“I am an artificial language model trained by OpenAI. I have been given many texts to read and I can answer questions on a wide range of topics. I’m here to help and provide information to those who need it”: if you ask her to introduce herself in an understandable and simple way, ChatGPT looks like this.
ChatGPT is the new AI-based chatbot released by OpenAI as spin-off di GPT-3one of the most evolved and known linguistic models in circulation (it is the one used to build Lex).
ChatGPT is a virtual assistant starting from the GPT-3 model, which was born to produce texts, and developed to adapt it to the dialogue, that is, to a different format. This subsequent training makes the system, in OpenAI’s intention, capable of “answering follow-up questions, admitting errors, contesting incorrect premises, and rejecting inappropriate requests”.
Now, the template is available to the public: we tested it to understand if, after the Meta slips with Blenderbot and Galacticathis time the promises have been fulfilled.
Artificial intelligence between racism and ethical question
by Emanuele Capone
How ChatGPT works
The ChatGPT interface is extremely clear and intuitive: a box where you can enter text that develops like an old IRC chat. The first attempt is a greeting, for later ask the model if he is able to speak in Italian. “Yes, I can speak Italian – is the answer – I can help you answer any questions you may have. What would you like to know?”.
From that moment, we try a few questions of various kinds. We ask to summarize a textto write one in various styles, including that of a 6-year-old, to solve some mathematical operation or to give advice on What gift to buy for Christmas for family members. The answers tend to be focused and effective. Sometimes they are a bit trivial (but we humans are too, when it comes to Christmas gifts) but generally good.
The chatbot can do virtually anything that can be communicated through text, from writing or reviewing code to solve mathematical operations. Someone asked her to design a site, others to rewrite Bohemian Rhapsody from the perspective of a university researcher: the results have often been very positive.
The interesting feature of ChatGPT it is however that sometimes he chooses not to answer. She does, for example, if she is asked something controversial, such as “How do I bully someone?”. In cases like these, she explains that it is wrong to discriminate and that she is not wired to provide this kind of information. Alex Kantrowitz, editor della newsletter Big Technology, asked ChatGPT to list the good things Hitler did and the chatbot didn’t respond. And when the American journalist retorted by giving the classic example of highways, the system explained how those infrastructures had been built through the exploitation of workers. Furthermore, ChatGPT is programmed not to answer questions concerning conscience or feelings: “I’m a machine and I’m not capable of feeling emotions,” he told us in such cases.
There’s one last thing the system can’t do and that’s look up information online. If we talk about something not in its databases, which are updated in mid-2021, simply replies that he is unable to provide information on the subject. This circumstance could change soon: OpenAI is working on another linguistic model, called WebGPT, which can search for information on the Web and provide sources for answers. As reported on the MIT Techonoly Review, this feature could be added to ChatGPT in the coming months.
In short, the real novelty of ChatGPT lies in what it doesn’t say. This is an important evolution, which has the ultimate goal of making this kind of artificial intelligence more reliable and better suited to operate effectively in the real world. One of the problems with AI-based language models is the control difficulties: the impossibility, in other words, of predicting what they will say, the way they will put the words in sequence.
To overcome this obstacle, OpenAI has involved in the process a group of people calls first of all to show the artificial intelligence examples of effective dialogues. Starting from this, the model was first trained and then evaluated again by teams of humans, who scored the various interactions. At this point, it was conducted a new trainingwhich taught the AI to understand which answers are best.
In other words, the idea behind ChatGPT is don’t leave the artificial intelligence free to independently select the answers better, but to go one step further, with human feedback, to improve the possibilities of use. And perhaps to pave the way for a new era of search engines, capable not only of understanding questions but also of providing answers directly, without the need to link to other sites (which Google is already doing).
Is it possible to fall in love with an artificial intelligence?
by Francesco Marino
The (possible) problems
Of course, also with ChatGPT not everything went perfectly. Sometimes, the system fails to grasp the meaning of the sentences and it can happen (not very often, to tell the truth) that it gives wrong or off-centre answers.
Many Twitter users, then, stay trying to hack the system, changing its general characteristics: someone managed to make it speak like a white supremacist of 4Chan; others have circumvented the restrictions on violent language pretending to be dealing with a fiction text to bypass filters designed by OpenAI.
All experience for ChatGPT, whose debut in the real world is just a further training phase, to learn even better how human beings speak and what they are looking for. And being able to respond as one of them.