AGI Isn't Here. You're Asking the Wrong Question.

Everyone wants to know if we’ve reached Artificial General Intelligence. Is GPT-4o or Claude 3 Opus secretly sentient? Did we finally build the machine that thinks?

No. And honestly, it’s a boring question.

It’s a distraction from the much more interesting reality: we’ve built tools that are superhuman at specific, valuable tasks. Worrying about AGI is like seeing the first steam engine and asking when it will write a sonnet. The point isn't that it can do everything a human can do; the point is that it can do a few things with more power than any human ever could.

The 'Sparks' of Intelligence Are Real

Let's be clear: what current models can do is astonishing. If you showed someone GPT-4o's real-time conversational ability just three years ago, they’d call it science fiction. The models are getting so good at mimicking human thought patterns that we call it “reasoning” because we don’t have a better word for it.

They excel at tasks that require synthesizing vast amounts of information or generating structured text based on complex rules. Here's where they smoke an average human:

• Code Generation: Need a Python script to scrape a website and dump the data into a CSV? A model can write that in 15 seconds, complete with comments and error handling. A junior developer would take 30 minutes. An expert might be faster, but the AI is available 24/7 for zero marginal cost.
• Summarization: You can feed Claude 3 Opus a 150-page technical paper and ask for a five-bullet-point summary for your boss. It won’t get bored, miss key details, or misunderstand the jargon. It just does it.
• Brainstorming: Stuck on naming a new product? Give a model three keywords and ask for 50 options. Ninety percent will be garbage, but five will be better than what your team would come up with in an hour-long meeting.

These aren't parlor tricks. This is genuine utility. The models are incredibly good at manipulating symbols and predicting the next most likely token. That process, scaled across a massive dataset, looks a lot like intelligence. But it isn't.

The Uncanny Valley of Reasoning

For all their power, models still fail in ways that reveal they aren't thinking—they're just completing patterns. Their failures are often subtle, weird, and deeply non-human.

First, they have no persistent memory or true learning ability. Each interaction is a blank slate. You can spend an hour teaching a model a new concept, but the moment that context window closes, it's gone. It's like talking to a genius with severe amnesia. It doesn’t build on past conversations to form a world model. It just has the context you provide right now.

Second, they have no physical grounding. An AI can describe the color red, but it has never seen it. It can write about peeling an orange, but it has never felt the zest on its fingers or the sticky juice. This lack of embodiment leads to hilarious and sometimes dangerous common-sense failures. It might suggest using a metal fork to get toast out of a plugged-in toaster because it doesn't understand electricity, metal, and the human nervous system on a physical level.

Third, they are bad at causal reasoning. They know that the rooster crowing and the sun rising are correlated, but they don't know which one causes the other. Ask a simple logic puzzle, and you can see the wheels come off.

Here’s a classic I tried on a top-tier model:

A man is looking at a portrait. Someone asks him whose portrait he is looking at. He replies, "Brothers and sisters I have none, but that man's father is my father's son." Who is in the portrait?

The correct answer is his son. The model often gets tangled, sometimes answering