When AI makes execution cheap, what kind of people do we actually need?

TL;DR

Falling execution cost does not resolve goal selection, value judgment or accountability. If anything the opposite: the stronger the execution, the larger the consequence of a wrong decision.

• What is interesting in Rieseberg's Hunger Beats Tenure is not "junior engineers can out-ship seniors." It is that as the external obstacles to action fall away, differences in internal motivation show up more directly in the output.
• There is experimental evidence that AI may raise everyone's average output quality while compressing the collective diversity of ideas (Doshi & Hauser, 2024). Each individual looks more productive; society's search space does not necessarily widen.
• The four capabilities answer four different questions: creativity asks whether a better possibility exists, independent judgment asks whether it is credible, self-direction asks who will turn it into action, responsibility asks how the result is made trustworthy and who carries the consequences. Drop any one and the loop breaks.
• Responsibility should not exist apart from power. Holding an employee accountable for an AI system's failure while denying them the authority to review, pause and correct it is not accountability — it is transfer.
• A closing caution: do not turn these four into a new workplace morality. Different roles need different capability profiles, and creativity and self-direction depend heavily on environment — someone without basic security or autonomy will struggle to sustain either.

On 27 October 2026, Anthropic engineering manager Felix Rieseberg published Hunger Beats Tenure: Advice for Junior Engineers [1]. He mentions having dropped the traditional coding test from his team's hiring — not because coding ability stopped mattering, but because the test no longer separates who will create more value in the new working environment. He observed engineers with relatively little experience using AI tools to reach a working prototype faster than seasoned developers, and even to push through projects that would previously have required cross-team coordination.

What is worth discussing underneath that piece is neither whether junior engineers can surpass senior ones, nor whether AI coding will replace traditional programming. It points at something deeper: when knowledge becomes easy to obtain, technology easy to use, and ideas easy to realize, where does the real difference between people come from?

Over the past few decades, modern education and corporate organizations built a relatively stable system for evaluating talent — degrees measure accumulated knowledge, years of service measure experience, job levels measure ability, task completion measures performance. That system formed in an era when acquiring knowledge was expensive, specialist skills were scarce, and coordination was costly. AI is changing part of the economic basis it rests on.

I increasingly think the most valuable people will not necessarily be those with the most knowledge, the most experience, or the greatest fluency with AI tools, but those who genuinely possess creativity, independent judgment, self-direction and responsibility. All four mattered before. In the AI era their value becomes more pronounced.

1. When execution stops being the bottleneck, human value gets reordered

Industrial-era organizations raised efficiency through division of labor: managers plan, specialists solve particular problems, executors complete standardized tasks. The knowledge economy raised the standing of specialists, but the underlying logic did not fundamentally change — firms decompose complex goals into manageable tasks, then coordinate resources through roles, processes, budgets and performance systems. One engineer owns an API, one designer owns the interface, one marketer owns acquisition; as long as each executes their piece correctly, the organization runs.

The core assumption is that specialist execution is a scarce resource requiring organizational allocation. AI is weakening that assumption. Today someone with basic technical judgment can use AI for requirements analysis, code, interface design, data processing, automated testing, even preliminary market research.

This does not mean firms no longer need specialists, nor that AI can reliably complete every complex task. The real change is that the initial threshold for execution is falling across many kinds of work — an idea that once needed several people to validate can now get an initial test from one; a plan that took weeks can reach a runnable version much sooner.

And as execution gets cheaper, new questions appear:

• If anyone can generate a business plan quickly, who can identify the problems actually worth solving?
• If anyone can ship a product quickly, who can judge whether that product has real value?
• If anyone can generate solutions in bulk, who can spot the errors and risks hidden inside them?
• If AI can autonomously execute more and more, who is accountable for the outcome?

