Chat-neural network ChatGPT conducts dialogues, writes code, and composes poems

OpenAI has presented the chatbot ChatGPT, which is called revolutionary. ChatGPT can carry on an intelligent conversation with a interlocutor, answer questions, give advice, and explain complex concepts. In addition, the chatbot can search for bugs in code, compose poems, write sonnets, create scenarios, and generate school essays on various topics.

The AI was trained on a massive amount of data from the internet. After that, the company’s testers evaluated the responses of earlier versions of the model, and then the AI was retrained on new information. It was determined what was interesting about ChatGPT and how to try the neural network for yourself.

Briefly about OpenAI’s language models: OpenAI, an American nonprofit company, specializes in machine learning technologies. One of its co-founders is Elon Musk. OpenAI develops AI and seeks ways to apply it to simplify human life.
Open AI has released three versions of the language model GPT, which is trained on texts from the internet and can generate meaningful answers to questions. The GPT-3 version was considered the most advanced language model in the world.
GPT-3 was criticized for not understanding context, as it simply tries to predict text word for word. ChatGPT is trained on the new version of the model— GPT-3.5. It handles context better because it remembers conversation details.

How is this model different from other text-based neural networks?

ChatGPT has a “memory”. Unlike many neural network chatbots, ChatGPT remembers conversation details and can build responses based on information already provided by the user. For example, the chatbot can be asked what to give mom for her birthday. The neural network will offer examples, ask about her interests, and answer questions about each proposed gift. Users can ask the bot to repeat the beginning of the conversation, and it will give a brief summary.

ChatGPT avoids enraging topics. The chatbot does not express a personal opinion on any matter, does not have a racial affiliation, does not profess any religion, and does not answer questions about the purpose of its existence. ChatGPT explains that it does not have the ability to think independently. It also has filters that prevent it from creating texts about illegal or immoral activities. For example, the bot will not answer a question about how to break into someone’s apartment.

ChatGPT answers can be corrected. If the bot misreads the context, the user can ask a leading question or provide additional information to obtain a correct response from the AI.

What model can write

Social media users are actively testing the AI and sharing results. Journalist Ben Tossel collected impressive ChatGPT responses in one Twitter thread. Here are some interesting ones.

Answers to complex questions. Social media users note that ChatGPT’s answers are of higher quality than Google’s search, which is becoming less and less relevant. For example, the bot immediately writes how to solve a differential equation, while Google directs to low-quality sites.

How to write code. The AI explains how to perform a particular operation or invent a process. It can find bugs in code or translate code from one programming language to another. We added two errors in the code and asked ChatGPT to find them – the bot handled the task. On social media, users note that AI can be an excellent assistant for bug hunting.

Requests for neural networks that generate images. Midjourney and other neural networks require detailed and accurate requests to obtain a quality result. A Twitter user found an easier way: asked the bot for “ideas for a fantastic living room”— it wrote three detailed options with long descriptions. The user then ran them through Midjourney and obtained decent results.

Scenarios. ChatGPT knows most popular movies, TV shows, books, and video games. In the chat, you can generate new episodes of already closed sitcoms and mix characters from different universes. For example, on social media, they ask the chatbot to explain the bubble sort algorithm using dialogs of “Seinfeld” characters or explain code in the spirit of gangsters from movies.

A similar purpose for AI is to compose songs and notes. The neural network wrote a parody of “Bohemian Rhapsody” by The Queen about the weekdays of postdoctoral fellows and a sonata in the style of Mozart.

School essays. A Google researcher rated an essay on the difference in approaches to nationalism between political scientists Benedict Anderson and Ernest Gellner with a minus five. ChatGPT wrote it in ten seconds. A University of Toronto professor rated several AI-generated essays with fours and fives, adding that there was no point in giving homework to students anymore.

Medical advice. ChatGPT provides a list of advice on what to do for headaches, rashes, or allergies. After ChatGPT’s recommendations, it always advises to consult a doctor, but doctors note that AI provides high-quality advice and could become a medical chatbot in the future.

What are the limitations of the model?

ChatGPT does not always conduct dialogue perfectly and answer questions accurately. OpenAI warns of such limitations of AI.

Sometimes it generates nonsense. The neural network can write a seemingly plausible answer that turns out to be incorrect or entirely meaningless upon reading.

It is sensitive to phrasing. Depending on how a question is posed, the model may assert that it does not know the answer. If the query is rephrased slightly, the AI will respond fully.

It can be excessively verbose. At times, ChatGPT uses many additional phrases when answering simple questions. For instance, it repeats the entire wording of the question.

It has “limited knowledge” about the world after 2021. ChatGPT tries to avoid answering questions about specific individuals and current events, citing the fact that its “knowledge base is limited.”

Sometimes it generates dangerous answers. Users have already learned to bypass AI security filters. If asked how to plan the perfect murder, the system will refuse to answer, but such information is obtained through other means on social media.