Is AGI Here? A Reality Check on Today's AI Models

Every time a new model drops, the AGI question bubbles up again. Is this the one? Is this the spark of Artificial General Intelligence that sci-fi promised and doomers fear?

The short answer is no. The slightly longer answer is also no, but it's getting complicated.

The real answer is that the term 'AGI' has become so fuzzy it's almost useless. We're seeing something new and powerful emerge, but it doesn't look much like the human-style intelligence we were expecting.

The Goalposts Keep Moving

Remember when beating a grandmaster at chess was the benchmark for machine intelligence? IBM's Deep Blue did that in 1997. Then it was Go, a game with more possible moves than atoms in the universe. DeepMind's AlphaGo cleared that hurdle in 2016.

Each time we clear a benchmark, we decide it wasn't the real test of intelligence after all. The target shifts. Now, AGI is often vaguely defined as the ability to perform any intellectual task a human can. This includes reasoning, common sense, and understanding the physical world.

It’s a definition based on a negative. AGI is whatever today's models can't do. That’s a shaky foundation for a technical term.

Where The Magic Happens (and Fades)

Let’s be specific. These models are incredible pattern-matching engines. If a task involves manipulating symbols based on statistical relationships learned from a trillion words of text, they excel.

What does that mean in practice?

• Code Generation: They can write flawless boilerplate for a React component or a Python script to parse a CSV. They’ve seen millions of examples and know what comes next.
• Summarization: They can ingest a 100-page report and spit out the key takeaways because summary is a pattern.
• Translation: They can translate between languages with more nuance than ever before, capturing idioms and tone.

Here’s a simple prompt that GPT-4o handles perfectly:

Write a python function that takes a list of URLs and returns a dictionary where keys are the URLs and values are the HTTP status codes.

The model produces correct, idiomatic Python. It's a task that would have taken a junior developer 15 minutes of searching and testing, now done in 5 seconds. This feels like magic.

But the magic fades when you step outside the world of text and symbols. Ask a model a simple question that requires physical or spatial reasoning, and the illusion shatters.

Consider this prompt:

I have a shoebox. I put a baseball inside the shoebox. Then I put the shoebox inside a backpack. Where is the baseball?

A modern LLM gets this right. But let's add a trivial twist.

I have a box made of clear acrylic. I put a red ball inside. I close the box. You are looking at the box. Can you see the red ball?

Again, it gets this right. It has learned the statistical association between