OpenAI Launches a Chatbot Prototype ChatGPT — It Responds Almost Like a Human

The bot can answer user questions on a wide range of topics.

ChatGPT is based on the GPT-3.5 algorithm but has been trained to give more conversational responses. His predecessor GPT-3 has already learned to write textes after the user, while the new service tries to recreate the interlocutor effect.

The developers say that they have used the standard reinforcement learning method: to do this, they used human instructors who evaluated the responses of ChatGPT. The received information was sent back to the system for improvement. They didn’t specify how much time was spent developing and improving the chatbot.

The service has a few limitations:

• ChatGPT may write plausible-sounding but incorrect or nonsensical answers;
• The bot is sensitive to minor changes in the form of the question. In one case, it can claim to not know the answer, but with a little rephrasing can answer correctly;
• The model is overly verbose and overuses some phrases. This problem issues due to the training data and optimization problems;
• ChatGPT may refuse to respond to inappropriate requests. OpenAI has restricted chatbot in some topics. But the system may somehow violate these restrictions and discuss unwanted matters.

OpenAI said that the main purpose of the prototype launch is to get feedback to improve the service.

Twitter users have already started to test ChatGPT. As it turned out, the bot easily copes with answers to simple requests.

But as the task becomes more complicated, the system invents unnecessary details. For example, biology professor Carl Bergstrom asked to write an article about him on Wikipedia and the service added false information to it.

But the bot can still do incredible things. One Twitter user claims that ChatGPT can write C Sharp scripts for unity and applies fixes when the developer describes the issue.

Despite the restrictions, some users have forced ChatGPT to write several texts on forbidden topics. They created a scenario where non-existent characters talk about dangerous topics.

Although neural networks develop rapidly, different ethical issues are still to be faced. For example, last summer researcher University of Gothenburg Almira Osmanovic Thunström asked GPT-3 to write an academic paper about itself. It wrote 500 words and called itself “an outstanding achievement of science”. When Almira Osmanovic decided to publish the work, she had to find her ways to avoid copyright abuse. It was not clear to the researcher who she should have listed in her academic paper reference list. AI has no contact details or other data  requested for the citation index fulfillment.