Since the end of November, everyone has been talking about ChatGPT (an acronym for Generative Pretrained Transformer). According to Greg Brockman, OpenAI's president, five days from its launch, more than a million people signed up to test it.
Many users have shared screenshots of their experiments, sometimes marveling at the tool's ability to generate answers so similar to human language, at others ironizing or criticizing for the extravagance or the obvious inaccuracies. You can ask the chatbot anything: to create short essays on any topic, create literary parodies, answer complex programming prompts, and much more.
However, we should keep in mind that it is an artificial intelligence tool, besides the fact it is still in a developmental stage. We wanted to test it ourselves with a few questions, but before presenting you some examples and our conclusions, here is a brief overview of what ChatGPT is and how it works.
ChatGPT is an AI prototype developed by OpenAI, the San Francisco AI company, that can understand human language and hold very complex conversations. This chat is based on the GPT-3 artificial intelligence language and is designed to serve in a variety of natural language processing tasks, such as text generation, translation, or code creation. It was developed to store a huge amount of data: 570 GB of text and a model of more than 175 million parameters.
As a result, ChatGPT is able to generate answers and texts of various paragraphs, switch the style (either simple or more sophisticated) and understand the context to provide the most suitable answers.
We are talking about a conversational model, which can answer questions and provide information. It is based on samples of texts taken from the Internet (books, newspaper articles, and web pages): the width of the samples with which AI is trained usually determines the accuracy of the result.
ChatGPT's sentences seem natural, have a construction and syntax indistinguishable from that of humans, and can respond in a highly accurate and context-relevant way.
The team working on the development of ChatGPT provided Gpt-3.5 with human-processed written responses as training data. Moreover, they employed a simulated reward and punishment technique, known as reinforcement learning, to prompt the model to provide better answers to the training questions.
When you log into the chatbot page to use the tool, in addition to giving some examples of what you can ask and its capabilities, it also informs users of its limitations.
In fact, the model is able to admit its errors, correct inappropriate premises, and state when it cannot answer a question. It can very accurately simulate human language, but obviously, since it is artificial intelligence, it cannot have emotions, feelings or personal opinions.
For example, we asked the chatbot if it was happy today or what did it think about becoming the new trend. And these were its answers:
You can ask it to create fictional and creative stories about a determined topic, whatever pops into your mind.
But the chatbot is also able to combine completely different styles, for example, we asked it to explain what SEO is with Doctor Seuss' narrative style.
Or write a poem about Location Intelligence in the same style.
It is capable of creating any type of poem or song. We tested it with some bizarre requests, such as explaining in rap rhythm the first law of thermodynamics or writing a pop song about Switzerland. And it gave us quite a good outcome.
Besides a more creative use, the chatbot can give you any business-related answer. The more detailed the question, the more accurate the answer will be. You can ask it for ideas for a new business, or how to open an e-commerce, which products to sell, or how to increase the traffic on your web-site.
However, we also mentioned earlier that ChatGPT presents some limitations that we cannot ignore.
Firstly, the chatbot sometimes gives incorrect answers, so before using or publishing any content is better to verify it. For example, we asked which one is the fastest mammal in the world. The answer was partially correct: it says that it is the cheetah, but it also adds that the second fastest mammal is the peregrine falcon. Thus, the peregrine falcon is not a mammal, but a bird.
Moreover, the chatbot uses data updated until 2021: after that date, this tool knows nothing about the world. Even if it says something after 2021, it is often incorrect. For example, when we asked what the current monarch of the United Kingdom is, this was the result.
We were impressed by its capabilities. Certainly, Chat GPT has proven to be suitable for many relevant use cases and has caused a lot of enthusiasm in the community. However, as we previously stated, this tool still has many limitations, such as a fixed knowledge base that is not updated beyond 2021 or the fact that it cannot give answers in real-time or browse the Internet, which clearly distinguishes it from search engines such as Google or Alexa. Also, precisely because the chatbot uses a very human-like language and is able to create well-structured and articulated sentences, it can also give completely wrong answers and present incorrect information as facts that sound authentic, original and totally plausible.
This might lead to problems regarding how to distinguish a real human text from one created by AI, with the risk, for example, of spreading fake news people will trust.
All in all, ChatGPT has huge potential, so for the moment, we remain confident that it will be upgraded and improved to provide more accurate and, especially, more reliable outcomes.