HOW CLOSE ARE WE TO AGI?
AI systems are rapidly becoming more capable, and their progress is moving so quickly that it is difficult to give a concrete answer to one question: How close are we to AGI?
Today’s AI can solve advanced mathematics and science problems, write software, generate realistic images and videos, assist researchers, and perform increasingly complex tasks across different fields.
But impressive capabilities do not automatically mean AGI. To understand how close are we to AGI, we need to look at what AI can already do, where it still struggles, and which capabilities could bring us closer to genuinely general intelligence.
The rise of increasingly capable AI also bring risks. The same systems that can support scientific research, healthcare, education, and productivity can also produce unreliable results or be misused. Understanding both sides is essential when discussing the road toward AGI.
What can AI actually do in 2026?
- Reasoning: AI can solve advanced mathematics, science and logic problems, pass professional exams, and answer highly complex questions.
- Coding: AI can write, debug and modify software, increasingly handling software-development tasks with limited human input.
- Scientific Research: AI is being used to generate hypotheses, analyse data, troubleshoot research problems, design proteins for example and discover more efficient algorithms.
- Multimodal Capabilities: Modern AI can work across text, images, audio and video, while increasingly interacting with computers and other environments.
- AI Agents: AI is moving beyond simply answering questions. Agents can use tools, navigate software, write code and perform multy-step tasks with less human supervision.
Almost a billion people now use general-purpose AI systems in their daily lives for work, learning or both. Companies are investing hundreds of billions of dollars to build the infrastructure to train and deploy them. In many cases, AI is already reshaping how people access information, make decisions, and solve problems, with applications in industries from software development to legal services to scientific research.
AI can increasingly perform tasks. But AGI would need to reliably manage goals.
Where does AI still Fail?
- Influence and manipulation: In experimental settings, AI-generated content can be as effective as human-written content at changing people’s beliefs. Real-world use of AI for manipulation has been documented but it is not yet widespread. This could change as AI capabilities improve.
- Cyberattacks: AI systems can discover software vulnerabilities and write malicious code. In one competition, an AI agent identified 77% of the vulnerabilities present in real software. Criminal groups and state-associated attackers are actively using general-purpose AI in their operations. However, it remains uncertain whether attackers or defenders will benefit more from AI assistance.
- Biological and chemical risks: General-purpose AI systems can provide information related to biological and chemical weapons development, including details about pathogens and expert-level laboratory instructions.
- Multi-step tasks and reliability: AI systems can still be derailed by simple errors during multi-step projects. They continue to generate false statements, often referred to as “hallucinations”, and cannot yet reliably integrate with robotic components to perform basic physical tasks such as housework.
- Language limitations: Performance also tends to decline when AI is so prompted in languages other than English, particularly those that are underrepresented in training datasets.
- Inconsistency: AI performance can also be inconsistent; for example, accuracy on mathematics problems can decrease when irrelevant information is inserted into the problem description.
General-purpose AI system can currently perform tasks such as:
| 🗣️ | <> | 🌅 | 🧮 | 🔬 |
| Fluent conversation in many languages | Write and debug code for well-defined software tasks | Create highly realistic images and short videos | Solve exam-style maths and science problems at graduate level | Contribute to scientific research through literature review and data analysis |
General-purpose AI systems cannot perform tasks such as:
| 📆 | 📄 | 🦾 | 🧮 | 🗣️ |
| Executing multi-day projects | Generating text without false statements | Performing useful tasks involving robotics, such as household work | Solving math and science problems that require heavy compositional reasoning | Perform as well in languages with significantly less digital presence than English |
How Close Are We To AGI: The Near Future
AI progress through 2030 is uncertain, but current trends remains consistent with continued improvement.
The key drivers of AI progress －compute, algorithmic improvements, and data － have grown rapidly in recent years. New inference-time scaling methods are also improving model capabilities after training.
Researchers cannot reliably predict when specific capabilities will emerge, and experts disagree about whether the rapid growth of AI inputs can continue. Some expects current training techniques to plateau, while bottlenecks in data, energy, or computing could limit future progress. Others believe progress could accelerate further, as AI systems are increasingly used to support AI research itself, potentially creating positive feedback loops.
SOME PREDICTIONS FOR THE YEARS AHEAD INCLUDE:
- For the most compute-intensive models, training compute has grown by around 5x per year. If this trend continued until 2030, these models could be trained with roughly 3,000x more compute than today.
- Training datasets have expanded from billions to trillions of data points, with an average increase of around 2.5x per year.
- AI systems could autonomously complete hours-long software projects by 2027-2028 and days-long projects by the end of the decade. However, these projections assume an 80% success rate, with may still be below the reliability required for autonomous deployment in many professional settings.
- AI agents already perform better than humans on some two-hours tasks, but their success rates are lower on eight-hour tasks. This is suggestive, but real-world AI involves additional bottlenecks, including ambiguous goals and the time required to determine whether an algorithmic improvement actually improves model performance.
- AI models could reach around 55% accuracy by 2027 and 75% by 2030 on undergraduate-level FrontierMath problems.
