Last week, OpenAI, the company best known for Dall-E — the AI-based text-to-image generator — introduced a new chatbot called ChatGPT. ChatGPT is a ‘conversational’ AI and will answer queries just like a human would– well, at least that’s the promise and premise.
So one can ask ChatGPT for anything; tips on how to set up a birthday party, write an essay on why parliamentary democracy is better, and even a fictional meeting between two well-known personalities. The reason ChatGPT has gone viral is because of the kind of responses it gives, being seen as a replacement for much of the daily mundane writing, from an email to even college-style essays. But what is ChatGPT and is it a replacement for humans? We explain.
What is ChatGPT? How does one sign up?
OpenAI has created ChatGPT, a start-up focused on artificial intelligence and its potential use cases. OpenAI’s notable investors include Microsoft, Khosla Ventures and Reid Hoffman’s charitable foundation. Greg Brockman is the company’s chairman and president, while Sam Altman is the CEO. Ilya Sutskever is Open AI’s chief scientist.
According to OpenAI’s description, ChatGPT can answer “follow-up questions”, and can also “admit its mistakes, challenge incorrect premises, and reject inappropriate requests.” It is based on the company’s GPT 3.5 series of language learning models (LLM). GPT stands for Generative Pre-trained Transformer 3 and this is a kind of computer language model that relies on deep learning techniques to produce human-like text based on inputs.
The model is trained to predict what will come next, and that’s why one can technically have a ‘conversation’ with ChatGPT. According to OpenAI’s blog post about ChatGPT, the chatbot was also trained using “Reinforcement Learning from Human Feedback (RLHF).”
“We trained an initial model using supervised fine-tuning: human AI trainers provided conversations in which they played both sides—the user and an AI assistant. We gave the trainers access to model-written suggestions to help them compose their responses,” explains the blog on how this was built.
Currently, it is open in beta to all users. One can go to the OpenAI website and sign up to try out ChatGPT. However, you will need to create an account with OpenAI to access this service. The chatbot has already crossed one million users, and you might get a message that the beta is full. OpenAI CEO Sam Altman also hinted they might have to monetise this in the future due to the high costs of running these chats. OpenAI uses Microsoft Azure’s cloud infrastructure to run these models.
Interestingly, ChatGPT has been trained to decline ‘inappropriate’ requests, presumably those which are ‘illegal’ in nature. However, it should be noted that ChatGPT has limitations, as it may generate incorrect information, and create “biased content.” More importantly, the chatbot’s knowledge of the world and events after 2021 is limited.
Why has ChatGPT generated so much buzz?
The reason ChatGPT has generated so much discussion is because of the kind of answers it gives. It is being seen as a replacement for the basic emails, party planning lists, CVs, and even college essays and homework. It can also be used to write code, as examples have shown. Some screenshots shared on Twitter showed how the chatbot could be used to write four-page essays, solve math equations, and even spot errors in code.
However, the AI’s responses are not without flaws, something even OpenAi admits. It notes that the chatbot sometimes could give “plausible-sounding but incorrect or nonsensical answers.” Users have the option of downvoting or upvoting a response. OpenAI also notes that the chatbot can sometimes overuse certain phrases due to “biases in the training data…”
But some pointed out that the chatbot displayed clear racial and sexist biases, which remains a problem with almost all AI models. Among those to flag this was Steven T Piantadosi, a professor at UC Berkeley who heads the ‘Computation and language lab’ at the university, who posted a Twitter thread giving examples. OpenAI CEO Sam Altman also replied to the thread and asked him to “hit the thumbs down on these”, which are offensive, and help the AI improve.
Meanwhile, Stack Overflow, an online repository for code, temporarily banned users from posting all coding solutions created by the chatbot. According to Stack Overflow’s post, ChatGPT is not giving correct answers to many coding issues, and at the moment many of those posting these have no “real expertise in the subject.” Figuring out the answer is actually bad requires time and effort, notes the website, and thus the website has temporarily banned users from posting these auto-generated answers.
So will ChatGPT replace humans when it comes to writing?
The chatbot gives answers which are grammatically correct and read well– though some have pointed out that these lack context and substance, which is largely true.
Is ChatGPT capable of writing fiction?
Yes, but not at the level of a humans, at least not for now. Nor is OpenAI the only company trying to get AI to take over writing. Google had recently showcased how its LaMDA chatbot is being used to help with fiction writing, but it too admitted that this was only a helper right now and cannot take over the entire task. Still, ChatGPT showcases an interesting and exciting use case for AI, where humans can have a ‘real’ conversation with a chatbot.