AI coding productivity is real, but the gains belong almost entirely to the engineers who went all in
Reports from inside engineering organizations put productivity gains anywhere from 2x to 100x. That range is not confusion in the data. It is the data. The gains are concentrated among a subset of practitioners who made AI tools central to their work, not peripheral to it.
The productivity numbers being reported from inside engineering organizations do not form a tidy consensus. They span an order of magnitude. Marc Andreessen puts leading-edge programmers at roughly 20x more productive than they were a year ago. Andrew Feldman, describing the top AI-assisted coders at Cerebras, says they have gone from being 10x contributors to 100x contributors. Nathan Labenz, reporting from a practitioner gathering, found the median answer was a 2x productivity gain, though with a pointed qualifier: productivity would drop to near zero without human oversight still in the loop. The range from 2x to 100x is not a sign that the evidence is confused. It is the evidence. The gains are real and they are not evenly distributed.
The structural pattern beneath those numbers is what several observers describe as a bimodal split. Peter, speaking on Latent Space, describes a subset of engineers who have gone fully in, invested the hours to learn the tools, and the productivity gap between them and those who have not done that is enormous. Dara Khosrowshahi at Uber offers a concrete shape to that split. Of the engineers at Uber who adopted AI coding tools, roughly 30% qualified as power users, and those users show clear differentiation in the number of diffs they produce. The majority of adopters are not in that group.
What makes this split more than a usual early-adopter curve is the speed at which the gap is widening. Feldman’s 10x-to-100x framing implies that the ceiling is still moving. Kyle Daigle at GitHub adds a platform-level data point that fits: GitHub is now doing more activity in a single month than it did across an entire prior year. That is not explained by a modest efficiency gain spread evenly across users. It implies a concentration of output at the top.
You you start to see more of a biodal distribution of engineers, right? You you start to see like, wow, there's there's this subset of of people that they they really get it like they're they're all in and they they've clearly invested the the hours needed to learn these tools and and how to be effective. >> And then there's sort of the the group of people that haven't done that. And that the productivity gap is just enormous. Peter
The gains are also appearing well outside the trained-engineer population, which complicates any reading that treats this as merely a professional productivity story. Andrew Wilkinson describes a CFO with no coding background who, within about a week of being introduced to Claude Code, built a sophisticated portfolio tracking tool. Jesse Genet reports that she had never even opened a terminal until six months ago, when she started building with AI coding tools. The barrier being lowered is not just speed for people who already knew how to code. It is entry into the activity itself.
Steven Bartlett draws a hiring implication from this that is worth taking seriously. He reports passing over entry-level candidates who lack AI proficiency, because a candidate with that proficiency in the same role is, by his estimate, a five- or ten-times more productive hire. If that judgment is spreading among hiring managers, the downstream effects for candidates without AI fluency could arrive faster than the standard lag between productivity shift and labor market response.
The practitioner evidence is not uniformly bullish in its texture. Labenz’s 2x figure comes with the caveat that human oversight remains load-bearing: remove the human and productivity drops to near zero, not to some lower baseline. That is a different claim than saying AI augments engineers. It is saying AI currently operates in a supervised mode where the human is still doing the work of verification. The 20x and 100x figures presumably come from practitioners who have internalized that oversight function to the point where it no longer feels like a constraint, but the distinction matters for any organization trying to replicate those results by simply deploying tools.
The implication is not that most organizations are failing to use AI. It is that the productivity gains now visible at the leading edge are not automatically realized by adoption. They are realized by a subset of practitioners who have made the tools central to how they work, not peripheral to it. Engineering teams that treat AI tooling as an optional add-on will produce results that look like the median. Teams that produce the outlier figures appear to have done something more deliberate. Whether organizations can manufacture that commitment at scale, rather than waiting for it to emerge organically among the 30% who self-select into power use, is the operational question the evidence raises but does not answer.