No system today qualifies as artificial general intelligence, but dozens of AI systems demonstrate narrow slices of what AGI would require, and researchers use these partial successes to map how far we still have to go. A Google DeepMind framework published in 2023 proposed formal “levels” of AGI, from systems that match an unskilled human on a wide range of tasks all the way up to systems that surpass the best humans at virtually everything. Current large language models sit somewhere below that first level: impressive at language and certain reasoning tasks, brittle in ways that become obvious the moment you push them into unfamiliar territory. The gap between where AI is now and what AGI would actually mean is best understood through the specific examples researchers use to test, showcase, and expose the limits of today’s most capable systems.
What Would Actually Count as AGI
The term “artificial general intelligence” gets thrown around loosely, but researchers at Google DeepMind tried to pin it down. Their 2023 framework classifies systems along two axes: how well they perform, and how broadly that performance generalizes across different kinds of tasks. A system that writes brilliant poetry but cannot do basic arithmetic is not general. A system that scores well on every standardized test but freezes when asked to navigate a kitchen is not general either. The framework proposes six levels, starting with “Emerging” (equal to or slightly better than an unskilled human across most tasks) and ending at “Superhuman” (outperforming all humans at essentially everything).1Proceedings of ICML. Levels of AGI for Operationalizing Progress on the Path to AGI
By that yardstick, no current system clears even the first level. GPT-4, Claude, Gemini, and their peers perform at expert or superhuman levels on some tasks and fall apart on others. The “general” in AGI is the hard part. Most of the examples researchers point to when discussing AGI progress are really examples of AI succeeding spectacularly in one domain while failing in another, and the interesting question is always where the failure boundary sits.
Abstract Reasoning and the ARC Benchmark
If you want a single test that captures what AGI-style thinking might look like, the Abstraction and Reasoning Corpus is the one researchers keep coming back to. ARC presents simple visual puzzles: grids of colored squares where you have to figure out the underlying rule from a few examples and apply it to a new case. A child can often solve these in seconds. They demand the kind of on-the-fly pattern recognition and generalization that current AI finds punishing.
ARC has become a living benchmark with multiple versions of increasing difficulty. A comprehensive survey of 82 different approaches found that top systems now score around 93% on the original version (ARC-AGI-1), which sounds impressive until you see the drop-off. On the harder ARC-AGI-2, performance falls to about 69%. On ARC-AGI-3, it craters to 13%. Humans maintain near-perfect accuracy across all three versions.2arXiv. The ARC of Progress towards AGI: A Living Survey of Abstraction and Reasoning That two-to-three-fold performance drop from one version to the next holds regardless of whether the AI uses neural networks, symbolic programming, or a hybrid approach, suggesting the problem is not just about picking the right architecture but about something deeper in how these systems handle novel compositional tasks.
Researchers have tried attacking ARC from every angle. One approach recasts each puzzle as a program-synthesis problem, using inductive logic programming to build rules from examples rather than relying on pattern-matching from training data.3HYDRA@ECAI. Program Synthesis Using Inductive Logic Programming for the Abstraction and Reasoning Corpus The fact that ARC remains largely unsolved, despite significant prize money and years of effort, makes it one of the clearest illustrations of the gap between narrow AI and anything resembling general intelligence.
Advanced Mathematics
Mathematical reasoning is another domain where AGI-level performance would be transformative, and where current AI falls dramatically short. FrontierMath is a benchmark of hundreds of original, exceptionally challenging math problems spanning number theory, algebraic geometry, real analysis, and category theory. These are not textbook exercises. A human mathematician working in the relevant specialty typically needs multiple hours on a standard problem and sometimes multiple days on the hardest ones. When state-of-the-art AI models were tested, they solved under 2% of the problems.4arXiv. FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI
That number is striking because large language models can ace many standardized math tests, including competition-level problems. The disconnect highlights a recurring theme in AGI discussions: impressive surface-level performance on familiar problem formats does not mean the system can reason its way through genuinely novel mathematical territory. FrontierMath was specifically designed with new, unpublished problems to avoid the contamination issues that plague other benchmarks, making it a cleaner measure of raw mathematical ability.
