Sooool repo explores how Sol plays games despite delayed model responses
The developer says Sol keeps watching new frames and choosing actions with up to eight model requests running in parallel.
TLDR
The Sooool developer says Sol chooses game actions from images rather than using code as a policy, with a research loop to handle decisions that arrive after the environment changes. The system runs up to eight model requests in parallel and simulates API delays with cheaper inference during development. The developer describes Sol playing Touhou, racing from camera frames and choosing movements in Asterix from screenshots; Sol chooses which lessons enter later prompts.
Combined views
—
7 Sources, first seen ago
