With the large language model GPT-3, OpenAI has set records in natural language machine processing. What's next?
Wojciech Zaremba is a co-founder of OpenAI, where he led AI-based robotics research for several years before OpenAI abandoned it in October 2022. According to Zaremba, this was because the path of training large AI models, so-called foundational models, with human feedback was more promising for OpenAI on its way to AGI than robotics research.
GPT-4 to GPT-X
Now Zaremba is giving a preview of how large language models like the upcoming GPT-4 could change the way we communicate. The OpenAI co-founder believes AI will have a similar or even greater impact on human communication than the Internet.
For example, future AI models could read a book and then have a conversation about its content and answer questions - customized to the interests and knowledge of individuals. This principle could also be applied to emails, news, and other forms of information, Zaremba writes.
AI as a life coach
Zaremba also describes an advanced AI thinking and writing aid that helps with thought sorting, detects errors in argumentation, and helps make messages "understandable and kinder".
He also finds that GPT is "quite empathetic" when properly prompted, so future conversational AI models could help with self-discovery. "It's clear that AI could become a fantastic life coach, meditation teacher, or therapist," Zaremba writes.
That AI would dissolve language barriers is so obvious that there's really no need to mention it, according to Zaremba.
First indications of GPT-4
Zaremba doesn't make his predictions in a vacuum: Last September, for example, OpenAI introduced an AI model that can iteratively summarize large books thanks to human feedback. OpenAI has been experimenting with this feature since GPT-2.
And the case of Google's conversational AI Lamda and Blake Lemoine shows that conversational AI could develop some persuasive power and thus be useful as a conversational partner. Translation AIs from many providers has long since arrived in everyday life.
It's possible that Zaremba's GPT wish list is also a preview of new features in GPT-4. OpenAI CEO Sam Altman already confirmed last fall that GPT-4 is under development.
GPT-4 will again be a text-only model, meaning it will not be multimodally trained like DALL-E 2, and it will be similar in scope to GPT-3, but will be trained with high-quality data and provide better algorithms and more accurate fine-tuning, Altman said.