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

Let's get this out of the way: AGI is not here. Not even close. The debate over whether GPT-4o or Claude 3 Opus shows 'sparks of AGI' is a distraction from the real story.

The real story is that these models have crossed a critical threshold. They are now so good at complex reasoning and tool use that they can automate entire classes of knowledge work that were safe just a year ago. That's a much bigger deal than arguing about a philosophical definition from the 1950s.

The old goalposts for AGI, like the Turing Test, are useless now. An AI can easily convince you it's human for five minutes. The new goalposts are fuzzier—long-term planning, self-improvement, understanding the physical world, creating truly novel ideas. On that front, we're still at the starting line.

So instead of asking 'Is it AGI?', let's ask a better question: What new capabilities have these models gained, and where do they still completely fall on their face?

The 'Holy Cow' Moments Are Real

There's no denying the magic. Give Claude 3 Opus a 200,000-token codebase for a messy old web app and ask it to identify security vulnerabilities, suggest refactors, and write the Jest tests for the new code. It will do it. It won't be perfect, but it'll get you 80% of the way there in minutes, saving a senior engineer a week of misery.

This is a step-change from just generating boilerplate. It's multi-step reasoning applied to a massive context. We're seeing models act as junior partners in complex creative and technical work. They can take a vague goal, break it down into steps, and execute those steps with decent accuracy.

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

Act as a marketing strategist. My company sells a high-end, subscription-based coffee bean delivery service. Our target audience is tech professionals aged 25-40 who value convenience and quality. Outline a three-month content strategy for our blog and social media. Give me five concrete blog post ideas with short outlines, and ten social media post concepts for Instagram, complete with image suggestions.

What you get back isn't just a list. It's a coherent plan that understands the persona, the product, and the channels. It demonstrates a theory of mind about the target audience. This is the kind of task that used to require a human with marketing experience. Now, a model can draft the entire strategy.

These models are also getting shockingly good at handling ambiguity. You can show GPT-4o a picture of a hand-drawn wireframe on a napkin and it will generate the React code to build it. It infers structure, understands intent, and translates a messy, real-world artifact into structured code. This ability to jump between modalities and concepts is where the 'sparks of AGI' talk comes from. It feels like real intelligence.

But it's a very specific, disembodied, digital kind of intelligence. The moment you introduce the real world, the illusion shatters.

Where It All Falls Apart

An LLM has never made a cup of coffee. It doesn't know what hot feels like. It has no concept of gravity beyond what it's read in text. This complete lack of physical grounding, or 'embodiment,' is its biggest limitation.

Ask a model to devise a plan to tidy a messy room, and it can give you a perfect, step-by-step list. But if you put that model in a robot and ask it to execute the plan, it would fail immediately. It wouldn't know how to pick up a sock, how much force to use to close a drawer, or that you can't put a book on top of a full glass of water. This is Moravec's paradox in action: what's easy for a five-year-old (navigating the physical world) is impossible for our most advanced AI.

This extends to long-term planning and memory. Models live in a perpetual present. Each interaction is a new world, a new problem, constrained by its context window. It can't remember what you talked about last week unless you paste it into the prompt. It can't have a long-term goal like 'learn to be a better programmer' and actively work towards it over months.

An agentic system can fake this by using a vector database for memory and breaking goals into tasks. But the core model isn't learning or growing. It's just a brain in a jar, re-booted for every query, with its 'memory' bolted on the side.

Finally, there's true novelty. AI is a phenomenal remix engine. It can combine concepts in fascinating ways, but it can't yet generate a truly new scientific theory or a fundamentally new artistic movement from scratch. It synthesizes and rearranges the vast archive of human knowledge it was trained on. It's an accelerator, not an originator.

Build Agents, Not Gods

The chase for AGI is a red herring for builders. You don't need a god-in-a-box to create immense value. The opportunity right now is in building highly specialized agentic workflows.

Instead of one AGI, think of a team of expert AIs:

• The Researcher: An agent that uses browsing tools to scrape the web for information on a topic.
• The Analyst: An agent that takes the raw data from the researcher, synthesizes it, and finds key insights.
• The Writer: An agent that takes the analysis and writes a clear, human-readable report.
• The Coder: An agent that can build a tool or script based on the report's recommendations.

By chaining together these specialized models and giving them the right tools and memory, you can build systems that automate incredibly complex tasks. This is the real frontier. It's less glamorous than building Skynet, but it's what's actually working today.

Stop waiting for AGI. The tools are already here to change everything. Go build a specialist.