The Principal-Thinker Problem and the Skunkworks

Einstein did some of his most important thinking while working at the Swiss patent office. The familiar version of the story makes this sound like an ingenious choice: unable or unwilling to tolerate academic conformity, Einstein found himself a quiet bureaucratic job and escaped into physics. The history is less tidy. Einstein wanted an academic position, tried to get one, and failed. The patent office was where he landed.[1]

What matters is what he later said about the arrangement. Near the end of his life he described the work of formulating patents as a blessing because it forced him to think in varied ways and provided useful stimulation for physics. He also saw an advantage in being outside academic life. An academic career, he wrote, put a young scientist under pressure to produce impressive quantities of work and therefore created a “temptation to superficiality.”[2]

The patent office was not without obligation. It gave him a demanding practical job requiring precision and mechanical understanding. On the other hand, it largely removed another set of pressures concerning what an ambitious young physicist should work on, how quickly he should publish it, and what recognizable scientific progress ought to look like. The useful distinction is between the different audiences present in the two working environments whose judgments a developing idea must eventually survive.

There is a familiar economic model that rhymes with this problem. In the principal-agent problem, a principal delegates work to an agent whose information or incentives may differ from the principal’s.[3] Intellectual organizations have a neighboring problem even when the person doing the thinking is diligent, loyal, and sincerely trying to produce something valuable. That person also learns which kinds of thought will be intelligible to the surrounding system, which questions attract useful engagement, and which ideas require fifteen minutes of translation before anyone can begin deciding whether they are good.

Call this the principal-thinker problem. The analogy is loose because the principal may not be a person at all. It can be a field made up of managers, colleagues, promotion criteria, grant committees, accepted models, standard analytical tools, and the accumulated expectations of everyone who will eventually have to understand the work. Its most consequential effects occur before formal evaluation, because people think in the presence of an anticipated audience.

Skunkworks matter because they alter that audience. They do not abolish judgment or substitute freedom for discipline. They change who gets to judge developing work, by what criteria, and at what stage. An immature idea gets a smaller jury before it has to become legible to the larger organization.

The Thoughts That Never Become Proposals

An idea rarely arrives as a finished object. It begins as a hunch, an anomaly, a mismatch, or something that seems worth another hour of attention. At that stage there may be a calculation to make, an experiment to try, a paper to read, a prototype to build, or several evenings that will disappear into finding out whether the thing is real.

The decision to spend that attention is already social. Someone who has worked inside an organization for a while develops a fairly accurate model of the people who will eventually hear the idea. If it fits the house language, the next step is cheap. If explaining it requires first explaining why the house language is part of the problem, the cost rises quickly. An idea can therefore disappear without ever being rejected; its originator can simply decide, consciously or otherwise, that another hour of thought is unlikely to repay the effort.

Experiments on collaborative idea generation support the basic intuition: anticipated evaluation can reduce not only the number of ideas people generate but the range of conceptual categories they explore.[4] The evidence cannot show us a private thought that never became identifiable even to the thinker, but it does show evaluation reaching into generation rather than acting only on finished proposals. That is enough to make the otherwise familiar office behavior more interesting.

At first some of the filtering is conscious. You know the quarterly planning meeting is not a promising place to question the metric on which the quarterly plan is constructed. You know which colleague can tolerate an unfinished idea and which one requires a conclusion before the premise has had time to become clear. You know that one line of thought will require a long translation into the vocabulary of the organization before anyone can decide whether it deserves attention.

None of this requires censorship. Much of it is ordinary social competence. The difficulty appears after enough repetitions, when strategic filtering becomes hard to distinguish from judgment. A thought once classified as expensive begins to feel unpromising, and a question that predictably produces no useful engagement stops feeling like an interesting question. The expected jury has reached backward into the search process.

Kuhn’s normal science describes part of the reason this can be productive. A shared paradigm tells a scientific community which puzzles are worth solving and allows knowledge to accumulate without reopening every foundational dispute. James March’s exploration-exploitation problem captures a neighboring organizational tendency: systems become very good at refining what they already know how to do because accumulated competence has real value.[5] The principal-thinker problem is narrower. It concerns what happens when the expected terms of later evaluation influence which candidate thoughts receive enough investment to mature.

The problem is not misaligned effort. It is aligned thought. An organization can sincerely ask for original thinking while maintaining an environment that teaches competent people, quite reasonably, which original thoughts are unlikely to be worth developing.

A Different Jury

Most people who have worked in a sufficiently large organization already know what skunkworks means even if they know very little about Lockheed. The proper name belongs to Lockheed’s Skunk Works, but the generic term has long since escaped it. Skunkworks are small, semi-independent groups operating near the edge of a larger organization, given unusual latitude to attempt work the main machinery would have difficulty performing.

