We Keep Asking 'Is It AGI?' — It's the Wrong Question

Everyone wants to know if we’ve reached Artificial General Intelligence. Is GPT-4o sentient? Does Claude 3 have a soul? It’s a fun debate for Twitter and philosophy class, but for developers trying to build things, it’s a total distraction.

The problem is the definition. We don't have a stable, universally agreed-upon definition of AGI. The goalposts are always moving. Ten years ago, playing Go at a world-champion level felt like AGI. Now it's a solved problem. We're chasing a mirage.

Let's ask a better question: what can these tools do that feels like a step-change in capability? And where do they still fall flat on their face? That’s where the interesting work is.

Where Models Feel Like Magic

Sometimes, the latest models are indistinguishable from a very smart, very fast human intern. The flashes of brilliance are in tasks that require synthesizing multiple domains of knowledge or modalities.

I recently gave GPT-4o a screenshot of a complex UI from a data visualization app. My prompt was just: "Build this in React with Tailwind CSS using dummy data." It didn't just replicate the layout. It correctly inferred the component hierarchy, created plausible-looking JSON for the dummy data, and even added comments explaining the state management.

This wasn't just pattern matching. It required understanding vision (the screenshot), front-end development (React), styling (Tailwind), and data structures (JSON). That's a generalist skill.

We see the same thing in complex code translation. You can throw a 500-line Python script that relies heavily on Pandas at Claude 3 Opus and ask for a performant Go equivalent. It will reason about data types, find idiomatic Go libraries for data manipulation, and structure the code with concurrency in mind. This is way beyond a simple syntax conversion.

Here’s what these magical moments have in common:

• Cross-Domain Synthesis: They blend distinct areas of expertise, like design, code, and data.
• Implicit Intent: They understand the goal behind the request, not just the literal words.
• High-Level Abstraction: You can operate at the level of ideas ("make this dashboard") instead of painstakingly specific instructions.

When you're in the flow with a model like this, the AGI question feels temporarily answered. It feels like a collaborator.

The Hard Limits That Prove We're Not There

But then, you ask it to do something simple that requires a shred of real-world experience, and the illusion shatters. The brilliant intern suddenly reveals they have never actually lived on planet Earth.

The most glaring failure is physical common sense. Ask a model for five ways to get a key that fell down a storm drain. It will give you textbook answers: use a magnet, a stick with tape, etc. Ask it to troubleshoot why your magnet isn't working, and it's lost. It has no concept of the weight of the key, the strength of the magnet, or the viscosity of the murky water. It's a database of facts without the physics engine of reality.

Long-term planning is another wall. You can't ask an LLM to "manage the launch of my new app over the next three months." It can generate a perfect-looking project plan. But it can't execute it. It can't remember a key decision from last Tuesday unless you painstakingly remind it in the context window. It has no persistent memory or agency. It lives one prompt at a time.

And then there's true creativity. Models are phenomenal at remixing and style transfer. "Write a sea shanty about API documentation." You'll get something clever and hilarious. But it's working from the patterns of every sea shanty it's ever read. It can't invent a new genre of music. It's the world's greatest cover band, but it doesn't write its own songs.

Forget 'General Intelligence', Seek 'Actually Solvable' Intelligence

The trap is thinking the goal is to build a perfect digital human. It's not. The goal is to build tools that solve problems. These models aren't AGIs, and they aren't going to be anytime soon.

They are, however, the most powerful thinking tools we've ever created. They are leverage for your own intelligence. They can draft code, summarize research, brainstorm ideas, and automate tedious crap, all at a scale that was science fiction two years ago.

So stop asking if the calculator is sentient. Stop worrying about the philosophical definition of AGI. Start asking, "What can I build with this that was impossible for me last year?" The answer is probably "a lot".