Seed IQ demos direct vision-to-action play in 3D Doom II

Fit_Transition8824 · reddit · 2026-07-29

Seed IQ shows a 3D Doom II agent that claims to act directly from vision, without a symbolic map or LLM replay. The post contrasts this with ARC-AGI-style 2D benchmarks and argues the system is already operating in a live 3D environment, perceiving depth, navigating obstacles, tracking threats, and adapting on the fly.

The author frames the demo as a step beyond benchmark overfitting: frontier labs can study public replays, but Seed IQ is presented as doing direct perception + kinetic action in 3D with no pretraining, no LLMs, no GPUs, and no classical ML. More demos are promised.

Original post →

More from Fun

Fun channel →