Research and development
CoburnicusLabs
The R&D lab funded by Think New. It builds AI from the ground up, in public, so we understand these systems from the first line of code.
Why a lab
We learn by building the whole thing
Most companies that sell AI services have never trained a model themselves. We wanted to know how these systems work all the way down, so Think New funds a lab that builds them from scratch.
The lab works in the open. Code, training runs and results are shared as they happen, and the work is explained in plain language for anyone who wants to follow along.
For clients, this matters most in private AI. Knowing how a small model is built, trained and tested is what lets us size one to your task and run it on your own hardware.
Project
AdamAI
An open, free language model grown from an empty file, in public. Every stage of its training is recorded, with snapshots and tests along the way, so we can see what it learns and when.
- Genesis
- 114M parameters. The first full model, trained from scratch.
- Exodus
- 340M parameters. The next generation, training now.
Project
RobotoCAD
A browser studio for designing real, buildable robots. Formerly CoburnicusCAD.
RobotoCAD lets you design humanoids, quadrupeds, crawlers and robot arms in the browser. Every part is real, printable geometry, so the robot you design is the robot you can build.
Designs can be assembled and simulated on screen, then ordered as a kit that ships with AdamAI and ROS 2 built in.
From lab to client work
How the research shows up in what we deliver
Right-sized models
Training models of different sizes ourselves tells us how small a model can be and still do a specific job well, which keeps your hardware and running costs down.
Honest evaluation
The lab tests its models on a schedule throughout training. We bring the same discipline to client models, measured on your own cases before launch.
Looking inside
AdamAI is built so its inner workings can be studied. That focus on understanding what a model has learned carries into how we test and explain client models.
Want a model like this for your business?
We build private, specialized models that run on your infrastructure. A 30-minute call will tell you whether one fits your needs.