So weak to strong synthetic data still biases towards strong. And strong to weak synthetic data biases towards strong. Sounds like we're on the cusp of some k

So weak to strong synthetic data still biases towards strong. And strong to weak synthetic data biases towards strong. Sounds like we're on the cusp of some kind of approach for unsupervised fine tuning, particularly with the trend towards MoE. I'd guess we're maybe only one to two generations of models away from that kind of unsupervised self-talk approach being wildly successful at advancing net model competencies.