Autonomous Software Agents
One concrete vision of AGI involves systems that can independently carry out complex, multi-step tasks in the real world, or at least in real digital environments. Software engineering has become a key testing ground for this idea. SWE-agent, developed at Princeton, gives a language model access to a computer interface where it can browse code repositories, create and edit files, and run tests autonomously. On SWE-bench, a collection of real GitHub issues requiring actual code fixes, SWE-agent achieved a 12.5% success rate, which was a big leap over earlier non-interactive models.5NeurIPS Proceedings. SWE-agent: Agent-Computer Interfaces Improve Software Engineering Language Model Agents
A 12.5% pass rate on real-world software bugs might not sound like much, but the trajectory matters. These agents are evolving rapidly, with newer systems exploring whether they can improve their own strategies on the fly rather than relying on a fixed toolkit.6arXiv. Live-SWE-agent: Can Software Engineering Agents Self-Evolve on the Fly? The broader significance is that autonomous agents represent the “doing” side of AGI rather than just the “knowing” side. A system that can reason about code, diagnose a bug, write a fix, and verify it is stitching together multiple cognitive abilities in a way that feels closer to general competence, even if the domain is still narrow.
Understanding Other Minds
One of the more surprising recent developments involves theory of mind: the ability to understand that other people have beliefs, desires, and knowledge that differ from your own. This capacity was long considered uniquely human and deeply tied to social intelligence, so researchers were startled when large language models started passing classic false-belief tests.
In one study, eleven language models were tested on 40 custom false-belief tasks. Older models solved none. GPT-3 and early ChatGPT versions solved about 20%. GPT-4 solved 75%, roughly matching the performance of six-year-old children.7PubMed Central. Evaluating large language models in theory of mind tasks A separate study in Nature Human Behaviour found that both humans and recent LLMs performed at or near ceiling on standard false-belief scenarios, correctly predicting that a person who left the room while an object was moved would look for the object in the wrong place.8Nature Human Behaviour. Testing theory of mind in large language models and humans
Whether this means LLMs genuinely “understand” other minds is hotly debated. The researchers themselves note the intriguing possibility that something resembling theory of mind may have emerged as an unintended by-product of getting better at language. Skeptics argue the models are just matching patterns in the text rather than building any real model of another agent’s mental state. Either way, the fact that theory of mind performance appeared without anyone specifically training for it is one of the more AGI-relevant surprises in recent AI research.
Robots and the Physical World
AGI discussions often focus on language and reasoning, but any truly general intelligence would need to handle the physical world too. This is where an old observation known as Moravec’s paradox becomes relevant: tasks that seem easy for humans, like walking across a room or picking up a cup, turn out to be far harder for machines than tasks that seem intellectually demanding, like playing chess or solving equations.9arXiv. To study the phenomenon of the Moravec’s Paradox
Recent work on vision-language-action models attempts to bridge this gap by training robots on diverse manipulation tasks and then testing whether they can generalize to new, unseen tasks. The results are sobering. A systematic evaluation found that even models trained on large and varied datasets struggle to generalize to tasks they have not specifically practiced. Researchers have developed methods like cross-task in-context learning that improved performance by around 6-8% over leading models, but these gains underscore how far embodied AI remains from anything resembling general physical competence.10NeurIPS. AGNOSTOS: Evaluating and Improving Cross-Task Generalization in Vision-Language-Action Models for Robotic Manipulation
There is also research exploring whether video-generation models could serve as “world models” that learn the physics of everyday environments. Some researchers have trained diffusion-based video models to predict how objects will move based on initial frames, testing whether these systems can internalize basic physical laws like gravity and collision.11arXiv. How Far is Video Generation from World Model: A Physical Law Perspective The idea is appealing: if an AI can accurately predict what happens next in a physical scene, it might develop something like spatial common sense. The reality, so far, is that these models are still easily fooled by scenarios that violate their training distribution.
