Zahradník: AI still fascinates and terrifies me. In a year, we overhauled the way we work

Petr Zahradník, CEO Qminers, was a guest on Money Maker, the podcast from CzechCrunch magazine hosted by Ondřej Holzman and Luboš Kreč, which regularly runs AI specials. Over ninety minutes they covered almost everything: a private chat that helps him with a family member's medical diagnosis, the risks AI poses to democracy, and why Europe needs its own top-tier model. He also explained how AI radically changed day to day operations at Qminers over the past year, and why so much of his energy now goes into a nonprofit that supports math education.

Petr Zahradník describes his relationship with AI as almost paradoxical: it's his livelihood, and yet it still catches him off guard every day. "It's my daily bread, but at the same time it's something that, in a way, shocks me," he says right at the start. AI has, after all, transformed his own work at Qminers too. After years spent mostly on management decisions, he found himself, as CEO, back to coding directly in a development environment, something that wouldn't have made sense even a year ago:

"Suddenly I'm back to opening an IDE and building things myself, because individual empowerment is on a completely different level now, and one person can achieve so much more. When you extrapolate that to what happens to the world, it's simply breathtaking."

An Artificial Civilization of Fifty Million

When it comes to risks, Zahradník waves off the most commonly cited doomsday scenario, AI turning against humans on its own. "For now, we still have the power to pull the plug and simply switch the data centers off." He's far more interested in what happens once advanced AI becomes a geopolitical weapon. Here he quotes Anthropic CEO Dario Amodei directly, who pictures advanced models as a brilliant nation of fifty million people that nobody quite knows what to do with:

"That's what Amodei calls it: a country of fifty million people that's brutally smart, in some sense capable of anything, and learning far faster than you are. What do you do with that? If Americans realized a state like that really existed somewhere, there might be an immediate invasion. Either way, they'd feel compelled to do something about it."

The moment governments realize they have something like that within their own borders, Zahradník doesn't expect a joint agreement, but rather a race. "We see the risks and we know it would be right to slow down, except there's no way to agree that everyone slows down together." Nobody wants to be the one who stops first.

But he's even more focused on everyday life than on geopolitics. He worries AI could become a tool for tearing society apart. He sees the risk in apps that simulate a virtual friendship or relationship and pull people away from real, lived ones, while nobody's checking whose interests that "friend" is actually serving.

"I see a fairly big risk that we could lose democracy. Two years ago I would have thought that was in some sense out of the question, today I no longer rule it out. Someone could quite easily start shifting opinions with a specific goal in mind, say, taking apart one particular civilization, the Western one, for instance."

He follows that up with an image that sounds straight out of a dystopia, except he thinks it's already happening: someone stays shut in at home with their virtual partner, missing nothing, because that virtual partner tells them exactly what they want to hear. The need to go out, talk to people, build relationships, even learn how to disagree, simply disappears. Then again, it isn't all dark scenarios, he adds with a touch of self-irony: most people today use AI for perfectly innocent things, "they open an app and type into the chat how to make dill sauce." People trust AI completely. And that trust is exactly what makes AI such an easy tool for quiet manipulation.

Then there's the economy. Petr warns that AI risks breaking the long-standing balance between labor and capital, since more and more money flows into data centers and less and less to people, and a universal basic income as a safety net will only work as well as states manage to actually tax the AI giants. His example is bluntly simple: "Google pays almost no tax in the Czech Republic, even though it makes enormous money off Czech people."

A Bit of a Ponzi Scheme

He sums up the economics of the big AI companies in a single word, or rather a single adjective: Ponzi-like (after the infamous Ponzi scheme).

"I think the frontier AI research labs are still a bit 'Ponzi-like.' It's incredibly expensive, and honestly it's not clear to me how the whole thing is supposed to get paid for. OpenAI lost more than twenty billion dollars last year alone, Anthropic is doing better, but the losses are still enormous."

The paradox, he says, is that this very cost structure makes it impossible for these companies to step off the treadmill: they're locked into investors and hyperscaler partners so tightly that stopping would mean shutting the whole business down. "That's the most interesting part. Nobody can actually stop it," he says.

