Is This AGI? A Sober Look at What AI Still Can't Do
New AI demos feel like magic, sparking claims of AGI. But underneath the impressive fluency, the models lack true understanding, a world model, and the ability to innovate. They are powerful simulators, not thinkers.
May 18, 2026 · 2 min read · SuperThinking team
The latest AI demos are stunning. GPT-4o's real-time, emotive voice conversation feels like something straight out of the movie Her. We see AI coding entire web apps from a sketch and acing graduate-level exams. It's easy to see why the question is bubbling up again: is this it? Is this AGI?
No. And the distinction isn't just academic hair-splitting. Understanding what these models are—and what they aren't—is critical to using them well and avoiding some serious pitfalls.
They aren't nascent general intelligences. They are simulators. Incredibly sophisticated, world-scale simulators of text, images, and sound, but simulators nonetheless.
The Seductive 'Sparks of AGI'
It's easy to see why people get swept up. When you ask a model to explain quantum mechanics in the style of a pirate, and it does so coherently, it feels like understanding. When it refactors your messy Python script into something elegant and efficient, it feels like reasoning.
Researchers from Microsoft even published a paper titled "Sparks of AGI" based on early access to GPT-4, noting its surprisingly human-like performance on a range of tasks. They argued its capabilities were so broad and deep that it was showing early signs of general intelligence.
And they have a point. The sheer scale of the training data means these models have encoded patterns of human thought, logic, and creativity. They can connect disparate concepts, mimic complex writing styles, and generate novel-sounding ideas because they've seen virtually every permutation of human expression on the internet. It's a masterful illusion of thought.
But an illusion is all it is. The performance is a brilliant feat of high-dimensional pattern matching, not a demonstration of comprehension. The model doesn't know what a pirate is, or what quantum mechanics entails. It knows what words typically follow other words in contexts discussing those things.
Where the Simulation Breaks Down
The cracks in the facade appear when you push the models beyond the patterns they were trained on. Their failures reveal a fundamental lack of what we consider intelligence.
First, they have no persistent world model. They don't understand cause and effect, the laws of physics, or the constraints of reality. Ask one to create a recipe for a cake made of iron. It might generate a plausible-sounding recipe, but it has no internal concept that iron is inedible or can't be creamed like butter. It's just remixing recipe-like text with iron-related words.
Second, they lack true agency or continuous learning. Each interaction is largely stateless. The model doesn't remember your last conversation or learn from its mistakes in a permanent way. It's not building a cumulative understanding of you or the world. All the