This is cool. Sometimes me and my kids will play a game of trying to create a short sentence that we feel has never been said before in history (ie. Dinner will be pretzels served on a Frisbee with electric eel sauce)

This should be a way to model those sentences. 

freetx

This project implements a least likely next token predictor using reinforcement learning. It leverages a pre-trained language model (GPT-2) to estimate token probabilities and trains a second model to predict tokens with the lowest probabilities, making it the "apex predator" of unlikely token predictors.

This project implements a least likely next token predictor using reinforcement learning. It leverages a pre-trained language model (GPT-2) to estimate token probabilities and trains a second model to predict tokens with the lowest probabilities, making it the "apex predator" of unlikely token predictors.

![](https://m.stacker.news/78875)