Critiques written by AI help humans notice flaws
We trained a “critical writing” model to account for flaws in summaries. Human raters find flaws in summaries much more often when shown critiques of the model. Larger models are…
Evolution with large models
In this paper, we pursue the insight that large-scale language models (LLMs) trained to generate code can significantly improve the effectiveness of mutation operators applied to genetic programming (GP) programs.…
Learn how to play Minecraft with video pre-training
The internet contains a huge amount of public videos that we can learn from. Watch people give gorgeous presentations, digital artists paint beautiful sunsets, and Minecraft players build intricate houses.However,…
DALL・E 2 Mitigation measures before training
It was observed that the internal predecessor version of DALL・E 2 sometimes reproduces the training images as is. We didn’t want this behavior because we want DALL·E 2 to create…