5 ms·
I really want to know what sorts of horrible rants copilot produces that makes male, female, israel, socialism and immigration need to be filtered out. Seems fr
by marcinzm 5y ago
I really want to know what sorts of horrible rants copilot produces that makes male, female, israel, socialism and immigration need to be filtered out. Seems from the list that it's memorized some great angry political rants which is rather impressive for something trained on source code.
- yunohn 5y agoI think it’s more about preventing it from ever going on a rant, to prevent any obvious “outrage”. Especially given the number of reactions to Copilot centered around poisoning OSS repos.
- occamrazor 5y agoI assumed that copilot is actually GPT-3 with additional fine-tuning, rather than a completely new model.
- onionisafruit 5y agoFrom what I can tell from the preview, this is right. I don’t have access to gpt-3 to compare, but copilot seems to have the same functionality that gpt-3 advertises when editing a plain text file.
- moyix 5y agoYes, they mention this in the Codex paper [1]: > Since Codex is evaluated on natural language prompts, we hypothesized that it would be beneficial to fine-tune from the GPT-3 (Brown et al., 2020) model family, which already contains strong natural language representations. Surprisingly, we did not observe improvements when starting from a pre-trained language model, possibly because the finetuning dataset is so large. Nevertheless, models fine-tuned from GPT converge more quickly, so we apply this strategy for all subsequent experiments. It's not completely clear exactly what relationship the Codex models and Copilot have to one another, but given that the Copilot model is internally named "Cushman" (going by the API URL), which is the same name as the faster of OpenAI's two Codex models, they're probably trained the same way. [1] https://arxiv.org/pdf/2107.03374.pdf https://arxiv.org/pdf/2107.03374.pdf