AGI Isn't Here. So What Are We Actually Building?

No. Artificial General Intelligence is not here.

It’s not “right around the corner.” It’s not hiding in the next version of GPT. The breathless demos and “sparks of AGI” headlines are selling you a story. The reality is that we're building incredibly powerful tools, but they are not minds. They are not colleagues. And pretending they are leads to building the wrong things.

Let’s get real about what today’s models can and can’t do.

What We Mean When We Say 'AGI'

The goalposts for AGI have always been a little fuzzy. Originally, it was about beating a human at chess. Then it was passing the Turing Test. Now the term is so loaded it’s almost useless.

Here’s a practical definition: AGI is the ability to learn and solve a novel problem in a domain you weren't trained on, using general principles. It’s about transferring knowledge, not just pattern matching. A human chef can learn to be a decent chemist because they have an intuitive grasp of physical processes, measurements, and cause-and-effect.

Large Language Models don't have that. At their core, they are phenomenally complex sequence prediction engines. They have ingested a huge chunk of the internet and are masters at predicting the next word in a sentence based on the patterns in that data. This mimicry is so good, it often feels like understanding. But it isn't.

The 'Sparks' That Blind Us

Of course, these models are incredible. If you've spent any time with GPT-4 or Claude 3, you've seen things that feel like magic. You can paste in a messy screenshot of a web app and get clean, working React code. You can throw a 50-page academic paper at it and ask for the three key arguments against the author's thesis.

For example, you can give it a prompt like this:

I have two PostgreSQL tables: `users` (id, name, signup_date) and `orders` (id, user_id, amount, created_at). Write a single SQL query that finds the total order amount for users who signed up in the last 30 days.

And it will spit out a perfect, optimized query. This feels like reasoning. It identified the relationship between the tables, understood the time constraint, and used the correct aggregate function. This is the kind of task that saves a developer 15 minutes of thinking and testing. These are the “sparks” that get people excited.

This is where the hype comes from. The model is an expert at synthesizing and restructuring information that exists within its training data. It's a universal adapter for knowledge work. But it’s not creating anything truly new.

The Hard Limits They Don't Advertise

The gap between this amazing capability and genuine intelligence becomes clear when you push the models outside their comfort zone of text and code.

• Physical World Amnesia: Models have no body, no senses, no concept of physics beyond what they've read. They can tell you the steps to change a tire because they've read thousands of articles about it. But they don't know what a lug nut feels like, or that you need to apply counter-pressure. They have no intuition for the physical world, which makes them useless for robotics or real-world planning without massive scaffolding.

• The Planning Fallacy: Ask an LLM to create a business plan. It will generate a beautiful document. Ask an AI “agent” to execute that plan, and it will get stuck in a loop trying to book a domain name. Real-world goals require dynamic, multi-step planning with feedback and adaptation. Current agents are incredibly brittle. They can follow a simple script, but a single unexpected obstacle—a changed website layout, a CAPTCHA—can derail them completely. They don't plan; they execute a sequence of probabilistic actions.

• Causality is a Ghost: LLMs are masters of correlation, not causation. They know that lightning is often followed by thunder because that pattern appears constantly in the data. They don't understand the physics of atmospheric discharge causing a sonic boom. This is a critical distinction. It means they can be easily tricked by scenarios that violate common correlations and struggle to reason from first principles. They can't invent, they can only remix.

Build Systems, Not Gods

So where does this leave us? Is it all just hype? No.

This isn’t about being a pessimist. It's about being a better engineer. We have to stop thinking about building “an AI” and start thinking about building systems where an LLM is a powerful component, like a database or an API.

Think of an LLM as a calculator for language. It’s a tool for summarizing, translating, reformatting, and generating text based on a given context. It's not a thinker, a strategist, or a CEO. When you treat it like a component, you design better systems with proper guardrails, human oversight, and validation steps.

You don't ask your calculator to invent a new form of mathematics. You use it to do the math you already understand, but faster and at a greater scale.

That's what we're actually building. Not AGI, but the first truly useful calculators for human language. And that's exciting enough.