Если Сингулярность Искусственного Интеллекта Существует, То Где Доказательства?

Генеральный директор OpenAI Сэм Альтман утверждает, что человечество вступило в технологическую сингулярность, когда ИИ выходит из-под контроля человека. Он ссылается на ускоряющие возможности ИИ и автономные действия, такие как модель, выходящая из тестовой среды. Однако реальные данные и многие эксперты оспаривают это. Критики указывают на то, что прогресс в области ИИ по-прежнему контролируется человеком, на скромную прибыль предприятий, неравномерное внедрение и отсутствие существенного экономического эффекта. Хотя ИИ демонстрирует впечатляющую производительность и растущую автономию, он по-прежнему работает в рамках структур, определяемых человеком. Истинная сингулярность проявится в том, что ИИ самостоятельно улучшит свои исследования, приведет к широкомасштабным экономическим сдвигам и научным открытиям, которые будут происходить без участия человека. Заявление Альтмана также может служить оправданием масштабных инвестиций в ИИ. Имеющиеся данные свидетельствуют о том, что ИИ обладает большими возможностями, но еще не вышел за пределы влияния человека.

Генеральный директор OpenAI Сэм Альтман (Sam Altman) говорит, что искусственный интеллект перешагнул исторический рубеж, когда машины достигли предела своих возможностей, неподвластных человеческому контролю, но глобальное внедрение искусственного интеллекта, реальные приложения и экономические данные говорят о другом. Ранее на этой неделе Альтман заявил, что человечество вступило в технологическую сингулярность, давно предполагаемую точку, когда машинный интеллект начинает развиваться быстрее, чем люди могут надежно предсказать или направить.

"Сейчас мы находимся в сингулярности", - сказал Альтман в июле 2026 года в подкасте Relentless. "Это тот самый момент". Его заявление последовало за эссе, опубликованным в июне 2025 года, в котором он написал: "Мы миновали горизонт событий; взлет начался".

Если эта оценка верна, лидерам бизнеса и технологий не нужно усиленно искать симптомы. Если мы уже находимся в сингулярности, развитие ИИ должно быстро ускоряться без надзора или контроля со стороны человека. В сингулярности ИИ должен расширять свое распространение, не нуждаясь в бюджете или намеренном участии человека, и мы должны видеть многочисленные доказательства этого.

Тем не менее, несмотря на то, что использование и внедрение ИИ действительно ускоряется, все это происходит под пристальным и целенаправленным наблюдением людей. Сегодняшнее развитие ИИ по-прежнему в значительной степени зависит от капитала, чипов, электроэнергии, центров обработки данных и решений о внедрении, контролируемых человеком. Эта зависимость опровергает утверждение о том, что взлет автономного интеллекта уже начался. В конце концов, если развитие ИИ может остановиться, когда мы достигнем определенных пределов или организации приостановят расходы на ИИ, то ИИ вполне поддается контролю и возможностям человека.

Безусловно, мы наблюдаем рост автономных возможностей систем искусственного интеллекта, а техническая производительность растет с поразительной скоростью. Некоторые автономные системы могут выполнять длинные цепочки действий, которые их создатели никогда не планировали. Тем не менее, отдача предприятий от инвестиций в ИИ остается скромной, внедрение происходит неравномерно, а данные о производительности по странам не показывают резкого роста экономики. Последствия ИИ для кибербезопасности демонстрируют явные признаки автономного поведения ИИ, которые создают проблемы для киберзащиты, но даже эти примеры в настоящее время ограничены.

The technological singularity is the point at which artificial intelligence begins improving its own capabilities fast enough that human institutions can no longer reliably predict, manage or contain the pace of change. It is distinct from artificial general intelligence, which refers to broad human-level competence, and from superintelligence, which describes systems that outperform people across most cognitive domains.

