Researchers train language models by asking people to rate them. A separate set of human judges preferred the model summaries, even to those written by humans. The team fine tuned its gpt three models to generate summaries that would please this a i judge. Some researchers think that language models might never chieve human level common sense as long as they remain solely in the realm of language.
In 2020, the artificial intelligence (AI) GPT-3 wowed the world with its ability to write fluent streams of text. Trained on billions of words from books, articles and websites, GPT-3 was the latest in a series of ‘large language model’ AIs that are used by companies around the world to improve search results, answer questions, or propose computer code.
However, these large language model are not without their issues. Their training is based on the statistical relationships between the words and phrases, which can lead to them generating toxic or dangerous outputs.
Preventing responses like these is a huge challenge for researchers, who are attempting to do so by addressing biases in training data, or by instilling these AIs with common-sense and moral judgement.
This is an audio version of our feature: Robo-writers: the rise and risks of language-generating AI
Hosted on Acast. See acast.com/privacy for more information.