Inside The Lab
Current Experiments
We are actively exploring the limits of personality replication and efficiency. Our biggest live experiment is the transition to the RWKV architecture for WillMe GPT 3.1.
By training a lightweight 163m parameter model on roughly 10 billion tokens of English and Swedish text, followed by heavily tuning it on exactly 53,000 real conversations with William, we are testing how well an O(1) inference model can maintain a natural human rhythm, slang, and specific persona lock-in over long contexts without the heavy compute costs of traditional transformers.