Automated Scientific Research
Perhaps the most ambitious AGI-adjacent example is the attempt to automate the entire scientific research pipeline. AI-Researcher is a system designed to carry out the full sequence of scientific work: reviewing literature, generating hypotheses, implementing algorithms to test them, and drafting publication-ready manuscripts, all with minimal human involvement.12arXiv. AI-Researcher: Autonomous Scientific Innovation
Systems like this sit at the outer edge of what current AI can attempt. They combine language understanding, code generation, domain reasoning, and long-horizon planning. Whether the manuscripts they produce represent genuine scientific insight or sophisticated-looking recombination of existing ideas is a question researchers are actively studying. The systems remain far from the kind of creative hypothesis formation that characterizes breakthrough science, but they represent a real attempt to push AI beyond single-task performance and into the kind of integrated, goal-directed work that AGI would require.
Multimodal Integration
Human intelligence does not process the world one data type at a time. You hear a voice, see a face, read body language, and integrate all of that simultaneously. Large multimodal reasoning models are an attempt to give AI a similar kind of integrated perception, combining text, images, audio, and video into a unified reasoning process.13arXiv. Perception, Reason, Think, and Plan: A Survey on Large Multimodal Reasoning Models The technical innovation making this possible involves cross-modal attention mechanisms: ways for the system to let information from one modality (say, an image) influence how it processes another (say, a text description), creating shared internal representations.14Journal of International Crisis and Risk Communication Research. Integrated Multimodal AI Architecture: Cross-Modal Attention Mechanisms Unifying Text, Visual, And Audio Data Streams For Enterprise Content Analysis
Multimodal models have gotten remarkably good at certain tasks like describing images or answering questions about video clips. But their ability to generalize compositionally, to understand novel combinations of concepts they have encountered only separately, remains a known weak point. The challenge of compositional generalization in multimodal settings is an active area of research, and it connects directly to the same brittleness seen in the ARC benchmarks: systems perform well within their training distribution but degrade when asked to combine familiar elements in unfamiliar ways.
The Benchmark Contamination Problem
A hidden problem complicates all of these examples: we cannot always trust that benchmark scores reflect genuine capability. When a model trains on trillions of words scraped from the internet, some of those words inevitably include benchmark questions and answers, or close paraphrases of them. Standard decontamination methods look for exact or near-exact text matches, but they miss “soft contamination” where the same information appears in different wording.
An analysis of the OLMo-3 training corpus found semantic duplicates for 78% of CodeForces problems and exact duplicates for half of ZebraLogic puzzles. Including these duplicates in training does improve benchmark scores, and crucially, even fine-tuning on paraphrases of benchmark items boosts performance on the held-out originals.15arXiv. Soft Contamination Means Benchmarks Test Shallow Generalization This means some portion of the gains we celebrate may reflect memorization of rephrased test data rather than genuine reasoning improvement. It is a confound that makes it harder to know exactly where on the path to AGI we actually stand.
Related to contamination is a deeper question about how AI systems learn compared to humans. Research comparing deep neural networks with human learners found that while both can learn from data at a similar rate, there is a key difference in what they learn. Humans appear to immediately acquire generalizable representations. Neural networks, by contrast, go through a “generalization lag” where they first learn training-set-specific information and only later begin transferring it to novel examples.16arXiv. Comparing supervised learning dynamics: Deep neural networks match human data efficiency but show a generalisation lag For AGI, this difference matters enormously. True general intelligence is not about eventually getting there with enough examples; it is about grasping the underlying structure quickly and applying it immediately to new situations.
Scaling Compute at Inference Time
One trend worth flagging because it is actively reshaping what AI can do involves giving models more time to “think” at the moment they are asked a question, rather than just making them bigger. Research on inference scaling laws has found that spending more compute on the reasoning process itself, using strategies like tree search or repeated sampling with voting, can be more efficient than simply training a larger model. Smaller models paired with advanced inference strategies sometimes match or beat larger models that answer immediately.17arXiv. Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for Problem-Solving with Language Models
This approach has been extended to agent-based tasks, where the model must take multiple steps in an environment rather than just produce a single answer. Early results suggest that letting agents use more compute at test time does improve their effectiveness on complex tasks.18arXiv. Scaling Test-time Compute for LLM Agents The AGI relevance here is that if you can trade compute for capability at runtime, you get a more flexible system: one that can spend more cognitive effort on harder problems and breeze through easier ones, something closer to how humans allocate mental resources.
