Games: Reinforcement learning milestones.

Mastering perfect and imperfect information systems.

The foundation of modern game intelligence was built on reinforcement learning, progressing from brute-force chess calculations to deep neural networks playing complex, real-time strategy games from self-play.

• ▸Deep Blue (1997): Defeated world champion Garry Kasparov in Chess by calculating 200M positions/sec.
• ▸AlphaGo (2016): Defeated 18-time world champion Lee Sedol 4-1 in Go by combining deep learning with Monte Carlo Tree Search^[41].
• ▸AlphaZero (2017): Mastered Chess, Shogi, and Go entirely from scratch via self-play RL, without using human games^[68].
• ▸AlphaStar (2019): Reached Grandmaster level (top 0.15% of active players) in StarCraft II, mastering imperfect information and real-time planning^[42].
• ▸AlphaDev (2022): Discovered faster sorting algorithms in assembly code, integrated directly into the LLVM libc++ library^[69].

Long-horizon planning in open-ended sandboxes.

Today's agents are moving beyond board games to master open-ended construction, resource gathering, and economic trade. These sandboxes serve as pre-deployment testbeds for physical robotics and agentic swarms.

• ▸Factorio Learning Environment: Evaluates long-horizon planning in open-ended logistics loops. Claude 3.7 Sonnet achieved a score of 29.1 in lab-play mode, establishing a new baseline^[19].
• ▸Minecraft Economies: Project Sid (2024) ran simulations of 1,000 autonomous LLM agents in Minecraft, where they self-organized labor, formed a trade economy, and established social structures^[44].
• ▸GTA V Mods (PedGPT): Deploying on-the-fly conversational LLM agents (like Llama 3.1) into GTA V, giving characters unscripted autonomous behavior and memory.
• ▸Oasis Neural Engine (late 2024): The first streamed generative world model by Decart AI, generating playable Minecraft gameplay frame-by-frame on-the-fly from keyboard and mouse inputs, bypassing traditional physics engines^[?].

Generative physics replacing traditional game engines.

Rather than hand-building physics grids, developers are utilizing generative world models to simulate continuous, interactive visual environments instantly.

• ▸Google Genie 3 (July 2025): Generates playable, interactive 3D worlds at 720p / 24 FPS with a 60-second temporal memory from a single text prompt^[43].
• ▸World Labs Marble (November 2025): Turns text, photos, and videos into persistent, explorable 3D environments.
• ▸Decart DOS 2.0: A streamed, real-time neural game engine that calculates pixels and game states continuously on-the-fly.
• ▸NVIDIA Cosmos & Causal Simulators (2026): Shifting toward causal, physics-aware world models (like Yann LeCun's AMI Labs) that explicitly model gravity, object persistence, and collisions to resolve physical hallucinations in robotics simulators^[?].