What’s the difference between ChatGPT and GPT-3?
Since the release of ChatGPT in late 2022, it has amassed an astonishing 100 million users worldwide. ChatGPT is a conversational AI that was developed by OpenAI. GPT, or Generative Pre-trained Transformer, is a language model that uses machine learning techniques to generate text. ChatGPT is able to answer almost any questions in a fluent manner without grammatical mistakes.
You may have heard of GPT-3 as well. ChatGPT and GPT-3 are almost the same thing, with some differences. We will look at the differences in this article.
1. ChatGPT is a fine-tuned model of GPT-3
GPT-3 is the third version of OpenAI's language model based on a type of neutral network architecture called a transformer. GPT-3 has 4 specific models or implementations, each with different capabilities. "text-davinci-003" is the smartest, most capable GPT-3 model. It's able to understand instructions written in natural text and give you an answer to your query.
GPT-3 models can be fine-tuned. This means you can feed your custom data to train a GPT-3 model. Remember that the GPT-3 model has already been trained on a large amount of data. This means you are not training a language model from scratch. Rather than that, you are using your own examples of questions-and-answers to tell a pre-trained GPT-3 model how to answer certain questions. The result is higher quality responses.
For example, let's say you want to generate dialog for characters in a children's book. You want the dialogs to follow a style that children would like. You can fine-tune a GPT-3 model with example dialogs that you prepared. Then, the result is a fine-tuned model that can create dialogs that are similar in style to the examples you provided.
ChatGPT is a fine-tuned model of GPT-3. It is the Davinci model (the most capable model in GPT-3) that has been fine-tuned for conversational responses. ChatGPT, as its name suggests, is optimized for having a conversation with a user, while GPT-3 is a general AI model that can perform a wide range of tasks.
2. ChatGPT's responses are longer, softer, and more friendly to users
While both GPT-3 and ChatGPT are capable of answering questions to an amazing degree of accuracy, based on our experience, ChatGPT gives more comprehensive answers to a user's question. It also says the answer in "softer" ways that are slightly more friendly to read.
For example, I asked the question: "What is the harm of binge-drinking?" to GPT-3 and ChatGPT. GPT-3 answered:
Binge drinking can lead to numerous short-term and long-term health risks. These include an increased risk of injuries, car accidents, alcohol poisoning, high blood pressure, stroke, heart and liver disease, and even death. Additionally, binge drinking increases the risk of developing mental health issues, such as depression and anxiety. In addition, it can lead to poor decision making and can put people in dangerous situations. It can also lead to social issues, such as strained relationships, academic failure, and legal problems.
Binge drinking, which is the consumption of a large amount of alcohol in a short period of time, can cause a range of harmful effects on the body and mind, including:
- Increased risk of alcohol poisoning, which can lead to seizures, coma, and even death.
- Risk of injury due to impaired judgement and coordination, such as falls, accidents, and alcohol-related violence.
- Damage to the liver, brain, and other organs, which can lead to long-term health problems, including liver disease, neurological disorders, and cancer.
- Negative impact on mental health, such as increased risk of depression, anxiety, and suicidal behavior.
- Higher likelihood of engaging in risky behaviors, such as unprotected sex or drug use.
- Increased risk of addiction and dependence on alcohol.
In short, binge drinking can have serious and long-lasting consequences for both physical and mental health, as well as for personal and social well-being.
You can append more instruction to GPT-3 to customize the response, such as "Tell me in an easy-to-read language." But, without further instructions, ChatGPT creates more friendly responses.
3. GPT-3 has an API access while ChatGPT does not (yet)
[Update: March 15, 2023- Now ChatGPT has an official API.]
API, or Application Programming Interface, is what allows an application to communicate with another, third-party application. For example, GPT-3 API allows your web server to ask a question to GPT-3 and get the answer from it.
So, if you want to build a GPT-powered application such as a chatbot, you can create a script that takes your user's input, send it to GPT-3 through its API, get the generated response from GPT-3, and display it to the user.
API is necessary to integrate GPT into an app. Currently, GPT-3 has an API that a developer can use to build an app, but ChatGPT doesn't. In order to use ChatGPT, you have to go to ChatGPT's website to use it.
The good news is that ChatGPT now has an API Waitlist. OpenAI has not given us a date when the API for ChatGPT will be launched, but you can join the waitlist and be notified when it becomes available.
4. GPT-3 can be fine-tuned but ChatGPT cannot be
Because ChatGPT is a fine-tuned model of GPT-3, you cannot fine-tune ChatGPT to suit your specific purpose. On the other hand, GPT-3 models such as the text-davinci-003 model can be fine-tuned using your custom-prepared example sets, as stated previously. This can potentially be useful feature for your business if you cannot generate desirable responses using GPT-3 or ChatGPT.
Note that fine-tuning task will cost tokens (which is charged to the business by OpenAI), and using your fine-tuned model consumes slightly more tokens than using a GPT-3 model that's not fine-tuned.
5. The pricing model is different
The pricing model is different between GPT-3 and ChatGPT. ChatGPT is currently considered a "research preview" and is free to use on the Open AI's website. Open AI recently released the ChatGPT Plus tier, which costs $20 USD per month. ChatGPT Plus offers faster speed and users can use it even when the demand is high.
GPT-3 pricing model is by tokens. Every question or instruction you give to the GPT-3 (called "prompt") and its response will consume a certain number of tokens. The most capable Davinci model costs $0.0200 USD per 1000 tokens. A token corresponds to approximately 4 characters in English. The question "What is the harm of binge-drinking?" and its answer as mentioned above used 116 tokens ($0.00232 USD). You can currently use GPT-3 on Open AI's Playground page or via API.
ChatGPT is a fine-tuned model of GPT-3's more recent models such as the text-davinci-3. It is fine-tuned to provide users with more comprehensive responses that are slightly more user-friendly than GPT-3. While ChatGPT doesn't have an API access (yet), GPT-3 has, so it can be integrated into a custom-built application. GPT-3 can be fine-tuned, so if you have a specific business need that cannot be satisfied by ChatGPT or GPT-3, it may be possible to fine-tune a GPT-3 model using your own example sets.