Is AGI Here? A Practical Look at What Models Can Actually Do

Let's cut to it: No, Artificial General Intelligence (AGI) is not here. Not even close. The debate is a distraction from the more interesting question: how powerful are the tools we have right now, and what are their real limits?

If your definition of AGI is a system that can reason, learn, and act autonomously across a wide range of tasks at or above human level, we're not there. What we have are incredibly sophisticated pattern-matching engines. They are phenomenal mimics, but they don't understand the world.

But that doesn't mean they aren't useful. In fact, their capabilities are weirdly spiky. They can write a sonnet, then fail a simple logic puzzle a five-year-old could solve. Understanding these peaks and valleys is key to actually building things with them.

The 'Sparks of Intelligence' Are Real

It's easy to see why people get carried away. The best frontier models, like GPT-4o or Claude 3 Opus, can do things that feel like genuine thinking. They exhibit what some researchers call 'sparks' of general intelligence, and it's not just hype.

For example, you can give it a messy, half-formed idea for a web app and get back surprisingly clean Python code.

# Prompt: "Make a simple Flask app that takes a city name,
# uses an external API to get the current weather, and displays it.
# Keep it all in one file for simplicity. Add some basic HTML styling."

A model can nail this. It will infer the need for the requests library, structure the HTML, handle basic form submission, and even add comments. It's synthesizing knowledge about programming, APIs, and web design. That's impressive.

They are also powerful reasoning engines for structured tasks. You can give a model a goal, a set of available tools (like functions it can call), and ask it to form a plan. It can break down