Technology lowered part of the cost of completing tasks without automatically resolving goal selection, value judgment or accountability. If anything the opposite — the stronger the execution capability, the larger the impact of a wrong decision. Future competition for talent will not center on execution speed, but on problem judgment, creative exploration and the delivery of trustworthy results.

2. Creativity: not more ideas, but different possibilities

The easiest illusion AI creates is that one now has unlimited creativity. Put in a problem, the model returns ten approaches; put in a business direction, dozens of products; put in a technical requirement, several architectures. But a count of ideas is not creativity.

In 2024, Anil Doshi and Oliver Hauser published an experimental study on generative AI and creativity in Science Advances [2]. AI assistance improved the quality and creativity ratings of participants' short stories, with the largest gains for participants who were weaker to begin with. At the same time, the AI-assisted stories became more similar to each other overall.

That experiment targets a specific creative task, so it cannot be extended to a claim that AI reduces innovation everywhere. But it surfaces something worth watching: AI may raise each person's average output quality without increasing the diversity of thought across the group.

If everyone relies on similar models, similar data and similar prompting, the resulting plans converge — firms adopt similar growth strategies, startups pick similar product directions, creators settle into similar modes of expression. Every individual looks more efficient, and society's exploratory space does not necessarily expand.

The creativity that matters is not generating a few more ideas than the model, but the ability to break out of how a problem is currently framed. When users complain that software is complicated, the usual move is to cut steps, improve the interface, or add automation. A creative person might ask a different question: why does the user need to perform this operation at all? Can the task be eliminated?

The first searches for a better answer inside a fixed problem; the second redefines the problem. Human creativity does not come only from recombining knowledge. It comes from lived experience, social observation, value preferences, and a sense of where reality contradicts itself. AI can help explore a great many possibilities, but why a particular person finds a particular problem interesting, why they consider some status quo unreasonable, why they are willing to challenge an accepted rule — those remain bound up with human agency.

3. Independent judgment: when answers are free, questioning them is the scarce skill

If creativity determines how many possibilities we can see, independent judgment determines whether we can form our own view among them.

Acquiring knowledge once required reading, searching, calculating and repeated study — obtaining an answer was itself labor. Now AI can produce, in moments, an explanation that is well structured, fluent and superficially coherent. That lowers the cost of cognitive work and introduces a new risk: we may grow accustomed to accepting a pre-organized answer instead of building our own understanding. The more natural the expression and the more complete the reasoning appears, the easier it becomes to mistake linguistic coherence for factual correctness.

In 2025, researchers including Microsoft Research published a CHI study surveying 319 knowledge workers on critical thinking while using generative AI [3]. Within that sample, higher confidence in AI was associated with less critical-thinking effort, while higher confidence in one's own ability was associated with more critical-thinking activity.

This is self-report research and should not be read as proof that AI necessarily erodes independent thought. But it is a reminder that convenience may change how people participate in thinking.

Independent judgment is not contrarianism, and certainly not a refusal to use AI. It is the ability to keep a necessary distance between external information and one's own assessment. Facing a plausible-looking conclusion, one can keep asking:

• What assumptions does this rest on?
• What facts support it? Is there contrary evidence?
• Under what conditions does it fail?
• If it is wrong, who bears the cost?

Someone with independent judgment does not consider a problem solved because AI produced an answer. They treat the answer as a hypothesis available for testing, not a conclusion that is reliable by default.

More importantly, independent judgment shows up not only as questioning others but as willingness to revise oneself. Holding a position regardless of what evidence appears is not independence; it is cognitive closure. Real independence needs both the nerve to not defer to authority and the capacity to accept evidence and admit error.

More people will be able to retrieve answers quickly. People who can judge whether an answer is credible remain scarce.

4. Self-direction: AI amplifies the will to act, not just technical capability

Self-direction may be the most underrated capability in the emerging picture.

