A superstar in the artificial intelligence (AI) scene, ChatGPT

March 12, 2024

The hottest potato in the recent artificial intelligence scene is ChatGPT. It seems like most people have at least heard the name and many have even tried using it themselves. Thanks to this, voices of anticipation and concern continue to pour out through various media outlets around the world, not only within the industry but also among the general public.

Unrivaled Extraordinary Presence

‍

3.5 years, 2.5 years, 10 months, 5 months, 2.5 months, ...

These are the times it took for Netflix, Facebook, Spotify, and Instagram, respectively, to reach 1 million users. However, not too long ago, a new superstar emerged, effortlessly surpassing the records of these renowned services. It's none other than OpenAI's ChatGPT, which surpassed them in just 5 days after its release.

So, why are people so enthusiastic about ChatGPT?

Of course, one reason for the enthusiastic response is that it was freely available for anyone to use like a chat service, leading to active virality through social media. However, even three months after its release in this rapidly changing era, it's still hot, which is quite remarkable. Let's take a closer look at what makes this remarkable entity so attention-grabbing, how it can be utilized, and what impact it might have.

Let's find more about this

First, let’s check the introductory article on the OpenAI blog that unveiled ChatGPT.

We’ve trained a model called ChatGPT which interects in a conversational way. The dialogue format makes it possible for ChatGPT to answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests.

As described by OpenAI, ChatGPT is a type of language model trained to interact with users conversationally. At first glance, it may seem like a simple chatbot, but its true capabilities become evident from the next sentence.It can "answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests."Isn't that surprising? Even reading this text again at the present moment, it becomes curious whether it's truly an accurate introduction of a machine.

‍

Where on earth did this remarkable entity emerge from?

Basically, ChatGPT is based on GPT-3.5. (If you're reading this, you probably already know) GPT (Generative Pre-trained Transformer) is a natural language generation model developed by OpenAI. In fact, GPT alone has been astonishing since its initial release, generating text that appears as if it were written by a human. GPT-3.5, released in 2018, uses over 1,500 times more parameters than the original GPT-1, achieving a remarkable performance improvement.

‍

However, the number of parameters in GPT-3.5 remains the same as GPT-3. Instead, significant improvements have been made through Reinforcement Learning from Human Feedback (RLHF). Thanks to RLHF, ChatGPT can not only understand user input and generate responses but also engage in natural communication as if conversing with a real person who has expertise in the relevant field.

‍

RLHF has already been applied to GPT-3 and its predecessor, InstructGPT, which can be considered a sister model to ChatGPT. InstructGPT, as its name suggests, is trained to generate outputs based on human instructions, much like today's protagonist, ChatGPT. However, there are some differences in data collection methods and other details, and this is precisely where the crucial difference from previous-generation models lies.

‍

To achieve this, OpenAI started by collecting training data. Human workers, commonly referred to as labelers, were tasked with creating scripts consisting of questions and answers. The conversational (response) format data created in this way is used to train the existing model through supervised learning-based fine-tuning.

‍

In the next step, a control group dataset is collected, and training of the reward model is conducted. To create a reward model for reinforcement learning, comparative data is needed to rank the model's responses. To collect this data, prompts and responses generated by the model are sampled, and human workers rank them using a reward model based on the ranking method, which facilitates reinforcement learning.

‍

Afterward, this process is repeated several times, eventually resulting in the final product, ChatGPT.

‍

So, what can it do?

‍

OpenAI describes the capabilities and provides usage examples of ChatGPT through its service webpage.

Additionally, it is known to be capable of performing a wide range of tasks, including Q&A, conversation generation and summarization, debates, translation, essay or scenario writing, category classification, programming, and code error correction, almost as well as humans. Moreover, being one of the hottest services recently, it has already undergone various tests by numerous people. A quick internet search reveals many interesting use cases.

‍

Looking at several examples, I've come to realize that the greatest feature and potential of ChatGPT lie in its conversational format. Even if users fail to provide appropriate prompts as before, ChatGPT can generate better results through a continuous exchange of questions and answers, resembling a conversation with a human. While this may seem natural given the methods used in ChatGPT's training, it's still remarkably impressive to see how effectively it functions (perhaps even better than expected) through dialogue with both machines and humans.

‍

Are there any limits and risks?

It is clear that ChatGPT is a significantly advanced model compared to its predecessors and has shown remarkable potential. However, it still has some of the limitations and issues demonstrated by previous models. The ChatGPT website's description also acknowledges these limitations, such as occasionally generating inaccurate information, producing harmful or biased content, and having limited knowledge of the world and events after 2021.

Additionally, it also suffers from the common problem of hallucination found in language models. Where they occasionally generate plausible sentences not based on facts. Of course, while ChatGPT has shown improvements by acknowledging mistakes and challenging inaccuracies or refusals compared to previous models, it seems that these issues have not been completely resolved.

‍

Moreover, these limitations entail significant risks a lot more than you might expect. It's because they stem from harmfulness and biases present in some of the training data. Realistically, there is no foolproof way to fully control all the data used in training large-scale models immediately.

On the other hand, it is also true that ChatGPT has shown an astonishing potential to the point of being intimidating. Particularly, the way it interacts through conversation and conducts tasks is reminiscent of collaboration between humans and machines. Just as we can achieve better results through collaboration than working alone, I believe ChatGPT has demonstrated the potential to maximize our abilities through collaboration with machines.

In the movie, The Iron Man communicates with the artificial intelligence assistant J.A.R.V.I.S., finding the desired information and solving problems together through dialogue. The emergence of ChatGPT has shown that the imaginative scenarios depicted in movies are not far from becoming reality. With this trend, it's not far-fetched to imagine that soon we'll be interacting with AI assistants to manage tasks, engage in creative activities, and enjoy daily life, and more. I look forward to seeing it would maximizes human potential.