AGI Isn't Here. Here's How We Know.

No, Artificial General Intelligence is not here. Not even close.

Every few months, a new model drops and the cycle repeats. Impressive demos, claims of 'sparks of AGI', and a tsunami of hot takes. It's easy to get caught up in it. When a model can write a React component from a blurry photo of a napkin sketch, it feels like magic.

But that magic is a specific, narrow kind. It's the magic of a system that has ingested a significant fraction of all human text and code. It is an autocomplete on a cosmic scale, a pattern-matcher so vast it can convincingly replicate the forms of human intelligence. But it's not the real thing.

To see why, you have to stop giving it problems it has seen before and start giving it problems that require a genuine, internal model of the world.

The Real-World Reasoning Test

LLMs have no hands, no eyes, and no body. They don't exist in the physical world and have zero intuitive grasp of physics, space, or causality. They've read about gravity, but they've never felt it.

Give any top-tier model, from GPT-4o to Claude 3 Opus, a simple physical reasoning puzzle. Something a ten-year-old could solve by just picturing it.

For example:

"I have a cardboard shoebox, a bowling ball, and a glass vase. I place the bowling ball inside the shoebox. I then place the shoebox on a small, wobbly table. I put the vase on top of the shoebox. Is this setup stable? What is likely to happen?"

Models will often get this wrong in subtle but crucial ways. They'll say something generic like, "This setup may be unstable because the table is wobbly." They identify keywords like 'wobbly' but miss the core, obvious physical reality: the bowling ball is heavy, the cardboard shoebox is weak, and it's going to collapse under the weight, likely before the table's wobbliness even matters.

They are reasoning from text associations, not from a mental simulation of the objects. It's a fundamental gap. They know the word 'heavy' is often associated with 'unstable', but they don't understand heaviness.

The Novelty and Abstraction Test

Another place these systems crumble is with truly novel problems. Not just a harder version of a coding challenge they've seen, but a problem with rules that don't exist anywhere in their training data.

Consider a simple logic game:

"In the game of 'Fleep', a 'glarp' can 'snorf' a 'flumph'. A 'flumph' can 'snorf' a 'glarp' only if the 'flumph' is blue. All 'glarps' are green. Given a green 'glarp' and a blue 'flumph', who can 'snorf' whom?"

This is trivial for a human. The glarp can snorf the flumph. The blue flumph can also snorf the glarp. It’s a two-way street.

LLMs often get tangled in these. They might fixate on one part of the rule set or incorrectly apply the conditions. Why? Because there's no corpus of 'Fleep' game logs to draw from. They can't fall back on pattern recognition. They have to reason from first principles, and their ability to do that is brittle.

This reveals the core of the issue: they are brilliant interpolators, but poor extrapolators. They can operate flawlessly within the space of their training data, but step one foot outside, and they're lost.

So, What Do We Actually Have?

This isn't to say today's models aren't useful. They are incredibly, transformatively useful. But we need to be precise about what they are. They are not nascent general intelligences.

They are universal simulators. Or maybe 'calculators for words'.

They can:

• Summarize and synthesize text: Condensing a 10,000-word report into 5 bullet points is a superpower. It's a task of pure textual manipulation, and they excel at it.
• Generate boilerplate code: Need a Flask server with a basic endpoint? An LLM can spit that out in seconds, saving you 15 minutes of typing. It's seen millions of examples.
• Brainstorm and reframe ideas: Stuck on a marketing angle? Ask an LLM for 20 different headlines. Most will be bad, but a few will spark a new direction. It's a great tool for unblocking creativity.

Think of it like a fantastically advanced search engine that synthesizes the results for you. It's a tool, not a colleague. A very powerful bicycle for the mind, but you're still the one steering.

Don't Wait for AGI

The path from today's models to AGI isn't a straight line. Simply scaling up the current transformer architecture with more data and more compute likely won't bridge the gap between pattern matching and genuine understanding. We're probably missing a fundamental architectural idea.

Maybe it requires embodiment, a true sensory connection to the world. Maybe it requires a completely different approach to memory or causal reasoning. Nobody knows.

So, stop worrying about whether the next model is AGI. The answer will be 'no' for a while yet. Instead, focus on what these powerful, weird, and deeply flawed tools can do for you right now.