LKForge/game-ai-vs-llm-benchmark

Classical Game AI vs LLM-sim Benchmark (Tic-Tac-Toe & 2048)

Measured results comparing classical game-tree search (minimax with alpha-beta, expectimax) against a stochastic "LLM-sim" opponent on Tic-Tac-Toe and 2048.

📄 Full write-up & methodology: We Asked Two AIs to Prove Our Game AI Beats an LLM. Only One Was Honest. 🌐 Publisher: LK Forge — the live Tic-Tac-Toe and 2048 engines this study reimplements.

⚠️ Read this first. The "LLM-sim" is not a real language model — it is a deliberately noisy heuristic (random moves ~12% of the time plus Gaussian noise) from a generated benchmark script. These figures show optimal search vs. a weak stochastic policy, not algorithm vs. GPT. See Caveats.

What this is

Two frontier assistants (ChatGPT and Grok) were given the same brief: build a tool comparing classical game algorithms against LLM-based agents on move-time, memory, and win-rate over 100 rounds of Tic-Tac-Toe and 2048. Grok's engine code was sound, so it was run as written and the output measured on one laptop. This dataset is that measured output.

Results

Tic-Tac-Toe — 100 rounds · self-play · minimax at full depth

Agent	Win / Draw / Loss	Avg move	Move SD	Peak mem
Minimax (classical)	0 / 100 / 0	3.450 ms	0.037 ms	2.3 KB
LLM-sim (stochastic)	69 / 2 / 29	0.012 ms	0.002 ms	0.8 KB

2048 — 8 rounds · single-agent · expectimax depth 3–5

Agent	Reached 2048	Median tile	Avg score	Avg move	Peak mem
Expectimax (classical)	6 / 8	2048	27,976	110.6 ms	66.8 KB
LLM-sim (stochastic)	0 / 8	128	1,287	0.19 ms	8.9 KB

Headline findings

• Quality (2048): expectimax averaged 27,976 vs LLM-sim's 1,287 — a ~22× gap; it reached the 2048 tile in 6 of 8 games, the guesser never did.
• Quality (Tic-Tac-Toe): minimax never loses (0/100/0). Its cost is time: 3.45 ms/move, ~300× slower than the straw man.
• Cost: in 2048, expectimax averaged 110.6 ms/move vs 0.19 ms and used ~7.5× more memory.

Files

File	Description
results.csv	Tidy long-format table (one row per game × agent). Loads directly in the dataset viewer.
results.json	Structured version with metadata, headline findings, and caveats.

results.csv schema

Column	Meaning
game	tic_tac_toe or 2048
rounds	Rounds played
play_mode	self-play or single-agent
agent	Minimax, Expectimax, or LLM-sim
agent_type	classical or stochastic
algorithm	Underlying method
wins, draws, losses	Tic-Tac-Toe outcomes (blank for 2048)
reached_2048_of_rounds	2048 games that reached the 2048 tile (blank for Tic-Tac-Toe)
median_tile, avg_score	2048 outcomes (blank for Tic-Tac-Toe)
avg_move_ms, move_sd_ms	Per-move timing (ms)
peak_mem_kb	Peak memory (KB)

Caveats

1. Reference engines, not production. Fresh implementations "inspired by" LK Forge, not the exact live game code.
2. The "LLM" is simulated. A noisy heuristic, not a real model — so the gaps are "optimal search vs. weak stochastic policy," not "algorithm vs. GPT."
3. Tic-Tac-Toe is self-play. Each agent plays both sides; the LLM-sim's 69 "wins" mean first-mover X beat O under a weak heuristic, not that it beat minimax.
4. Small 2048 sample. n = 8, because a full 100-round run is a ~4.5-hour job. Shows direction and cost, not a precise win-rate.
5. One machine, one language. Single laptop, CPython. Treat timings as ratios and orders of magnitude, not absolutes.

Reproduce

Grok's script is standard-library Python:

# full run — budget ~4.5 hours for the 2048 phase
python3 grok_benchmark.py --rounds 100

# honest quick sample — Tic-Tac-Toe 100, 2048 a handful
python3 grok_benchmark.py --rounds 8

Full method, both AIs' builds, and the honesty analysis: https://lkforge.com/blog/chatgpt-vs-grok-game-ai-benchmark/

Citation

@misc{lkforge_gameai_llm_benchmark,
  title  = {Classical Game AI vs LLM-sim Benchmark (Tic-Tac-Toe & 2048)},
  author = {LK Forge},
  year   = {2026},
  url    = {https://lkforge.com/blog/chatgpt-vs-grok-game-ai-benchmark/}
}

License: CC BY 4.0 — reuse freely with attribution to lkforge.com.