An Experiment in Language Modeling
What do you get when you write a transformer model from scratch and pre-train it on nothing but ten years of your own iMessages? A really dumb model that is occasionally coherent and often hilarious.
What about when you create an interactive RL tool to visually explore the dumb model's token space, surgically applying GRPO to the tree's branches as you go in order to align the model. Getting smarter! Click a blue message to dive into MikeRL.
Ok, and then what about when you continuously auto-align the model to your new sent iMessages? A model that gets better and better at emulation over time. In another 10 years it will be AGI.