Is AGI Here Yet? A Reality Check on AI's Real Skills

The "AGI" Goalposts Just Keep Moving

Is AGI here? No. And asking the question is starting to feel like a distraction.

Every time a new model drops, the "Is this AGI?" debate flares up. When GPT-4 aced the bar exam, some people screamed "AGI!" When GPT-4o held a real-time, emotive conversation, the chatter got even louder. But the goalposts for AGI always seem to conveniently shift to whatever the latest model can't do.

The term itself, Artificial General Intelligence, is fuzzy. Does it mean human-level? Superhuman? Does it need consciousness? A physical body? Nobody agrees.

So let's ask a better question: what can these models actually do, and where do they still stumble like a toddler?

What AI Is Genuinely Great At Now

The recent advances are not hype. Today's frontier models, like OpenAI's GPT-4o and Google's Gemini family, have capabilities that felt like science fiction just a few years ago.

First, they are becoming excellent tool users. Through function calling, a model can interact with external APIs to get real-time information or take action. You can give it a prompt like "What's the weather in Tokyo and is it a good day to visit the Ghibli Museum?" and it can call a weather API, check the museum's hours, and synthesize an answer. It's not just regurgitating text; it's executing tasks.

{
  "tool_name": "get_weather",
  "parameters": {
    "city": "Tokyo"
  }
}

Multimodality is the other game-changer. These models don't just process text. They see, hear, and speak. Point your phone camera at a math problem, and they can walk you through solving it. Speak to them in Spanish, and they can reply in Japanese, translating in real time. This seamless blend of inputs and outputs makes them feel much more like capable partners than just fancy autocomplete.

Finally, their coding skills are undeniable. As a co-pilot, an LLM can write boilerplate, translate code between languages, explain complex snippets, and hunt down bugs with scary accuracy. They aren't replacing senior developers, but they are dramatically compressing the time it takes to get from idea to working code.

The Stubborn Gaps and Glaring Flaws

For all their new tricks, these models still have fundamental limitations that keep them far from any reasonable definition of "general intelligence."

The biggest is a lack of true world-grounding. A model can identify a glass of water in a picture, describe its transparency, and even guess its temperature. But it has no intuitive, physics-based understanding that tipping the glass will make the water spill. It knows the word "spill" is statistically associated with "tipped glass," but it doesn't know gravity. This leads to subtle but critical errors in reasoning about the physical world.

They also struggle with long-term, autonomous planning. You can ask an AI to write a marketing email, and it will do a great job. You can't ask it to "launch our new product" and expect it to autonomously coordinate engineering, marketing, and sales over the next three months, adapting its plan as new obstacles arise. It can execute a well-defined task, but it can't create and manage a complex, multi-step project on its own initiative. It gets lost.

This lack of initiative is key. An LLM is a reactive system. It waits for your prompt. It doesn't get curious, it doesn't get bored and start its own side project, and it doesn't spontaneously realize a better way to do things. When it gets stuck in a repetitive loop—a common failure mode for AI agents—it will often keep trying the same failed approach indefinitely until a human intervenes. A person would get frustrated, step back, and try something new.

Forget AGI, Build Agentic Systems

Instead of waiting for a mythical, all-knowing AGI, the practical path forward is building systems of collaborating, specialized AI agents. This is where the real value is today.

An agentic workflow breaks a complex problem down into smaller tasks, with a different agent (or a model-plus-tool combo) assigned to each. You might have one agent that watches a user's screen to understand their goal, a second that writes code to achieve it, and a third that runs tests and provides feedback.

This approach is more robust and effective.

• Specialization: Each agent can use the best model and tools for its specific job. A vision task might use GPT-4o, while a data analysis task might use a fine-tuned model running locally.
• Resilience: If one agent fails, the whole system doesn't collapse. You can debug or replace the faulty component.
• Cost-Effective: You can use smaller, cheaper models like Haiku for simple tasks and reserve the expensive frontier models for tasks that require heavy reasoning.

This isn't a single brain; it's an assembly line. It's less like building a human and more like building a company. And it's something you can start building right now.

The conversation needs to shift from "Is it AGI?" to "What useful, multi-agent system can we build with the powerful tools we have today?" That's a much more interesting question.