“Engineers are becoming sorcerers” | The future of software development with OpenAI’s Sherwin Wu

The productivity gap between AI power users and everyone else isn't just widening , it's creating two entirely different species of workers. At OpenAI, engineers now deploy fleets of AI agents that can knock out code reviews in minutes instead of hours, while most of the business world still struggles to get positive ROI from basic AI deployments. This isn't a temporary learning curve. It's the emergence of what Sherwin Wu calls "sorcerers" , developers who wield AI tools with such fluency that they operate at superhuman scale.

Wu, a product manager at OpenAI who works directly with the company's developer platform, has a front-row seat to this transformation. His team has reached the point where "we've basically reached that point now" of using AI agents as core infrastructure rather than experimental tools. But the picture he paints is more complex than the Silicon Valley hype machine suggests. While OpenAI's own engineers are becoming productivity wizards, the rest of the world is still figuring out how to make AI pay for itself.

The Deployment Reality Check

The dirty secret of enterprise AI adoption is that most companies are burning money. Wu observes that "a lot of companies actually have negative ROI on their AI deployments," though he's careful to note the measurement challenges. The pattern is becoming clear: organizations rush to deploy AI tools without understanding how to integrate them effectively into existing workflows. They're buying hammers when they need to learn carpentry.

This disconnect stems from what Wu identifies as a fundamental misunderstanding about where value comes from. "Silicon Valley is a bubble. Software engineering is a bubble. Most people in the world, most people in the US are not software engineers, are not very AI pilled," Wu explains. The result is that best practices developed by AI power users don't translate to mainstream business contexts. Companies copy the tools without copying the mindset.

The stress factor compounds the problem. Wu notes that even at OpenAI, "it happens all the time" that agents break mid-workflow, creating anxiety about wasted time and broken processes. If the people building these tools struggle with reliability, imagine the friction for organizations trying to bolt AI onto legacy systems and untrained workforces.

The Coming Software Cambrian Explosion

But Wu's most provocative argument isn't about current deployment failures , it's about what happens when the barriers finally fall. He predicts a "boom where anyone can build software for anything," leading to "100x more of these startups" than we've seen before. This isn't hyperbole. As AI tools become more accessible and reliable, the number of people who can create functional software will explode exponentially.

The early signs are already visible in how AI startups are organizing around vertical niches. Instead of building horizontal platforms, successful AI companies are diving deep into specific industries where they can understand unique workflows and pain points. Wu sees this as a preview of what's coming: thousands of micro-software companies serving hyper-specific use cases that were previously too small to justify custom development.

This vertical specialization strategy also addresses the platform risk that haunts every AI startup founder. When asked about OpenAI potentially crushing competitors by building competing features, Wu's response is refreshingly direct: "the market is so big and so massive. Like, I actually think, you know, startups should just not overly think about where OpenAI or these labs are going." The implication is that OpenAI will focus on infrastructure and foundational capabilities, leaving endless vertical opportunities for specialized players.

Preparing for the Productivity Divide

The strategic question isn't whether AI will transform how work gets done , it's whether your organization will be among the sorcerers or the muggles. Wu's observations suggest three critical preparation areas.

First, invest in AI literacy before AI tools. The companies succeeding with AI deployment aren't necessarily the ones with the biggest budgets, but those that understand how to redesign workflows around AI capabilities. This means training people to think in terms of AI-assisted processes, not just AI-powered features.

Second, watch for the vertical software explosion. As development becomes more accessible, expect a flood of niche solutions that can outcompete general-purpose tools in specific contexts. The companies that thrive will be those that can quickly evaluate and integrate specialized AI tools rather than trying to build everything in-house.

Third, prepare for the reliability threshold. Wu's team represents what's possible when AI agents become dependable enough to be core infrastructure rather than experimental additions. When that reliability arrives for mainstream users , and Wu suggests it's coming soon , the productivity gap will accelerate dramatically.

The transformation Wu describes isn't just about better software development. It's about the fundamental reorganization of how value gets created in a world where code-writing becomes as common as spreadsheet-building. The early sorcerers are already emerging. The question is who will join them before the gap becomes unbridgeable.