Vision models beat Wordle from screenshots alone: OCR plus reasoning lands correct guesses
maximelabonne · x · 2026-10-06
Maxime Labonne shared capability demos of vision models playing games from screenshots only.
- Wordle: the model receives only screenshots, OCR reads the board state, and reasoning produces the correct next guess "very consistently"; adding text input doesn't help — pure vision is enough.
- Quick, Draw!: works purely on image input, but guesses from a predefined word list rather than being open-ended, and is quite reliable.
He plans a follow-up demo that also accepts user inputs.
More from Fun
- Can You Still Heist the Mona Lisa After the Singularity? AI Circle Debates Post-AGI Value — tszzl · 2026-10-06
- Matt Shumer builds Hogwarts live on Spawn with AI in a multiplayer world — mattshumer_ · 2026-10-06
- Running 5 email-reading agents means 5 duplicate alerts for the same suspicious login — jaivinwylde · 2026-10-06
- Company that solves Navier–Stokes but can't ship a working chat UI — Sauers_ · 2026-10-06
- Dev finds his agent files overnight issues to public repos, where bots auto-fix and merge them — kevinkern · 2026-10-06
- AI demo still 'impresses every time' — no way a human gets this — nrehiew_ · 2026-10-06