Most companies are looking at AI through the wrong end of the telescope. They see tools like Claude or ChatGPT and think, “Great, I can fire half my engineers, cut my R&D budget, and build the exact same product for pennies.” This is a defensive, race-to-the-bottom mindset. If you are only using AI to shrink your cost center, you are completely missing the actual alpha of this technological shift.

To win, you have to look at AI through the lens of Jevons Paradox.

In economics, Jevons Paradox states that when technological progress increases the efficiency with which a resource is used, the rate of consumption of that resource actually rises because of increasing demand.

Apply this to software: AI is rapidly driving the cost of writing code to near zero while pushing the speed of creation toward infinity. If building is suddenly dirt cheap and blindingly fast, the winning move isn’t to build the same amount of software with fewer people. The winning move is to build 100x more product faster in the right direction — to iterate rapidly, explore adjacent markets efficiently, and ship features at a velocity that suffocates your competitors.

But to do that, you don’t need fewer people. You need different kinds of people. Perhaps, even more of them.

The Myth of the Autonomous AI

We have not yet reached a ubiquitous level of AGI. There is no AI on the market today that you can point at a business problem, walk away, and come back to a finished, resilient product that serves all consumer and business markets in a self-learning way.

I say this as someone possibly in the top 1% of users in agentic engineering, consuming billions of tokens every month purely on “vibe coding” with 1,000 contributions a month — aggressively deploying and steering models to build software. This amount of consumption gives you higher-fidelity intuitions around where today’s boundaries are. People bragging on X about “vibe coding an iOS app over the weekend” leave out the most critical variable: they were the ones providing the idea, taste, feedback to the plan, the course corrections, and the final reviews.

AI right now is a massively powerful engine beyond our imaginations, but it is not yet self-learning and self-improving.

Capabilities are incredibly “spiky.” A frontier model can spit out flawless Python logic for a complex backend system in minutes, but it cannot evaluate whether a user interface actually looks good. It is completely tone-deaf when it comes to understanding and generating the messy, human nuances of bad pronunciation or thick accents — a massive bottleneck if you are trying to build resilient voice AI agents.

To bridge the gap between an “impressive AI demo” worthy of a social media post and “production-ready software” to be given to your enterprise customers, you often need heavy, manual, human intervention, testing, and iteration.

The Death of the Specialist and the Rise of the Generalist

Because AI is spiky and constantly shifting, building a team of narrow specialists could lead to a dead end. A developer who only knows how to write React components is actively being commoditized.

The only true alpha a software startup has today is human capital — specifically, individuals with high curiosity, high agency, and high energy. You need technical generalists who don’t wait for a Jira ticket. You need people who will figure out what needs to be built, relentlessly beat the AI into producing the best code possible, recognize when the model has gone off the rails, and provide feedback and iterate around the “spiky” areas where the AI models currently fail or have blind spots.

This isn’t just a fringe theory. Boris Cherny shared that Anthropic’s Claude Code team today is actively hiring for these exact kinds of adaptable, high-agency generalists more so than narrow specialists. It is the only way to maintain resilience and productivity inside an R&D department while the underlying models change every few months, if not weeks.

The Hard Truth About Your Org Chart

If you want to achieve 100x velocity, you have to make hard, uncomfortable decisions about your talent pool right now.

You must aggressively onboard smart, high-agency generalists who view AI as a lever to multiply their own output, if not automate it entirely through first principles and a builder mindset. Simultaneously, you have to either rapidly upgrade or part ways with the seasoned specialists who expect to do one highly defined task for the next five years — tasks that are very likely to get automated and improved by AI. Those roles are already automated; the people holding them just don’t know it yet.

Until we hit true AGI — at which point all of this advice becomes obsolete — the companies that win won’t be the ones who used AI to shrink their payroll. They will be the ones who armed a small, elite team of high-agency generalists with infinite AI leverage, and out-built the rest of the industry.