What Are ChatGPT Plugins?

According to OpenAI, ChatGPT plugins are specialized tools designed for language models, emphasizing safety while enabling ChatGPT to access real-time information, perform computations, and interact with third-party services.

These plugins are crucial because the basic version of ChatGPT is constrained by its training data, which only extends to 2021. Additionally, the model struggles with certain tasks, such as solving complex mathematical problems.

As of April 21, twelve plugins have been launched, many of which integrate with services like:

• Wolfram
• Zapier
• Expedia
• Instacart

Think of ChatGPT plugins as similar to apps in an App Store, with ChatGPT representing the device itself.

Key Takeaways from Greg Brockman’s TED Talk

Here are four main insights from his presentation:

1. Automating Workflows with ChatGPT

   The talk began with a demonstration of how to streamline workflows that involve multiple tools. The showcased workflow included:

   • ChatGPT suggesting a meal post-TED
   • DALL-E creating a visual representation
   • A retrieval plugin for saving the conversation
   • Using Instacart to order ingredients
   • Posting the meal idea on Twitter via Zapier

Here’s the image generated by DALL-E:

2. Fact-Checking with Web Search

   Large language models, including ChatGPT, occasionally generate inaccurate information or make logical errors. ChatGPT plugins can address this issue!

By using the Browsing plugin along with a Chain of Thought technique, ChatGPT can verify its responses against information available online. For instance, during the demo, Brockman asked ChatGPT to check the date difference between two posts: "Unsupervised Sentiment Neuron" and "Learning from Human Preferences."

3. Enhancing Data Analysis with Code Interpreter

   The third key point was particularly fascinating. Brockman demonstrated how the Code Interpreter plugin can automate and enhance the data analysis process.

One of its standout features is the ability to parse CSV files, making it a valuable tool for data scientists looking to expedite their analyses. The workflow involved:

• Parsing and understanding column data in a CSV file
• Proposing various charts for data visualization, such as histograms and trendlines
• Adjusting the charts upon request

ChatGPT provides both the results of the code and the execution outputs, facilitating easy debugging and modifications. Here’s the word cloud generated during the live demonstration:

4. Formulating Medical Hypotheses

   The final highlight from Brockman's talk was a remarkable story about a dog whose life was saved thanks to ChatGPT's medical insights.

GPT-4 proposed a potential medical issue for the dog, which was initially misdiagnosed by one veterinarian. A second vet later confirmed GPT-4's suggestion, validating the AI's capabilities.