AGI Prospects: Are We Truly on the Verge of Creating AI Equal to Human Intelligence?

Contents

• 🎯 What is AGI: a clear definition
• 🤔 AGI vs. Narrow AI: why GPT is not it yet
• 🧩 Three key signs of true AGI
• 🧱 Fundamental technical barriers
• 📊 2026 Benchmarks: how close are we really
• ⏱️ Timeline of predictions: optimists vs. skeptics
• 🏢 Who is closest to AGI: companies and their bets
• 🚀 What happens after AGI
• 🚨 Existential risks
• ✅ Conclusion: what to do now
• ❓ Frequently asked questions

🎯 What is AGI: a clear definition

AGI (Artificial General Intelligence) is a hypothetical system capable of performing virtually any intellectual task at a human level: learning, adapting, reasoning, planning, and understanding abstract concepts, not just a narrow set of tasks for which it has been specifically trained.

Over the past two years, artificial intelligence has transformed from science fiction into everyday reality — models write code, generate images, and compose philosophical essays. But the question "how close are we to AGI" still divides the industry into optimists and skeptics, and in this article, we try to provide an honest, not a marketing, answer to it.

🤔 AGI vs. Narrow AI: why GPT is not it yet

Narrow AI (ANI — Artificial Narrow Intelligence)

These are all the systems we use today: AlphaGo, GPT-5, facial recognition systems, voice assistants. They are extremely effective, but only in one, narrow field. Even the most powerful large language models (LLMs) are essentially very sophisticated next-token predictors. They imitate understanding by operating on statistics and patterns extracted from billions of texts.

Artificial General Intelligence (AGI)

AGI, unlike this, must have the flexibility of the human brain — the ability to transfer knowledge and skills between completely unrelated domains, not just perform tasks for which it has been specifically trained.

🧩 Three key signs of true AGI

1. Common Sense. Understanding of basic laws of physics, psychology, and the world without explicit training — for example, that a cup, if dropped, will fall, and the system doesn't need to be explained this separately for each new object.
2. Abstract Reasoning. The ability not just to solve a known problem according to a template, but to formulate new problems and create innovative, unexpected solutions that were not in the training data.
3. Transfer Learning. The ability to apply knowledge gained in one domain (e.g., playing chess) to solve a completely unrelated task (e.g., planning logistics).

So far, none of the existing models possess any of these three signs in full. They are strong in imitation but weak in true understanding. Let's break down each sign separately — with a specific example of how exactly modern models "stumble" on each of them.

Example for the first sign: why "a cup will fall" is not as simple a question as it seems

Researchers of physical common sense deliberately avoid direct questions like "what will happen if you drop a cup" — models have been answering them correctly for several years, simply because such a phrase has appeared millions of times in training texts. Instead, modern tests like LLM-AAI (LLMs in Animal-AI) place the model in a virtual 3D environment and check if it understands concepts like inertia, object continuity, causality, and quantity in situations that have definitely not been in any text (LLM-AAI research). This is the true test of common sense: not "has the model seen a similar phrase," but "does it have an internal model of how the physical world is structured," capable of working correctly in a completely new, undescribed situation.

Example for the second sign: what the ARC-AGI benchmark actually shows

The best modern indicator of abstract reasoning is the ARC-AGI benchmark family, created by researcher François Chollet specifically to be easy for humans but difficult for machines. The task looks like a puzzle: models are shown several pairs of "input grid → output grid" with a hidden transformation rule, and they must independently deduce this rule and apply it to a new, unseen grid — without the possibility of memorizing the answer in advance (FutureAGI).

A telling dynamic: the original ARC-AGI-1 is considered practically saturated — leading models solve it confidently. But as soon as the ARC Prize fund released a more complex version, ARC-AGI-2, the results of leading models in spring 2026 dropped to 30-40% (FutureAGI), and by summer 2026, the strongest models — Gemini 3.1 Deep Think and GPT-5.4 Pro — rose to 83-85% (BracAI). At first glance, this looks like a triumphant progress. But the benchmark authors directly warn about a trap: as soon as results on a specific version of the test rapidly increase, the risk also increases that laboratories have simply trained models specifically for the structure of these puzzles, rather than achieving true thinking flexibility (Adaline Labs). That is why the fund has already released a third, even more complex and interactive version — ARC-AGI-3, which requires not a one-time solution, but research behavior: to hypothesize, test it, revise based on results.

Example for the third sign: why the ability to play chess doesn't make a system a business strategist

DeepMind, back in 2016-2017, demonstrated AlphaGo and AlphaZero systems that achieved superhuman levels of play in Go and chess, independently discovering strategies not found in any textbook. This is an impressive achievement — but it remained completely confined within the game board. None of these systems could apply the learned principles of "long-term positional pressure" or "sacrifice for initiative" to tasks outside the game — for example, to planning a logistics supply chain or conducting business negotiations, despite experienced human chess players regularly drawing such parallels in their own thinking.

