RL-trained 4B VLM aces GeoGuessr, with code, env, dataset and evals to be open-sourced
ariG23498 · x · 2026-09-02
Developer adithyask previewed a project showing that reinforcement learning on a 4B-parameter VLM is enough to ace the geography guessing game GeoGuessr.
The team promises to open-source everything needed to reproduce it end-to-end: code, RL environment, dataset, training setup, and evals — "dropping soon." ariG23498 quipped that world-champion geolocator georainbolt's "time is up."
More from Fun
- After a loss, I spent my newborn daughter's first week expecting the worst — justalexoki · 2026-09-02
- John Wickachu: AI mashup video turns John Wick into a Pokemon trainer — Brave-Wishbone-3650 · 2026-09-02
- Will GPT-6 Ship Before GTA-6? AI's Speed Leaves Games Behind — hyhieu226 · 2026-09-02
- A packed mobile staircase is a 'surprisingly literal alignment problem' — misovalko · 2026-09-02
- Dev builds a terminal 'black hole' that warps your code until you take a break — petewoodbridge · 2026-09-02
- AI circles trade favorite definitions of slop: material divided by meaning — jordiponsdotme · 2026-09-02