Have We Hit AGI Yet? A Sober Look at GPT-4o and Friends

Every time a new flagship model drops, the question bubbles up again. After the GPT-4o demo, with its real-time, emotive voice and uncanny awareness, the shouts got louder: "This is it. This is AGI."

Let's cut to the chase. No, it's not.

But that's the wrong question. It's a distraction that keeps us from asking the more important one: what can we build with this incredibly powerful new tool that feels, for the first time, like a real partner?

The Seductive 'Sparks of AGI'

It's easy to see why people get swept up. The new wave of multimodal models performs feats that were science fiction a few years ago. You can have a spoken conversation with an AI, show it a live video of your coding problem, and have it talk you through the fix. It feels intelligent.

When a model can look at a chart of your website's analytics and you can ask, "What's the most surprising thing here?" and it gives you a genuinely insightful answer, that feels like reasoning. When it can write 300 lines of clean, functional Python from a one-sentence request, that feels like expertise.

These moments are the "sparks" people talk about. They are tasks that require a synthesis of different skills: vision, language, code, and a massive repository of world knowledge. The models are no longer just text-in, text-out. They are becoming general-purpose problem solvers for a specific set of digital tasks.

For example, I recently planned a weekend trip by talking to the ChatGPT app. I showed it a screenshot of a map, a picture of a menu, and told it my budget. It stitched everything together into a coherent itinerary. This workflow was impossible a year ago. Today, it's trivial.

This is where the magic is. Not in some philosophical notion of consciousness, but in the radical expansion of what's possible in a human-computer interface.

Where the Simulation Breaks Down

For all their brilliance, these models are still brittle simulators of intelligence, not the real thing. Once you know where to push, the illusion shatters pretty quickly.

The biggest giveaway is their lack of a true world model. They don't understand physics, causality, or logic from first principles. They've just read every book and webpage that describes them, and they are masters of statistical pattern matching.

This leads to fundamental errors that a human child wouldn't make. Ask a model a trick question about object permanence or a simple logic puzzle phrased in an unusual way, and it will often confabulate a plausible-sounding but completely wrong answer. It parrots the form of reasoning without the substance.

Here are the other big limitations:

• Inconsistency: You can ask the same complex question twice and get one brilliant answer and one mediocre one. Rephrasing a single word in your prompt can send the model down a completely different, and often worse, path. True intelligence is far more robust.
• No Continuous Learning: A model's knowledge is frozen in time. It doesn't learn from your conversations. The next time you start a chat, it has no memory of you or your previous interactions outside the current context window. It can't grow or adapt with you.
• Lack of Agency: This is the big one. The model doesn't want anything. It has no goals, no curiosity, no drive. It is a passive tool that executes a command: predict the next most likely token. Without internal motivation, you can't have general intelligence. It will never get bored and decide to invent something new on its own.
• Confabulation (aka Lying): When they don't know something, models don't say "I don't know." They make things up with incredible confidence. This is a direct result of their design. They are built to generate plausible text, not to verify truth.

These aren't minor bugs to be patched. They are fundamental architectural limitations of the current transformer-based approach. Getting past them will likely require a new paradigm, not just a bigger model.

The Wrong Debate

So, why does the AGI distinction even matter to us as developers?

Because fixating on AGI makes us evaluate these tools against a mythical, sci-fi standard. It's like refusing to use a calculator because it can't write a novel. We should be judging them by what they enable, not by what they aren't.

Forget Artificial General Intelligence. Think of it as an Artificial Specialist Intelligence. We now have access to a universal intern that is an expert in nearly any digital domain. It's a world-class programmer, a decent copywriter, a tireless data analyst, and a pretty good graphic designer, all rolled into one API call.

The right mindset isn't "how do I build a conscious machine?" but "how can I automate specific, high-value tasks by chaining together these powerful specialist skills?"

This is where agentic workflows come in. You give the model a high-level goal, a set of tools (APIs, a file system, a web browser), and a feedback loop. The model can then plan and execute a series of steps to achieve the goal. It's not AGI, but it's an incredibly effective way to automate complex work.

Don't wait for AGI. The tools we have right now are more than powerful enough to build things that were previously unimaginable. The question isn't whether the machine is thinking. The question is what you're going to build with it.