At this point, ChatGPT has taken the world by a storm; thus far, it has been used by over five hundred million people worldwide. There has been lots of conversation regarding this innovative chatbot, and while I’m not a machine learning expert, I decided to research more into the development of ChatGPT and test out the algorithm for myself!
ChatGPT is a platform created by the company, OpenAI. They focus on developing various artificial intelligence based technologies; in fact, ChatGPT isn’t their first chat tool, but rather a significant improvement on their previous chatbot, GPT-3. Due to ChatGPT’s unprecedented prowess, there have been concerns as to whether it can “replace” human writing, or be used to plagiarize in academic settings. Hence, I’ll be providing a summary of the technology that powers ChatGPT, as well as my take on its potential applications.
Some Helpful Pre-Knowledge
Before understanding the ins and outs of ChatGPT, it’s first necessary to understand what machine learning and deep learning are.
Source: Singapore Computer Society
As summarized by the diagram, deep learning is a subfield of machine learning that uses neural networks (visualize interconnected neurons, inspired by the human brain) to train computers on information; in other words, computers are “learning” the information and from their past mistakes to make more informed predictions.
Deep learning can be broken down even further, though. Supervised and unsupervised learning are sub branches of deep learning, for instance. Supervised learning involves using classified datasets so that computers can predict outcomes accurately, while unsupervised learning involves analyzing and clustering unlabeled datasets. There’s also reinforcement learning, in which developers “reward” desired outcomes and predictions. This is what ChatGPT uses, and we’ll explore this concept later on in the article.
Another term to know is natural language processing, or NLP – a field in which computer scientists program computers to process and analyze large amounts of natural language data. Deep learning, neural networks, and other algorithms are applied to NLP. And considering that ChatGPT is a chatbot that processes and responds to human language, NLP plays a significant role here.
So… How Does It Work?
ChatGPT employs the NLP process; first, we’ll go over the specifics of NLP, and then go into how ChatGPT really stands out.
Natural Language Processing
NLP can be broken down into smaller steps, all of which contribute to a computer first processing, understanding, and emulating human language.
First, a computer must process language by breaking it into separate pieces (known as “tokens”). When comprehending a language, we humans do the same, as well. Through sounding out words and sentences while reading (especially as young children) we’re breaking it up into smaller syllables, and then understanding the smaller bits in the context of the entire sentence. A computer does this using algorithms to segment a paragraph into sentences. With these sentences, it normalizes words into their root form; “intelligent” becomes “intelligen” and “plays” becomes “play.” Lastly, it removes common or filler words such as “a”, and “the.” Lastly, the computer relates words to one another, determining what the noun, subject and adjectives are as well as what they describe. For a more technical understanding of the processing phase, I recommend checking out this summary and this article on part of speech tagging.
The processing phase is finished, and all the fancy machine learning algorithms now come into play! To achieve full understanding of a sentence, a computer must understand the semantic information (specific meaning of an individual word), syntax information (understanding the parts of a sentence, which is mainly covered in preprocessing), and context information (the backstory of a sentence). There are a plethora of algorithms that can accomplish this, using principles of deep learning.
For instance, word embeddings (representation of words as numeric vectors) are created based on the distributional hypothesis, which states that words appearing within a similar context possess a similar meaning. Using a neural network, these word embeddings are pre-trained on a task where the objective is to predict a word based on its context. Specifically, a long short term model (LSTM) is employed as the computer attempts to “fill in a blank” with the most probable word given the surrounding context; this requires lots of input data, as the computer “learns” which word is more probable. This process gets extremely detailed, so if you’re looking to get a better understanding of word embeddings, I recommend this read.
Source: Molly Ruby, Medium.com
ChatGPT, Transformers, and Reinforcement Learning
ChatGPT incorporates the same principles of machine learning as described above while improving upon the LSTM. The LSTM is limited since a model is unable to value certain words more than others; for instance, based on common examples, both “loves” and “hates” can fit the context of reading. Also, the input data is only processed sequentially, and the computer doesn’t examine the complexities and correlations that could exist within the text as a whole.
This is where transformers come in. Introduced in 2017 by Google Brain, transformers use a self-attention mechanism that enables them to process all input data simultaneously in order to find correlations between different words (which would help determine the context). The theory behind a transformer is explained in depth here, and I think it’s worth reading.
ChatGPT extended the transformer architecture through incorporating an encoder to process the input sequence and a decoder to generate the output sequence. The encoder will take a sentence as input and then output a sequence of numbers – the vector representation of the words. This serves as the input for the decoder, which will then output a sentence.
Source: Nechu BM, Medium.com
To increase its accuracy, ChatGPT has incorporated a novel way to use human feedback in its model – known as Reinforcement Learning from Human Feedback (RLHF). This is described in a 2022 paper from OpenAI.
First, supervised learning was utilized as a supervised training dataset was created; each input had a known output to learn from. In fact, these inputs were collected from users. GPT-3 (ChatGPT’s predecessor), was trained on this data to create a supervised fine tuning model.
After, a reward model was trained in which the model input is a series of prompts and responses, while the output is a value, referred to as a “reward.” This type of model is the precursor to a Reinforcement Learning Model, where the model produces outputs to maximize its reward. The model is presented with a random prompt and returns a response; based on the reward model, a value is then calculated for the prompt and response pair. This value is then fed back into the model, allowing it to produce better outputs. Collectively, this process is known as developing a “policy,” or strategy, that the machine uses to achieve its goal (in this case optimal outputs). This method is known as Proximal Policy Optimization (PPO).
To recap, ChatGPT is incredibly innovative as it uses the reward system – similar to how one would train their dog to play fetch. If the dog gets the stick, it gets a treat; otherwise, it doesn’t. From this, the dog will learn what the appropriate behavior is, as it is being rewarded for the correct behavior. Similarly, through returning a positive scalar value, the ChatGPT model is learning what the best output is.
