AI coding tools are splitting the engineering workforce into two productivity tiers, and the gap is already enormous

Marc Andreessen puts the number plainly: leading-edge programmers are roughly 20x more productive than they were a year ago. Andrew Feldman, describing his top coders at Cerebras Systems, goes further, saying they have gone “from being sort of 10x guys to being 100x guys.” These figures sit at the high end of a range that multiple observers are now measuring, but the lower end of that range is itself striking.

Nathan Labenz surveyed a room of working AI practitioners and found the median self-reported productivity gain was two times. The framing around that number deserves as much attention as the number itself. Labenz notes the group put it this way: if you were not there, your productivity would drop to close to zero. The implication is not merely that AI helps. It is that for people who have built workflows around these tools, removal of the tools would be close to catastrophic.

Patrick Collison, Stripe’s chief executive, offers a structural version of the same observation. A single engineer, he says, can now do what two full teams of engineers could do two years ago. Collison also describes one engineer at Stripe who orchestrates 16 agents simultaneously from a single screen, with most of the relevant pull requests coming from that one person alone. Kyle Daigle, describing output at GitHub, states the organization is now doing more in a month than it did in an entire year the year before.

The median answer was basically two. In other words, people felt like they're getting two times as work done thanks to AI. But that was also framed in an interesting way where it was like, but note that as of today if you were not there, your productivity would drop to close to zero. Nathan Labenz

The gains are not uniform, and that unevenness is itself the story. Dara Khosrowshahi, Uber’s chief executive, reports that roughly 30 percent of the company’s engineers who used AI coding tools were power users, and those power users show clear differentiation in the number of diffs produced. A person identified only as Peter, observing the engineering workforce more broadly, describes a stark bimodal distribution forming: engineers who have invested the hours to learn these tools effectively on one side, and those who have not on the other, with an enormous productivity gap between them. Steven Bartlett, who remains active in hiring, says he now screens explicitly for AI proficiency in entry-level candidates, having concluded that a proficient candidate in the same role is a five to ten times more effective person.

The distribution of who benefits extends well beyond software engineers. David Sinclair reports his laboratory is doing work that would have taken 160 years and, in his words, “quite literally billions of dollars,” on a $10,000 budget. Dylan Patel describes a single person completing a research project that would have taken a team of 200 economists a year. Andrew Wilkinson recounts pointing a colleague who had never coded in his life toward an AI coding tool. That colleague, Wilkinson says, built something within about a week, a result Wilkinson describes as “insane.” Jesse Genet reports that she had never even opened a terminal until six months before, yet built her own tooling through AI assistance. The barrier that once separated people who could build software from those who could not is dissolving in real time.

Daniel Priestley, citing a claim attributed to Spotify, says the streaming company’s best developers have not written a line of code since December. That claim is secondhand, Priestley citing Spotify rather than speaking from his own observation, and should be read accordingly. It nonetheless fits the direction of every first-party data point in this picture. The gains documented across organizations, in other words, are real and measurable, even as the anecdotal ceiling sits dramatically higher.

What the evidence describes is not a productivity tool that helps everyone a little. It is a capability that rewards prior investment, separates practitioners from non-practitioners, and compounds over time. Employers who are not yet measuring the gap between their AI-enabled and non-AI-enabled workers are likely underestimating how wide it has already grown. The question for any organization that depends on technical output is not whether to take this seriously. It is whether the gap is still narrow enough to close.