AGI Isn't Here. So What Did We Build Instead?

No, we don't have Artificial General Intelligence. Not even close.

But for the first time, it feels like a serious question to ask. Models like GPT-4o and Claude 3 Opus can write elegant code, analyze charts you upload, and even hold a reasonably coherent spoken conversation. It's an uncanny leap from the clunky bots of just two years ago.

When a tool gets this good, our brains want to anthropomorphize it. We start seeing intention, understanding, and maybe even consciousness. This is a trap. What we've built isn't a nascent mind; it's a universal pattern-matching machine. And knowing the difference is the key to actually using it well.

The 'Holy Cow' Moments Are Real

Let's be clear: these models are ridiculously capable. They can do things that were squarely in the domain of sci-fi a few years ago. You can now throw a blurry screenshot of a web app dashboard at an AI and get back working React code to replicate it. That's wild.

GPT-4o's ability to reason across modalities—text, images, audio—in real-time is a genuine breakthrough. You can show it a math problem on a piece of paper and have it walk you through solving it, verbally. It can look at a live video of your room and comment on your decor.

On the pure text and code front, the context windows have become massive. With Claude 3, you can drop in a 150,000-word novel or an entire codebase and ask it questions. For example, you can give it a complex open-source library and ask:

Based on the entire codebase provided, identify the primary data structures used for user session management. Point me to the exact files and line numbers where these are defined and manipulated.

Getting a correct, detailed answer to that in seconds is a superpower for any developer. It's an incredible accelerator for comprehension and summarization. It feels like having a brilliant, infinitely patient intern who has already read everything.

These models excel at synthesizing and transforming information that follows predictable patterns, which, it turns out, covers a huge amount of professional work. Writing marketing copy, refactoring code, translating languages, summarizing meetings—they are world-class tools for these tasks.

But the intern is an amnesiac. And has zero common sense.

The Glaring, Obvious Holes

The illusion of intelligence shatters the moment you push the models outside the web of patterns they were trained on. They don't understand the world; they've just memorized a compressed version of how humans talk about the world.

Here are the big tells:

• No Real-World Model: They have no intuitive grasp of physics, space, or causality. A classic test is a simple riddle: "I have a bag, and inside the bag is a box. Inside the box is a book. What is inside the bag?" While top models usually get this now, they used to fail spectacularly. A slightly harder spatial problem can still trip them up. They don't reason from first principles; they find the answer that statistically looks most like correct answers they've seen before.

• Logical Brittleness: They can solve complex calculus problems but then fail a simple logic puzzle that a child could solve. The reasoning is shallow. Give it a slightly weird premise and watch it go off the rails. For example:

There are three brothers: Alex, Ben, and Carl. Alex is not the oldest. Carl is two years younger than the oldest brother. Ben is not the youngest. Who is the oldest?

Top models might get this right, but they can also reason themselves in circles, claiming Ben is the oldest because Carl is younger than the oldest, and Alex isn't the oldest, therefore Ben must be. The path it takes to an answer is often nonsensical, even if the final answer is right.

• No Persistent Learning: This is the biggest one. You can't teach it something new. Not really. Every conversation you have starts from a blank slate. It doesn't remember your previous interactions, your preferences, or the feedback you gave it yesterday. An AGI would learn and grow over time. LLMs are permanently frozen at the end of their last training run.

This is why they feel less like a colleague and more like a tool. An amazing tool, but a tool nonetheless.

A Better Mental Model: The Universal Simulator

So if it's not a thinking machine, what is it?

The best description I've heard is that it's a universal simulator. It was trained on a vast corpus of human-generated text and images, and its core function is to calculate the most probable next word or pixel given a sequence. It is simulating the kind of text that would appear next in its training data.

When you ask it to write Python code, it simulates a developer. When you ask it for a recipe, it simulates a food blogger. When you ask it to be a Socratic tutor, it simulates one. It's the ultimate role-player, a master of stylistic imitation without any underlying understanding of the roles it plays.

Thinking of it this way is more productive. Don't ask it if your business idea is good—it has no concept of what "good" means. Instead, ask it to "act as a skeptical venture capitalist and list three major risks in this business plan." You're channeling its pattern-matching ability toward a useful simulation.

It's a mirror of our collective knowledge, a powerful summarizer, and a creative tool for brainstorming. But it's not your peer. It doesn't know what it's saying. Use its strengths, but be deeply aware of its strange, inhuman limitations.