from Multimodal
- Pika H3 Flex Turbo LoRA cuts video sampling from 6 steps to 4 with a single drop-in file — Illustrious_Row_9971 · 2026-10-30
- ComfyUI Updater adds one-click Core and Frontend updates with rollback from the Settings dialog — Inuya5haSama · 2026-10-30
- Ideogram 4.5 image editing on fal now edits regions at original resolution without downscaling — JenniferHli · 2026-10-30
- After AI artist 'Young Soul' goes viral, 'Old Soul' emerges as a new AI musician persona — bennash · 2026-10-30
- Creating an anime scene with AI video: keeping the same location across shots is still hard — Alive_Winner_8440 · 2026-10-30
- 'Hairy Potter' AI video goes viral: hairy legs in heels strutting through Hogwarts — denisvolin · 2026-10-30