ChatGpt, the strength of the chatbot is also its biggest flaw

While it hasn’t released all the details about its text-generation software’s new interface, OpenAi has shared some information in a company blog post. The ChatGpt development team fed Gpt-3.5 human written responses as training data and then used a simulated reward and punishment technique known as reinforcement learningto encourage the model to provide better answers to the training questions.

Christopher Pottsa Stanford University professor, explains that the method used to help ChatGpt answer questions, which OpenAI had already presented in the past, seems to represent a significant breakthrough to help AI handle language in a more understandable way. “It’s truly remarkable – says Potts, although the technique could make his job more complicated -. It got me thinking about what I will need in my courses that have short answer assignments”.

According to Jacob Andreasan assistant professor of AI and language at the Massachusetts Institute of Technology (MIT), the system seems destined to expand the pool of people able to use Ai’s linguistic tools: “It is presented with a familiar interface that causes you to apply a mental framework that you are used to applying to other – human – agents you interact with”.

Recurring problems

Despite the great potential, however, ChatGpt also shows that it has i defects that haunt text generation tools.

Over the past couple of years, companies like OpenAi have demonstrated how AI algorithms trained with massive amounts of images or text can be capable of extraordinary feats. However – since they imitate man-made images and text purely statistically – rather than actually learning how the world works these programs are also prone to invent facts and reproduce hatred and prejudice. These problems are also present in ChatGpt, as evidenced by the first users who encountered how the service is capable of convincingly invent nonsense about a given topic.

While ostensibly designed to prevent users from making them say nasty things or recommend illegal or nasty actions, ChatGpt can still reproduce horrible prejudices. Users have also proved that its control systems they can be circumvented: For example, asking the program to generate a script for a movie about conquering the world is a way to circumvent its refusal to respond to a direct request. “It’s clear they’ve tried to put in some guards, but it’s pretty easy to drop them – underlines Andreas -, and this still seems like an unsolved problem“.