Gary Marcus Roasts LLMs: Commercial Models Still Can't Play Tic-Tac-Toe Properly
GaryMarcus · x · 2026-08-04
Gary Marcus reshared a post from three years ago, pointing out that commercial large language models (LLMs) are still terrible at chess and cannot even beat serious players without relying on external tools.
The referenced chat logs show that even when explicitly asked to play tic-tac-toe using an optimal strategy, GPT-4-Turbo makes basic logical errors and loses the game. This adds to the ongoing debate about whether LLMs possess actual reasoning capabilities or are merely mimicking text patterns.
More from Fun
- User Recalls First GPT-3.5 Jailbreak: A Massive Rant Roasting Social Media — RileyRalmuto · 2026-08-05
- The 8 Virtues of Vibe Coding: Ignoring Code and Skipping Tests? — yangyi · 2026-08-05
- Testing 4 Top AI Models on 'Does God Exist?': Huge Differences — Ibz04 · 2026-08-05
- AI Video Demo: First-Person Night Forest Run Twists into an Indoor Set — fofrAI · 2026-08-05
- Felony Bench Tracks AI Company Incidents: OpenAI Leads with 5 Points — felpix_ · 2026-08-04
- XBOW Launches AI Security Challenge: Throw Your Exploits at Us — moyix · 2026-08-04