Recently the company that developed GPT-3, OpenAI, published its newest language AI, aptly named ChatGPT.
The reception from the general public has been amazingly positive and writing from personal experience, the model itself is quite deserving of its applause.
ChatGPT is what someone of a younger age would consider to be JARVIS from the popular Marvel franchise, due to its extremely adaptable and conversationally sound responses.
Even to adults and those in the know, the responses generated from ChatGPT are quite hard to distinguish from an actual human. Only with close inspection and the sometimes overtly technical explanation, can one deduce that the responder is in fact an AI.
Let’s delve a bit into context.
What is ChatGPT?
ChatGPT is a language model based on GPT-3, published by the research laboratory OpenAI, which also published the aforementioned GPT-3. It can respond to any and all questions you ask, most likely with perfectly structured sentences and sound explanations, making it the perfect tool for conversation simulation, story generation and text transformation.
As for the downsides, OpenAI themselves stated that the only problems with ChatGPT currently are its over-detailed, long-winded explanations to things as well as being a bit too paranoid of giving an answer to a possibly controversial question.
Computationally, it is less robust than GPT-3, yet it’s much more convenient and accurate in its responses. Besides the few occasional flaws, ChatGPT is near perfect as a conversational AI assistant.
A sample interaction with ChatGPT:
Q: ChatGPT, I forgot how many cities there are in the continent of Asia and I am too lazy to Google it. Can you give me an answer?
ChatGPT: The continent of Asia is quite large and diverse, so it is difficult to give an exact number of cities without more specific parameters. However, I can tell you that there are many large and well-known cities in Asia, and the number of cities in a continent can be in the thousands depending on how small of a population size is considered as a “city”.
For example, China alone has around 662 cities and towns; India, which is also one of the most populous countries in the world has around 7,935 towns and cities and Japan also has at least 700 major cities, to name a few examples from the continent.
Please let me know if you need any more information.
What is GPT-3?
GPT-3 is a widely recognized language model published by OpenAI on June 11th, 2022. It was trained on 175 billion parameters, and at that time, was considered as the largest neural networks (neural networks and language models go hand in hand) created to this day. GPT-3 could generate, transform, paraphrase, summarize and basically everything else when it came to text.
You could ask it a question, give it a prompt, give it coding problems or ask about the weather for the week or ask it to translate, change the text’s narrative tone or just plain talk to you like old friends and it would give you a mostly cohesive answer with relevant information or the correct persona. Of course, small bits of mishap happen here and there but the model is generally accurate and helpful.
For a bit more info, GPT-3 is the 3rd generation of the GPT series (Generative Pretrained Transformer) by OpenAI.
The keynote, though, is that GPT-3 is based on prompts. It was designed to accept prompts and generate outputs based on said prompts. This means that it did not exactly ‘discipline’ itself to be more conversationally capable or dialogue-friendly. It just generated what it considered to be the most accurate body of text as an answer to the prompt.
What is a language model?
Language models are algorithmic computer processes that collect and process tremendous sizes of data, learn from said training and then finally generate results that make sense based on said data and the user-given prompt.
In less wordy terms: If you ask an NLP (Natural Language Processing) model a question, it will formulate an answer based on its existing knowledge to give you the best answer it can provide. The data was taken from humans, so the response will be human-like. Depending on the engineers that made the model, it may be more inclined towards a specific answer compared to the alternatives.
So what is the difference between ChatGPT and GPT-3?
To properly realize the relationship between ChatGPT and GPT-3, the story must be recounted chronologically and neatly.
OpenAI first made a bunch of generative pretrained transformation models, their latest being GPT-3. This took an unfathomably large amount of data and multitudes of cutting edge hardware to properly train and utilize. It was known as the largest NLP model at the time.
Then, OpenAI started exploring other ‘creative’ venues besides writing such as art with DALLE-2 and music/audio with Whisper. Each one of these models were such a hit that the general public was highly aware of them, with conversations about AI and its capabilities sparking one after the other on reputable media platforms, only increasing the relevancy of this now crucial topic of sophisticated AI generations.
In this way, OpenAI can be considered a pioneer in AI and Machine Learning, as they have set the foundations for further development and public interest in the subject matter for this generation. This does not take away from the past contributors’ tremendous work, of course. In fact, it builds on it.
Furthermore, with increased public support and an abundant budget, OpenAI was set to release GPT-4 sometime in the near future (this was during December). However, before they have done so, they released GPT-3.5, the middle child.
GPT-3.5, now commonly referred to as ChatGPT, is a smaller yet more conversationally specialized version of GPT-3. It doesn’t handle general text transformation and generation like GPT-3, but it does handle anything and everything related to a chat. The model was trained for the sole purpose of continuing a conversation naturally with deeper understanding of the chat’s context, structure and tone. Thus, it makes for a much more accurate and faster model, compared to the general purpose GPT-3, when it comes to chatting.
It is also considered by OpenAI to be more aligned, meaning it is much more in-tune with humanity and the morality that comes with it. Its results more constrained and safe for work. Harmful and highly controversial utilization of the AI has been forbidden by the parent company and it is moderated by an automated service at all times to make sure no abuse occurs.
To summarize, OpenAI is a general text transformation tool capable of completing various text-based tasks mostly accurately while ChatGPT is an finetuned, smaller version of OpenAI that’s specialized for chat-form dialogue.
- ChatGPT was released on January 27th , 2022.
- 117 million parameters.
- Finetuned using ‘reinforcement learning with human feedback’ methodology.
- Trained on conversational data.
- Intuitive user interface suitable for those interested.
- Works with a chat format (requests/questions).
- Much more aligned.
- Not as robust as GPT-3, but more accurate in its own method.
- GPT-3 was released on June 11th, 2020.
- 175 billion parameters.
- Intended to be finetuned.
- Is the basis of all the models.
- Trained on general written media such as scripts, books, articles.
- Works with any type of general text/prompt.
- Not too accessible for general public.
- Prone to error.
- Very robust, able to handle most prompts. Can be just as accurate as ChatGPT, requires better prompting expertise.
So, to summarize, ChatGPT is similar to a kitchen knife, while GPT-3 is similar to a Swiss knife.
The capabilities are similar yet the effectiveness is different, depending on the application, one is preferred to the other.
But, given enough time and care, GPT-3 is able to be utilized much more effectively and in-broad range than ChatGPT.
If you are someone who’s interested in all the hype about ChatGPT, please go to this link to try it out yourself.
If you are more inclined towards learning the ropes to be able to better utilize the amazing tool that is GPT-3, please go to this link to start immediately.
All you have to do is create an account for both. Once in, you’ll realize that ChatGPT’s dashboard is quite simple.
You simply input some text in the text entry area and start conversing away.
As for GPT-3, you will be able to try out the tool on OpenAI’s Playground section.
As you can see, the dashboard has quite a few number of buttons and dials compared to its younger off-shoot. But this is all for good reason.
You are able to choose between text generations models (for code generation, long-form, short-form content et cetera), control the generation’s creativity, when and where to stop, how long of a text to generate, set a threshold for generation quality and more.
In summary, ChatGPT is a more publicly available, fine-tuned, simpler text generation tool, whereas GPT-3 is the model that ChatGPT based off of. The aforementioned two are able to handle text similarly, but ChatGPT is much more specific in its generations (keyword: Chat). Whereas, GPT-3 is used for a wider variety, although it comes at a cost of complexity.
Try them out both and see for yourself!
Thank you for reading.