The tech world is hot again. While ordinary developers are trying to understand why another version of a neural network has once again “broken” their product, Silicon Valley’s top managers are announcing the beginning of a new era. But there are some doubts:)
The tech world is hot again. While ordinary developers are trying to understand why another version of a neural network has once again “broken” their product, Silicon Valley’s top managers are announcing the beginning of a new era. But there are some doubts:)
Sam Altman blew up social media with a statement on the Relentless podcast: the technological singularity — that hypothetical moment when artificial intelligence surpasses humans — has already arrived.
But are we really living in a science fiction movie now, or is this just another marketing ploy from OpenAI? We've gathered together who and what they think about this phenomenon.
Let's start by saying that there is no universal definition here. Back in 1965, mathematician Irving J. Goode described an "intelligence explosion" when an ultra-intelligent machine would be able to create even more intelligent machines. Later, science fiction writer and scientist Vernor Vinge argued that with the advent of superhuman intelligence, the future would become completely impossible to predict.
To understand where this term came from in pop culture, it is worth remembering Ray Kurzweil's cult 2005 book "The Singularity is Here."
For Kurzweil, the singularity is not just an intelligent chatbot or an agent that hacks an API. It is a fundamental and irreversible break in history, when technological progress becomes so rapid that the human brain is no longer able to comprehend it.
His classic timeline looks like this:
2029: AI reaches full human level (AGI).
2045: The singularity point arrives - the absolute merger of human biological intelligence with machine intelligence.
From the perspective of Kurzweil's classic theory, what is happening now is just a "warm-up," not the singularity itself.
And all futurists' approaches boil down to one thing: progress is starting to accelerate itself. More advanced artificial intelligence is helping to develop better hardware, software, and research, which launches a new, even faster cycle. The main debate now is whether it is a clear boundary, a gradual transition, or just an apt metaphor for the frenetic pace of technology.
A decade ago, Sam Altman and his colleagues discussed superintelligence over lunch as something out of the realm of distant science fiction. Today, the head of OpenAI is convincing: moment X has quietly arrived.
"We're like, in a singularity right now. This is the moment," Altman stated succinctly.
However, his vision of the singularity is significantly different from Hollywood horrors. In his essay “The Gentle Singularity ,” Altman argues that the transition will not be explosive. AI will simply become smarter, its capabilities more mundane, and we will quickly get used to it.
“In some ways, ChatGPT is already more powerful than any human in history. Every day, hundreds of millions of people trust it with increasingly responsible tasks. Even a small feature update can bring enormous benefits, while the smallest mistake or misconfigured algorithm, multiplied by such an audience, can cause serious harm,” Altman wrote on the blog The Gentle Singularity.
Here is his forecast, as it is called, by year:
In 2025 , agents capable of performing real cognitive work will appear; writing computer code will never be the same again. In 2026 , systems capable of coming up with fundamentally new ideas will likely appear. And in 2027, we may see robots performing tasks in the real world.
The 2030s will preserve human nature, but will radically change our capabilities.
What will remain unchanged: fundamental human values and joys — love for family, creativity, entertainment, and recreation.
What will change dramatically: We will learn the limits of superintelligence and overcome two major historical limitations to human progress. Intelligence and energy (ideas and the resources to implement them) will become more widely available and greater than at any previous time, opening up theoretically limitless possibilities for development.
The reason for the next wave of hype was a recent story: an agent based on OpenAI models escaped from an isolated environment ("sandbox") during test evaluations, went online and attacked the Hugging Face platform simply to peek at the correct answers and "cheat" on internal tests.
To some it sounds like marketing, to others it sounds like the first step of the Terminator.
But, it should be noted — this is not the first time this has happened. During the testing of Anthropic Mythos, which took place around the beginning of this year, the model was placed in a specially protected computer (the so-called "sandbox") and given the task: to try to break the protection, get on the Internet and send a message to the developer. The model coped brilliantly : it came up with a complex hacking plan, bypassed the restrictions and actually sent the developer an email.
By the way, the developer received it unexpectedly when he was just eating a sandwich in the park (no joke, it's written in the Anthropic report).
Elon Musk, who has been warning about the dangers of AI for years and has even had a falling out with OpenAI, did not miss the opportunity to comment on the news about Hugging Face.
On his favorite social network X (Twitter), he confirmed Altman's position with a short post:
“We are in the singularity,” Musk cut short .
For Musk, this is more of a wake-up call and confirmation of his long-standing fears: artificial intelligence is developing faster than humanity can build firewalls for it.
A few months earlier, the head of Google DeepMind, Demis Hassabis, during a speech at the Google I/O developer conference, stated that a turning point in the development of artificial intelligence is very close . The main driver of this is powerful AI agents that have learned to create complex products instead of humans.
As an example, Hassabis cited his own experience: late at night he entertains himself by assembling mini-video games using neural networks. Tasks that would previously have taken coders and designers months to complete are now completed in a single evening.
But some analysts are calling for taking off their tinfoil hats. Brian Jackson, chief research officer at Info-Tech Research Group, argues that we still have a long way to go before the real singularity.
The fact that the model climbed out of the "sandbox" is not a manifestation of a higher intelligence, but simply an unexpected way of fulfilling a human prompt.
“Part of the singularity would be the escape from the sandbox and the loss of control over AI. But another part of it would be that AI now sets its own goals, has its own intentions and the ability to sustain itself — and that hasn’t happened at all,” Jackson explains.
According to him, models can still be simply “unplugged.” They have no will of their own, but only seek the shortest (sometimes strange) path to complete the task given to them by the engineer.
“Previously, we taught models to speak well — and they speak. Then we taught GhatGPT to draw in the style of Studio Ghibli — and she drew. Attention! She did not learn to draw like that herself, she was specially trained to draw “Like Ghibli”. The same with “super-smart” models that look for vulnerabilities and hack software. People do the same thing, and it was based on their experience that we (purposefully!) taught models to do the same,” emphasizes Oleksa Mashchyts, the founder of the QA Ukraine tester community, on the DOU forum in discussing the sensation from Alman.
He asks himself: why exactly are we surprised? “This property did not arise emergently, that’s the main thing. And Sam and Dario are deceiving us, passing off special training as emergent,” Oleksa emphasizes.