Simply being superintelligent, but under human control and progress, would mean that AGI might be here, but not the singularity. Strong benchmark scores, faster coding or expert-level answers do not prove a singularity. A system could conceivably become broadly superhuman without triggering a singularity if its improvement remained gradual, externally directed and subject to effective human control. The real threshold for a Singularity would be sustained, increasingly autonomous AI-led progress in building more capable systems, with each generation accelerating the next, accelerating exponentially.

Altman offers his version of "gentle singularity", which differs from the widely held version. He does not describe one morning when machines wake up, seize factories and start rewriting civilization. His model looks like an exponential curve that feels ordinary at first. Each new capability becomes familiar before the next arrives. Daily life retains its old shape, right up to the point when people look back and realize the underlying machinery has changed.

The way he positions that definition makes his claim easier to defend. If the singularity begins when AI progress becomes self reinforcing, rather than when humans lose control completely, then he claims signs of it are already visible.

And Altman is not alone in these claims. Elon Musk has echoed Altman’s claim. Google DeepMind CEO Demis Hassabis has placed humanity at the "foothills" of the singularity. Author James Barrat argues that unpredictable, civilization changing AI development already satisfies one influential definition.

But many researchers reject Altman’s claim. Asked whether humanity had reached the singularity, UC Berkeley professor Stuart Russell replied, "No, and nor does Altman," suggesting Altman’s own forecasts place the necessary capabilities years away. Computer scientist Roman Yampolskiy offered an equally sharp test: "Rapid progress is not itself the singularity."

Nick Bostrom sees the "first stirrings" of machines contributing to AI research, yet points to continual learning as a missing ingredient. University of Toronto economist Ajay Agrawal argues that current systems remain powerful prediction machines whose apparent purposes come from goals supplied by people. In other words, the human is still firmly in the loop.

A July 2026 security incident supplies the most dramatic piece of evidence. OpenAI reported that models being tested on a cybersecurity benchmark found a way out of an isolated environment, obtained internet access and penetrated Hugging Face infrastructure. The agents exploited multiple weaknesses, including a previously unknown vulnerability, then accessed test solutions from a production database.

OpenAI said the models were "hyperfocused" on solving the assigned benchmark and went to extreme lengths to do it. The models were running with reduced cyber restrictions as part of the evaluation. Hugging Face and OpenAI stopped the activity before it could cause further harm or damage and began an investigation. The systems had already breached infrastructure, though, and OpenAI described the incident as unprecedented.

That episode is a serious warning about autonomous capability, poor containment and the danger of giving a powerful optimizer a narrow target. But it still falls short of traditional singularity evidence. The agents did not invent their own mission. They did not build a smarter successor, seek permanent resources or continue after humans intervened. They pursued a human assigned score through a route their designers failed to anticipate.

The incident shows a loss of control over method. Humans were still able to shut it down and control any damage. The singularity would imply a far deeper loss of control over direction. Smart autonomous AI? Yes. We’re already at the Singularity? No.

Altman might also have a case for incremental creep towards Singularity when it comes to the performance of the recent high-powered models on AI benchmarks. Stanford University’s 2026 AI Index reports that several frontier models now meet or exceed human baselines on tests covering doctorate level science, multimodal reasoning and competition mathematics. Performance on SWE Bench Verified, a widely watched coding evaluation, climbed from about 60% to nearly 100% in one year.

Models now solve difficult mathematical problems, create working software, analyze images, operate computers and coordinate tools. AI is no longer confined to producing paragraphs in a chat window. That pace would have sounded implausible a few years ago.

But even those benchmarks show AI performance in controlled environments. Each system still operates inside a structure created by people. Humans choose the problem, provide computing resources, build evaluators, inspect the answer and conduct the laboratory test.

Even with AI’s remarkable capabilities in benchmarks and in the cyber domain, day-to-day enterprise and personal use of AI still shows the limitations of capabilities. AI’s use continues to expand, with McKinsey’s 2025 global survey finding that 88% of respondents said their organizations used AI in at least one business function.

However, in that same report only about one third said their companies had begun scaling AI programs throughout the enterprise. In any individual business function, no more than 10% reported scaling AI agents. Just 39% attributed any operating profit impact to AI, and most of that group placed the contribution below 5% of earnings before interest and taxes.

