3 ms·

Or. ChatGPT will overtrain on it's own data and go to shit the way google search did

• visarga 4y ago
  Training on its own data is a tradition already. For example RLHF example pairs rated by humans are generated by the model. So even our best models trained on their own outputs + rating from human labellers. The internet is a huge rating machine, AI will distill this signal and improve even while ingesting its own text.
• andyjohnson0 4y ago
  Meta-ChatGPT's loss function optimises for ChatGPT generating training data that maximises the shittyness of Google's LLM.