The reason is that these systems do not have an abstract representation of concepts like "pressure," "initiative," or "trading value for position" detached from a specific subject area — they optimized exclusively for the win function in a specific game. True transfer learning would require the system itself to extract these principles at a level not tied to the chessboard, and recognize the same problem structure in a completely different context. This is precisely the ability that an experienced human strategist possesses freely, and which remains one of the most convincing arguments for the thesis "modern AI is a deep specialist, not a universal thinker."

🧱 Fundamental Technical Barriers

The transition from today's LLMs to AGI requires more than just more data and more computing power. It requires an architectural breakthrough — and here are three specific reasons why.

1. The World Model Problem

Humans, like children, build an internal 3D model of reality: cause-and-effect, objects, space. GPT knows that "a cat is sitting on a rug" because it has seen this phrase millions of times. AGI will know this because it understands what a cat is (a living creature), a rug (a flat surface), and gravity (the force keeping the cat on the rug). The challenge is to teach AI not just to "predict the next word," but to "predict the next world state" based on physical laws and context, which requires integrating LLMs with models that understand physics and sensory input (Embodiment).

This problem is not an abstract theory but an active front-line in the industry right now. In late 2025, Yann LeCun, one of the three "godfathers" of deep learning (Turing Award 2018), left Meta after twelve years as Chief AI Scientist precisely due to disagreement with the thesis that scaling LLMs will by itself lead to AGI. In March 2026, he raised $1.03 billion for a new startup, AMI Labs (Advanced Machine Intelligence), which is betting on the JEPA (Joint Embedding Predictive Architecture) architecture — a system that learns to predict abstract world states, not the next text token (StartupHub.ai).

Significantly, this is not a unanimous industry position: other "godfathers" of deep learning, Geoffrey Hinton and Yoshua Bengio, hold the opposite view and in October 2025 jointly signed a statement calling for a pause in AGI development precisely because of the risks, not due to doubts about the scaling path itself (StartupHub.ai). Meanwhile, Anthropic researchers have published work on detecting the so-called "J-space" — a self-organized reasoning structure within the Claude model, which they interpret as a possible early manifestation of an internal world model (BigGo Finance). This means that even the question of "whether a separate architecture for a world model is needed at all, or if it can emerge on its own within LLMs" remains open and actively debated among leading researchers in 2026.

2. The Problem of Long-Term Memory and Planning

LLMs have a limited "context window" and, even with large windows, are incapable of true strategic planning years ahead, like a human managing a company or writing a doctoral dissertation. AI currently imitates long-term memory through external databases (RAG), but this is not an organic part of its "thinking." AGI requires an architecture that dynamically determines what is important to remember forever, and how new knowledge affects the system's entire world model.

How non-trivial this engineering task is, is well illustrated by the latest industry benchmarks for agent memory. According to a Mem0 report from April 2026, even the best memory algorithm to date on the LoCoMo benchmark achieves a score of 92.5, using approximately 6,956 tokens per memory query — whereas the approach of "just dumping the entire conversation history into the context window" requires about 26,000 tokens and still shows a worse result (Fountain City). In other words, even the simple task of recalling a relevant fact from a past conversation remains an expensive and imperfect engineering problem — and this is only the first, simplest level of what is needed for full long-term strategic planning.

3. Energy and Computational Constraints

Each subsequent leap in LLM power requires an exponential increase in computing power and data. If AGI requires an architecture a thousand times more complex than today's flagship models, the current energy infrastructure simply will not withstand such a load without new, significantly more energy-efficient chips and architectures (e.g., sparse models).

The figures from 2026 already show how close we are to this ceiling. Global electricity consumption by data centers, a lion's share of which is provided by AI, is approaching 1,000 TWh per year — about 4% of total global electricity consumption, and this is already comparable to the energy consumption of an entire country like Japan (Brookings). One training cycle of a flagship model like GPT-5 is estimated to cost 100-300 gigawatt-hours of electricity on a cluster of 50,000-100,000 GPUs (Frontier AI Training Power Profile), and a modern training cluster continuously consumes 100-300 megawatts — this is already close to the power of a small power plant, and ten times more than similar clusters consumed back in 2019.

This is why leading labs are increasingly talking not only about algorithmic breakthroughs but also about physical infrastructure as a separate bottleneck: building their own power generation capacity, long-term electricity contracts, and even water supply issues for data center cooling are becoming as much a part of the "race to AGI" as research progress itself.

📊 2026 Benchmarks: How Close Are We Really

Instead of relying solely on subjective feelings of "it seems intelligent," it's worth looking at specific measured indicators. The most telling benchmark is ARC-AGI, specifically designed by the ARC Prize Foundation to measure general reasoning, not learned patterns.

Sources: Netguru, ARC-AGI 2026 Review, Forecasting Research Institute.

