We are a small, independent research group working on fundamental AI research. The initial focus is on ARC, test-time compute, active inference, and system2 reasoning.

Research priorities

• Main priority: Beat ARC
• Secondary priorities:
• Research test-time-compute, active inference, system2 reasoning
• Learn to train and fine-tune LLMs, and develop the infrastructure to do so

Strategy

• Funding is secured for at least the next 3 years.
• Within that timeframe, try to beat the ARC benchmark (>85% on the private test set) to get broader recognition.
• This should open new financing and recruiting opportunities. (hopefully elevating the lab from GPU-poor to GPU-middle-class to GPU-rich)
• Move on to more challenging benchmarks

Values

Intellectual curiosity, high intrinsic motivation (obsession), perseverance. Fast experimentation process

We are driven primarily by the desire to reverse engineer intelligence and build machines according to those principles.

Regarding AI safety: we believe that after reaching a certain level of capability, AI safety will become a critical issue and is likely to constitute an existential threat. But first we want to carry out the capability research and re-assess as we go. We will have a process where AI safety can be made the top priority.