Microsoft and UC Berkeley built Gorilla, an open-source AI that learns to use APIs across different domains without hallucinating. It can handle multi-step tasks like booking travel or finding restaurants, but it's not true AGI—it still needs human data and struggles with creativity and common sense.
Questions this video answers
August 4, 2023.
Microsoft and UC Berkeley.
An improved version of LLaMA 7B.
API Bench.
Released open-source on GitHub.
Chapters
What if an AI could autonomously browse the web, learn from new tools, and even doom scroll through Reddit? Sounds crazy, right? Well, it's not as far fetched as you might think. In fact, there's a new AI model that can do all of that and more. It's called Gorilla, and it might be the closest thing we have to an AGI or Artificial General Intelligence. So Gorilla is a large language model that can provide the appropriate API calls for any task or question you give it. API stands for Application Programming Interface, and it's basically a way for different software applications to communicate with each other. For example, if you want to book a flight online, you might use an API that connects you to different airlines and travel websites. Gorilla is not just any LLM though. It's a super powered LLM that can interact with a massive number of APIs from different domains and platforms. It can use APIs from machine learning hubs like Torch Hub, TensorFlow Hub and Hugging Face, as well as APIs from cloud services like AWS, GCP and Kubernetes. And it can do all of this without being pre programmed to do so. It can learn how to use new tools and adapt to changes in real time. It is actually a joint project between Microsoft and UC Berkeley and it was recently released as an open source project on GitHub. So how does Gorilla compare to other LLMs like GPT4 or ChatGPT? Well, there are some key differences that make this AI model stand out. First of all, Gorilla is more accurate and reliable than other LLMs when it comes to generating API calls. It can produce the correct syntax arguments and outputs for any API call you ask for, and it also reduces the hallucination errors that other LLMs often make, which means it doesn't make up things that are not true or relevant. Secondly, it is more flexible and adaptable than other LLMs when it comes to using tools. It can handle changes in documentation, updates and versions of APIs without breaking down or losing functionality. Plus, it can learn from new sources of information on the fly, such as web pages or documents. Lastly, it is more powerful and versatile than other LLMs when it comes to performing tasks because it can handle complex tasks that require multiple API calls or multiple steps of reasoning, and tasks that span across different domains or platforms. So essentially it can do pretty much anything you can think of that involves using tools on the Internet. Here are some use cases for travel if you tell Gorilla to book a flight from New York to Paris for the cheapest date in August, it will make the API call to connect you with travel sites like Expedia or Kayak for food. If you want a pizza with pepperoni and mushrooms from Domino's, tell Gorilla and it will make the API call to link you with food delivery services such as uber eats or DoorDash for shopping. If you're looking for wireless headphones on Amazon that have noise cancellation and good reviews, tell Gorilla it will make the API call to connect you to the shopping platform. And these are only basic things Gorilla can help with, but it can do much more than that, and you can test it out by using their Colab notebook or by installing their CLI tool. Now let's see how it actually functions. So it is built on an improved version of llama 7b, and it fine tunes itself with a dataset named API Bench. This set has thousands of special commands known as API calls from various machine learning platforms like torch hub, TensorFlow hub, and hugging Face. These commands help Gorilla understand and work with different tools to produce the right results. Sometimes it needs more info, so it has a system that fetches documents from the Internet or other places. For instance, if Gorilla comes across a tool it hasn't seen before or a new version of a command, it will search for guides or examples online. After finding them, it updates what it knows and produces the right command. Essentially, it has three key the Llama system, the API Bench data, and the document Fetching system. If you're a developer or a researcher who wants to use Gorilla for your own projects, you'll be happy to know that it's very easy to get started. You can also use Gorilla's Spotlight search feature, which is a web based interface that allows you to search for any task or question and get the corresponding API call. Now the big question is, is Gorilla an AGI, which stands for Artificial General Intelligence and it's the holy grail of AI research? Well, to be honest, I don't think it's an AGI yet. It still has some limitations and challenges that prevent it from being truly general and intelligent. For example, it still relies on human generated data and documentation to learn how to use tools. It also still needs human guidance and feedback to improve its performance and reliability. And it still can't handle tasks that require creativity, emotion, or common sense. However, it is definitely a step closer to AGI than any other LLM out there. It has shown an impressive ability to learn from new sources of information, adapt to changes in tools, and perform complex tasks across different domains and platforms. It has also shown a remarkable reduction in hallucination errors, which is one of the major obstacles for LLMs to achieve AGI. Basically, gorilla is like teaching an A how to fish instead of giving it fish. It gives the AI the ability to use tools to solve problems instead of hard coding solutions for specific problems. And by doing so, it opens up a whole new world of possibilities for what AI can do. So what do you think? Is Gorilla the dawn of AGI, or is it just another hype? Let me know in the comments below, and if you enjoyed this video, please hit the like button and subscribe to my channel for more AI content. Thanks for watching and I'll see you in the next one.
