In recent years, artificial intelligence (AI) systems like ChatGPT have made their way into the mainstream. Implemented by researchers and businesses alike, these AI systems use complex algorithms to generate human-like conversation. However, there is an underlying concern regarding the technology’s factual accuracy, as it is prone to generating false information—commonly referred to as “hallucinations.”
To mitigate instances of AI hallucinations, OpenAI announced its efforts to enhance ChatGPT’s mathematical problem-solving capabilities, utilizing two types of feedback: “outcome supervision” and “process supervision.” Outcome supervision provides feedback based on the final result, while process supervision involves input for each step in a chain of thought. After evaluating both models using math problems, OpenAI found process supervision to be more effective in improving accuracy.
This development in AI technology is crucial for the future of artificial general intelligence (AGI). OpenAI acknowledges the need for further investigation to understand the complete implications of process supervision in different domains. If the observed outcomes hold true in broader contexts, process supervision could provide a valuable combination of performance and alignment compared to outcome supervision.
Although OpenAI’s ChatGPT is not the only AI system that experiences hallucinations, recent incidents like the Mata v. Avianca Airlines case demonstrate the importance of addressing this issue. In that case, lawyer Steven A. Schwartz relied on ChatGPT as a research resource, only to discover that the information provided by the chatbot was entirely fabricated.
Similarly, Microsoft’s AI generated inaccurate figures for companies like Gap and Lululemon while examining earnings reports during a demonstration of its chatbot technology in March. These occurrences represent real-life challenges that highlight the need to improve AI systems’ factual accuracy.
As OpenAI continues its endeavor to enhance ChatGPT’s mathematical capabilities and reduce hallucinations further, they have publicly released the complete dataset of process supervision. This fosters collaboration and encourages research into enhancing the accuracy and utility of these AI systems.
In conclusion, the advancements in AI technology, such as those implemented by OpenAI, hold significant potential for improving fact-checking and generating reliable information. However, understanding the broader implications of process supervision and other methods of mitigating hallucinations is essential for the responsible and accurate deployment of AI systems in the future.
Source: Cointelegraph