Traditional organizations drive individual action through external mechanisms — managers set goals, assign tasks, fix delivery dates, then evaluate completion. In that environment, someone without much initiative can still work steadily on the strength of clear assignments. But once AI can absorb a large share of concrete execution, the competitive edge of someone who only waits for instructions keeps shrinking — AI also takes clear instructions, and executes many standardized tasks continuously and quickly.

One important difference between people and AI will be whether a person can identify, on their own, a reason to act. A self-directed person does not only ask what task they should complete; they ask why it should be done, whether it actually matters, and whether something else deserves the effort more.

A concrete example: a product engineer is asked to build a feature. The execution-oriented response is to work out how to ship it on time. A self-directed person might first study why users need the feature, whether a simpler solution exists, and whether the request has drifted from the real commercial goal. The first creates value by completing the task; the second may extend into finding problems, improving decisions and creating opportunities.

Self-Determination Theory offers a useful explanation. Deci and Ryan identify autonomy, competence and relatedness as key to understanding intrinsic motivation [4] — when a person can endorse their own goals, feel their capability growing, and maintain meaningful connection with others, sustained self-directed motivation is more likely to form.

So self-direction is not workaholism, and not a demand for unlimited overtime. It means a person understands why they are acting, and remains willing to push a meaningful goal forward without external supervision.

This matters especially now, because technology has shortened the distance from idea to action. People once had plenty of ideas that technology, capital or team constraints prevented them from testing; some of those constraints are weakening. As the external obstacles to action fall away, differences in internal motivation show up more directly in the output.

Given the same AI tools, one person completes the tasks their manager assigns; another identifies a long-ignored problem, builds a solution, and pushes it into real use. Their technical resources may be similar and the value they create entirely different. In that sense AI is not only a productivity tool but an amplifier of the human will to act.

Self-direction should also include the ability to choose not to act. Initiative does not mean every idea must be built immediately — mature self-direction includes restraint, prioritization and abandonment. Recognizing what is worth doing, and what is not, is itself a scarce capability.

5. Responsibility: when AI can complete the task, who carries the outcome?

Of the four, responsibility is the easiest to overlook and probably the most critical.

AI can generate code, analyze markets, draft contracts, set strategy, and run increasingly complex workflows automatically. But a system completing a task does not mean it has assumed responsibility for it.

Suppose an AI agent generates a trading strategy, calls the trading API and submits an order. If a faulty risk assessment causes a serious loss, "the AI generated it" does not explain anything. We still have to answer: who decided to deploy this system? Who set its permissions? Who verified its safety? Who has the authority to pause it? Who carries the duty to monitor, correct and handle incidents?

These questions show that technical execution capability and accountability mechanisms are two different things. AI can raise capability to act without automatically establishing an effective accountability regime — and as AI systems gain autonomy, the boundaries of responsibility need to get clearer, not vaguer.

Responsibility here is not simply being conscientious, working hard, or owning up when something breaks. It is considering consequences before acting, identifying risk during execution, and being willing to follow up, verify and correct after the result lands.

A responsible engineer does not conclude a system is safe because the code passed AI-generated tests. A responsible product manager does not stop caring whether real users got value because the feature shipped. A responsible manager does not ignore the long-term problems a metric may be creating because the team hit it.

More importantly, responsibility should not exist apart from power. If an employee cannot review the AI's decisions, cannot pause a dangerous process, cannot change a wrong objective, yet must carry the consequences of the system's failure alone, that is not an accountability mechanism — it is a transfer. Real accountability requires decision authority, risk-control capability and the scope of blame to match each other. Responsibility is not only a personal quality; it needs institutional support.

The people worth trusting are not those who never make mistakes, but those who surface mistakes early, stay alert to risk, and push to get problems resolved. As AI multiplies an individual's execution capability, that awareness matters more.

6. The four form a loop, and dropping any one breaks it

Each capability answers a different question:

Together they form the full path from finding a problem to creating value: find the problem, form a judgment, act, verify the result, then revise your understanding against reality. AI can participate deeply in that process; people still carry responsibility for the goals, the key judgments, and the outcomes within their authority.

