Have We Reached AGI? A Reality Check on GPT-4 and Claude

AGI is not here. It's not even on the horizon. Anyone who tells you otherwise is selling something.

Yes, models like GPT-4 and Claude 3 Opus are astonishingly capable. They can write elegant code, draft legal arguments, and summarize dense medical research in seconds. They feel like a conversation with a brilliant, knowledgeable entity. But they aren't thinking. They are incredibly sophisticated pattern-matching machines.

And that difference is everything.

What Even Is AGI?

Let's get the definition straight. Artificial General Intelligence (AGI) isn't just a very smart chatbot. It's an AI that can understand, learn, and apply its intelligence to solve any problem, much like a human being. Not just text, not just images, but novel, abstract challenges it's never seen before.

A calculator is a genius at arithmetic but can't recommend a good book. That’s narrow AI. Today's LLMs are also narrow AI, just with a much, much wider narrowness. Their domain is human language and the patterns within it.

AGI would possess common sense. It would understand cause and effect in the real world. It could form a long-term plan and adapt it when things go wrong. Today’s models can’t, not really.

Where Today's Models Shine (And Fool Us)

Make no mistake, what we have is magical. You can give a model a fuzzy idea and watch it blossom into a functioning Python script. I do it every day.

For example, asking Claude 3 Opus:

Write a python script that takes a folder path as an argument. It should find all .jpg files in that folder, resize them to 1024px on the longest side while maintaining aspect ratio, and save them to a new 'resized' subfolder. Use the Pillow library.

Within seconds, you get a perfectly workable script. It imports the right libraries, handles file paths, does the math for resizing, and includes error checking. This feels like intelligence. It feels like your request was understood.

But what’s happening under the hood is an act of high-dimensional statistical retrieval. The model has processed billions of lines of code from GitHub and documentation. It has seen this pattern, or patterns very similar to it, thousands of times. It's not reasoning about image processing; it's predicting the most probable sequence of text that follows your prompt, based on its training data.

This is an incredibly useful capability. But it's also a convincing illusion.

The Cracks in the Facade

The illusion shatters when you push the models outside the dense cloud of their training data. They lack a world model, a genuine understanding of how things work.

Ask a model a simple common-sense riddle:

I put a book on a table. Then I moved the table to the other side of the room. Where is the book?

It will get it right: the book is on the table. But it's not because it understands physics or object permanence. It's because this exact relationship has been described countless times in its training data.

Now try something slightly more abstract that relies on inferring constraints, like this classic:

I have two US coins in my hand that total 30 cents. One of them is not a nickel. What are the two coins?

Many powerful models fail here. They get stuck on "one is not a nickel" and can't make the logical leap that the other coin can be. The answer is a quarter and a nickel. The quarter is the one that's not a nickel. This requires a small but crucial step of genuine reasoning that pattern-matching struggles with.

Here are the other big tells:

• Zero Physical Intuition: An LLM can write a beautiful essay on gravity, but it has no internal concept of it. It has never dropped something. This lack of embodied experience means its understanding is paper-thin.
• Confabulation (aka Hallucinations): When they don't know an answer, models don't say "I don't know." They generate statistically plausible but completely false information with absolute confidence. They invent legal precedents, create fake URLs, and cite non-existent studies. This is a direct result of being a text-prediction engine, not a knowledge engine.
• Poor Long-Range Planning: Ask an LLM to outline a 12-chapter book, and it will do a great job. Ask it to write that book, ensuring perfect continuity of plot points and character arcs across 80,000 words in a single go, and it will fail miserably. It loses the plot, literally. Its context window is a form of short-term memory, not a framework for sustained, goal-oriented reasoning.

So, Are We Close?

No. And simply scaling up the current approach—more data, more GPUs, more parameters—is unlikely to get us there. It will make the models better text predictors, but it won't magically grant them common sense or a true understanding of the world.

We've built the world's most impressive parrots. They can repeat and remix human knowledge in breathtaking ways, creating immense value in the process. We should be excited about that. These are powerful tools that can augment our own intelligence.

But the path to AGI probably doesn't look like a bigger GPT-5. It likely requires a fundamental architectural shift. Maybe it involves new types of neural networks, integration with physical robotics for embodied learning, or a breakthrough in symbolic reasoning we haven't had yet.

For now, use the tools we have for what they are. They are incredible productivity enhancers and creative partners. Just don't mistake a great mimic for a thinking mind.