Is AGI Here? Who Cares. Let's Get Real About AI.

Every time a new model drops, the question floods social media: "Is this AGI?" It’s a boring question. The goalposts for AGI move constantly, and the debate is more philosophy than engineering. It's a distraction.

The better question for anyone building things is: What can this tool actually do for me today, and where are the sharp edges I need to watch out for? Let's look at the real-world capabilities and the persistent, frustrating limitations of models like GPT-4o and Claude 3.5 Sonnet.

The 'Sparks of AGI' Argument

To be fair, you can see why people get excited. The latest models have moments of genuine brilliance that feel like a step-change. When you watch GPT-4o take a screenshot of a website and write the React code to replicate it, that feels like magic. It’s pattern matching at a god-like level.

These models can plan. You can give them a high-level goal like, "Refactor this Python script to be more modular and add error handling," and they can break it down into steps: read the file, identify functions, create new files for classes, rewrite the main script to import them, and add try/except blocks. For a self-contained task, the results can be shockingly good.

Where they excel is in translation between domains. This is the core of their power.

• Image to Code: "Here's a drawing of a UI. Make it HTML and Tailwind CSS."
• Natural Language to SQL: "Get me the total sales for our top 5 customers last quarter."
• Code to Explanation: "Explain this gnarly regex in plain English."

This ability to fluently move between human language, programming languages, and visual information is a superpower. It's why they feel intelligent. But this fluency masks a profound lack of actual understanding.

The Hard Reality: Where It All Breaks Down

The gap between a brilliant demo and a reliable production system is where the AGI talk dies. Today's models fail in ways that a truly general intelligence wouldn't. Their failures aren't about lacking knowledge; they're about lacking a coherent world model.

First, they have no physical grounding. A model can tell you how to change a tire, step by step, with perfect clarity. But it has never felt the weight of a lug wrench, the stubborn resistance of a rusted nut, or the slightly wobbly feel of a car on a jack. This lack of embodied experience means it can't reason from first principles about the physical world. It can only regurgitate and remix text that describes that world. This leads to subtle but critical errors when dealing with real-world mechanics or robotics.

Second, their memory is a joke. Yes, context windows are huge now—a million tokens! But a large context window isn't the same as memory. It's like being able to read a very, very long page of notes you can never add to. Models constantly forget earlier instructions and lose the plot in complex, multi-step tasks. Try to get an agent to refactor an entire 10-file codebase. It will start strong, then get lost, forget the overall architecture, and start introducing circular dependencies. It can’t hold a mental map of the system.

Third, they are maddeningly inconsistent. The same prompt can produce a brilliant answer one minute and a lazy, incorrect one the next. They hallucinate API endpoints that don't exist. They get stuck in repetitive loops. You simply cannot trust them for mission-critical, unattended tasks. An AGI you have to babysit isn't very general or intelligent.

A Better Framework: The Universal Simulator

So, if not AGI, what are they? A better mental model is to think of them as universal simulators. They can simulate the patterns of any text-based task you throw at them. They can simulate a helpful programmer, a snarky critic, a SQL database, or a JSON formatter.

This framing is more useful because it puts the responsibility back on you, the user. Your job is to define the simulation you want to run. This is what prompt engineering is really about: providing a clear specification for the