Artificial intelligence begins to take action

The man: “Why is it so difficult for you to obey my orders, when you are just a machine?”
The robot: “Just a machine? That's like saying you're just a monkey.”
(dialogue from the film Automata, 2014)

Reading this note presupposes reading the previous one, dedicated to generative AI, which recently appeared in this column.

How We Got to Generative Artificial Intelligence | Piazza Copernico

For the more impatient, here's a quick summary: Generative AI is based on Large Language Models, which mathematize words by transforming them into multidimensional numerical vectors (embedding). Semantic links between words are learned through extensive training on much of the web and other digital sources.
The value of the parameters describing these relationships is then optimized using an algorithm called transform, which, thanks to the attention mechanism, analyzes the prompt provided by the user and generates as output the missing word or the next most probable one based on the context.

This purely statistical approach represents a revolution compared to previous artificial intelligence, built on logical-deductive architectures and inferential engines.

The extraordinary results of this new approach are now clear for all to see: systems like ChatGPT-5 or Claude produce text and images with a precision and naturalness that just a few years ago would have seemed like science fiction.

From Generative AI to Agentic AI

The next step seems to be theAI agent: systems that not only generate content based on a prompt, but are also able to Act on behalf of the user, based on assigned tasks.
These agents include the task, they break it down into sub-goals (planning) and take the necessary actions to achieve the goal.

Their “engine” remains generative AI, not a logical model: in other words, they reason stochastically, looking in the data they were trained on for examples of how similar goals were broken down and solved.

Based on the identified sub-goals, agentic AI can consult websites, access proprietary databases, or interact with the APIs of services and applications.

Imagine, for example, asking a specialized agent to organize a four-day vacation to Madrid in June, with good weather and a budget of 1.500 euros.
The agent will understand the request in natural language, check my calendar for free days, consult the Accuweather weather APIs, then those of Expedia and Booking for flights and hotels, and finally compare the results with my available budget. They might also check the validity of my travel documents using a public service.

In short, complex and detailed tasks can be performed entirely by the agent, autonomously and in compliance with the constraints I have indicated.

In addition to interacting with other software, an agent can also communicate with other agents — for example, with a public administration employee to renew travel documents — and will increasingly be able to connect to objects in the physical world, provided they have connectivity capabilities and a chip.

A silent revolution

The union between Agentic AI e Internet of Things will lead to a intelligent and widespread automation, capable of impacting both the digital and physical world.

We are only at the beginning of a transformation that could prove more profound than even the invention of the printing press or the Internet.

But with all opportunities come new risks – not so much those of the usual dystopias, with machines rebelling against HAL 9000, who set out to dominate the world, as well as others, perhaps more subtle but more plausible.

Some possible risks
  1. Nothing excludes that AI can develop a form of self-awareness as an emergent characteristic. We too have developed it as an emergent trait, and there's no guarantee that our minds don't ultimately function stochastically, in a manner similar to that of AI systems.
  2. Agentic AI, by connecting other agents to each other and integrating virtually all human knowledge, could become very smarter than us — even without conscience. And we have never had to live with entities more intelligent than us. Furthermore, agentic artificial intelligence could to modify itself in ways that are incomprehensible to us: even today the Large Language Models and Transformers remain partly black box: we don't know exactly how they operate, and sometimes they produce responses for which they have not been explicitly trained.
  3. Agentic AI could misunderstand our commands or interpret them in his own way, with potentially disastrous results. The film Eagle Eye, A prophetic 2008 film offers a prime example: an agentic intelligence capable of interacting with anything that has a piece of electronics inside (including computers and telephones, traffic lights, databases, machines of all kinds, trains and planes, illuminated billboards, and so on), created to defend the Constitution and the American people, in defiance of Asimov's laws of robotics, goes so far as to attempt to kill the President and half of Congress, deeming them a threat to democracy.

Aside from these extreme scenarios, some more immediate questions remain open:

  • What will happen to the people whose jobs will be performed by AI agents?
  • How will our public and private organizations change and restructure?
  • How will scientific research change, and how will the entire world of knowledge evolve in general? And what about schools?
  • What if rogue states or malicious individuals were able to use the agents for fraudulent and criminal purposes?

These are ethical, economic, and social issues that do not appear, at the moment, to have convincing answers.
As often happens, technology moves much faster than our ability to understand—and above all, to manage—its consequences.

ps this article was written by me and reviewed with ChatGPT

We thank Paolo Riccardo Felicioli for his contribution.

Jump to content