Interacting with AI chatbots like ChatGPT can be fun and sometimes useful, but the next step in everyday AI goes beyond answering questions. This means that the AI agent does the work for you.
Major technology companies, including OpenAI, Microsoft, Google, and Salesforce, have recently announced or announced plans to develop and launch AI agents. They argue that these innovations will bring new efficiencies to the technical and management processes that underlie systems used in healthcare, robotics, gaming and other businesses.
A simple AI agent can be taught to respond to standard questions sent via email. The more advanced can book airline and hotel tickets for transcontinental business trips. Google recently demonstrated to reporters Project Mariner, a browser extension for Chrome that can infer text and images on the screen.
In the demonstration, agents helped people plan meals by adding items to a shopping cart on the grocery chain’s website and finding substitutes if certain ingredients were not available. Human involvement is still required to complete the purchase, but agents can be instructed to perform all necessary steps up to that point.
In a sense, you are an agent. You take action in the world every day in response to what you see, hear, and feel. So what exactly is an AI agent? As a computer scientist, I offer the following definition: AI agents are technological tools that can learn a lot about a given environment and then solve problems or perform specific tasks in that environment with a few simple prompts from humans.
Rules and Goals
A smart thermostat is an example of a very simple agent. Our ability to perceive our surroundings is limited to a thermometer that tells us the temperature. When the room temperature drops below a certain level, the smart thermostat reacts and turns up the heat.
A familiar predecessor to today’s AI agents is Roomba. For example, a robot vacuum cleaner learns the shape of a carpeted living room or the amount of dust accumulated on the carpet. Then take action based on that information. Your carpet will be clean in just a few minutes.
Smart thermostats are an example of what AI researchers call simple reflex agents. It makes decisions, but those decisions are simple and based on what the agent perceives at that moment. A robot vacuum cleaner is a goal-based agent with a single goal: to clean all accessible floors. The decisions the robot makes – when to turn, when to raise or lower the brushes, when to return to the charging base – are all aimed at achieving that goal.
Goal-based agents succeed simply by achieving their goals by any means necessary. Goals can be achieved in a variety of ways, some of which may be more or less desirable than others.
Many of today’s AI agents are utility-based. This means putting more consideration into how you will achieve your goals. Weigh the risks and benefits of each possible approach before deciding how to proceed. They can also consider competing goals and decide which one is more important to achieve. Go beyond goal-based agents by selecting actions that take into account the user’s unique preferences.
making decisions and taking action
When tech companies talk about AI agents, they aren’t talking about chatbots or large-scale language models like ChatGPT. Chatbots that provide basic customer service on websites are technically AI agents, but their perception and behavior are limited. A chatbot agent can recognize the words the user types, but the only action it can take is to reply with text that provides the user with an accurate or informative response.
The AI agents that AI companies refer to as having the ability to take actions on behalf of the people and companies that use them represent a significant advance over large-scale language models like ChatGPT.
According to OpenAI, agents will soon become tools that allow people or businesses to run independently for days or weeks without needing to check progress or results. Researchers at OpenAI and Google DeepMind say the agent is another step toward “strong” AI, i.e. artificial general intelligence, or AI that surpasses human capabilities in a variety of domains and tasks.
The AI systems people use today are considered narrow AI or “weak” AI. A system may be good at one domain, perhaps chess, but when thrown into a game of checkers, that same AI has no idea how it will behave because the skills are not translated. Artificial general intelligence systems can better transfer skills from one domain to another, even if they have never seen the new domain before.
Is it worth the risk?
Are AI agents ready to revolutionize the way humans work? This will depend on whether technology companies can prove that their agents not only perform the tasks assigned to them, but are also equipped to overcome new challenges and unexpected obstacles as they arise.
The use of AI agents depends on people’s willingness to give them access to potentially sensitive data. Depending on the tasks your agents will perform, they may need access to Internet browsers, email, calendars, and other apps or systems. It is relevant to the given task. As these tools become more common, people will need to consider how much of their data they want to share.
If an AI agent’s systems are breached, personal information about your life and finances could fall into the wrong hands. Are you willing to take this risk if it means the agent can save you some of your work?
What happens if the AI agent makes a wrong choice or a choice the user disagrees with? Currently, AI agent developers are keeping humans in the loop so that people have a chance to verify the agent’s actions before a final decision is made. In the Project Mariner example, Google does not allow agents to make final purchases or accept the site’s terms of service. The system will provide you with up-to-date information and the opportunity to deselect agents you have not approved.
Like any other AI system, AI agents also have biases. These biases can arise from the data the agent was initially trained on, the algorithm itself, or the way the agent’s output is used. Keeping humans informed is one way to reduce bias by allowing them to review decisions before acting on them.
The answers to these questions will likely determine the popularity of AI agents and how well AI companies can improve them once people start using them.
This article is republished from The Conversation under a Creative Commons license. Read the original article.
Image credit: Ant Rozetsky on Unsplash
