Is AGI Here? A Sober Look at GPT-4 and Claude 3

Let’s get this out of the way: No, Artificial General Intelligence (AGI) is not here. Not in a box in a lab, not in the API you're using, not anywhere.

The problem is, 'AGI' is a terrible metric. It's a fuzzy, philosophical goalpost that keeps moving. Every time an AI model accomplishes something we thought was uniquely human—writing poetry, passing the bar exam, generating code—the definition of AGI mysteriously shifts to include whatever it can't do yet.

A better question is: are these models displaying capabilities that were pure science fiction just three years ago? And can you use those capabilities to do your job better?

Yes. Absolutely.

The AGI Checklist: Where We Stand

Instead of chasing a vague definition, let's break down the supposed components of general intelligence and see how models like GPT-4o and Claude 3 Opus actually perform. The results are a strange mix of genius and idiocy.

Multi-Step Reasoning: They're getting good at this. You can give them a complex scheduling problem with multiple constraints, and they can often spit out a perfect solution. For example: "Plan a team offsite for Tuesday. Alice can't make it before 10 AM, Bob has a hard stop at 3 PM, and we need a 90-minute core session plus two 30-minute breaks." They usually nail this.

But they fail spectacularly with physical or spatial reasoning. Ask one to solve a simple coffee-cup-and-ball puzzle, and it gets confused. It has no internal model of the physical world. It's all text, all the way down.

Creativity: This is a tricky one. They are incredible synthesizers. Ask for ten startup ideas combining SaaS and pet care, and you'll get ten plausible, well-reasoned concepts. They can write a sonnet in the style of Shakespeare about your dog. This is powerful creative leverage.

What they can't do is generate a truly novel idea from a blank slate. They operate on the patterns in their training data. They can combine 'A' and 'B' in a thousand interesting ways, but they can't invent 'C'.

Common Sense: This is their biggest blind spot. A model knows that fire is hot because it has read billions of sentences stating that fact. It does not understand hotness. It has never felt a burn or the warmth of a campfire. This lack of embodied experience leads to strange, illogical errors that a five-year-old would never make.

Long-Term Planning: Forget about it. Their memory is limited to the context window. While that window is getting huge (200k tokens for Claude!), it's not true memory. An AI can't decide on a goal today and work towards it for a week, remembering its previous steps and learning from mistakes. Agentic systems are trying to hack this with external memory stores and complex looping, but it’s a bolt-on solution, not a core capability.

Spikes of Superintelligence, Valleys of Stupidity

The most important thing to understand about using today's AI is its wildly uneven capability profile. One minute, it can refactor 500 lines of spaghetti code into a clean, elegant Python class—a task that might take a human developer an hour. The next minute, it will confidently tell you that a horse has six legs.

This is the spiky frontier. We're not dealing with a machine that is a 7/10 at everything. We have a machine that is a 12/10 at some tasks, and a 1/10 at others. The skill is learning to identify the spikes.

Here's a concrete example. I recently asked Claude 3 Opus to help me write a script to process some log files. It produced a brilliant, efficient script in seconds. I then asked it to add a comment to a specific line. It got confused, apologized, and rewrote the entire script from scratch, hallucinating a function that didn't exist.

Peak competence followed immediately by abject failure. This is the normal user experience.

How to Actually Use This Stuff

Given this spiky profile, treating the AI like a junior employee is a mistake. A junior employee is generally consistent. An AI is a brilliant but deeply unreliable savant. Your workflow needs to reflect that.

• Brainstorming Partner: It’s an amazing tool for getting unstuck. Use it to generate outlines, titles, counterarguments, and different angles on a problem. It lowers the activation energy for starting.
• Code Generation: Use it to write boilerplate, simple functions, and unit tests. Always assume the code is broken until you've tested it yourself. Never, ever ship AI-generated code to production without a human review.
• First Drafts, Not Final Drafts: Let it produce the messy first 80% of a document, email, or report. Your job is to do the final 20%—the part that requires nuance, fact-checking, and true understanding of the audience.
• Rubber Duck: Stuck on a problem? Explain it to the AI. The act of articulating the problem often reveals the solution. The AI's (often wrong) suggestions can also spark the right idea in your own head.

The Real Question

So, no AGI. But we have something new: a tool that can manipulate language and ideas with incredible speed and scale. It's a calculator for words. And just like the electronic calculator didn't make us dumber at math, these models won't replace thinking. They will augment it.

The debate about AGI is a distraction from the real work. The work is figuring out how to build new workflows, new products, and new ideas on top of a tool that can instantly do things that used to be difficult and slow.

That’s a much more interesting problem to solve.