Has AGI Arrived? A Pragmatic Look at What AI Can Actually Do

The debate over Artificial General Intelligence (AGI) is mostly a waste of time. It’s a semantic argument that distracts from the real question: are these tools capable of superhuman performance on specific, valuable tasks right now?

The answer is yes. And also, no. It depends entirely on what you’re asking.

Trying to label GPT-4o or Claude 3 Opus as 'AGI' or 'not AGI' is like arguing whether a souped-up Formula 1 car is a 'true vehicle' because it can’t go off-road. It’s the wrong framework. Instead, let's look at what these things can actually do, where they shine, and where they fall flat on their face.

The 'Sparks of AGI' Are Real

There are moments using the latest models that feel like magic. You give it a messy, half-formed idea, and it returns something more coherent and useful than you could have produced in the same amount of time. These are the 'sparks of AGI' people talk about.

Take code generation. Give a model like GPT-4o a screenshot of a simple web app and a one-sentence description, and it can spit out the full HTML, CSS, and JavaScript in seconds. It won't be perfect, but it's a 90% solution that gets you past the blank-page problem. It understands context from an image and translates it into functional code. That’s wild.

Another superpower is synthesis. You can dump 100 pages of dense financial reports into Claude 3 and ask, “What are the top three risks mentioned that are specific to the South American market?” It won’t just find keywords; it will read, understand, and synthesize the information into a concise, human-readable summary. For knowledge workers, this is a genuine force multiplier.

We're seeing this across modalities. The new models can watch a video of a sports game and explain the rules, listen to your voice and detect your mood, or turn a rough sketch on a napkin into a digital illustration. This cross-domain 'understanding' is what makes people reach for the AGI label.

These models excel at tasks that rely on pattern recognition, translation, and summarizing vast amounts of learned information. They are autocomplete on steroids, but the steroids are working really, really well.

Where the Illusion Shatters

For all the magic, the ghost in the machine is still a statistical echo. The illusion of understanding shatters the moment you require robust, multi-step reasoning or a consistent internal world model. The models don't know anything; they predict the next most likely token.

Give one of these models a simple logic puzzle that requires tracking relationships between multiple entities. For example, “John is taller than Dave. Dave is shorter than Mike but taller than Sarah. Who is the tallest?” The model often gets tangled in its own associative logic and gives a wrong answer with complete confidence. It can't hold and manipulate a persistent state of 'the world' in its head.

Another major failure point is planning. Ask an LLM to outline a plan for a complex project, and it will do a great job. Ask it to execute that plan, making adjustments based on real-world feedback over time, and it collapses. It has no memory beyond its context window and no agency. It can’t learn from a mistake in one conversation and apply that learning to the next one an hour later.

This is why agentic systems are so hard to build. The core LLM is like a brilliant but amnesiac consultant. It gives great advice in the moment but forgets who you are and what you were talking about five minutes ago. You have to build immense scaffolding around it to give it memory and the ability to act on its own outputs.

They also have no real concept of physical reality. An LLM can describe the physics of a bouncing ball perfectly, but it can't intuitively reason about why a tower of blocks might fall over if you remove the bottom one. This lack of grounded, embodied knowledge is a massive gap between current AI and anything resembling general intelligence.

Forget AGI, Build a Task-Specific Superhuman

So, who cares about the AGI label? It's a distraction. What we have are powerful reasoning engines that are incredibly good at certain tasks and comically bad at others. The smart move isn't to wait for AGI; it's to identify and exploit these 'spikes' of superhuman capability.

You don't need AGI to automate drafting legal contracts. You need a model with a massive spike in understanding legal language and precedent.

You don't need AGI to triage customer support tickets. You need a model that can classify intent and sentiment with 99.9% accuracy.

This is the real work of building with AI today. It's about finding a narrow, well-defined problem where an LLM's pattern-matching abilities and vast knowledge base can be applied with a human in the loop to handle the edge cases and sanity checks. We are building powerful tools, not colleagues.

Stop asking, 'Is it intelligent?' and start asking, 'Is it useful for this specific task?' The answers you get will be much more productive.