Overview of ChatGPT: AI’s new revolutionary conversational tool by OpenAI

ChatGPT is a conversational chatbot which is developed by OpenAI and released on 30 November 2022. Within a week of its launch, ChatGPT had approximately 1 million users, setting a new record for any AI model to attain this number.

What exactly is ChatGPT?

ChatGPT belongs to the InstructGPT model’s community(Large Language model), InstructGPT models are generally fine-tuned on GPT-3(famous transformer model) using reinforcement learning with the data collected from human feedback. Despite the fact that InstructGPT still makes simple errors, they are constantly improving with human feedback.

Most people consider ChatGPT to be the Google and Stack overflow killer because of its advanced features such as text summarization, suggestion generation, creating content, code reading, code snippet generation and debugging, explaining and tutoring and many more.

Why is ChatGPT gaining so much popularity?

ChatGPT is a fine-tuned GPT-3 model with 1.3 Billion parameters whereas GPT-3 uses 175B parameters as stated by OpenAI.

ChatGPT can act as your personal virtual assistant which is free of cost. Additionally, it can also read and comprehend your code and create customized code that fully explains it.

Methods used to train ChatGPT

OpenAI researchers started by training the pre-trained GPT-3 model using supervised learning techniques with an enormous amount of data. Numerous samples of the human prompts are collected from the conversations between the user and AI assistant so that human AI trainers are asked to write down the correct output for fine-tuning the model. They fine-tuned the model with 12,725 labeled data.

For more information, refer to this article by OpenAI.

They sampled human prompts and generated multiple model outputs for each prompt. The outputs are then ranked by a labeler. The resulting data is used to train a Reward model with 33,207 prompts and 10 times more training samples using different combinations of the ranked outputs.

More human prompts are then sampled and used to fine-tune the supervised model using the Proximal Policy Optimization algorithm (PPO), a Reinforcement Learning algorithm from Human Feedback (RLHF). The prompt is fed into the PPO model, the Reward model generates a reward value, and the PPO model is iteratively fine-tuned using 31,144 prompts data.

More about PPO here

The following figure represents all the steps used to train ChatGPT model.

Pro’s and Con’s of ChatGPT

Pro’s:

1. It is faster in responding to your questions

2. It has the ability to generate code snippets(it doesn’t have the ability to write an entire code.)

(P.S: When it was asked how to perform transfer learning? it generated the following output)

3. It can answer your inquiries in the most appropriate way possible without requiring you to spend a lot of time searching online.

4. It can be used to generate SQL queries, automate data preprocessing and clean scripts, debug the code snippets and explain where and why it caused errors, and many other data science oriented tasks.

5. It can be used for AR scene construction with some support from additional tools like WebAR.

Con’s:

1. It’s still a prototype and has some bugs to fix.

2. Considering the fact that it is a pre-trained model which can only provide the information based on what it has been trained on, it fails to fetch the latest information from the internet.

3. It is not able to see or manipulate visual information as it is a text-based AI assistant. So it cannot create images.

   
   

Limitations of ChatGPT

1. As stated by OpenAI, ChatGPT still has its limitations, where it can write plausible-sounding but incorrect or nonsensical answers.

2. ChatGPT is responsive to changes in input phrasing or repeated attempts at the same prompt. For instance, the model might claim to not know the answer if the question is phrased one way, but with a simple rephrasing of the question, it might be able to respond accurately.

3. The model frequently employs unnecessary words and phrases, such as repeating that it is a language model developed by OpenAI (These issues arise from biases in the training data and well-known over-optimization issues.)

4. Although they’ve worked to make the model reject inappropriate requests, there are still moments when it’ll take negative instructions or behave inimically. Although they anticipate some false negatives and positives for the time being.

5. When the user provides an uncertain query, the model should ideally ask clarifying questions. Although their other existing models typically infer what the user intended instead.

*** You can now leverage ChatGPT with its Google extension. ***

Happy Learning!!!