Is AGI Here? A Sober Look at Today's AI

No, AGI isn't here. But asking the question is a distraction. Let's look at what models like GPT-4o and Claude 3 can actually do, where they fail, and the better questions we should be asking.

June 11, 2026 · 2 min read · SuperThinking team

A sleek humanoid robot staring intently at its own reflection in a mirror.

No. AGI is not here.

That was easy. But it’s also the wrong question. The debate over Artificial General Intelligence—a machine that can understand or learn any intellectual task that a human can—is mostly a distraction for developers and builders.

It’s a moving goalpost, a philosophical rabbit hole. What matters is what these tools can do today, right now, in your terminal. And what they can do is both genuinely stunning and comically limited, often in the same session.

What 'Almost-AGI' Looks Like in Practice

If you showed someone GPT-4o five years ago, they'd call it AGI without blinking. It can look at a picture of your whiteboard, transcribe the diagram into Mermaid.js syntax, write the Python code to implement the logic, and then explain it all back to you in Spanish. That feels like magic.

We're seeing models that can ingest a 200-page PDF of a company's 10-K filing and instantly answer nuanced questions about financial risk factors. I recently fed Claude 3 Opus a messy, undocumented legacy codebase and asked it to refactor a key module and add unit tests. It did it in about 90 seconds, saving me half a day of tedious work.

These are not parlor tricks. This is high-level cognitive labor being automated. Look at this example with the OpenAI API, where you can send an image and a question directly:

from openai import OpenAI
client = OpenAI()

response = client.chat.completions.create(
  model="gpt-4o",
  messages=[
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "What is the key takeaway from this chart?"},
        {
          "type": "image_url",
          "image_url": {
            "url": "https://your-server.com/chart.png",
          },
        },
      ],
    }
  ],
  max_tokens=300,
)

print(response.choices[0])

This ability to reason across modalities—text, images, audio—is a massive leap. It's the closest we've come to a flexible, generalist tool. It feels like you're working with a brilliant, hyper-fast intern who has read the entire internet.

But that intern has some very, very weird blind spots.

A complex business process diagram with arrows and boxes drawn on a transparent whiteboard.

The Glaring Gaps That Prove We're Not There

For all their power, today's models have no real-world grounding. They are masters of syntax and pattern matching, but they don't understand things in the way a human does. They simulate understanding, often flawlessly, but the simulation is brittle.

They lack persistent memory and the ability to learn continuously from interaction. Each conversation starts from a near-blank slate, armed only with its training data and your context window. You can't ask it to