Kelly Johnson’s original Skunk Works is useful because his operating rules show how misleading the language of latitude can be. Johnson eventually codified fourteen rules. The program manager was to have nearly complete control of the program. Both contractor and customer were to maintain strong but very small project offices. The number of people connected with the project was to be restricted “in an almost vicious manner.” Drawings had to be easy to change, reporting kept to a minimum, costs reviewed every month, specifications agreed in advance, funding kept timely, outside access controlled, and performance rewarded without making the number of subordinates supervised the measure of managerial importance.[6]

This was not a relaxed working environment. Technical standards, schedules, secrecy, cost, and personal responsibility were severe, and the U-2 team eventually worked roughly 65-hour weeks.[6] The important question is what those demands were attached to.

Johnson insisted that engineers and draftsmen sit no more than about fifty feet from the assembly floor. When manufacturing encountered a problem, the people building the airplane could take it directly to the people designing it. Engineers could make pencil changes on drawings, and problems that might otherwise have spent days moving through an organization could be resolved in hours. The arrangement shortened the physical and organizational distance between evidence and the people authorized to act on it.[6]

For someone well suited to that kind of work, Johnson’s rules could feel less like deprivation than relief. The airplane remained brutally difficult. What disappeared was much of the surrounding work of satisfying people who were neither solving the technical problem nor directly accountable for the result. The appeal is easy to understand if you would rather be judged by the airplane than by an organizational performance about the airplane. Johnson had changed the jury.

A conventional large project exposes developing work to many legitimate audiences. Engineering management, finance, procurement, inspection, the customer, senior management, and adjacent organizations all have concerns worth having. The problem is not that these audiences are foolish or malicious. Each additional audience, however, creates another requirement for an immature idea to be translated into criteria that audience already knows how to evaluate.

The Skunk Works compressed those interfaces. Authority stayed close to the technical object, the customer maintained a small counterpart organization, and designers, manufacturing personnel, and decision makers could interact directly. Developing ideas still had to survive criticism, but for longer they were being criticized by people whose criteria were closely tied to the thing being built.

Skunkworks therefore do not protect ideas from criticism. They can protect ideas from having to satisfy too many different kinds of criticism too early. An unfinished idea often needs a demanding audience; what matters is whether that audience can judge what the idea is trying to become.

The Outside

Banias gives us a useful test of the mechanism because Banias was not a skunkworks.

The Intel team in Haifa that developed it had just come off Timna, another processor designed in Israel. Timna was cancelled very late after the economics of the product deteriorated, and the team took the cancellation hard. Many of the same people then moved onto Banias, a mobile processor project that began from a different premise than Intel’s dominant desktop architecture. The local jury therefore had some history of its own before the architectural argument even began.[7]

The prevailing processor story was easy to understand: faster clock speed meant a faster processor. Intel had spent years making that relationship true often enough that frequency had become both an engineering objective and an extraordinarily legible public metric. The Haifa group was designing for laptops, where power consumption and heat changed the problem. Its architecture accepted lower clock speeds in exchange for doing more useful work with each cycle.[7]

Rony Friedman, one of the engineers, later described the position plainly: “frequency was king,” while the Israeli team was “on the outside.” Their problem was not merely to build the processor. They had to persuade the rest of Intel that a chip carrying a lower number on the industry’s most visible performance metric could be the better architecture. Banias became the Pentium M.[7]

The case is useful precisely because the alternate jury emerged inside an ordinary product organization. The Haifa site had its own technical history, local working relationships, and enough separation for a different frame to survive, but Banias ultimately had to win over Intel rather than remain protected from it. Geographic distance did not guarantee epistemic independence. It provided one of the conditions under which a local audience could keep an initially awkward idea alive long enough for the evidence to improve its reception.

The principal in this story is difficult to identify because there was no executive issuing an order that engineers must believe in GHz. The pressure came from prior engineering success, product strategy, customer expectations, marketing, technical habits, and the convenience of a single number everybody understood. A model can discipline thought without an enforcer.

Which Thoughts Should Be Expensive?

Banias also shows why the obvious response to the principal-thinker problem—make unusual ideas easier to pursue—is inadequate. GHz was not a stupid metric. It had earned its authority. For years higher clock speeds had accompanied meaningful performance gains, and a great deal of engineering and commercial knowledge had accumulated around that relationship. The danger appeared when a successful measure became easy to substitute for the more difficult judgment it had once approximated.

Organizations rarely need to say that an idea is too unfamiliar. They have better and usually more legitimate words. The proposal is not ready. It lacks clarity. It does not fit the strategy. The risk is too high. Those judgments are often correct, and organizations could not function without making them. The principal-thinker problem appears when some of those labels begin measuring the cost of translating an idea into the incumbent frame rather than a weakness in the idea itself.

Difficulty of reception is weak evidence. It attaches just as readily to a bad idea as to a frame the current vocabulary cannot yet hold. When legibility becomes a practical substitute for judgment, the cost of understanding the new frame can quietly become part of the case against it.

