Hina Javed Malik
Despite their impressive capabilities, artificial intelligence (AI) chatbots like ChatGPT remain highly unpredictable and challenging to control; they frequently deviate from the intended path, disseminating false information or rambling statements that can only be described as absurd. This peculiarity is currently being alluded to as simulated intelligence “mental trips,” and OpenAI has, at last, reported that it’s taking care of business.
The company that makes ChatGPT has shown a new way to stop hallucinations. In a process called “process supervision,” AI models are taught to reward themselves for each correct step of reasoning when they get an answer. This is not the same as “outcome supervision,” which is the current method of distributing rewards when a correct conclusion is reached.
Because the approach follows a more human-like chain of thought, researchers believe that process supervision could lead to AI that is easier to explain. OpenAI adds that reducing hallucinations is a crucial step toward the development of AGI, or intelligence that could comprehend the world as well as humans.
Numerous mathematical examples are provided in the OpenAI blog post to demonstrate the accuracy enhancements brought about by process supervision. The company, on the other hand, says that it is “unknown” how well process supervision will work outside of math, but they will investigate its impact in other areas.
OpenAI has been unmistakably cautioning clients against indiscriminately believing ChatGPT right all along, with the simulated intelligence bot’s connection point introducing a disclaimer that peruses, “ChatGPT might deliver incorrect data about individuals, spots, or realities.”
The organization has recognized these deficiencies in its report: “Even the most cutting-edge models have the propensity to fabricate false information when faced with uncertainty. Because a single logical error can derail a much larger solution, these hallucinations are especially problematic in domains that require reasoning in multiple steps. Enhancing one’s capacity for reasoning necessitates recognizing and reducing hallucinations.
However, some experts and critics argue that OpenAI’s current measures are insufficient and that more regulation and transparency are required.
