Why OpenAI hired human contractors to improve GPT-3
It is possible to get around this problem, but it doesn’t come with the same scalability and, worse, relies on something that is not technically advanced: human input. The smaller language models that are fine-tuned by actual human-written responses will generate less biased text.
Further complicating the situation, models such as OpenAI’s GPT-3 do not always produce text that is particularly useful. This is because they are trained to \”autocomplete sentences\” based on an enormous trove of text scraped off the internet. The models do not know what the user wants and how to respond. OpenAI stated that the models were not aligned to their users.
This idea could be tested by seeing what happens when you pare down the models and add a little human input in order to make those neural networks more…humane. OpenAI recently contracted 40 human consultants to guide the behavior of GPT-3.
Source:
https://www.theregister.com/AMP/2022/04/09/openai_gpt3_contractors/
Leave a Reply