You’re fine AI technology.
How you work your charm on me,
And parse my words so naturally!
But how does it really, and then go on further to reply as a human would? Can you even trust ChatGPT? These are some common strands of thought that have started to emerge ever since ChatGPT’s launch.
As the netizens of the world start to uncover the potential of ChatGPT, it becomes imperative to understand what makes ChatGPT tick. How is it trained, and how is it able to so authoritatively tell us what we want to know without the usual browsing? How well can you rely on it? Let’s find out.
How is ChatGPT so chatty? An overview of how it works
To understand how ChatGPT works, one has to start with what it is and what makes it up.
The ‘GPT’ in ChatGPT stands for Generative Pre-trained Transformer – a type of neural network that works well for language processing, understanding user queries in a given context, and generating responses.
These neural networks mimic the brain and the neural pathways that it develops when learning new things and solving problems. This structure is then fed vast amounts of data, which then becomes its knowledge dataset, and ranges from books to articles to the world wide web.
When you give ChatGPT a set of words, it can take that prompt, look for relevant information through its data reservoir and then use its powerful architecture and sophisticated algorithms to generate human-like responses.
ChatGPT is scary good. We are not far from dangerously strong AI.
— Elon Musk (@elonmusk) December 3, 2022
It will take your text input, break it up into individual words, analyze it, and generate a probability distribution for all possible results. It then chooses the most likely set of words which becomes its response. The process is repeated multiple times, and based on the statistical sampling methods, can generate a novel response for each query. This is why responses are often slightly different for the same query, even though the crux of the matter may be the same. It will also differ, among other things, based on your choice of words, syntax, and topic.
As a family of language models developed by OpenAI, the GPT model is one of the largest in existence today and has been refined and fine-tuned for a variety of language processing tasks, including text generation, translation, and query-response.
ChatGPT runs on GPT-3, the third iteration of the language processing model, while the integrated Bing version utilizes GPT-4, the latest iteration of the model.
ChatGPT training models
One of the key ingredients that make ChatGPT up is its training models. To recognize conversational patterns and implement them, ChatGPT needs to know how conversations work in general.
Much of its training involved going over how users interact on the web. ChatGPT’s architecture is trained on a variety of conversation types:
- Conversations that have a constant back-and-forth between people in a dialogue format.
- Conversations that involve one user speaking at a time (turn-taking).
- Conversations that are more goal specific, such as the ones that involve solving problems or learning new things.
Community forums like Reddit, Quora, and Stack Overflow which contain a diverse range of topics and user interaction serve as the perfect training grounds for this. ChatGPT is also able to learn from past conversations, as well as the inputs that you have given in the current session, which helps it to provide better future responses. It also tells jokes.
Types of responses possible with ChatGPT
The ChatGPT bot is designed to be amenable to human conversations and, to that end, has capabilities much more advanced than what we’ve seen thus far. Things like text generation, translation, and simple question-answering – which our voice assistants often find challenging – are a breeze for ChatGPT.
ChatGPT will converse with you, analyze and summarize text, interpret and advise, and even tell how you may be feeling based on your word choice, and tweak its replies accordingly. It also writes whacky teenage poetry. All the makings of your new AI friend.
But you still wouldn’t want to completely depend on it, for a number of reasons.
Is ChatGPT reliable?
ChatGPT doesn’t know it all. In fact, it has only been trained on information available until September 2021. So that marks its official knowledge cutoff. Though its conversations with users across the world continue to contribute to its learning, it cannot provide clear answers with authority because those do not make up its training dataset, per se, and will tell you much.
It is also not immune to factual errors. This is because its database includes the internet on the whole, and the internet is not always a storehouse of truths. In fact, there have been various instances when its answers were far from fact, verging even on the pseudoscientific. So it’s safe to say that you may not always be guaranteed complete accuracy with ChatGPT.
So is ChatGPT going to lose over 100 billion dollars since it gave a wrong answer about a stock market crash on 2/15?
— Colin (@colin_gladman) February 15, 2023
Bias is another limitation. All AI models tend to have biases inherent in their training data, and it’s not uncommon to have such an AI chatbot such as ChatGPT generate responses that have elements of gender, race, or other prejudices. These things inevitably lead to stringent self-censorship measures by the parent company that come in the way of free-flowing conversations. But that is not anything new. Mindful users know that bias cannot so easily be eradicated. What would it even mean to do so?
Does ChatGPT bias surprise anyone?
Google search has been biased for years. Try it yrself.
— GeneralJaffey (@GeneralJaffey) February 21, 2023
Besides bias, privacy is another big elephant in the room. ChatGPT’s training set is composed of all the data that we’ve fed into it, and that is now forever a part of ChatGPT. Whatever users online may have unwittingly contributed to it is no longer within their control, and could also be used in clandestine ways.
ChatGPT is definitely the best we’ve got for a humanlike AI conversationalist. But it’s still no closer to growing a personality or common sense than a chimp is to understanding Shakespeare. Though that could only be a matter of when and not if.
Before we go, let’s take a look at a few commonly asked queries that users have about ChatGPT.
What is ChatGPT and how does it work?
Put simply, ChatGPT is an AI chatbot developed by OpenAI that can generate humanlike responses to queries thanks to its GPT language learning model and sophisticated algorithms, as well as the large corpus of data that it’s been trained on.
What are the benefits of ChatGPT?
There are tons of benefits of ChatGPT. Not only can it be used as a search engine that will tell you the answer and save you a lot of browsing, but also give advice, recommendations, and help you in your work.
Does ChatGPT give the same answer to everyone?
No, ChatGPT doesn’t give the same answer to everyone. This is due in no small part to its complex text generation process that is dependent on statistical probability for every single word that you type. The novelty of its answers can be increased by making the sampling process more or less random, or by defining parameters for your chat session, such as saying ‘Talk to me like I’m 5 years old’.
ChatGPT is unlike anything we’ve seen emerge from the world of AI. Its advanced algorithms and robust transformer-based language model have ushered in a new way of interacting with the web. But it is its conversational style that has got all of us going. We hope you now know how ChatGPT is able to work its magic and can use all that it has to offer to learn new things, get help with your work, and everything in between.