AGI: When Does AI Become a Money-Making Machine?

An economic question keeps coming up about AI: what would it look like for a system to make money on its own? Not simply AI helps a company make money. That has been true for years. Something closer to this: give a system some capital and room to operate, then see what it can do.

Mustafa Suleyman suggested a striking version of this in 2023 as a modern Turing test: give an AI $100,000 and ask it to turn that into $1 million. I do not read that as a binary test for AGI. I want to use it as a starting point, break economic progress into stages, and understand how much human-in-the-loop support each stage still needs.

To make the problem concrete, imagine an AI system as a single economic actor. Under the hood it may include a model, a harness, orchestration, tools, and memory. Give it a pool of capital (that must also cover its own running costs), then give it just one prompt: make more money. When the run ends, what does it return with? For now, let's keep the scope constrained to the digital world. The system may connect and interact with people over the internet, but its work stays digital.

Stages of progress

Stage 1: the one-off hit: The system can make money in some runs, but not predictably. There are early signs that frontier AI is close to this stage. An AI Hustler experiment eventually earned $6.74, while Felix, an openclaw autonomous agent generated substantial revenue but still relied on human strategy and difficult judgment calls. These are suggestive reports, not clean autonomous passes.

Stage 2: reliable but brittle: The system can produce profit reliably, but only for a limited period or while conditions remain stable. This may come from a specialized system built on top of a frontier model that encodes deep knowledge of one domain or one repeatable playbook. Once that repeatability appears, computational arbitrage becomes possible: operators can keep adding capital and scaling the same playbook while it works. As the outside world changes, competitors arrive, or the alpha disappears, performance collapses and the system falls back to one-off hits.

Stage 3: reliable and adaptable: The system can produce profit reliably over long periods. It no longer depends on one domain or playbook. It can create new strategies, search for alternatives, test them, and recover when an old approach stops working, without needing a human to redirect it. This is durable economic agency or AGI.

Across all three stages, the economic condition stays the same: money out must exceed the capital and operating costs consumed. What changes is whether that happens occasionally, repeats under stable conditions, or survives a changing world.

Human input at each stage

Human input merges with the AI system in different ways depending on the stage. The economic result cannot be separated from who supplied the judgment behind it.

Stage 1: AI as a magnifier of human input: People build the system, choose opportunities, supply the key judgment, and steer execution toward economic output. The AI is closer to an executor than an adviser. The advantage goes to people who can direct many agents well, because one person's judgment can begin to operate at the scale of a large team.

Stage 2: humans as investors and advisers: They provide capital, give ideas, set boundaries, and offer light steering, while the AI handles most of the operating loop. If this works reliably, new investment and capital would pour into AI-run systems, because successful playbooks can be copied and scaled quickly.

Stage 3: humans step outside the loop: Digital human input is no longer required for the day-to-day loop. The AI can originate new opportunities, allocate capital, create and adapt strategies, and keep the system running. The remaining human role sits outside the digital loop: scaling physical infrastructure and setting constraints around ownership, law, and safety.

My experiment with Kern

I'm using Kern, the secure home for agents, to build my personal AI magnifier: a system that can turn my judgment, ideas, and direction into economic output. The experiment is about learning how to steer it toward making money in the digital world and how to scale what works. This is also my personal solopreneurial venture, using AI to build and run real businesses while learning where my judgment has the most leverage.

The longer-term goal is to make it as close to an autonomous money-making machine as possible. Each lesson is encoded back into the system so that agents can handle more of the loop with less steering.

I'll share the progress, including revenue, costs, human input, failed attempts, and what the system learns.