Listen to this story
Notwithstanding the limitations of ChatGPT, AI intelligent agents that break the boundaries of training are being tested on the LLM model. An AI intelligent agent is designed to perform tasks autonomously and intelligently without human intervention. Using various techniques such as machine learning, NLP and computer vision to analyze, reason and make decisions, AI agents are programmed to perceive the environment and take actions to achieve a goal.
There is an increased application deployment of these agents on GPT-4 and GPT-3.5 platforms. Thanks to unprecedented autonomy, intelligent agents are also pushing the movement closer to AGI. With multiple applications, here are some of the AI agents that users have been experimenting with.
Former director of AI at Tesla, Andrej Karpathy from OpenAI, called AutoGPT the “next frontier of prompt engineering”. Considered a significant step in the field of AGI, AutoGPT has got more GitHub stars than Pytorch. AutoGPT is a tool to make LLM fully autonomous by allowing GPT-4 to prompt itself. It enables the model to run tasks automatically and helps with the optimisation of business processes, business ideas and test cases. However, users need to have a basic programming language to set up the same.
Complementing GPT-4, which doesn’t have access to real-time information, (except through plugins), AutoGPT can be used to scout the internet for information. It even has memory retrieval and can store files and summarize them. Though still under development, AutoGPT, an open source tool, is already being used by developers for code generation – it can not only develop, but debug and self-improve codes.
AgentGPT is a tool that helps train and build conversational AI models with ease. It allows you to take AutoGPT and use it in the browser.
You can run any task with AgentGPT by asking a query. It “thinks” and comes up with an execution plan via step-by-step processes and generates the desired results by scraping data from the internet. However, the model can only return text-based output.
A task management system running on Python scripts, BabyAGI uses OpenAI and Pinecone APIs to “create, prioritize, and execute tasks”. Working on the function of creating tasks based on the results of the previous task and a predefined objective, the Python script uses OpenAI’s NLP to design new tasks, and Pinecone to save and access task results for context. It is warned that running this script will result in high API usage and since OpenAI and Pinecone APIs are used, a proper setup of these APIs should be done.
Do Anything Machine (DAM)
DAM is an AI agent that achieves your to-do list, and helps track and prioritize your task list. Built on GPT-4 and still under training, DAM needs to be provided with the tasks to be executed. For this to seamlessly work, the user will have to give access to all applications required for running the task. For instance, in the example below, access is given to the user’s Notion and Subslack platforms to run the request successfully.
An agent to run the internet, HyperWrite AI enables a “personal assistant” for web browsing. You can ask the agent to accomplish tasks via a single command. In the sample below, the user prompts the assistant to order a pizza, which it accomplishes through steps set by the assistant itself.
And although some users are calling the application “sick”, strategy and product designer, Pete Trainor, has referred to the application as a “possible accessibility game-changer”. HyperWriteAI is not yet released to the public but has opened up registrations for early access.
As the name suggests, MemoryGPT is like “ChatGPT with memory”. Unlike ChatGPT, the application can remember past conversations which can be referenced in the future, and possibly have personalised conversations. It will not only remember conversations from a single session, but forever. The possible uses range from being a coach, friend, or simply an assistant that can run your business. Users have even tried experimenting with it by teaching it complex role-playing games.
Available on the ChatGPT-3.5-Turbo model, MemoryGPT is free and requires no API. Within four days of its launch, MemoryGPT received over 6k signups and 17k conversations.