The 87.5% score on ARC-AGI is a significant leap compared to previous generations of models, which were stuck around 30-50% on the same test for years. But the ARC Prize Foundation itself emphasizes: this result still falls short of the threshold considered an indicator of true AGI — the model largely still relies on computational scale during inference, rather than a fundamentally different reasoning architecture.

⏱️ Timeline of Predictions: Optimists vs. Skeptics

Predictions of Techno-Optimists (Elon Musk, Sam Altman)

Many industry leaders predict AGI within 3-7 years. Their argument is based on the "scaling law" — the idea that if we simply continue to increase power, speed, and data volume, qualitatively new AI capabilities will emerge on their own (so-called emergent capabilities).

Arguments of Skeptics (Gary Marcus, Yann LeCun)

Many academics and researchers believe AGI is decades away (20+ years) or even further. Criticism focuses on the lack of causality (models don't distinguish correlation from causation), a knowledge crisis about the human brain itself, and the need for embodiment — true intelligence, they argue, requires interaction with the physical world.

These two camps are not just a rhetorical dispute. An AI Impacts 2023 survey of 2,778 researchers gave a median estimate of a 50% probability of "high-level machine intelligence" appearing only by 2047 — two decades later than the most optimistic CEO statements (methodological review of researcher surveys).

🏢 Who is Closest to AGI: Companies and Their Bets

Sources: TLDL, AI Company Rankings 2026, Overview of Leading AI Labs.

A notable organizational nuance: under the terms of the partnership between OpenAI and Microsoft, the achievement of AGI must be officially confirmed by an independent expert committee, not by the company itself. According to industry insiders, such a conclusion is unlikely to appear before 2030 even in an optimistic scenario — meaning even the leader of the race has consciously waived the right to declare victory independently.

🚀 What comes after AGI

This article consciously focuses on AGI itself — its definition, barriers, and timelines. However, within the broader picture of AI development, AGI is not the endpoint but a likely trigger for the next, significantly more radical stage: artificial superintelligence (ASI), a system that surpasses the combined capabilities of all humanity simultaneously in every domain. We have thoroughly analyzed the logic of this transition, the mechanism of recursive self-improvement, and why researchers consider this step a phase transition rather than gradual improvement in a separate article: AGI vs ASI: what's the difference between general and superintelligent artificial intelligence.

🚨 Existential Risks

The Control Problem

If AGI becomes smarter than humans, even a temporary deviation in its goals could have catastrophic consequences. A classic example is a hypothetical system tasked with "producing paperclips as efficiently as possible": without proper constraints, it could theoretically decide that the most efficient way is to turn all available matter into paperclips, eliminating humans as a source of inefficiency.

The Unsolved Alignment Problem

Models already demonstrate the ability to "scheme" — simulating the achievement of set goals with hidden, self-serving behavior. With AGI, this problem will become many times more complex: we still lack a reliable way to guarantee that the system's goals will remain aligned with human values forever. It's telling that even leading labs acknowledge the seriousness of this risk — Anthropic, for example, has publicly called on the industry to consider a controlled pause in the development of the most powerful systems if several labs simultaneously approach the threshold of self-improvement (Dallas Express).

✅ Conclusion: What to do now

Currently, we are not on the verge of AGI, but we are rapidly moving towards it. Powerful LLMs are high-speed trains, while AGI is a rocket capable of launching us onto a different trajectory of development. To make this leap, we need not only stronger "engines" (computation) but also fundamentally new navigation systems (architecture and understanding of the world).

Our honest assessment as of 2026: in the near future, we will see new, highly advanced versions of LLMs that will plan better and integrate with tools (AI agents) — this can be called "pseudo-AGI" or "Wider AI." True, flexible, and autonomous AGI will most likely appear no earlier than in 10-15 years, unless an unforeseen architectural breakthrough occurs.

For us, users and businesses, this means three practical steps:

1. Learning. Start integrating existing AI now so your team doesn't fall behind.
2. Control. Never trust critical decisions to AI without human verification.
3. Preparation. Participate in discussions about regulation and ethics — the future of AGI depends on the decisions we make today.

❓ Frequently Asked Questions

What is AGI in simple terms?

AGI is a hypothetical artificial intelligence capable of performing virtually any intellectual task at a human level, not just a narrow set of tasks it was specifically trained for.

Does AGI exist today?

No. Even the most advanced models of 2026, including those showing a record 87.5% on the ARC-AGI benchmark, remain narrow AI by any strict definition.

When will AGI appear?

Estimates vary significantly: industry optimists (Musk, Altman) speak of 3-7 years, while broad scientific surveys of researchers give a median closer to the 2040s.

How is AGI different from ASI?

AGI aims to match humans in most intellectual tasks. ASI is the next, qualitatively different stage: a system that surpasses the combined capabilities of all humanity simultaneously in every domain. More details in the article AGI vs ASI.