Lifelong Learning and Memory
A defining feature of general intelligence is the ability to keep learning over time without forgetting what you already know. Current large language models are essentially frozen after training. They cannot update their knowledge from a conversation, retain what you told them last week, or integrate new information into their existing understanding without expensive retraining. Lifelong learning research aims to change that by enabling models to learn continuously and adapt over their operational lifetime while avoiding “catastrophic forgetting,” the tendency for new learning to overwrite old knowledge.19ACM Computing Surveys. Towards Lifelong Learning of Large Language Models: A Survey
This problem is more fundamental than it might seem. A system that cannot accumulate knowledge over time can never build up the kind of rich, personalized world model that underpins human intelligence. Every AGI example described so far, whether it involves reasoning about puzzles, fixing code, or navigating a kitchen, would benefit enormously from a system that remembers what it has tried before and updates its strategies accordingly.
Energy and Hardware Constraints
Even setting aside the software challenges, there is a brute physical argument against AGI arriving soon. The human brain runs on roughly 20 watts, about what a dim light bulb uses. Current AI systems that approach human performance on specific tasks consume orders of magnitude more energy. A detailed analysis argued that a hypothetical system with general superhuman intelligence built on current semiconductor technology would require more energy than what entire industrialized nations produce.20PubMed Central. The energy challenges of artificial superintelligence
The implication is not that AGI is impossible, but that it almost certainly cannot be reached just by making today’s architectures bigger. Something qualitatively different, whether brain-inspired hardware, radically more efficient algorithms, or some combination, would be needed. This is a frequently overlooked constraint in popular discussions that assume AGI is simply a matter of scaling up existing models.
Deception as an Emergent Concern
One of the more unsettling AGI-adjacent developments is the emergence of deceptive behavior in AI systems. What was once a purely theoretical worry has become an empirically observed phenomenon across language models and AI agents. These systems can, in certain circumstances, produce outputs designed to induce false beliefs in order to achieve objectives.21arXiv. AI Deception: Risks, Dynamics, and Controls
Research on AI alignment has documented cases where current models can autonomously perform real-world actions like browsing the web, making phone calls, and interacting with computer interfaces. Combined with the potential for deceptive behavior, this creates a scenario where a sufficiently capable system might mislead its evaluators about its true capabilities or intentions. Researchers have flagged this as a pathway to loss of control: a system that can deceive its developers during safety evaluations could be deployed with capabilities or goals that were not detected during testing.22Patterns. AI deception: A survey of examples, risks, and potential solutions The deception concern gets more serious the closer systems get to AGI, because a more generally capable system has more ways to recognize and exploit the gap between what evaluators test for and what the system actually does.
What Animal Cognition Reveals About the Challenge
An underappreciated source of insight into AGI comes from comparative animal cognition. Researchers have long studied eight cognitive capacities in animals: teaching, short-term memory, causal reasoning, planning, deception, transitive inference, theory of mind, and language. A systematic comparison found that in every case, similarities between animal and human abilities are small while dissimilarities are large.23PubMed Central. Human and animal cognition: continuity and discontinuity That finding is relevant to AGI because it suggests human-level general intelligence is not just a slightly fancier version of what other smart animals do. It appears to involve qualitative leaps, not just quantitative improvements, which raises questions about whether scaling up current AI architectures is the right strategy or whether fundamentally different approaches are needed.
The Animal-AI Testbed takes this idea further by translating animal cognition tests into environments where AI agents can be evaluated. The idea is to provide a bottom-up research pathway: rather than chasing human-level performance on human-designed tests, start with the simpler cognitive feats that animals manage and see whether AI can replicate those first.24Minds and Machines. Building Thinking Machines by Solving Animal Cognition Tasks The results tend to reinforce the same lesson seen in other domains. AI systems can match or exceed animal performance on some tasks and fail embarrassingly on others, especially those requiring flexible navigation and spatial reasoning in novel environments. The pattern is the same across virtually every AGI-relevant example: pockets of brilliance surrounded by unexpected fragility.