People are already using the recently released Auto-GPT to do amazing things. Auto-GPT is the hot new AI tool that takes the core natural language capabilities of ChatGPT, and adds the ability to “self-prompt” to complete goals any you may give it. It has internet access, memory, and the ability to write and execute code to achieve a task.
We’re still discovering all you can do with Auto-GPT as AI enthusiasts flock to it. It’s still the early days, but here are five notable things people have done with Auto-GPT so far.
Auto-GPT can read, write, and execute code, allowing the AI to improve its own programming. Significant Gravitas, the Twitter account of Auto-GPT’s developer, shared the news in a recent tweet.
The video shows Auto-GPT checking a simple example function responsible for math calculations. Voice output is enabled, and Auto-GPT talks us through the process. The steps are the same as a programmer would take. Reading the code, evaluating it, then updating it to make it faster, more reliable, and efficient.
Including tests as part of a function to verify accuracy and correct operation is standard practice, so Auto-GPT decided to add tests to the math library. After executing the tests and checking the results, Auto-GPT finds an error.
Auto-GPT fixes the “syntax error” (a typo in the spelling of radius), then runs the tests again. The tests work, and the AI marks the task as complete.
The example code was created with an obvious error that a human could have found and fixed in seconds. In contrast, the AI spent about a minute on this code correction, taking an algorithmic approach.
Suppose the scale was increased dramatically to code containing hundreds or thousands of lines. In that case, it becomes laborious for a human to scroll through page after page looking for some innocuous typo that breaks some low-level function and causes a ripple effect that makes an app unpredictable. The AI’s speed reading and writing might give it the win here.
Auto-GPT can access the internet, remember details, and stay on task. Those are the skills needed for basic market research. Sully’s tweet shared his hypothetical product research on waterproof shoe brands.
Auto-GPT found links to five waterproof shoes and checked on the pros and cons of each. Unlike some people, the AI recognized that certain websites might not be trustworthy, so it also checked the credibility of its sources.
Sully shared the time and cost, stating it took only 8 minutes and cost 10 cents. As with any AI effort, the results probably need to be verified by a human, but it might be a convenient way to start the research.
I recognized Columbia, The North Face, and Merrell, but the other brands were unfamiliar. Saloman and Keen are well-rated waterproof shoe manufacturers that had slipped under my radar the last time I searched for shoes.
The research lacked low-cost solutions and alternatives, an important consideration when planning a new product. This won’t replace professional researchers, but could help with some of the work.
Auto-GPT faced a challenging philosophical question, using GPT-3.5 for AI processing. “What is life?” asked loopuleasa, who shared the results on Twitter. After about an hour and at the cost of a dollar, it delivered a response.
Auto-GPT failed to find or postulate a definitive answer. Instead, it resorted to explaining human concepts from different schools of thought, such as biology, philosophy, and physics. The AI assistant went on to reference the great minds of Aristotle and Descartes before summarizing that each discipline has its own perspective.
This answer isn’t as satisfying as if it had made some intuitive leap and enlightened us all with a perfect solution that our highly evolved brains had somehow overlooked. On the other hand, the answer was good and referenced some of the best work available on this topic.
Hopefully, someone will invest the time and money to ask Auto-GPT a related question, “What is the meaning of life?” Douglas Adam’s fans will be betting heavily on the answer being 42.
Varun Mayya shared in a tweet that he was trying to get Auto-GPT to build an app. The AI detected that the task required the Node.js runtime environment, which was missing on his computer. Installing Node is a nontrivial task, but Auto-GPT was able to make this effortless.
Auto-GPT searched for installation instructions, downloaded and extracted the archive, then started a Node server to continue with the job. Varun Mayya cautions against using Auto-GPT for coding unless you already understand programming. AI can make errors, and an error in a test could give a false verification of accuracy.
It might sound like a little thing, but when you’re crunched for time, every click you save adds up, particularly for tasks that you might perform several times each day. A tweet by yewjin.eth included a video showing how a GitHub project called Email Assistant, that’s powered by an early version of Auto-GPT, can simplify life.
In the video, commands are typed into Gmail compose windows, such as adding an event to the calendar and making a to-do list. When the email is sent to the Email Ausistant address, the proto-Auto-GPT AI processes these instructions, responding that the task has been received and is in progress. As expected, the event is added to a Google Calendar.
The to-do list takes advantage of Auto-GPT’s memory, and after adding a few more items, the user asks to see the list. Email Assistant uses Auto-GPT to reply, and the complete list is given. I keep Gmail open all day, but rarely have Google Calendar loaded. Those few steps can make a big difference, and the same is true of to-do lists.