The Landscape on the Language Model Side

Yannick Kilture trained a version of GPT 3-ish on sort of fine-tuned on 4chan. He published that through Huggingface as GPT 4chan, which did what you would imagine it would do. That since got taken down, but there's some great folks at Huggingface who are really thinking hard about how do we do this in a beneficial way while actually mitigating risks. And so Huggingface has been developing policies so that they can be more proactive rather than reactive in the future.