Meta’s Horizon Worlds shows the opposite problem. In 2022, internal data showed that many of the people building Horizon were not spending much time in it. Meta’s vice president of the metaverse asked why the team did not love the product enough to use it constantly; a later memo proposed holding managers accountable for weekly use and told the organization to make it its mission to “fall in love” with Horizon Worlds. Meta later retreated sharply from the VR-centered social-world thesis those memos were trying to protect.[8]

Requiring builders to use their own product is perfectly sensible. Dogfooding increases contact with evidence that abstractions and dashboards can hide. Asking people to make loving the product part of the mission is different, because it risks converting an unfavorable reaction to the product into a failure of alignment by the person having the reaction. Meta was not a failed skunkworks. The protection had attached to a conclusion rather than to an unfinished search, and the late jury—the users, usage patterns, and accumulated product evidence—was being given less authority to change the verdict.

Banias and Horizon illuminate opposite errors. An incumbent frame can make a live alternative too expensive because it performs poorly against criteria built for another problem. A favored thesis can become too cheap when the organization begins insulating it from evidence that ought to count against it. Skunkworks are useful because they delay some judgments, not because they guarantee survival for whatever is being protected.

Hybrid Air exposes a different failure. PSA Peugeot Citroën placed its compressed-air hybrid project inside a secret, urgent, highly autonomous group, and those conditions helped the team explore a radical technology. The same separation later made it harder to reconnect the work to commercial knowledge and to the larger organization, a problem researchers studying the project called the “Robinson Crusoe effect.” The search had been protected successfully; the difficulty came when the result had to make the trip back.[9]

Skunkworks address only the first of those problems, and they do so by changing the sequence of evaluation rather than abolishing it. Developing work receives one jury first and the larger jury later, after it has acquired enough evidence, machinery, and explanatory structure to survive a broader test. The arrangement works only if both juries retain their proper authority: the early one must be capable of letting the unfamiliar survive, and the later one must remain capable of saying no.

The principal-thinker problem is difficult to notice from inside because people become skilled at recognizing which half-formed thoughts are likely to justify the time, translation, status, and attention required to develop them. Some of that filtering is exactly what expertise is supposed to accomplish, while some reflects the local cost structure of the environment. After enough adaptation the two feel the same. The idea that has been priced out no longer feels suppressed or forbidden; it simply feels unpromising, which means the jury has done its deepest work before anyone notices that there was a jury.

Notes

[1] American Institute of Physics, “Einstein Exhibit — Formative Years III”, on Einstein’s unsuccessful efforts to obtain a university position and his subsequent employment at the Swiss Patent Office. See also John Stachel, “Einstein’s Discovery of Relativity”, for the employment and personal context surrounding those years.

[2] American Institute of Physics, “Einstein in the Bern Patent Office”. The page reproduces Einstein’s retrospective observation that a practical profession was a “salvation” for someone of his type and that academic pressure toward scientific production created a “temptation of superficial analysis.”

[3] Michael C. Jensen and William H. Meckling, “Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure,” Journal of Financial Economics 3, no. 4 (1976): 305–360.

[4] Xiang Zhou et al., “Exposure to Ideas, Evaluation Apprehension, and Incubation Intervals in Collaborative Idea Generation,” Frontiers in Psychology 10 (2019). Across the reported experiments, evaluation apprehension was associated with reductions in both the number of ideas generated and the number of idea categories explored.

[5] For Kuhn’s account of normal science, paradigms, and puzzle solving, see the Stanford Encyclopedia of Philosophy, “Thomas Kuhn.” For the organizational exploration/exploitation distinction, see James G. March, “Exploration and Exploitation in Organizational Learning,” Organization Science 2, no. 1 (1991): 71–87.

[6] Lockheed Martin, “Kelly’s 14 Rules,” for Johnson’s formal operating rules. For the U-2 development environment, staffing, work practices, and close physical relationship between engineering and production, see Peter W. Merlin, Unlimited Horizons: Design and Development of the U-2, NASA, 2015.

[7] Anand Lal Shimpi, “A Prelude to Success,” on the Timna cancellation and the transition of the Israeli design team to Banias; AnandTech, “Intel’s Pentium M on the Desktop — A Viable Alternative?”, on the separate Israeli design effort and the low-power premise of Banias; and David Shamah, “Intel’s Israelis Make Chip to Rescue Company from Profit Plunge,”, for Rony Friedman’s recollections that “frequency was king” and that the team was “on the outside.” Intel’s contemporary public framing of Banias as a high-performance, low-power architecture appears in Paul Otellini’s Fall 2001 Intel Developer Forum keynote.

[8] Alex Heath, “Meta’s Flagship Metaverse App Is Barely Used by the Employees Building It,”, The Verge, October 7, 2022, reporting Vishal Shah’s internal memos, including the weekly-use accountability plan and the instruction to “fall in love” with Horizon Worlds. For Meta’s later separation of Quest VR from Worlds and its shift of Worlds toward mobile, see Samantha Ryan, “Our Renewed Focus in 2026,”, Meta Horizon OS Developers, February 19, 2026.

[9] Caroline Aubry et al., “Managing Skunkworks to Achieve Ambidexterity: The Robinson Crusoe Effect,” European Management Journal 39, no. 2 (2021): 214–225. The study examines PSA Peugeot Citroën’s Hybrid Air project and argues that secrecy, urgency, and autonomy aided technological exploration while contributing to later difficulties transferring and exploiting the work within the parent organization.