AGI Isn't Here. Here's What AI Can (and Can't) Actually Do.

Every few months, someone on Twitter declares the latest LLM shows “sparks of AGI.” It’s a tempting idea. When a model generates flawless code or a poetic email, it feels like something intelligent is on the other side.

It isn't. Not in the way we mean when we talk about general intelligence.

AGI, or Artificial General Intelligence, isn't just about being very good at a lot of tasks. It's about the ability to reason, learn, and apply knowledge to solve novel problems—problems far outside its training data. What we have now are phenomenal pattern-matching machines. They've ingested a huge chunk of the internet and are masters at predicting the next word based on that data. This is incredibly useful, but it’s not thinking.

Let’s cut through the hype and look at what that actually means for you, the developer building with these tools.

Where Models Shine (and Fool Us)

Modern LLMs are amazing at tasks that rely on recognizing and reconstructing patterns. They feel intelligent because the patterns they've learned are the very fabric of human knowledge.

Think about the things they nail almost every time:

• Boilerplate Code: Need a Flask server with a basic /health endpoint? Done. A React component with a state hook? Easy. They’ve seen thousands of examples of this exact pattern.
• Data Transformation: Converting JSON to CSV, or writing a SQL query from a plain-English request, is trivial for them. It's a structured translation task they can execute flawlessly.
• Summarization: Condensing a long article or a transcript is a core strength. This is pattern recognition at its finest: identifying key sentences and concepts and rephrasing them.
• Style Mimicry: You can ask for a blog post in the style of a 1940s radio announcer, and it will do a shockingly good job. It's not creative; it's identifying the statistical properties of that style and applying them.

These are not small feats. They save us hours of work. The reason they work so well is that the solution is already embedded in the training data, waiting to be retrieved and assembled. The model isn't reasoning from first principles; it's completing a familiar sequence.

The Cracks in the Facade

So where does it fall apart? The moment you ask the model to do something that requires genuine reasoning or understanding of the world. The kind of thing a human child can often figure out.

I’m not talking about trick questions or philosophical paradoxes. I mean simple, practical problems that require a step-by-step mental model of a situation.

Take physical reasoning. You can ask a model, "I put my keys in the fridge. I then walk to the living room. Where are my keys?" It will answer correctly. This question exists all over its training data. But add a layer of indirection:

I have a blue box and a red box. I put a small metal ball inside the red box. Then, I put the red box inside the blue box. I shake the blue box vigorously. Where is the ball?

The correct answer is "Inside the red box," which is itself inside the blue box. Most models get this. But watch what happens when we push it.

Okay, now I open the blue box and take out the red box. I place the red box on the table. Then I turn the blue box upside down. Where is the ball now?

A human immediately knows the ball is still in the red box on the table. The blue box is now irrelevant. But many powerful LLMs will get confused, often claiming the ball fell out of the blue box. They don't have a persistent mental model of the objects and their relationships; they just have a cloud of probable word associations.

Strategic planning is another massive weakness. Give an LLM a goal with multiple conflicting constraints and watch it tie itself in knots. For example, planning a project timeline where one developer is only available on certain days, a server needs a two-day setup period, and marketing materials have to be ready before the server goes live. The model will often create a plausible-sounding plan that violates one of the core constraints.

It can't "hold" the constraints in its head and reason through the implications. It can only generate text that looks like a plan, based on all the project plans it has seen before.

Build for the Tool You Have

None of this is to say that today's AI is useless. Far from it. These models are some of the most powerful tools we've ever had. But a power tool designed for cutting is useless for hammering.

Stop waiting for or designing for AGI. Instead, understand the tool you actually have: a world-class pattern-matcher and text generator.

Use it as a super-powered autocomplete for code and writing. Use it to summarize, translate, and reformat. Chain it with other tools and your own code to build powerful workflows. But always assume its reasoning is shallow. The moment a task requires a true world model, strategic foresight, or common sense, that's where a human needs to be in the loop.

Treat it like an incredibly smart, fast, and knowledgeable intern who has read everything but understood nothing.