Omni-Model Absolutism
From the AI Research Compass wiki
Omni-Model Absolutism: The future belongs to a single natively omni-modal model that sees, hears and speaks in real time; everything else is transitional.
Beliefs
Omni-Model Absolutism is Any-to-Any Imperialism after the empire has been declared. Its creed is that there should be exactly one model in the product: it listens to raw audio, watches the camera, reads the screen and answers in voice, text or image, all with conversational latency. Cascades of speech recognizer, language model and speech synthesizer are seen as a historical embarrassment, like a phone call routed through three interpreters.
It sits at the very top of the chart, just right of center, as far up as a position can go. The Monolith stance is absolute, and the near-central horizontal position reflects that it treats voice, vision and text as equal citizens. Its paradigm level is Orthodox: one natively omni-modal model is now what the frontier ships, even if much of its promise is still being worked out.
Arguments
- End-to-end audio modeling preserves tone, emotion, interruptions and background sound that a speech-to-text cascade throws away.
- Single-model latency can approach human conversational turn-taking, which pipelines struggle to match.
- Shared representations let skills learned in text, such as reasoning and knowledge, carry over directly into speech and vision.
Critiques
- Omni training is extraordinarily expensive, which limits the doctrine to a handful of well-resourced labs.
- Safety and evaluation are harder when every modality can be both input and output in real time.
- Pragmatists note that cascades remain easier to debug, customize per language and swap out component by component.
Representative works
Relationships
Friends
How to get it
Land on this cell in the full or pro test: x ≈ +0.57, y ≈ +1.00, paradigm ≈ −0.71.