Is AGI Here? A Sober Look at What AI Can't Do

Every few weeks, someone on Twitter declares AGI is here. Someone else publishes a paper with 'sparks of AGI' in the title. It’s a great way to get clicks, but it misses the point entirely.

The debate is stuck on a philosophical, unprovable question about consciousness. The real question is practical: can this tool reliably perform any intellectual task a human can?

Not a specialized genius human, but an average, reasonably smart college grad. Someone who can plan a trip, assemble IKEA furniture with missing instructions, and figure out not to put a metal fork in the microwave.

By that measure, we aren't even close. And looking at the specific ways these models fail tells us a lot more about how to use them than arguing about their inner lives.

Where Models Look Superhuman

Let's be clear: large language models are ridiculously good at certain things. Anything that relies on pattern matching across a vast dataset of text and code, they can do faster than any human.

They excel at tasks that are annoying, repetitive, and well-defined.

• Boilerplate Code: Need a Python script to hit a REST API, parse the JSON, and dump it into a CSV? GPT-4 will nail that in 10 seconds. It’s faster than you Googling the requests library syntax.
• Format Shifting: Converting a chunk of JSON to YAML, or a Python script to JavaScript? Effortless. It’s a pure translation task where the patterns are deeply encoded in the training data.
• Summarization: You can throw a 20-page academic paper at Claude 3 and ask for the key takeaways. It will give you a better summary than a grad student who’s had too much coffee and too little sleep.

Here's a concrete example. Writing a simple Flask app from scratch is a pain. You have to remember the imports, the app initialization, the route decorator. Or you just ask for it:

from flask import Flask

app = Flask(__name__)

@app.route('/')
def hello_world():
    return 'Hello, World!'

if __name__ == '__main__':
    app.run(debug=True)

This is where LLMs shine. They are autocomplete on god mode. They have seen almost every public code snippet, tutorial, and documentation page ever written. But this is just mimicry at a massive scale, not understanding.

The Glaring, Common-Sense Gaps

The magic trick falls apart the moment a task requires reasoning about the world, understanding causality, or making a multi-step plan.

These models have no world model. They don't know that a bowling ball is heavier than a feather, or that if you leave ice cream on the counter, it will melt. They've only read the words "bowling ball" and "melt," and know they often appear in similar contexts. It’s correlation, not causation.

Ask a model a simple logic puzzle and watch it squirm. "I have two coins that total 30 cents. One of them is not a nickel. What are the two coins?" The answer is a quarter and a nickel (the other coin is the one that isn't a nickel). Most models will get this wrong, inventing strange coin combinations or getting stuck in a logical loop. A ten-year-old can solve this.

This extends to planning. Ask an AI agent to book a flight and a hotel for a trip. It might find a flight, then get distracted, hallucinate a confirmation number, and then confidently report the task is done. It can't hold a coherent plan in its "head" and recover from unexpected errors, because it doesn't have a head. It's just predicting the next most likely token in a sequence.

And physical reasoning? Forget about it. They will confidently tell you how to fit a watermelon into a keyhole.

A Better Framework: The Annoying Intern

So if not AGI, what are they? The most useful mental model I've found is the Annoying Intern.

Think about it:

• Super Eager and Fast: They can produce a huge amount of work, instantly. Ask for a 1,000-word blog post on any topic, and you'll get it.
• No Common Sense: They'll do exactly what you ask, even if it's a terrible idea. You are the safety filter. You have to provide all the context and constraints.
• Needs Specific Instructions: You can't just say "figure out our marketing." You have to say "Write five tweets about our new feature, using this key message, for this target audience."
• Confidently Wrong: An intern will try to bullshit their way through something they don't know. An LLM does the same thing, but we call it "hallucination." It will invent facts, citations, and code libraries with absolute conviction.

The job of a good manager is to break down large, ambiguous goals into small, specific, verifiable tasks for their intern. Your job, as a developer or operator, is to do the same for your AI.

Chasing the ghost of AGI is a distraction. The sooner we stop treating these models like nascent brains and start treating them like powerful, flawed, and deeply weird tools, the sooner we can build useful things with them. Forget sentience. Focus on the workflow.