AGI is Not Multimodal

4 Jun 2025 · The Gradient – Substack Global

AGI framing shapes vendor claims and policy ambitions - understanding the debate helps APS practitioners scrutinise those claims critically.

Key points

Summary

This essay from The Gradient argues that current multimodal generative AI systems - despite their apparent generality - are not on a credible path to human-level artificial general intelligence. The author contends that scaling modular networks across text, image, and other modalities produces capable but shallow systems that cannot perform sensorimotor reasoning, motion planning, or social coordination. The piece advocates for embodiment-first AI research as a more principled alternative. It is a theoretical position piece rather than a report on new capabilities or policy developments.

Implications for Australian agencies

Implications are AI-generated. Starting points, not advice.