Have We Reached AGI? It's the Wrong Question to Ask

Every time a new model drops, the AGI question bubbles up again. Is this the one? Did OpenAI or Google finally build a machine that thinks?

It's a fun debate for Twitter, but for builders, it's a total distraction. The real question isn't whether a model is "general intelligence." The real question is: where is it superhuman, and where is it dumber than a bag of hammers?

Because right now, AI is both.

The Superhuman Intern

The best way to think about models like GPT-4o or Gemini 1.5 is as a superhuman intern. They have capabilities that no human could ever match, combined with a baffling lack of common sense.

First, the superhuman parts. Their context windows are a game-changer. Gemini 1.5 Pro can digest a million tokens. That's not just a big number; it means you can drop an entire codebase into the prompt and ask it to find bugs, write documentation, or refactor a core module. A human would need days or weeks to load that much context into their brain.

Their multi-modality is another superpower. You can show GPT-4o a screenshot of a web app and have it write the React code to build it. You can have it watch a video of a physics lecture and generate quiz questions. This fluid jump between domains—vision, code, text, audio—is incredibly powerful. No human has expert-level, instantaneous recall across all these modalities.

For example, you can give it a messy, uncommented Python script and a vague goal:

# This is my script for processing user data
def process(d):
    res = []
    for i in d:
        if 'email' in i and '@' in i['email']:
            if i.get('active', False):
                nm = i.get('name', 'N/A')
                res.append({'name': nm, 'email': i['email']}) 
    return res

Your prompt: "Refactor this to be more Pythonic and handle edge cases better. Use type hints."

The model will instantly spit out a cleaner, safer, and more readable version without needing a 30-minute meeting to clarify requirements.

This is the intern at its best: infinitely fast, with perfect memory of every textbook and Stack Overflow post ever written. For well-defined, self-contained tasks, it's an incredible force multiplier.

Where the 'G' in AGI Falls Apart

But this intern has zero life experience. It has never existed in the physical world, never had to plan a project over a week, and has no real memory of your last conversation.

This is where the "general" part of AGI completely breaks down. Models lack three critical things:

• Embodied Reasoning: They don't understand the physical world. Ask a model a simple question like, "If I put my socks in a drawer, then put my shoes in the same drawer, and then take the socks out, are my shoes still in the drawer?" It will likely get it right, but it's reasoning from text patterns, not a mental model of physical objects and containment. This is why AI-powered robots are still so clumsy.
• Long-term Agency: An LLM can't manage a goal over time. You can't tell it, "Your job for the next month is to refactor our user authentication service. Report back with progress every day." It lives in a stateless world of prompts and responses. Each interaction is a cold start. Agentic frameworks try to fake this with loops and external memory, but it's a brittle simulation of true persistence.
• Robust Causal Intuition: Models are correlation machines. They are incredibly good at identifying what text should come next based on the trillions of examples they've seen. They are terrible at understanding why things happen. This makes them unreliable for tasks requiring deep, first-principles thinking, like debugging a novel systems-level problem or designing a new type of database.

They can write a perfect five-paragraph essay on the causes of the French Revolution, but they can't invent a new way to bake a potato because they have no causal model of heat, starch, and moisture.

Forget AGI, Aim for Practicality

So, no, we don't have AGI. We have incredibly powerful pattern-matching engines that can perform isolated intellectual tasks at a superhuman level.

Instead of waiting for a mythical AGI, we should focus on what I call "practical AGI." This means building systems that intelligently chain together the superhuman capabilities of today's models while using human oversight to patch over their weaknesses.

This looks like multi-agent workflows where one AI writes code, another reviews it for security flaws, and a third generates documentation, all with a human in the loop to handle the high-level strategy and common-sense checks. It's about building a team where the models are your specialist interns, not your CEO.

The real breakthrough won't be a single, all-knowing model. It will be the clever systems we design to orchestrate these powerful but flawed tools to do useful work.