We Don't Have AGI. We Have Something Weirder.

Everyone wants to know if we’ve reached Artificial General Intelligence. If you show someone GPT-4o having a real-time, emotive conversation, they’ll say it’s here. If you ask that same model a simple logical question it gets wrong, they’ll say it’s a million miles away.

Here’s my take: asking “is it AGI?” is the wrong question. It’s a poorly defined, moving goalpost that distracts from the weirder, more interesting reality. We haven't built a synthetic human mind. We've built a universal intelligence simulator, and that’s a very different thing.

The "Sparks of AGI" Are Real

Let's be clear: the latest frontier models from OpenAI, Google, and Anthropic are shockingly capable. They can reason across modalities in ways that feel genuinely intelligent. You can show Gemini a video of a magic trick and it can explain how it was done. You can have a conversation with GPT-4o, interrupt it, change topics, and it keeps up, even changing its vocal tone to match the mood.

This isn't just pattern matching. The ability to perform zero-shot chain-of-thought reasoning is a legitimate spark of something new. Ask a model to solve a complex logic puzzle it's never seen before, and it can often lay out the steps to the solution. That feels general.

Here are some things that look a lot like general intelligence:

• Cross-Domain Knowledge Transfer: You can teach it a concept from physics and ask it to apply that concept to a problem in economics. It works surprisingly often.
• Tool Use: The function-calling APIs are a game-changer. Models can decide, based on your prompt, to call an external tool, get data, and then use that data to answer your question. That's not just retrieving information; it's actively seeking it.
• Complex Instruction Following: You can give it a multi-step task with weird constraints, and it can hold the entire instruction set in its “head” and execute. It’s like a brilliant, amnesiac intern.

If your definition of AGI is simply a system that can perform a wide variety of intellectual tasks at a human level, you could argue we're already brushing up against it.

But that definition misses the point.

Where the Simulation Breaks Down

The intelligence we see is a performance. When the curtain is pulled back, you find an alien architecture that lacks the fundamental pillars of what makes our intelligence truly “general.”

First, embodiment. Models have no body. They don't understand gravity in their gut. They've read every book on how to ride a bike, but they have no muscle memory, no sense of balance. This isn't a trivial missing piece; it's the foundation of common sense. We know a glass will shatter if we drop it because we've experienced a world of physical consequences. An LLM only knows this because it's in the training data.

Second, they have no persistent memory or continuous learning. Every interaction with a model starts from a sophisticated form of amnesia. Yes, the context window is its short-term memory, but it doesn't learn from your conversation and permanently update its core knowledge. A junior developer you hire gets a little smarter every single day. A language model gets a factory reset with every API call (fine-tuning notwithstanding, which is more like creating a new specialized model).

Finally, and most importantly, they lack agency. An LLM has no goals of its own. It has no desires, no curiosity, no drive to go do something. It is the ultimate passive oracle, waiting for a prompt. It will never wake up in the morning and decide to invent a new type of battery because it's bored. Its entire existence is reactive.

This is the core difference. Human intelligence is a proactive, embodied, and continuous process of learning and acting on the world to achieve goals. LLM intelligence is a stateless, disembodied simulation of that process, triggered by a user.

A Better Goal: Capable, Specialized Agents

So, if we don't have AGI, what do we have? We have the most powerful reasoning engine ever created. Instead of waiting for it to magically become a self-aware human-in-a-box, the real work is to use this engine as the core of agentic systems.

An agent is a system that uses an LLM to perceive its environment, make decisions, and take actions to achieve a specific goal. It’s the difference between asking a model to write a Python script and giving it a goal, a file system, and a terminal to run the script, debug it, and get to a working solution on its own.

This is where things get practical and exciting. We can build agents that do things like:

• Automated Software Development: An agent with access to a codebase, a terminal, and a testing suite can be tasked with