If you follow AI news, you’ve probably noticed two stories running side by side. In one, AI is so powerful that it may wipe out humanity within a few years. In the other, AI is wildly overhyped and the money pouring into it is a bubble about to burst. Even Sam Altman, the CEO of OpenAI, has said investors are “overexcited” about AI.

Both stories can’t be fully true at once. Either AI is about to become all-powerful, or it’s a lot less impressive than the valuations suggest. I work with tech companies on content and strategy, and I kept hearing smart people repeat one headline or the other. So I went looking for a more grounded view.

One of the most level-headed takes I came across was from Denis Stebunov, CEO of the software development company ivelum, who recently recorded his thoughts on these headlines. His view, which I’ll lean on throughout this piece, is that the doomsday scenario and the bubble panic both miss what’s actually going on with AI, and what we should expect from it over the next couple of years.

Why doom sells

It’s natural to feel uneasy about AI. It’s new, it moves fast, and few people fully understand how it works. Anxiety about the unknown is a very human reaction. But Stebunov argues there’s more to the doomsday narrative than ordinary nerves.

The first factor is old and well documented: bad news sells. Open any news site and the headlines are about wars, hurricanes, crashes, and crimes. Plenty of good things happen too; they just don’t get the clicks. “Bad news drive more engagement,” as Stebunov puts it. “People are more likely to click on something that sparks anxiety.” A study published in Nature Human Behaviour analyzed about 105,000 headline variations tested on the news site Upworthy and found that each additional negative word in a headline raised the click-through rate by about 2.3%.

The second factor is less obvious. The warnings about AI’s dangers don’t only come from random people online. They come from the leading AI companies, which talk to the public and to governments about risks and regulation. In Stebunov’s view, raising these concerns has become part of the PR strategy of the two largest AI labs, OpenAI and Anthropic. A product that might be dangerous is, by implication, a very powerful product. “It’s a bold move,” he says, “and it’s clearly working.”

The third factor is cultural. Most of us grew up on The Terminator and The Matrix, which paint a vivid picture of machines taking over. More hopeful visions exist, Star Trek among them, but they don’t stick the same way. So when someone mentions very advanced AI, most people picture Skynet, a story they already know.

The real risks today

None of this means AI is harmless. Stebunov has concerns of his own, and the most concrete is cybersecurity.

The threat is no longer hypothetical. In November 2025, Anthropic reported that it had disrupted a cyber-espionage campaign by a state-sponsored group that used its AI model to attack roughly 30 organizations. The AI carried out 80 to 90% of the tactical work on its own, according to the report, with humans stepping in mainly to approve the big decisions.

But the same technology works on defense. In July 2025, Google said its AI agent Big Sleep had found a critical vulnerability in SQLite, a widely used open-source database, that attackers were preparing to exploit. Google called it the first time an AI agent had directly foiled an exploitation attempt in the wild.

“It’s more like a cat-and-mouse game,” Stebunov says, “which probably will be accelerating.” Attackers get faster, defenders get faster, and organizations that don’t keep up will pay for it. That’s a serious problem, and a very different one from the end of humanity.

The extrapolation trap

In fairness, the doomsayers are mostly not talking about today’s AI. They mean future systems that will be far more capable, and their argument goes: look how much progress AI made in the past few years, and imagine where it will be in a few more.

“This is a classic extrapolation,” Stebunov says. You take the past trend and assume it will continue at the same pace, or faster. That’s a reasonable starting point, but only if you understand where the past progress came from.

What actually happened since 2018

The technology behind today’s chatbots goes back to 2017, when researchers at Google introduced the transformer architecture. In 2018, OpenAI built on it with GPT, the generative pre-trained transformer. That was the scientific breakthrough.

What followed, in Stebunov’s words, was “not a series of scientific breakthroughs, but rather scaling of the same idea.” The industry discovered that the bigger the model and the more it’s trained, the better it performs. That’s where most of the money went: bigger models, more compute, more data centers.

There were also two important advances in how the models are used. First, they learned to use tools. Second came agents, which run the model in a loop with the right tools and context so it can work through a task step by step. “That was big,” Stebunov says. “Agents can now perform real tasks, and this has already changed how people work.”

Still, the research itself produced no new fundamental breakthrough. Ilya Sutskever, one of OpenAI’s co-founders, has made the same point. He has described 2012 to 2020 as the age of research and 2020 to 2025 as the age of scaling, and argued that making models a hundred times bigger won’t transform everything. In his view, the field is heading back to research.

What to expect in the next two years

If most recent progress came from scaling, the question is how long scaling can continue. Not forever, Stebunov says, though nobody knows when it will slow down. “Maybe it will plateau in the next year, or maybe not.”

A genuine new breakthrough would change the picture, and he isn’t aware of one on the horizon. Without it, his forecast is refreshingly unexciting: “We’ll probably have more or less the same, just better, and with more practical applications.”

In practice, that means steady improvement rather than leaps: models that keep getting better at reasoning, coding, and following instructions. The biggest gains will come from putting the existing models to work in narrow, well-defined areas where agents can handle real tasks. And much of the change will come from more companies figuring out how to use what already exists, rather than from smarter models arriving.

That’s a long way from superintelligence by 2028, and a long way from a technology about to collapse.

What if the bubble bursts?

The bubble worry isn’t made up. In August 2025, Sam Altman told reporters: “Are we in a phase where investors as a whole are overexcited about AI? My opinion is yes.”

The numbers explain the nerves. According to Futurum Group, the five largest US cloud and AI infrastructure companies planned to spend roughly $660 to $690 billion on capital expenditure in 2026, nearly double the year before. OpenAI ended 2025 with about $20 billion in annualized revenue, and Futurum estimated th