It also means we should not simply sort people into "uses AI" and "doesn't." The more interesting distinction is this: some treat AI as a tool for completing tasks faster, some treat it as a tool for extending cognition, exploring the unknown and creating new value, and some go further and ask how to make those capabilities work durably inside a reliable institution. The third group's long-term value is the one to watch.

7. Firms need to rethink what a good employee is

If the above holds, traditional evaluation systems need to change.

Firms have tended to reward steady execution, domain experience, process fluency and predictability. Those remain important, especially in safety-critical systems, financial infrastructure and complex engineering. But a firm that values only those metrics may miss the other kind of talent: people who find problems, challenge unreasonable assumptions, seek out opportunities and are willing to own outcomes.

The difficulty is that many organizations say one thing and reward another — they emphasize innovation when hiring and obedience when managing; they want employees to think independently while routing every significant decision through layers of approval; they encourage new ideas without providing a cheap way to test them; they demand accountability for outcomes without granting matching decision rights. The genuinely proactive people eventually lose momentum against repeated organizational friction.

In 1991, James March discussed the tension between exploration and exploitation in organizational learning [5]: organizations need to exploit existing knowledge for efficiency and to explore new possibilities to avoid being trapped by accumulated experience. That tension may sharpen in the AI era — AI greatly improves the reuse and execution of existing knowledge, while whether an organization is willing to try new directions still depends on culture, power structure and incentives.

Microsoft's 2026 Work Trend Index also reports that within its survey and analytical framework, organizational culture, manager support and talent practices correlate strongly with the AI benefits employees report [6]. These are correlational results and should not be read causally, but they are a reminder that the productivity value of AI is not determined by individual capability alone.

• An organization that wants independent thinking has to let people disagree.
• An organization that values creativity has to provide a bounded but real space for experiments.
• An organization that wants self-direction has to give clear goals and genuine autonomy, not ever more detailed task lists.
• An organization that emphasizes responsibility has to build clear authority, verification and accountability.

The strong firms of the future may not be the ones with the most employees, the most process, or the most deployed AI agents, but the ones that give people with judgment and initiative enough room to create value, while the organization carries the governance responsibility that is properly its own.

8. Education: from training people to answer questions to training them to ask

This change reaches education too.

One central task of modern education has long been to transfer existing knowledge and teach standardized problem-solving. That model produced a great many specialists and underwrote modern technical progress. But when AI can explain knowledge, generate code, perform calculation and organize answers quickly, we have to reconsider how education defines human ability:

• If a student finishes homework quickly with AI but cannot explain why that solution was chosen, what have they actually learned?
• If someone can build an application with AI but cannot judge its security, reliability or long-term maintenance cost, how should that capability be assessed?
• If someone holds a great deal of knowledge but has never found a problem, formed a hypothesis or carried out independent practice, has the knowledge become the ability to create value?

Education cannot stop at knowledge transfer and skills training; it needs to cultivate curiosity, critical thinking, creative practice, self-directed learning and a sense of responsibility.

This does not mean foundational knowledge matters less — the opposite. Without it, independent thinking degrades into ungrounded opinion; without professional training, creativity stays as imagination unconstrained by reality. The question is not whether to learn, but what the learning is ultimately for: knowledge should not serve only exams and assignments, but help a person understand the world, form judgments, solve problems and keep correcting their own understanding.

Education should put students through more complete cycles of exploration — find a problem in reality, form a hypothesis, attempt a solution, take external feedback, recognize errors, and finally own the work. AI can be the research assistant, the programming tool, the discussion partner and the experimental infrastructure in that process, but the goal should not be students who depend ever more on AI's answers. It should be students who keep independent judgment and the drive to create even after being handed powerful tools.

