Anthropic trains large-scale generative models and does safety research on those models.

Judge as of 2022. True if, in 2022, Anthropic was training large-scale generative (language) models and publishing or conducting safety research on those same models, shown for example by 2022 papers on large language models of its own training (such as work on HH-RLHF, red teaming, Constitutional AI, or interpretability). Mostly true if both activities are documented but the safety research used mostly smaller or other models. Mixed if only one of the two activities is documented for 2022. Mostly false if it trained only small models, or only did research on other organisations' models. False if it neither trained large generative models nor did safety research on them in 2022.

Set before the research began.

1. E1 Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned arXiv Own record

   “First, we investigate scaling behaviors for red teaming across 3 model sizes (2.7B, 13B, and 52B parameters) and 4 model types: a plain language model (LM); an LM prompted to be helpful, honest, and harmless; an LM with rejection sampling; and a model trained to be helpful and harmless using reinforcement learning from human feedback (RLHF).”

2. E2 Training a helpful and harmless assistant with RLHF Anthropic · 12 April 2022 Own record

   “We apply preference modeling and reinforcement learning from human feedback (RLHF) to finetune language models to act as helpful and harmless assistants.”

3. E3 Anthropic's core views on AI safety Anthropic · 8 March 2023 Own record

   “We trained the first version of our headline model, Claude, in the spring of 2022, and decided to prioritize using it for safety research rather than public deployments.”

E2, dated April 2022, is Anthropic's published RLHF research, in which it fine-tuned language models to act as helpful and harmless assistants. E1 is a red-teaming paper (arXiv 2209) that studies safety behaviour across language models of up to 52B parameters, including RLHF-trained models. E3 is an official Anthropic statement that it trained the first version of Claude in spring 2022 and chose to use it mainly for safety research. All three sources are Anthropic's own published research or official records, and together they show it trained large generative models in 2022 and did safety research on them.