Is he worried about an actual collapse? Not really. "I'm a big believer in markets. I think this kind of race is simply the only possible way forward, and every so often it means we have to go through a 'dark valley.' That's just how it is."

As a counterweight to the closed American and Chinese models, he points to open, so called open weights models, mostly out of China (Qwen or Kimi), which a company or an individual can download and run independently of any single central provider, much like today's internet infrastructure runs on Linux rather than one commercial product.

In this, he sees Europe as having overslept. "It makes no sense for Europe to depend on other people's AI breakthroughs, just as it makes no sense to depend on someone else's electricity or someone else's batteries." And just as he thinks it's logical for the Czech state to want to own a nuclear power plant, even though that's more expensive and less efficient than leaving energy purely to the market, it makes sense for Europe to have its own top tier model, expensive as it may be. Counting on America to always act as a reliable ally, in any case, doesn't strike him as much of a plan after the past year: "Someone far worse than Trump could come along. Even in Europe, forces much worse than we can imagine could rise to power."

The Toughest Ring in the World

Qminers, which builds a fully autonomous algorithmic trading system, has been trading on world markets since 2014, and Petr Zahradník describes his own business as a constant fight in an arena where the rules get tougher every year.

"With Qminers, we're in the toughest ring that exists anywhere in the world, and we can see that ring getting more efficient, harder to succeed in every single year. So you keep doing more, faster, better, but sometimes that still means, like last year, that you end up slightly worse off, because everyone around you improved even faster."

Last year, he says, was the first year the company didn't grow, in fact it slipped slightly, a combination of a weaker first half against the competition and a strengthening koruna, since Qminers' costs are in crowns while its earnings come in dollars. Even so, the company stayed solidly profitable, and this year is on track to be a record one, if not a dramatic leap.

What's changed the most over the past year, he says, is how the work itself gets done. The team used to spend most of its time figuring out the answer to a well framed question, today AI handles much of that for them, and effort has shifted toward framing the question in the first place:

"These days you're basically a 'master' of agents: you come up with a good question, let an agent plan out the work, review it to make sure it's not off track, then say, fine, go do your job, and move on to the next question."

That sounds like a clean win, but Zahradník readily admits a side effect that surprised even him: it's unexpectedly exhausting, and it makes training new people much harder. If a junior only ever learns to check what an agent produced and never tries writing the code themselves, they may never learn to ask the right questions, the one skill Zahradník believes stays essential in the AI era. That's why Qminers still deliberately leaves juniors with more manual work than seniors, even at the cost of being slower.

In a market where deeper pocketed rivals can pour tens of millions of dollars into GPUs and servers, Qminers also has to pick its battles more shrewdly than before. "We have to be smarter about where our niche is," he sums up, adding that over the past year the company has dropped some directions outright, so it could put more weight behind the ones where it has a real shot at winning.

A Safe Bet

The last major topic of the conversation is Matika Česku, a nonprofit Zahradník co-founded and talks about with the same seriousness as his businesses. It rests on a simple idea: in an increasingly automated world, the ability to think in a structured way will hold its value, and math is the most fundamental training ground for that ability. The problem it's trying to fix, though, is uncomfortably concrete:

"We're losing teachers at an alarming rate. We're heading straight for a cliff, and it's an easy one to model. I think that within five years, half of all children won't have a math class taught by someone who actually studied math or how to teach it."

Education is going to have a hard time in the AI era, in more ways than one. That's why his advice to today's teenagers isn't to pin down a specific career in advance, since nobody knows what will even still exist in three years, but rather to build on two things he believes will stay valuable no matter which way AI develops: the ability to think critically, and the ability to maintain real relationships and care for other people.

"Cultivating structured thinking is a safe bet."

As for the role of the teacher, Zahradník believes it won't become obsolete, just transformed, from someone who transmits knowledge into someone who motivates. Here a family anecdote works as a pretty good litmus test for two different approaches to education: his father likes to joke that classrooms will one day just need AI assistants and someone with a whip, meaning supervision and discipline. Petr, though, thinks only the other path, the inspiring one, will actually work. With a sigh, he adds that as a society we may first have to go through the worse version before we get there.