1 paper
Dale Schuurmans, Hanjun Dai, Francesco Zanini
We show that autoregressive decoding of a transformer-based language model can realize universal computation, without external intervention or modification of the model's weights.…