Announcing a new engine - Incipit

[adamtwiss]

Announcing a new engine - Incipit

Incipit is a new AI-developed engine, developed in a way that is designed to try and prevent it from using other engines' code or data.

Coda helped break new ground - it was one of the first agentically developed strong engines. It started out as 'GoChess', a project whose only purpose was to help me experiment with, and get better at AI-driven development. It got stronger over time, and it evolved into what is now a top engine on CCRL.

I appreciate that AI-driven development has been a challenge for the Computer Chess community. It's disrupted a craft/hobby that many people took great pleasure from, and this has elicited some strong feedback. Some of this is about AI, but parts of Coda's development also leaned heavily on studying other engines' ideas, and this triggered further emotions.

I've tried to really hear this feedback - and appreciate the views across the community. Over the past few weeks I've been experimenting with a new engine - where I've explored what an AI-developed engine would look like under much stricter rules. Specifically, I wanted to see what could be achieved if I blocked the agents working on it from looking at any other engines' code, data or artifacts (including my own engines) and used NNUE networks trained purely on self-play data (bootstrapped from material + mobility only). This project became Incipit.

As to where the name came from - Coda was an acronym - Chess Optimised Developed Agentically - always owning the fact it was LLM-developed - but Coda is also a musical term to describe the end of a piece of music. Conversely, an Incipit is the beginning of a piece of music - to indicate it's a fresh start, and not based upon anything that came before.

Whilst my aim has been to be completely clean of any other engine's code - I can't control what data was used to train the LLMs themselves. That said - with Incipit the goal is to ensure that during development, no code or data, and no ideas are taken from studying other engines. There's more info on this in the project Readme and enforced in the Claude.md. I am happy to talk further about the approach taken (find me on Discord/GitHub).

The first public release is now ready. It's still based upon a relatively limited (14B positions) self-play training dataset which constrains its strength. My own testing now puts it at roughly 3600 CCRL 40/15 strength. My aim is to make sure I give back, so all code is open-source (GPL), and all training data is openly available and licensed under a broad Creative Commons license. Incipit is only a few weeks old so don't expect anything radically new or novel. However, it already uses a few less common choices, such as 32 king buckets (one per square with symmetry, factorised), and a linear readout from the first hidden layer straight to the output, which fixed hidden-layer saturation: wrong-way evals in decided positions fell from 10–20% to about 1%.

Coda lives on, and I'll release 0.9.5 shortly - but in recent weeks I've been putting most of my efforts into Incipit to see how far this approach can go.

Incipit is available on github with prebuilt universal Windows, Mac (ARM/Intel) and Linux (ARM/Intel) binaries avaialble at https://github.com/adamtwiss/incipit. Feedback always welcome.