9. Don't turn a talent standard into a moral one

There is a tendency worth guarding against in any discussion of creativity, self-direction and responsibility: it is easy to read these as a new workplace morality in which good people must be perpetually proactive, perpetually enthusiastic, perpetually innovating. That reading is dangerous.

Not every role requires highly autonomous exploration, and not everyone has to become a founder. Steady execution, long accumulation, professional caution and good collaboration remain indispensable to a functioning society. The real point is that different work needs different capability profiles: in highly standardized work, reliability and consistency may matter more than creativity; in startups, product development and exploration of unknown problems, creativity and self-direction come to the front; where life safety, financial safety and significant social impact are at stake, independent judgment and accountability mechanisms are irreplaceable.

Nor can we ignore the environment any capability depends on. An employee without basic security, under unreasonable evaluation, with no autonomy, will struggle to sustain high creativity or self-direction. Demanding responsibility without granting power only manufactures a new unfairness.

So the problem organizations really need to solve is not only how to find people with these capabilities, but how to create an environment in which those capabilities develop rather than get consumed.

10. What matters most may be recovering human agency

Back to Rieseberg's piece. What draws me to Hunger Beats Tenure is not that a young engineer built a prototype over a weekend with AI, nor the question of whether coding tests should be dropped. It is the possibility it demonstrates: a person's value may depend less on how much organizational standing they have accumulated, how many resources they control, or how many levels of authorization they hold. If they can find a real problem, form their own judgment, and use tools to produce a credible result, they have a chance to matter beyond their title and tenure.

Underneath that is a question about human agency. Industrial society emphasized discipline and division of labor; the knowledge society emphasized specialization and information processing. The AI era forces a re-examination: as more cognitive and execution work goes to machines, where should people put their time and capability?

My answer is that people should value the capabilities that make them the subject of their own action — staying curious about the world rather than only accepting problems others define; forming independent judgments rather than handing thinking entirely to a model or an authority; exploring new possibilities rather than waiting for instructions; owning outcomes rather than pushing errors onto tools, processes or other people.

Ideally technology gives people more freedom to act, not less. But that potential does not realize itself:

• Someone who only uses AI to complete tasks others assign may become more efficient without becoming more independent.
• An organization that only uses AI to accelerate existing processes may cut costs without producing real innovation.
• If a society pursues only generation speed while neglecting creativity, independent judgment and accountability, we may simply repeat existing ideas and existing mistakes more efficiently.

So I am increasingly convinced that what is most worth cultivating is not only fluency with AI, but creativity to find new possibilities, independent judgment to tell right from wrong, self-direction to drive change, and responsibility to carry the result.

AI can give a person far greater execution capability. But what they ultimately do with it, why it is worth doing, and what they are willing to answer for, is what determines how much value the technology creates. As strong tools become universal, the lasting differences between people will show up more and more clearly in those choices.

The future belongs to those who can not only complete tasks but define meaningful goals and own the outcomes. Because what determines a person's value has never been only what they can do, but why they do it, and what they are willing to answer for in the world they help create.

References

[1] Felix Rieseberg (2026), Hunger Beats Tenure: Advice for Junior Engineers. https://felixrieseberg.com/hunger-beats-tenure-advice-for-junior-engineers/

[2] Anil R. Doshi & Oliver P. Hauser (2024), Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content, Science Advances. https://doi.org/10.1126/sciadv.adn5290

[3] Lee et al. (2025), The Impact of Generative AI on Critical Thinking, CHI 2025. https://doi.org/10.1145/3706598.3713778

[4] Edward L. Deci & Richard M. Ryan (2000), The "What" and "Why" of Goal Pursuits, Psychological Inquiry. https://doi.org/10.1207/S15327965PLI1104_01

[5] James G. March (1991), Exploration and Exploitation in Organizational Learning, Organization Science. https://doi.org/10.1287/orsc.2.1.71

[6] Microsoft (2026), Work Trend Index: Agents, Human Agency, and the Opportunity for Every Organization. https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization