In 2025, Artificial General Intelligence (AGI) has moved from the pages of science fiction to the top of boardroom agendas. While the AI we use today (like ChatGPT or image generators) is considered “Narrow AI”, AGI is something else. Narrow AI (or Weak AI) refers to artificial intelligence systems designed for a single, specific task or a narrow range of tasks, excelling within that limited domain but lacking general understanding or consciousness. Artificial general intelligence (AGI) refers to the hypothetical intelligence of a machine that possesses the ability to understand or learn any intellectual task that a human being can.
As it is the beginning of 2026, there are less and less arguments about if it is possible to be developed, but there are conversations about how it can be measured when it does.
The difference between AGI and AI
Unlike AGI, which could theoretically learn to do any task that an average human could do, the AI that is mostly used today is a weak type of AI. This means the AI can only accomplish the tasks that programmers and developers train it to perform.
What exactly is AGI?
Current systems have specializations (e.g., a model that is great at coding but cannot plan a vacation without a specific prompt). AGI would possess:
- Cross-Domain Learning: The ability to learn a skill in one area (like chess strategy) and spontaneously apply its logic to another (like business negotiation).
- Reasoning & Common Sense: Understanding “why” things happen, not just predicting the next word in a sentence.
- Autonomous Planning: The capability to set its own goals and execute multi-week projects without human hand-holding.
But the definition about what AGI is different for everyone. There are many definitions of what Artificial General Intelligence (AGI) is, because the term encompasses complex, subjective, and highly debated concepts like “intelligence,” “consciousness,” and “generality”. This lack of a single, universally accepted definition stems from several factors: The term sits at the intersection of three very different worlds: Engineering, Economics, and Philosophy.
Because “Intelligence” itself isn’t a single physical property (like length or melting point), different groups define AGI based on what they are trying to achieve or protect.
1. The Engineering View: “The Generalist”
To a computer scientist, AGI is about adaptability. Most AI today is “Narrow AI” which, as mentioned for example, can play chess or translate text, but it can’t do both.
- Key Metric: Generalization (transferring knowledge from one domain to another).
2. The Economic View: “The Worker”
Companies like OpenAI often use a definition tied to productivity. This version is less about “thinking” and more about “replacing.”
- Key Metric: Revenue and labor replacement.
3. The Philosophical View: “The Mind”
This group argues that a calculator can be “smart” but isn’t “intelligent.” They focus on the internal experience of the machine.
- Key Metric: Reasoning, common sense, and self-awareness.
The “AGI Levels” Comparison
Because the definitions are so messy, researchers at Google DeepMind proposed a framework to help everyone speak the same language. It categorizes AI by Generality (how many things it can do) and Performance (how well it does them).
Why is the focus on the confusion
The reason the definition of Artificial General Intelligence (AGI) is so inconsistent is that it isn’t a technical specification; it is a “moving goalpost” shaped by different professional perspectives.
- Safety: If we define AGI as “economically valuable,” we might ignore the risks of a machine that is smart but lacks human ethics.
- Policy: Governments need to know when a machine reaches a certain threshold to pass laws (like the EU AI Act).
- Timelines: When an expert says AGI is “2 years away” and another says “50 years away,” they are usually using two different definitions.
Why is AGI gaining popularity right now?
The surge in popularity is driven by a massive “race to the finish line” among tech giants and several key 2025 breakthroughs:
The “Agentic” Change
In 2025, there was a move from “Chatbots” to “AI Agents.” These systems can now use computers, navigate websites and complete multi-step workflows. This feels like a “lite” version of AGI, making it look like the real thing is just around the corner.
Aggressive Timelines from Tech Leaders
It has become more interesting because of the equivalent to the Oppenheimer’s fissile material – bold predictions from industry CEOs. Sam Altman (OpenAI) and Dario Amodei (Anthropic) have both signaled that AGI-level capabilities could emerge as early as 2026 or 2027. This has created a “Manhattan Project” level of urgency and investment (estimated at over $100 billion this year alone), but this time, there are more than one Oppenheimers in the game that are interested and are leading it.
The Reasoning Breakthrough (o1 and o3 models)
Newer models (like OpenAI’s o3) have shown a massive leap in “reasoning” scores. By spending more “thinking time” on a problem before providing the answer, these models are solving complex math and physics problems that previously stumped AI, making the gap between machine and human intelligence look much smaller.
The Geopolitical Race
AGI is being viewed as the ultimate strategic asset. National governments are treating the development of AGI as a matter of national security, similar to the space race, because the first nation to achieve it will likely have an economic impact that will not be capable of being conquered or vanquished and immerse military advantage.
There are rumours about it being available in 2029: The Rationale Behind the Prediction
The year 2029 is a legendary date in the AI world because it is the specific year predicted by Ray Kurzweil, a world-renowned futurist and Director of Engineering at Google. For over 25 years, Kurzweil has maintained that by 2029, AI will achieve “human-level intelligence,” successfully passing a valid Turing Test and matching the intellectual capabilities of a skilled human across all domains.
Kurzweil’s prediction is based on a mathematical theory called the Law of Accelerating Returns.
- Exponential Growth of Compute: Kurzweil tracks the amount of computing power you can buy for $1,000. He argues that computational power doesn’t grow linearly ($1, 2, 3…$) but exponentially ($2, 4, 8…$). He calculated decades ago that around 2029, $1,000 of computing power would be roughly equivalent to the processing power of a human brain (estimated at $10^{16}$ calculations per second).
- The “Double Exponential” Effect: It’s not just the hardware getting faster; the algorithms are getting more efficient. When you combine faster chips with smarter software, the timeline for AGI “collapses,” bringing the date closer than traditional logic suggests.
- The Turing Test Milestone: Kurzweil specifically predicted that by 2029, a computer would be able to sustain a conversation with a human to the point where the human could not tell they were talking to a machine-not just for a few minutes, but in a deep, nuanced, and emotional way.
To paraphrase Anders Arpteg from the latest episode from the AIAW E173 Podcast on the answer to the question – what is actually AGI and why would AGI change anything?
AGI would happen when we actually have an AI that can be on par with the average coworker. It should be able to replace an average coworker, even create powerpoints. He continues to share that: long term, people will abuse AI and there need to be AI to regulate AI. People will abuse it and because AI is too stupid and will do whatever the human tells them to do, it first needs to be regulated until we have AGI.
Who Else Supports the 2029-2030 Timeline?
While Kurzweil was once considered an outlier, several major figures in 2025/2026 now agree that the end of the decade is the “likely” window:
- Sam Altman (OpenAI) has stated that AGI is possible “this decade” and recently suggested that we are now on a predictable path toward it as we scale up training data and compute.
- Elon Musk (xAI): Known for even more aggressive timelines, Musk has recently predicted that AI will be “smarter than any one human” by next year (2026) and “smarter than all humans combined” by 2029 and mentioned that on his social media account.
- Metaculus (Forecasting Community): The aggregate “wisdom of the crowd” on forecasting sites like Metaculus has seen the predicted date for AGI plummet. In 2020, the community predicted 2045; by late 2025, the median prediction has hovered around 2029–2032.
What is the “Missing Piece” for 2029?
To reach the 2029 goal, Kurzweil and others believe we need three final shifts:
- Contextual Memory: Moving from “sessions” to “long-term memory” where AI remembers you and the world over years.
- Common Sense Reasoning: Moving beyond statistical word-guessing to understanding “cause and effect.”
- Longevity Escape Velocity: Kurzweil also predicts that 2029 is the year we reach “Longevity Escape Velocity,” where for every year you live, science adds more than one year to your remaining life expectancy through AI-driven medical breakthroughs.
Is AGI going to be a reality in the near future?
If we look at the current state of technology in late 2025, we are in a “Schrödinger’s AGI” situation. If it is or if it isn’t happening or if it happens it will happen in the near or in the far future, depending on how it is measured.
But if we look at the bigger picture as it is now, the answer is: It depends entirely on which definition it is looked at.
