System 1 Speedcuber: a Rubik's cube agent
A speedcubing agent built on Jev and Laya ModernBERT-421M, with real-time 3D state, temporal memory and calibrated move probabilities instead of a solver.
# System 1 Speedcuber: Autonomous Cognitive Agent Studio ⚡🎲 ### Powered by Laya ModernBERT 421M + TypeSafe Jev [](https://huggingface.co/convaiinnovations/laya) [](https://typesafe.ai) [](https://laya.studio) [](https://vitest.dev) [](https://www.typescriptlang.org) [](https://github.com/astral-sh/uv) [](LICENSE) An open-source, real-time **System 1 Cognitive Agent Harness** for 3D Rubik's cube manipulation. Powered by local **ModernBERT-large (421M)** weights via [Convai Laya](https://huggingface.co/convaiinnovations/laya) and [TypeSafe Jev](https://typesafe.ai). Runs **100% locally** on your machine with **sub-350ms neural inference**, zero cloud API dependencies, and zero hallucinated tokens. --- <p align="center"> <video src="https://github.com/user-attachments/assets/09f7ce06-9bd1-418b-920d-da0150f3d844" controls="controls" width="100%" poster="assets/thumbnail.jpg"></video> </p> <p align="center"> <em>⚡ Watch the Autonomous System 1 Agent observe the 3D cube matrix, evaluate algorithm probabilities with Laya 421M, and actuate moves in real time.</em> </p> --- ## 💡 Why This Exists: Beyond Cold Math & Trivial Routing Following the release of frontier non-autoregressive **System 1 models** like Jev and Laya in September 2026, most community use cases stopped at pedestrian text routing—ticket categorization, email triage, and spam filtering. Meanwhile, traditional Rubik's cube solvers (Kociemba, IDA*, group-theory coset solvers) brute-force optimal 20-move paths. **They solve puzzles with cold mathematics, but they do not think like humans.** Humans don't run 20-ply graph search trees in their heads. A human speedcuber uses **hierarchical stage decomposition** and **intuitive pattern recognition**: 1. Observe the current 3D spatial state and recall w