Reasonable Person's AGI

The definition of AGI has changed over the last year. As the probability of “takeoff” decreases, the real test for AGI will be the ability to automate the bottom 5% of work.

Last year, the debate over AGI’s imminent arrival was raging. For a while, phrases like “we are 1000 days away from AGI” were commonly heard in Silicon Valley — only to be met by heated debate about whether it was in fact 500 days or 1500 days away. There were even false rumors spreading online that AGI had already been achieved, which OpenAI had to deny.

Leopold Aschenbrenner published a popular essay that summer which articulated the path to AGI: Continued scaling of AI to bigger and more powerful models would ultimately lead to an AI model that itself could do AI research, and thus one that could recursively self-improve, leading to a “takeoff” or “singularity” event. “By 2025/26,” he wrote, “these machines will outpace many college graduates. By the end of the decade, they will be smarter than you or I; we will have superintelligence, in the true sense of the word.”

Today, the narrative has changed. You can see this in recent commentary from three of the most prominent figures in AI: Ilya Sutskever, Sam Altman, and Andrej Karpathy. In December 2024, Ilya Sutskever predicted that scaling laws, the very basis for the AGI takeoff hypothesis, have reached their limits and that “pretraining as we know it will end.” In June 2025, Sam Altman published a blog describing a more “gentle singularity,” and describing AGI as more of an event horizon than an event. Later that same month, Andrej Karpathy contended that we are in for a “decade of agents,” rather than “AGI in 2027.”

Silicon Valley has held onto the term AGI, but its meaning has subtly shifted. You rarely hear the term “takeoff” used in conversation anymore. Instead, the new obsession is post-training and reinforcement learning — mechanisms that allow LLMs to “think” for longer, and that allow them to develop task-specific mastery. This new AGI is all about automation — the idea is that we can grind out serious progress from today’s models without considerable jumps in baseline intelligence. The softer version of this hypothesis involves the automation of white collar jobs, while the stronger version predicts the automation of both white collar and blue collar jobs, via further innovations in robotics.

I call this new version of AGI “reasonable person’s AGI,” because it eliminates the sensationalism around superintelligence, while retaining most of the practical, world-changing stuff: The economic abundance, the need for wealth redistribution, and the impacts onto social structures, etc. Reasonable person’s AGI has strong explanatory power — it explains the jobs people are taking, the startups they are building, etc. But even as this new narrative has become dominant, few people have come to terms with its full implications. For example, if progress is likely to come from domain specific applications of this technology, rather than exponential compute scaling, do we still need to build multi-million GPU clusters and add 90 GW of energy to the US power grid over the next 5-10 years? And could this new AGI take a decade or two, rather than a few years to accomplish?

I’d like to propose an “AGI test” that can help us track and map the progress of AI over the coming decades. I call it the 5% rule: In any domain where AI can beat the bottom 5% of practitioners, billions of dollars in value will be created. To some, it may sound underwhelming, and yet I think its economic impact will in fact be radically overwhelming. Today, there are five domains where AI clearly passes this test: Writing, coding, customer service, therapy, and companionship. The first four compete with existing jobs, the last one competes with existing social structures.¹

My contention is that the AGI dream (AI is better than everyone at their jobs) will give way to the AGI reality (AI is better than the least effective people at their jobs, creating an ever-rising floor for competence). The “aha” moment in each industry will happen at roughly the 5% threshold, this is the point at which AI feels like magic. Depending on the degree of difficulty of jobs, the average degree of competence, and the availability of data, different jobs will fall at different times. Once a domain crosses the 5% threshold, we will get a gently increasing of the floor, year after year, model after model.

Economically, this “5% test” is likely to have the most impact in industries where the quantity of work product is constrained by the supply of workers. For example, in coding, there are not enough workers to fill the total number of job openings. So a bottom 5% software engineer is in fact immediately economically valuable. In other industries, say medicine, there is likely to be limited demand for more bottom 5% doctors, and so the copilot model (supercharging existing practitioners) is likely to be a more popular distribution mechanism, at least to start.

The 5% rule is a lower threshold than reasonable person’s AGI. In the immediate term, it may offer the simplest way of forecasting which industries are most likely to go through the most radical changes. For entrepreneurs, one implication is that rather than trying to beat the metaphorical Garry Kasparov, it may be more straightforward to simply outperform Dwight Schrute.

One unanticipated consequence of this is that the average baseline for competence is likely to improve. Humans will compete with AI, and they too will get better. This should be additive to GDP. And if indeed these changes happen more gradually, then it will give humans a chance to adapt, shifting into areas where they have a relative advantage vs. AI, and learning to work collaboratively with AI for maximum impact.

1. It would take an entire essay to unpack the impact of AI on friendship and social structures. I will not make that attempt here, although I anticipate it will be one of the most important longstanding impacts of today’s LLM technology.