The Argument for “it’s happening”
If the definition for AGI is competence across all human domains, many experts argue we are in the “Emerging AGI” phase right now.
- Performance: AI now scores in the top 10% on the Bar Exam, medical licensing exams, and complex coding competitions. It can translate 100+ languages, compose music, and write software better than the average human.
- Multimodality: AI is no longer just “chats.” It sees, hears, and speaks. It can look at a broken engine through a camera and explain how to fix it. This “general” ability to handle any type of data is a core requirement of AGI.
- The Scaling Laws: So far, every time we add more data and more computing power ($GPU$s), the AI gets smarter. There is no evidence yet that we have hit a “wall.”
The Argument for “it’s not happening yet”
If we look at the definition of AGI as reasoning like a human, many scientists (including Meta’s Chief Scientist Yann LeCun) argue that we aren’t even on the right path yet.
- Lack of a “World Model”: Current LLMs are “Large Language Models,” not “Large World Models.” They know that the word “apple” usually follows “red,” but they don’t know that an apple will fall if you drop it unless they’ve read a description of gravity.
- Reliability & Hallucination: A true “General Intelligence” shouldn’t confidently invent facts. Humans have a “self-correction” mechanism and logic that current AI lacks.
- The “Wall”: We are running out of human-generated text on the internet to train these models. If AI can’t learn to “think” without just memorizing the internet, it may plateau.
The 2026 Reality: “Agentic AI”
Instead of debating if AGI “is” or “is not,” the industry in 2026 has moved toward Agentic AI. This is the middle ground.
- Chatbots (2023): You ask, it answers.
- Agents (2025-2026): You give a goal (“Plan my 10-day trip to the Nordics and book the flights”), and the AI uses tools, navigates websites, and executes the task autonomously.
- AGI (Future): The AI can invent a new way to travel to the Nordics that doesn’t exist yet.
How things look like at this moment
We are currently in Level 1 (Emerging AGI). Whether we reach Level 2 (Competent AGI) is what the 2026–2029 window is all about. The future will tell if the AI can do 50% of human jobs.
Some interesting things about AGI
- AI models passed the Bar Exam and SATs with ease by memorizing vast amounts of data. This has been seen happening in the last several years. In 2025, researchers introduced “Humanity’s Last Exam” (GPQA Diamond) which was a test so difficult that even non-expert humans with Google access can’t solve the questions, but top-tier experts in those specific fields can. But the catch was that models like Gemini 3 and GPT-5.2 are seen to be scoring at “excellent” levels on these reasoning tasks.
- As mentioned before, the most significant step in 2025 wasn’t a smarter chat window; it was the rise of AI Agents. Unlike a chatbot that just talks, an AI agent can.
- While AI is getting better at reasoning, there is still one “Final Boss” benchmark: ARC-AGI-2. This test uses simple color grids and puzzles that a 5-year-old human can solve instantly but that require genuine abstraction (learning a rule you’ve never seen before). In late 2025, a small startup called Poetiq stunned the industry by scoring 54% on this benchmark, significantly outperforming giants like OpenAI and Google (who hovered around 15–20%). This suggests that AGI might not come from “more data,” but from a completely new type of mathematical architecture.
AGI as the Ultimate Strategic Frontier
The debate over whether AGI is “happening” depends less on the code itself and more on the lens through which we view intelligence. For the functionalist, AGI is already arriving in the form of systems that outscore humans on the Bar Exam and manage complex enterprise workflows. For the architectural purist, we remain miles away from a machine that possesses a true “world model” or the common-sense reasoning of a child. As we move toward the 2029 milestone famously predicted by Ray Kurzweil, the industry has shifted its focus to Agentic AI – a middle ground where machines act as autonomous coworkers rather than just chatbots. Whether AGI is defined as an economic replacement for a knowledge worker, a scientific leap in cross-domain reasoning, or a philosophical milestone in consciousness, the consensus for 2026 is clear: we are no longer waiting for a single “spark” to ignite a distant machine mind. Instead, we are building the structural and regulatory foundations for a future where the definition of intelligence is as diverse as the humanity that created it.