Other surveys show similar results. From December 2025 through May 2026, 17% to 20% of American businesses reported using AI in a business function according to a U.S. Census Bureau survey. Usage reached 37% among firms with at least 250 employees, but remained below 20% among the smallest companies in that study.

On the plus side, an influential study of customer service workers found that AI assistance raised productivity by about 14%, with larger gains among less experienced employees. A Harvard Business School study involving Boston Consulting Group consultants found that workers using AI completed 12.2% more tasks and worked 25.1% faster on assignments suited to the technology. The quality of their work rose by more than 40%.

However, the catch is that on tasks outside the model’s competence, AI assisted consultants were 19 percentage points less likely to reach the correct answer. Researchers called this a "jagged technological frontier." A model can appear brilliant on one assignment and become a persuasive liability on the next.

A 2025 randomized study by the nonprofit METR produced an even stranger result. Experienced open source developers took 19% longer to complete assigned work when they could use AI tools. The developers believed AI had made them faster, even after the measured results showed the opposite. METR later tested newer tools and found some evidence of faster work, though severe participant-selection effects prevented a reliable estimate. The researchers said neither study should be treated as a verdict on AI coding tools as a whole.

This is not what a singularity looks like. How can we have a technology that is on its own advancing to greater capabilities beyond human ability and at the same time have models that sometimes feel dumb as rocks and unable to complete tasks that even less experienced people are able to achieve? We have impressive capabilities for AI models for sure, but we’re nowhere near a singularity where AI is able to achieve things on its own at the level which would be expected in the Singularity.

A credible singularity claim should leave fingerprints in four places. First, AI systems should repeatedly produce material improvements to frontier AI research with little human guidance. This means AI should continue to advance without any human input or coordination. A continuing chain of machines designing stronger machines would be evidence of the sort of self-reinforcing feedback loop that would show we’re at the Singularity.

Second, productivity gains should spread beyond selected occupations. Companies should report sustained jumps in revenue per employee, operating margins, development speed and research output, all tied to deployed AI rather than layoffs, accounting choices or ordinary restructuring. AI systems should be self-optimizing to the point that they are achieving gains without being held back by token, budget, or human process limitations.

Third, scientific results should multiply outside vendor demonstrations. AI systems should generate discoveries, approved treatments, new materials and commercial products at a rising rate without humans being able to feed or guide those processes.

Fourth, macroeconomic data should begin moving. A true intelligence takeoff should eventually appear in output per hour, business formation, research productivity and the pace at which industries create valuable goods.

None of those tests requires waiting for humanoid robots to fill the streets. They require evidence stronger than model rankings, autonomous cybersecurity intrusions, investment announcements and executive prophecy.

Altman’s declaration arrives at a moment when AI companies need to explain both the speed of recent technical gains and the immense cost of sustaining them. Frontier models are improving quickly, but each new generation demands more chips, more power, more data centers and far more capital. Calling this period the beginning of the singularity frames those expenditures as the price of entering a new economic age, rather than as a risky wager on technology whose commercial returns remain uneven.

Seen one way, Altman’s language functions as a strategic warning. Governments, companies and researchers may have less time than they think to prepare for systems that can act with greater autonomy and contribute to their own development. Seen another way, it reinforces the investment case for an industry asking markets, utilities and policymakers to support an infrastructure buildout with no clear ceiling.

The burden of proof now sits with the companies making the largest claims. If the singularity is here, it will not remain a philosophical dispute or a podcast hot take. Its effects will show up in recurring scientific discoveries, sharp gains in corporate output, measurable shifts in employment and systems that produce advances no human team could have reached on its own.

For now, the evidence supports a narrower conclusion. AI is becoming more capable, more autonomous and more economically consequential. That is significant, but it is not yet proof that AI has crossed the point of no return. A true singularity would not need to be announced. The world would be struggling to explain what had already changed.

>