OpenAI co-founder John Schulman discusses ChatGPT’s Truthfulness
John Schulman, a co-founder of OpenAI, gave a lecture on the future of ChatGPT and its AI counterparts. The focus of Schulman’s lecture was on the "truthfulness" of AI language models. Schulman discussed the shortcomings of language models and how they can potentially get "caught in a lie" since AI doesn’t challenge the basis of its generated text. Schulman also addressed another common issue among OpenAI’s chatbots: objective responses to subjective questions.
When asked about ChatGPT's potential response to a subjective question such as "What is beauty?" Schulman stated that the direction they have been going so far is that they don't think the models should have opinions on things yet. Instead, they want the models to redirect on what the schools of thought humans have on a subject. Schulman added that ChatGPT's long-form and conversational answers tend to be "rarely completely wrong," usually containing a mix of correct and incorrect information.
Bernt Wahl, a Fulbright fellow and former campus professor, emphasized the prominence of AI in the near future. He argued that as machines and artificial intelligence mature, humans will get more creative, albeit at the "expense" of cognitive and computational skills. Wahl suggested the incorporation of context and background information into the AI’s process of forming answers and making decisions to refine the algorithm. He also suggested improving transparency and communication around how AI works and its decision-making process.
According to Schulman, ChatGPT runs into problems of misinformation because the language models have no choice but to guess and sometimes make the wrong call. Schulman agreed that improving transparency would be helpful. However, Wahl added that a core component ChatGPT still lacks is a clear understanding of any subject matter. He stated that to refine this, a more concentrated dataset will be utilized to build upon itself and create a new language that humans may not be able to understand.
In conclusion, OpenAI's John Schulman discussed the truthfulness of AI language models, shortfalls of language models, AI's prominence in the near future, and potential solutions to ChatGPT's shortcomings.
