It was 10 p.m., and the sun had long since set. I was exhausted after the flight and drifting into sleep. Google Maps — yet again! — had chosen not the highway but a narrow road through the fields. And narrow roads through the fields in northern France are nothing if not winding.
The rented Toyota "saw" every speed limit sign along the way but failed to distinguish between regular signs and those meant only for trucks or for rainy conditions. As a result, it started beeping like crazy every few minutes, convinced I was a reckless speeder.
I wouldn't be surprised if I found out it had reported my 'crimes' to the local police.
Booking had listed the villa’s address incorrectly, and we spent a solid half hour just trying to find the place. The app flatly refused to translate the owner's messages with instructions for locating the keys into English.
It was weird to feel like a caveman and a citizen of a high-tech world at once. I am afraid AI agents will make us all feel the same way.
Here are two facts about AI agents, bots, autonomous vehicles (AVs), and voice assistants:
- They are hard to build.
- Everyone keeps trying to capitalize on them.
But do we really need them? Do you need a map app that rarely gives you the best route? Do you need a voice assistant that hallucinates too often?
An accurate paper map is more useful than a map app that can’t tell a highway from a goat trail.
I’m as old as personal computers. Engineers have told us for all these years that humans are error-prone and costly while machines are infallible and cheap. Turns out, machines fail too — and they’re anything but cheap.
The critical 10%
In 2018, an Uber car in autonomous mode struck and killed a woman pushing her bicycle across a highway near Phoenix, Arizona.
When asked about this tragic accident, tech CEOs often frown in irritation and say, "It was just an isolated accident."
But this accident, along with many others (thank God, less fatal), shows that algorithms aren’t infallible. But if both people and machines make mistakes, why do we need pricey algorithms when we already have 8 billion people?
Tech geeks tell us that self-driving cars will make way fewer mistakes and save many lives. But it’s not that simple.
The quote from a study:
“It turns out that humans are remarkably safe drivers. NHTSA data show that Americans drove around 3.2 trillion miles in 2022. The 42,795 traffic deaths in 2022 translate into a fatality rate of about 1 in 100 million miles driven (1.35 to be exact, down slightly from 2021), and this includes the drunk, drowsy, and distracted drivers.”
As of today, no data indicates that AVs can be safer drivers than humans.
Humans make mistakes, algorithms glitch. Why do we need to replace one imperfect system with another flawed one? Apple, General Motors, and Ford have discontinued their self-driving car programs for a good reason.
Every single minute, 250 natural machines are born — perfectly tuned for life on this planet.
So why are we racing to replace them with artificial ones? Don’t we have more urgent problems for algorithms to solve?
Decades ago, technology used to make complex tasks simpler.
- The telegraph was faster than handwritten letters.
- The telephone was more convenient than the telegraph.
- The Internet had clear advantages over the telephone.
In the 21st century, engineers try to make algorithms think like humans. But the law of diminishing returns (or the complexity curve) got in the way.
Building an algorithm that can perform 70% as well as a human is simple and inexpensive. But as you push for greater accuracy, both complexity and cost rise so sharply that you start to question whether it’s even feasible — or economically worth it. The last 10% looks like science fiction.
Increasing investment in algorithmic complexity doesn’t lead to a proportional gain in effectiveness. Pour millions into development — and your algorithms might only get a fraction of a percent smarter.
It's like pedalling harder and harder on an exercise bike for hours to burn just a few calories.
Nike, Adidas, and Under Armour attempted a simpler problem – to automate sneaker manufacturing – but failed. They spent years trying to teach robots to handle the soft and stretchy materials – a skill that a human worker can master just in a few weeks.
Will Alexa or ChatGPT be able to become your HAL 9000 from "2001: A Space Odyssey"? That’s still questionable.
The idea of the asymptotic difficulty curve is explained well here, and if you’d like to dig deeper, take a look here, here, or here.
Tech entrepreneurs and enthusiasts are so eager to bring the future closer that you can’t help but feel sorry for how uncomfortable they must be in the present. Could it be that they want to populate our planet with robots because they’re not good with people?
Machine-made products in a human-made world
There are three more reasons why AI agents might not become the next shiny new thing.
1. Machines make their own mistakes
The problem isn’t that algorithms make mistakes. It’s that they do it differently. And that’s a serious flaw in a world designed around human error.
We’ve designed this world around our own imperfections. For instance, we’ve created many rules and laws, but we know that there are always exceptions. Machines take them literally.
Teaching algorithms how to follow the rules is simple but not easy. Teaching them how to break the rules when it makes more sense – that’s neither simple nor easy.
2. High productivity is a self-licking ice cream cone
Entrepreneurs promise us that AI agents and co-pilots will boost our productivity. But this feels like a solution in search of a problem.
Globally, only about 23% of employees are engaged at work, which means a striking 77% of the world's workforce is either disengaged or actively disengaged.
Of this, approximately 62% are "not engaged" (simply going through the motions), and 15-18% are "actively disengaged" (basically, they hate their jobs).
Start building a healthy culture in your company for free and get a 100% discount for co-pilots. You just won't need them.
**
Bosses: “We’re bringing in AI agents!”
Employees: “Great! We’ll get the boring tasks done faster and finally focus on the fun stuff!”
Bosses: “No, you’ll just get even more boring tasks.”
**
3. Progress is not about what engineers invent, it's about what people need
According to Forbes, “of today’s 2.1 million active patents, 95% fail to be licensed or commercialized.”
Of course, there are many reasons for that. One of them is that human needs drive technological progress, not innovation alone.
Engineers can develop new technology, products, or materials, but if people don't feel they need them, the inventions will go to waste.
Do people really need AI agents and co-pilots? Nobody knows so far.
Conclusion
If you're going to join this AI agent race, think twice. The speed of the game is accelerated by VCs, but they have their own strategies. Dive deeply into customer needs – and you’ll never lose.