Is It AGI Yet? A Sober Look at GPT-4o and Friends

Everyone's asking if we've reached AGI. It's the wrong question.

Artificial General Intelligence isn't a light switch that flips from 'off' to 'on'. It's not a single event. It's a spectrum of capabilities, and we've been sliding along that spectrum faster than anyone predicted.

The real question isn't "Is it AGI?" The real question is, "What can these things do today that feels like general intelligence, and what's still just a clever party trick?"

So let's get real about what models like GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro can and can't do. No hype, just the facts.

The 'G' is Getting General

The most mind-blowing progress is in the 'General' part of AGI. A few years ago, an AI that was good at writing code was useless at analyzing poetry. A model that could summarize a medical paper couldn't plan a vacation. They were specialists.

That's over. The top models are now true generalists. You can throw wildly different tasks at the same endpoint and get shockingly good results. One minute, you're asking it to refactor a Python script to be more idiomatic. The next, you're uploading a screenshot of a weird-looking plant in your garden and asking for care instructions.

This cross-domain reasoning is the new superpower. I recently gave GPT-4o a terribly formatted text file of sales data, a vague goal of "show me what's interesting," and the instruction to write Python code to analyze it. It correctly inferred the data columns, cleaned the messy text, imported Pandas, performed a meaningful analysis, and generated a plot. This required language understanding, data sense, and coding skill all at once.

That feels pretty general to me.

Where It All Falls Apart

For all their brilliance, these models have a critical blind spot: the real world. They have no grounding in physical reality, and it shows.

They lack persistent memory and a consistent model of the world. Ask one to help you plan a multi-day project. It will give you a great outline for Day 1. When you come back to plan Day 2, it often has no stable memory of what it told you before. It might contradict itself, forget key constraints, or hallucinate progress.

Common sense is another huge gap. A model can write a dissertation on Newtonian physics, but it might also suggest you can boil water faster by adding an ice cube. It knows the text associated with concepts, but it doesn't understand them in a causal, physical way.

We see this constantly when trying to build agents. Getting an LLM to call a series of tools in the right order is a nightmare. It will try to use a tool's output before it has even called the tool. It will forget a critical piece of information it gathered three steps ago. It's like working with a genius who has zero object permanence.

A Practical Scorecard

Instead of a binary yes/no, let's score the models on a few practical 'AGI-like' skills. Here's my take:

• Cross-Domain Synthesis: A. Can connect ideas from different fields, like using a biological metaphor to explain a software concept.
• Complex Instruction Following: B-. Good with a few constraints, but add too many steps or negative constraints ("do X but don't do Y") and it gets lost.
• Autonomous Planning: D. Can generate a plausible-looking plan, but cannot reliably execute it or adapt to unexpected outcomes without heavy scaffolding from code.
• Physical Common Sense: F. Has none. It's all just statistical correlation. Don't ask it for advice on wiring a plug.
• Learning From Feedback: C+. You can correct it within a single session, but it doesn't truly learn for the next conversation. It's just using your correction as new context for a bit.

As you can see, it's a very mixed bag. We have A-level performance in some areas and complete failures in others.

The 'Sparks' Are Not the Fire

Microsoft's "Sparks of AGI" paper was a landmark, showing that GPT-4 exhibited unexpected, emergent behaviors that weren't explicitly trained. The model showed glimmers of theory of mind, self-correction, and abstract reasoning. These are definitely sparks.

But a spark isn't a fire. Having these capabilities emerge in a sandboxed, conversational context is incredible. It is not the same as having a self-motivating, continuously learning intelligence that can form its own goals and operate in the world.

The current architecture of transformers, which predicts the next token in a sequence, is fundamentally reactive. It responds to your prompt. It doesn't get bored and decide to learn a new skill on its own.

So no, we don't have AGI in the way sci-fi imagined it. What we have is a powerful general problem-solving tool that is both more capable and more limited than we realize. The trick is to stop worrying about the label and start getting really, really good at understanding its weird, spiky profile of genius and idiocy.