- Autonomous vehicles are expected to remain limited in 2030, with cultural resistance, infrastructure requirements, and regulatory barriers among the challenges. Small and medium-sized enterprises － which employ around 60% of workers globally － may face additional barriers to AI adoption, including limited technical expertise, insufficient computing infrastructure, and high integration costs.
AI developers are betting that computing power will remain a major driver of progress, with companies announcing hundreds of billions of dollars in data-centre investments. For example, Meta and OpenAI have announced plans involving approximately $65 billion and $500 billion in investment, respectively.
Whenever AI capabilities will continue improving at the same pace as they have in the recent years remains difficult to predict. Between now and 2030, progress could accelerate, slow down, or plateau. What is much easier to predict is that AI development will continue to be one of the most important technological forces shaping the next decade.
AGI: General-Purpose
- General-purpose AI refers to systems that can perform a wide range of tasks rather than being designed for one specific function or domain. These tasks can include generating text, images, video, and audio, understanding information, solving problems, writing code, and taking actions on a computer.
- AGI would take this general-purpose capability much further. Instead of simply performing many different tasks, an AGI system is generally expected to learn, reason, adapt, and apply knowledge across a broad range of domains at a level comparable to human intelligence.
- Modern-general purpose AI systems are largely based on deep learning, which uses large amounts of computational resources and training data to help models learn complex relationships and abstract patterns.
- Developing leading general-purpose AI systems has become extremely expensive. Training and deploying them requires vast amount of data, specialised expertise, and large-scale computational infrastructure. Building a leading system from scratch can now cost hundreds of million of US dollars.
- Recent progress has increasingly come only from making models larger, but also from post-training techniques and additional computation during use. This shift is important for AGI research because increasingly capable systems may be able to reason for longer, use tools, adapt to new tasks, and potentially contribute to the development of future AI systems themselves.
There is no universally accepted definition or benchmark for AGI. Different researchers and organisations use the term to describe different levels of generality, autonomy, reasoning and adaptability.
Sources:
- International AI Safety Report
- ChatGPT – suggestions and double check
- Gemini – images
Prefer listening? Press play to hear this page.
HOW CLOSE ARE WE TO AGI IN SPECIFIC CAPABILITIES?
Current AI systems are already approaching some capabilities associated with AGI, particularly language understanding, coding, image and video generation, mathematical reasoning, research assistance, and computer use. AI agents are also becoming increasingly capable of completing multi-step tasks and using external tools. However, being highly capable across many individual tasks is not necessarily the same as having general intelligence. Current systems can still struggle with reliability, long-term planning, unfamiliar situations, and transferring knowledge between different domains.
WHICH CAPABILITIES ARE IMPROVING FASTEST?
Some of the fastest improvements are happening in reasoning, coding, multimodal understanding, and AI agents. Models are becoming better at processing different types of information, solving increasingly complex problems, generating and debugging code, and completing tasks that previously required multiple tools or human intervention. Another important trend is inference-time scaling － giving models additional computation while solving problem. This allows some systems to spend more time reasoning about difficult tasks rather than relying only on what they learned during initial training.
WHICH PARTS OF AGI REMAIN DIFFICULT?
Several capabilities remain particularly challenging: reliable long-term planning, learning continuously from experience, understanding ambiguous real-world situations, robust reasoning, and operating autonomously for long periods of time. AI systems can also make mistakes, struggle when conditions differ from their training, and have difficulty judging when they do not know something. For AGI, high performance on individual benchmarks is not enough － a system would need to remain capable and reliable across a much broader range of unpredictable situations.
ARE CURRENT BENCHMARKS ACTUALLY MEASURING GENERAL INTELLIGENCE?
Not on their own. Most AI benchmarks measure specific capabilities such as mathematics, coding, language understanding, knowledge, or reasoning. They are useful for comparing models, but success on a benchmark does not automatically demonstrate generate intelligence.
There is also growth concern that increasingly capable AI systems may eventually perform well on benchmarks because they have encountered similar problems during training. This makes it harder to determine whether a model has learned a general capability or is benefiting from exposure to familiar patterns.
For AGI, evaluation would likely need to measure breadth, adaptability, learning, reasoning, autonomy, and performance on genuinely unfamiliar tasks, rather than relying on a single score.
WHAT EVIDENCE WOULD CONVINCE US THAT WE’RE GENUINELY APPROACHING AGI?
There probably won’t be one single benchmark or moment that proves we have reached AGI. Strong evidence would instead come from several capabilities appearing together. A convincing sign would be an AI system that can learn new tasks with little instruction, transfer knowledge between very different domains, reason reliably in unfamiliar situations, plan and execute complex long-term goals, use tools effectively, and improve its performance through experience.
Perhaps most importantly, such a system would need to perform these tasks reliably in the real world, not just in controlled demonstrations or benchmark environments. In other words, the strongest evidence for approaching AGI would not simply be that AI can do more things. It would be that AI can learn, adapt, reason, and apply what it knows across many different situations with increasing autonomy and reliability.
Read more about AI tools and AGI Examples here:
[…] How Close Are We To AGI […]