Over the past few months, ChatGPT has captured my attention through numerous forums, articles and AI experts. I am fascinated by this technology and learning more about it everyday. With this introductory post, I hope to share my knowledge and insights on ChatGPT - how it works, its capabilities and limitations, potential pitfalls, and future prospects.
Note - Just for fun, I did an experiment to write two more versions of this post - one that was written by ChatGPT itself using a rough outline I provided, and another also written by ChatGPT, but imrprovised from my original post. Curious to see the results of the experiment? Head over to the meta post Blog about ChatGPT in three different ways.
What is ChatGPT?
ChatGPT is a chatbot capable of having human-like conversations using AI and NLP (Natural Launguage Processing). This bot can answer your questions on a wide range of topics, write code, summarize text and explain concepts in a simple way. ChatGPT was developed by OpenAI, the creators of DallE-2, the popular AI text-to-image generator and Whisper, the powerful speech recognition system.
ChatGPT is currently in research preview and is free for everyone to use. Just a couple of weeks ago, OpenAI announced ChatGPT Plus, a subscription plan available for $20 per month.
ChatGPT under the hood
ChatGPT is built on top of the GPT 3.5 series of large language models that generates human-like text. Woah! That’s a lot of buzz words - let’s break it down one by one.
A large language model (LLM) is a mathematical function (probabilistic distribution) applied on a sequence of words. Given a prompt with a few words (or technically “tokens”), they can predict the most likely next word to complete the sentence. GPT (Generative Pre-trained Transformer) is one such family of LLMs which is trained on huge volumes of text data and generates conversational text using supervised learning and reinforcement learning. It uses the Transformer machine learning model that provides differential weighting to determine the significance of each part of the input data. This technique helps in priroritizing human-like responses over purely linguistically valid responses.
GPT uses human trainers throughout the learning process - first to train the model using labelled conversations (hence the term pre-trained) and then to rank the responses generated by the model, creating reward models to fine-tune the responses in an efficient way. GPT 3.5 was trained on Microsoft Azure AI supercomputing infrastructure.
GPT has a wide range of Generative AI applications such as chat bots, summarizing texts, translating content in different languages and generating new content and code. ChatGPT is one such application of GPT that interacts with the user in a conversational way.
Get started with ChatGPT
It is pretty simple to start, follow these three steps:
- Go to https://chat.openai.com/chat on your browser - use your desktop, laptop or phone
- Log in with an Open AI account. If you don’t have an OpenAI account, create one using Sign up button
- Once you login, ChatGPT will show you a few examples, capabilities and limitations. It will also show a text box at the bottom. This is the main interfac wHere you can interact with ChatGPT.
That’s it! You can type in anything into the text box and ChatGPT will respond with an answer, comment or a question. You can keep on asking questions and it will remember the context from your previous question and answer the next one based on that. Have fun with it!
The words that you type into the text box are called prompts and play a critical role in determining the quality of your interaction with ChatGPT. Well-designed prompts generate meaningful responses from ChatGPT. The more context and examples you provide, the richer the response will be. Here are a few resources that will help you create good prompts - Learn GPT - the best ChatGPT examples, Awesome ChatGPT prompts and 100 best ChatGPT prompts.
The Power of ChatGPT
ChatGPT can perform a variety of tasks. For example, it can..
- explain complex concepts in a conversational teaching tone - for eg:, try this prompt what is bernoulli’s equation
- generate any text content - write a 200-word summary of mobile technology platforms
- generate code for a given problem - for eg: write a Go program to check for palindrome. The code will be so complete that if you copy into an IDE, it will most likely just compile successfully.
- debug a given piece of code, for eg: you can say this code is not working like i expect — how do i fix it? along with some code that manages error channels in Go. It will ask you clarifying questions and tell you where the problem is.
- summarize text into concise form - you can give it a huge paragraph of text and say summarize this into 100 words, or write a summary of the book Homo Sapiens
- be a creative writing partner that can help write books, article or blog posts, Create a title and plot summary for a new historical fiction novel
The Pitfalls
Despite its incredible power and potential, ChatGPT has its own limitations and pitfalls that could cause potential issues. Like all AI bots, its knowledge is limited to the statistical regularities of its training data and does not have human-like understanding the world in an abstract complex form. Hence, quite often, ChatGPT will confidently present false or invented information as facts that sound plausible.
It is also vulnerable to the algorithmic bias in the training data which manifests in how it responds to prompts about describing people. Its training data set in the initial release was also constrained to the events that occurred after 2021, so it would give incorrect answers about recent events. OpenAI team has been diligent about refreshing the model as recent as 30 Jan 2023.
As the old adage goes A fool with a tool is still a fool. Anyone who uses ChatGPT should exercise caution in how the information provided by ChatGPT is used. Do not blindly trust its responses; we should verify its accuracy before using it anywhere.
The Future
ChatGPT (and LLMs in general) can fundamentally transform the way we apply AI to solve problems in the real world. As Sam Altman, the CEO of OpenAI said in an interview with Reid Hoffman, companies will move away from the current paradigm of starting with a domain-specific models and applying NLP on it, to a new paradigm where they will start with large language model like GPT-3 and apply domain-specific knowledge on top of it. This increases the possibility of applying AI in almost every domain.
LLMs and Generative AI could transform how we search for information and create content and art. Microsoft has announced AI-Powered Bing and Edge and is exploring ways to integrate GPT-3 into its Office suite of applications. It can be a powerful tool to explore ideas, generate content and write fluently and professionally with minimal literary expertise.
Conclusion
What does this mean to all of us? Does Generative AI pose a risk to artists, authors and software developers? Well, no.
Andrej Karapathy, who led the computer vision team at Tesla, tweeted that GitHub copilot has accelerated his coding so much that 80% of the code he writes is using GitHub co-pilot.
As Satya Nadella shared in a recent interview at the 2023 World Economic Forum, if Andrej, who is one of the best AI developers could get that much leverage and productivity boost, why wouldn’t every developer adopt this technology and take advantage of it? And I agree.
I myself did an experiment to understand how I can use ChatGPT in my day-to-day and it has completely changed the way I work. You can read about my experiment and its results at Blog about ChatGPT in three different ways.
Generative AI technologies like ChatGPT are here to stay and they can help us. We must learn how to use it and take advantage of it to improve our writing skills, coding skiils and overall productivity, albeit beware of its limitations. The benefit and productivty boost far outweights the risks associated with it.
So let’s not ignore or run away from this technology. Let’s embrace it, learn and use it carefully.
comments powered by Disqus