OpenAI presents ChatGPT-AI model that allows users to have conversations – IT Pro – News

OpenAI has trained a new model called ChatGPT. That is an AI model that users can chat with. Among other things, users can ask follow-up questions to the ChatGPT model. OpenAI provides a free demo of the chat model.

ChatGPT is an optimized version of a GPT-3.5 model, OpenAI writes in a blog post. The company trained that model at the beginning of this year. ChatGPT can answer follow-up questions and admit errors, according to OpenAI. The model will also receive some security measures, which should lead to ‘a significant reduction in harmful and misleading answers’.

The chat model is based on InstructGPT, which is specifically trained to follow instructions and provide detailed answers. OpenAI used for that reinforcement learning based on human feedback. The model is currently still in a research phase. During that period, users can try the model for free by creating an OpenAI account.

OpenAI shares several examples of conversations with ChatGPT. In an example, a user asks the AI ​​model for help debugging code. In another example, ChatGPT explains Fermat’s little theorem and in a third example, the model writes a proposal letter for a new local resident. In a fourth example, a user asks the chat model for advice about breaking into a home, but that request is rejected.

OpenAI does write in its blog post that ChatGPT currently still has some limitations. Some of the model’s answers may “sound plausible,” according to the company, but are actually incorrect. According to the company, this is difficult to improve, because there is no concrete source of truth is when training based on reinforcement learning. At the same time, adjusting the model to be more careful can also result in correct answers being withheld.

Furthermore, ChatGPT may occasionally react to ‘malicious instructions’ or display ‘biased behaviour’. The company uses a moderation API to prevent that sort of thing, but warns that false positives and negatives can occur. OpenAI will continue to improve the model in the near future based on user feedback.