[ad_1]
“This work takes an important step in the right direction,” says Douwe Kiela, a researcher at Hugging Face, an AI company that works on open-source language models. He suggests that the feedback-based training process can be iterated over many rounds, further improving the model. Leike says OpenAI can do this based on customer feedback.
InstructGPT still makes simple mistakes, sometimes giving irrelevant or meaningless answers. For example, if a false claim is made, it will consider the lie true. And because it’s trained to do what people want, InstructGPT will produce a much more toxic language than GPT-3 if manipulated.
Ehud Reiter, who works on text rendering artificial intelligence at the University of Aberdeen in the UK, welcomes any technique that reduces the amount of misinformation produced by language models. However, he notes that for some applications, such as AI that provides medical advice, no amount of lies is acceptable. Reiter questions whether large language models based on black-box neural networks can guarantee user security. As such, it favors a mix of neural networks and symbolic artificial intelligence, with hard-coded rules constraining what a model can and cannot say.
Whatever the approach, there is still a lot of work to be done. “We are not close to solving this problem yet,” Kiela says.
[ad_2]
Source link