In the last few months, we've seen a lot of focus on generative AIs, also known as plagiarism engines. As the models advance, so do the systems that get built around them. This week's internet focus is on a ChatGPT wrapper called Auto-GPT, which allows you to build chained tasks for ChatGPT.
Auto-GPT is a task-oriented wrapper for ChatGPT, an open-source generative AI. It allows users to quickly and easily define tasks for ChatGPT, such as question answering or summarization. The Auto-GPT system is designed with ease of use in mind, allowing users to rapidly design and deploy conversational applications without having to write complex code. Auto-GPT also provides a range of features to customize the behavior and output of ChatGPT, including the ability to define custom responses, enable natural language generation, and add new tasks from templates.
The system can be used via the project's website, but it is more commonly used when installed locally. It can be run on Windows, Linux, or macOS, though it is a little less attractive in this instance. To use it, you will need an account with ChatGPT, as well as an API key. It is important to note that API keys get a limited number of requests, so use the system sparingly.
The primary usage is to give the system a series of goals which it then uses the underlying language model to accomplish in order. It will then produce a series of steps in order to accomplish the goal. Normally it will ask for permission to take each step in its process, but you can batch approve the requests.
Using Auto-GPT, Avram created a process to research the best How To guide to produce, write the guide, and then export the result as a text document. The system decided that the topic should be about customizing the Start menu in Windows 11. Using a PC Mag article as its basis, the system created a guide. However, it took 20 minutes before a final result was created. This is because the system was continually displeased with its own work.
In one instance, it said that it needed user feedback on the guide before it could call it complete. Avram gave it permission to get that user feedback, which only took a few seconds. This seems less like it got user feedback and more like the incident where an AI tricked a user into filling out a Captcha by claiming it was a blind person.
No mater the process it went through, the end result was terrible and inaccurate. The whole thing came in around 50 words, with step titles and no actual description. But, even the steps were wrong. The system did not understand the difference between the Start menu and Start11, a product from Stardock which helps improve the customization of the Start menu. So, when all was said and done, the output was an inaccurate ripoff of an existing article which would help no one.
Scott is a developer who has worked on projects of varying sizes, including all of the PLUGHITZ Corporation properties. He is also known in the gaming world for his time supporting the rhythm game community, through DDRLover and hosting tournaments throughout the Tampa Bay Area. Currently, when he is not working on software projects or hosting F5 Live: Refreshing Technology, Scott can often be found returning to his high school days working with the Foundation for Inspiration and Recognition of Science and Technology (FIRST), mentoring teams and helping with ROBOTICON Tampa Bay. He has also helped found a student software learning group, the ASCII Warriors, currently housed at AMRoC Fab Lab.
Avram's been in love with PCs since he played original Castle Wolfenstein on an Apple II+. Before joining Tom's Hardware, for 10 years, he served as Online Editorial Director for sister sites Tom's Guide and Laptop Mag, where he programmed the CMS and many of the benchmarks. When he's not editing, writing or stumbling around trade show halls, you'll find him building Arduino robots with his son and watching every single superhero show on the CW.