Some Tests I Ran…
So, here’s the part where we circle back to the title of this post. Everyone’s been talking about how ChatGPT is providing a breeding ground for student cheating and plagiarism. However, I wondered to what extent this is true; as ChatGPT is pulling from a database and synthesizing information, is it able to author creative, striking insights required to earn a passing grade in the first place?
To demonstrate, I’ll be testing ChatGPT on AP English Language and Composition exam essay prompts. When scoring high in the AP English Language Essay, merely making a statement is not enough; rather, to earn a good score, this statement has to carry nuance and sophistication and has to be supported by evidence and commentary.
For example, take this argument essay prompt from the 2013 exam:
“For centuries, prominent thinkers have pondered the relationship between ownership and the development of self (identity), ultimately asking the question, “What does it mean to own something?”
Plato argues that owning objects is detrimental to a person’s character. Aristotle claims that ownership of tangible goods helps to develop moral character. Twentieth-century philosopher Jean-Paul Sartre proposes that ownership extends beyond objects to include intangible things as well. In Sartre’s view, becoming proficient in some skill and knowing something thoroughly means that we “own” it.
Think about the differing views of ownership. Then write an essay in which you explain your position on the relationship between ownership and sense of self. Use appropriate evidence from your reading, experience, or observations to support your argument.”
Source: 2013 FRQs
I then inputted this prompt into ChatGPT, and after regenerating the essay response a few times, I pasted the final essay below:
In my opinion, the relationship between ownership and the development of self (identity) is complex and multifaceted. While it is true that owning objects can sometimes have negative effects on a person’s character, as Plato suggests, it can also have positive effects, as Aristotle suggests. Similarly, while ownership of intangible things such as skills and knowledge can help to shape one’s sense of self, it is also true that simply possessing these things does not guarantee personal growth or development.
For instance, owning material possessions can lead to feelings of pride and a sense of achievement, which can contribute positively to a person’s sense of self. On the other hand, excessive attachment to material possessions can lead to feelings of greed and envy, which can have a negative impact on one’s character. In addition, owning too many things can lead to feelings of overwhelming responsibility and stress, which can also have negative effects on a person’s sense of self.
Similarly, becoming proficient in a skill or knowing something thoroughly can provide a sense of fulfillment and a deeper understanding of oneself. However, simply having these abilities does not guarantee personal growth or development. For example, a highly skilled artist may still struggle with feelings of self-doubt and insecurity, while someone who is not skilled in a particular area may have a strong sense of self and confidence.
In conclusion, I believe that the relationship between ownership and the development of self (identity) is complex and depends on the individual’s attitudes, motivations, and circumstances. While ownership can have both positive and negative effects on one’s sense of self, it is ultimately up to the individual to use these experiences to grow and develop in positive ways.
While ChatGPT does formulate a thesis, its ideas are copied from great philosophers; ChatGPT provides a summary of information, but doesn’t elaborate on insightful ideas. If I were to hand this essay to my English teacher, I’m confident I would receive a failing grade.
However, that doesn’t mean that ChatGPT can’t be useful in improving our writing. For instance, as I’m constantly looking for new topics to write about, ChatGPT has provided some inspiration. I asked it to “provide short story ideas on coming of age.” Pasted below is the response:
- A young person discovers their own identity while navigating their first love and the challenges of high school.
- A young person must find their place in the world after the sudden loss of a loved one.
- A teenager’s world is turned upside down when they find out they have superpowers, and must learn to control them.
- A young person sets out on a journey of self-discovery after feeling disillusioned with their upbringing and culture.
- A group of friends come of age and learn about themselves and the world around them during a summer of adventures.
- A young person learns to stand up for themselves and their beliefs, despite opposition from their community.
- A teenager grapples with the consequences of a reckless act and learns to take responsibility for their actions.
- A young person must navigate the complexities of family dynamics and cultural expectations while trying to pursue their dreams.
Overall, I was impressed; although these ideas are vague, they can serve as a valuable starting point for a short story or poem, enabling me to brainstorm further and develop a concrete plot.
Looking into the Future
None of this should take away from the academic concerns for ChatGPT. While the chatbot is a pretty bad way to cheat on English assignments, it should not be used to cheat or plagiarize at all. ChatGPT will play an inevitable role in our future; yet, this puts us in a critical place to figure out what that “role” may look like. Can ChatGPT be used to help students summarize information for reports? Or is that also considered “cheating”?
Personally, I think that ChatGPT can be a valuable asset to human communication, provided that it’s used in a deliberate and proper manner. The information synthesized by the chatbot cannot be fully trusted, meaning that the algorithm should be used wisely. Furthermore, ChatGPT does not create anything original; only humans can do that. Yet, communicating these original ideas is a challenge. Thus, depending on the prompts that we feed the algorithm, ChatGPT has the potential to help us become better communicators through providing insights on how to use the human language. As a writer, I envision ChatGPT helping me prepare first drafts; often, my ideas are at the tip of my fingers as I’m typing, but I can’t find the appropriate words and structures to give them life. Through experimenting with different prompts and inputting these fragments of ideas into the algorithm, ChatGPT can help me find the words I’m looking for and improve my communication.
Overall, ChatGPT is quite an interesting innovation. As the algorithm improves and more NLP tools surface, I think it’s important to keep one thing in mind: while technology can enhance our existing intellectual abilities, it can never serve as a replacement.
Sources
Although I linked some resources to learn more about ChatGPT throughout this article, I thought I would share a couple more resources that helped me learn about the algorithms behind the chatbot, namely this YouTube video and this Medium article.
Thanks for reading this article! Feel free to leave your comments below, and for more content like this, take a look at the Parea column.
