We Don't Have AGI, We Have Incredible Simulators

Everyone wants to know if we have AGI yet. Is this the one? Did GPT-4o or Claude 3.5 Sonnet finally cross the threshold into true, general intelligence?

It’s the wrong question. It’s a philosophical distraction.

The only question that matters is: can the AI do a complete, economically valuable job, from start to finish, without a human steering it at every turn? Can it replace a full-time employee, not just a feature in an app?

To figure that out, let's stop talking about consciousness and start talking about jobs. I propose two brutally practical benchmarks: The Junior Developer Test and The Small Business Owner Test.

The Junior Developer Test

A junior developer doesn't just write code. They are given a ticket in Jira, often with vague requirements, and are expected to figure it out. This involves several steps that go way beyond code generation.

First, there's understanding ambiguity. A ticket might say, "The login button looks weird on mobile. Fix it." What does "weird" mean? An experienced human knows to check for alignment, padding, and text wrapping. They might ask a clarifying question or just make a sensible executive decision.

Current models are okay at this, but they have no real-world context. If you feed GPT-4o the component code and the screenshot, it will likely fix the CSS. But it won't know about the company's design system or the implicit rule that all buttons must have a minimum tap target of 44px unless you tell it explicitly.

Second, there's the workflow. A junior dev needs to pull the latest from the main branch, create a new feature branch, write the code, run the local test suite, commit the changes with a properly formatted message, push the branch, and open a pull request. This involves a chain of tools: Git, the command line, VS Code, a testing framework.

This is where AI agents today completely break. They can generate the shell commands, but they can't execute them in a stateful environment, observe the output, and debug when something goes wrong. What if there's a merge conflict? What if a dependency fails to install? The model's context window is not the same as a persistent terminal session.

Finally, there’s self-correction. When tests fail, a human developer investigates. They read the error logs, form a hypothesis, and try a fix. It's a loop of trial and error. LLMs can look at an error message you provide and suggest a fix, but they can't run that loop themselves. They are brilliant one-shot problem solvers, but they lack the agency to persist through a multi-step debugging session.

So, can AI pass the Junior Developer Test?

• Write a function based on a clear prompt? A+. 10/10.
• Refactor a messy file? B+. It usually does a good job.
• Complete an end-to-end task from a vague Jira ticket? F. Not even close.

The Small Business Owner Test

Let’s raise the stakes. Forget a single role; can an AI run a simple online business? Imagine a dropshipping store for, say, niche coffee accessories.

This requires a wider range of skills. You need to identify trending products, generate marketing copy, create product images, run social media ads, and handle customer support emails. It's a fantastic test of multi-modal, long-term strategic thinking.

Models like Claude 3.5 Sonnet and GPT-4o are fantastic at the individual tasks. You can ask them to "Write five funny tweets about our new self-heating coffee mug" and they'll nail it. You can use Midjourney or DALL-E 3 to create slick product photos. You can even have an AI read an angry customer email and draft a polite, helpful response.

But who is the conductor of this orchestra? Who decides the ad budget for the week? Who notices that the ads featuring a blue mug are outperforming the red one and decides to double down?

This requires long-term memory and goal-directed planning. The AI would need to maintain a persistent state of the business—sales data, ad performance, customer sentiment—and make strategic decisions based on a top-level goal: "Maximize profit."

It would need to connect disparate domains. The text from a positive customer review should inform the next batch of ad copy. A sudden spike in search traffic for "pour-over kettle" should trigger a new ad campaign. Humans do this synthesis naturally.

AI systems, in their current form, do not. They are passive tools awaiting a prompt. They don’t have goals of their own. They don't have a nagging feeling that they should check the ad spend before the weekend.

We Have Simulators, Not Agents

The fundamental gap isn't intelligence in the academic sense. It's autonomy. Today’s models are incredible simulators. They can simulate the process of writing code. They can simulate a creative marketer. They can simulate a helpful customer service rep.

But they are not true agents. An agent perceives its environment, makes decisions, and takes actions to achieve a goal. A simulator responds to an input with a plausible output. That’s a world of difference.

The real danger isn't that AGI is going to wake up and take over. It's that we'll get stuck using incredibly powerful tools as if they were just slightly better search engines.

Your job isn’t getting replaced by AGI tomorrow. But it can be massively amplified by using these simulators for what they're good at: instantly handling the well-defined, isolated tasks that make up 80% of your work. Let the AI simulate the first draft, and you do the real work of steering, correcting, and connecting the dots.