Are We Really Ready for the Next Step in Autonomous AI Agents?

I have always enthusiastically supported the rapid evolution of artificial intelligence. Ever since I started Stanford's first courses on machine learning and neural networks in 2016, I realized we were entering a new revolutionary era, much like what had happened 20 years earlier with the advent of the Internet. I cannot forget those fantastic years of transition from adolescence to maturity, when between 1990 and 1995, the world was about to change completely thanks to the web.

I experienced incredible and exciting moments when, amidst the crackling sounds of the first 4,800 bps modems, I tried to connect for the first time to BBS and, later, to the Web through early browsers—just to slowly see a few documents and low-resolution images appear online. Even then, I clearly imagined the positive potential of that revolution, which is why, unlike many of my peers at the time, I was among the first to dive into discovering that new world, taking my first steps into the era of the First Digital Revolution brought about by the introduction of the first personal computers and game consoles of the 1980s.

In the last two years, however, we have witnessed incredible progress in AI. With the release of ChatGPT in November 2022, AI has transformed how we interact with technology and, by extension, how we live our reality. Similarly to what happened back then, after extensively experimenting firsthand, reading, and attending events on the subject, I have no doubt that we are facing another significant shift that, as before, few people have truly understood ??.

Despite my unwavering positivity toward technology—of course, when adopted in the right way—I find myself reflecting with a certain unease on the perhaps overly simplistic introduction of AI agents capable of acting autonomously from our personal computers. For instance, at the end of October, Anthropic introduced a new feature called "computer use," and not long after, OpenAI announced plans to launch "Operator," a research preview and developer tool set for release in January 2025. Additionally, Google recently revealed the development of "Project Jarvis," an AI system designed to assist users in executing daily tasks through its Chrome browser.

Imagine software that can act entirely on our behalf, performing tasks without our direct supervision. Sounds fascinating, right? Yet, I can’t help but think about the potential risks if all this unfolds too simplistically, as it seems to be happening. We are, in fact, handing these agents unprecedented access to our digital lives. For someone like me—both a passionate enthusiast and a professional in the field—if I had to use a metaphor, it's like shifting “from read-only permissions to write permissions.” And we all know how delicate and potentially risky this transition can be in computing.

My concern isn’t primarily about the controlled versions officially distributed by big players, but rather about the potential for malicious alterations. What would happen if these agents were manipulated through viruses or malware? They could steal our personal data, access our bank accounts, sabotage our work or that of our companies, or even publish private or false information on social media, with devastating consequences for our lives.

Using another metaphor, this time a classic one, it feels as if we are opening Pandora's box without being truly prepared to handle the consequences. I firmly believe we should proceed with caution. Introducing these agents without doing so in truly controlled environments—such as limited sandboxes—or without implementing stricter controls, like requiring a series of user confirmations for more operational actions, could potentially result in enormous damage.

Innovation is essential, and if we analyze history without "cognitive biases," it is undeniable that our evolution has been strongly supported by it. However, it should never completely endanger the safety of individuals and society. It is also normal for side effects to occur, as innovation remains the product of human creativity, inherently "subject to and shaped by" its use.

Nonetheless, I believe that by temporarily filtering out some major unresolved issues, such as the effects of climate change and wars—often amplified by social and human dynamics such as greed and the pursuit of power—we can still say that, as of today, the "glass is still half full." In other words, the long-term benefits of technological innovation have so far outweighed the negative consequences on a global scale, even though they have not yet been equitably redistributed across the planet.

But can we say the same about the future? My concern this time isn’t just about the risks we’ve always faced due to the limits of human nature. It is also rooted in the fact that, over the past 50 years, we have accelerated innovation so much that, to draw a parallel with movement, we are now traveling at the "speed of a meteorite."

To clarify this concept further: if we compare the speed of innovation to that of transportation, we see that it took roughly 4,800 years from the use of the first wheel for transport (Mesopotamia, 3000 B.C.) to reach speeds of around 100 km/h with steam trains in the mid-1800s.

Then, in just over 100 years, we managed to decuple that speed, reaching around 1,000 km/h by 1950, thanks to the first supersonic aircraft (the Bell X-1 broke the sound barrier in 1947).

To decuple speeds further, it took even less time. By 1969—just 20 years later—we sent orbital rockets into space, traveling at over 11,000 km/h.

Finally, in 2018, NASA's Parker Solar Probe set a historical speed record, reaching about 586,000 km/h—roughly 50 times faster than 50 years earlier.

What I mean is that if, until the early 20th century, we were moving at relatively manageable speeds that allowed for corrections (with the notable exception of nuclear power, where we've been fortunate so far), today we are accelerating to speeds where any misstep could no longer be corrected in time, leading us drastically off course with catastrophic global consequences.

In analogy with a famous slogan from Pirelli’s advertising in the 1990s, "Power is nothing without control," even as a self-proclaimed “Techno-optimist,” I can’t help but say: "Today, technology can truly be devastating without control."

This is a crucial moment where we must balance our enthusiasm for technological progress with deep reflection on its ethical and practical implications.

So I Ask Greg Brockman Mustafa Suleyman Andrew Ng Dario Amodei What do you think? Are we truly ready for this leap? Like you, I believe it’s time to ask ourselves: how can we ensure that AI’s advancement becomes an ally and not a boomerang for humanity’s future?

Join the conversation: #ArtificialIntelligence #AutonomousAgents #EthicalAI #AIInnovation #DigitalRevolution #ResponsibleInnovation

Davide Ciliberto - November 16, 2024

P.S. For data enthusiasts like me :-)