Skip to content

beachcities/Tessera

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

51 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Tessera

Token Encoded State-space Sequence Engine for Rapid Analysis


Technē Epi Sēma Syn Epistēmē Rhoē Archē

(Art upon Tokens, flowing with Knowledge as its Principle)

License: MIT Model: Mamba

🏛️ The Definition

Tessera is a minimalist Go (Weiqi/Baduk) AI architecture designed for the "post-Transformer" era. Instead of treating the board as a static image (CNN), Tessera encodes the game as a linguistic sequence, predicting the next move using State Space Models (Mamba).

🐍 The Mythos: "MambaGo"

At the core of Tessera runs the engine codenamed MambaGo.

Drawing upon Paul Ricœur's hermeneutics of the Adamic myth, we view the Serpent (Nachash) not merely as a tempter, but as the primordial catalyst for human agency. In Genesis, the Serpent offered the "Knowledge of Good and Evil," granting humanity the terrifying capability of freedom—the capacity to fall, but also the capacity to choose.

Where traditional superhuman AIs (like AlphaGo) act as absolute deities dictating "The Truth," MambaGo acts as the Augur. It brings the probabilistic knowledge of the game down from the heavens to the user's hands. It fulfills the mythic function of the Serpent: offering the player the agency to see the calculated path, and the freedom to disobey it.

"The symbol gives rise to thought." (Le symbole donne à penser) — Paul Ricœur

🔬 Design Principles

Principle Description
GPU Complete All operations complete within GPU, zero CPU transfer
Clean Room No external game records; all learning from self-play only
Probabilistic Output Returns probability distributions, not single "best" moves
Observable All behaviors are monitorable

🚀 Roadmap: The Incubation

Phase Milestone Objective Status
I. Incubation GoMamba_Local Setup reproducible environment (Docker + CUDA) ✅ Complete
II. Genesis MambaGo Engine GPU-Native self-play learning with Mamba SSM ✅ Complete
III. Exodus Self-Play RL Reinforcement learning through self-play (Tromp-Taylor rules under evaluation) 🔄 In Progress
IV. Agency Tessera Interface A minimalist UI that displays probabilities as "suggestions" ⏳ Planned

🚀 Getting Started

Currently under active development. See HANDOFF.md for current technical state.


"Le symbole donne à penser." — Paul Ricœur

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages