The latest GPT-4 technology is used by the OpenAI company to try to understand the neural process of the previous version GPT-2 , adding an intelligent control mechanism for research.
ChatGPT is an artificial intelligence tool developed by OpenAI that has caused quite a stir in the digital world. Although many experts have talked about the capabilities of AI chatbots, there are aspects that are beyond human comprehension. For this reason, the company responsible for the development of this generative AI has launched a tool that allows a better understanding of how language models work, with the aim of increasing transparency in their use.
To better understand how chatbots work, OpenAI has released a tool based on ChatGPT-4, the latest version of its large language AI. The tool makes it possible to produce and score explanations in natural language of the behavior of the neurons of the GPT-2 model, with the aim of improving its transparency.
Large language models like ChatGPT are made up of neurons that detect specific patterns in the text the user types. From these patterns, they create a response by following a series of steps. In the case of GPT-2, an explanation on the subject is first generated, showing relevant text sequences of the model. The tool then analyzes each time a neuron fires and, with the help of GPT-4, explains and predicts the behavior of neurons.
The tool then reuses GPT-4 to check that the neuron simulated before and the real neuron from GPT-2 are similar. This way, you can compare how closely the explanation of the tool’s behavior matches the actual behavior of the AI.
With this tool, William Sanders, one of the project managers, intends to “anticipate what the problems will be with an AI system.” At a time when the European Union is regulating the use of AI, this tool could help to comply with the rules and avoid the ban on the continent.
Generative AI and the concerns it brings
The concern for artificial intelligence has been reflected in numerous films that show a dystopian future in which technology is revealed against humanity. This has led many to fear that the advent of chatbots like ChatGPT could trigger this situation. Therefore, OpenAI has developed a tool that allows a better understanding of the functioning of language models.
In recent months, chatbots have been on everyone’s lips, as they have proven to be able to answer questions and carry on conversations just like a person would. However, behind these systems there is a team of experts dedicated to training chatbots so that they improve their knowledge and do not have biases or misinformation.
Although how chatbots work has been explained, there are still details that are beyond human comprehension. In an interview, James Manyika, Google’s senior vice president of technology and society, talked about an AI system that learned a language without having been trained to do so. This phenomenon is known as ‘the black box’ because it cannot be explained how it occurs.