Unverified50% confidenceBenchmarkExact time
EfficientZero (2021) combines the MuZero search framework with self-supervised learning, achieving human median-level performance on the Atari 100k benchmark for the first time using only data equivalent to 2 hours of human gameplay.
1
Sources
50%
Confidence
Long-term
Relevance
8/15/2026
First Seen
Sources
蒙特祖玛复仇强化学习现状：从探索难题到通用智能
Related Claims
UnverifiedThe Atari 100k benchmark was proposed by Kaiser et al. in 2020, allowing agents only 100,000 steps (approximately 400K frames, equivalent to 2 hours of human gameplay) of interaction.77% similarUnverifiedThe author estimates that a small game like Gomoku with complex rules and special effects can be completed into a publishable product through 2–3 days or even a week of iterative AI collaboration, even with zero coding background.62% similarUnverifiedA user attempted to use Coze to create an AI comic series, and the system indicated it would consume more than 10,000 credits, far exceeding the daily free quota.62% similarUnverifiedAI tools have made it possible to batch-complete over a dozen account analyses in one hour, compared to the two analyses that used to take a full day.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/750639API
curl https://kongchang.com/api/v1/knowledge/claims/750639MCP
get_claim(id=750639)