[A lot of my readers seemed to like this post a lot. So, I figured I'd post it here as well. If anybody here is familiar with orgs that more or less run on a "Compton model," I'd love to hear about them in the comments. As I talk about in this piece, DeepMind is one prominent—and promising—example. Enjoy:)]

 

If you’re interested in the structure of scientific institutions, we’re living through remarkably exciting times.

This past week I was corresponding with Gerald Holton, whose 1952 work I covered in my piece When do ideas get easier to find?. Holton, now 100 years old, is obviously spending less time actively working and keeping up with the fields in which he was prolific in his heyday. Holton asked what people like me, Stuart Buck, and others in the progress movement do, and what the research and communication of ideas like “progress studies” and “good science” looks like.

“I have often thought about this concept, how it could be described and possibly measured, and how to defend it against skeptical opponents.” It was clear from his comments that Holton (and others researching these problems 70 years ago) did not see any reasonable path to the managers and administrators of scientific organizations incorporating this research into how they structured their organizations.

I felt lucky (and frankly a bit giddy) to be the one to share with him just how far this rough idea, which he’d been thinking about for over 70 years, had come. So, I told him about some of the fascinating research going on in the space. And, just as importantly, that there was now a community of chief science officers and scientific philanthropists who were not just open to this kind of evidence, but eager to consume it and learn from it.

I was fortunate to enter the field just as this wondrous new ecosystem of experimental new science organizations was emerging. But interactions with individuals like Gerald Holton constantly remind me that intelligent people have been thinking and writing about these problems for at least 100 years. And much of their writing is evergreen, written in a way that still sheds light on how to think about structuring scientific institutions today.

In this piece, I’ll detail a letter Karl Compton, then Professor of Physics at Princeton, wrote to Science in 1927. In the letter — which impressed many people with its clear accounts of the role of a university, the best way to select research topics on a university level, and how to administer departments themselves — Compton made himself known as not just a great physicist, but also a high-level administrative thinker. Members of the MIT Corporation would cite his impressive thoughts on these topics when they would make the odd choice of offering this Princeton physicist who didn’t consider himself extremely industrially-oriented the role of President of the Institute a few years later.

His thoughts are quite interesting because for many the present structure of university research departments is the most natural organizational arrangement. Most professors are given their own separate pot of money, access to their own grad students/postdocs/RAs, and are more or less in charge of their own time. But Compton, in his letter to Science, highlights a pretty natural alternative organizational structure that really doesn’t get as much consideration as it should.

Meeting in the Radiation Laboratory at the University of California, Berkeley in March 1940 to discuss the 184-inch cyclotron. From left: 1939 Physics Laureate Ernest O. Lawrence, 1927 Physics Laureate Arthur H. Compton, Vannevar Bush, James B. Conant, Karl Compton, and Alfred Loomis. Source: Lawrence Berkeley Laboratory Public domain via Wikimedia Commons via nobelprize.org

The Context

At the time Compton wrote his letter, the US was just beginning to take the role of research and research universities as seriously as Europe had historically done.

 

Early in his letter, Compton takes the strong stance that research should be considered a vital university function. While he knows the readers of Science agree with him, it is clear that he is advocating for an opinion that is far from unanimous.

The three great functions of a university are to train young people in the art of living, to guide in the search for truth and actually to engage in pursuit of truth. The first two of these are universally agreed upon and endorsed, but not so universal and whole-hearted is the recognition of the third — the research function of an institution of higher learning. In some quarters the research aspirations of a university meet with distinct disapproval as encroaching upon the supposedly more serious business of the institution. In other quarters they are viewed with grudging or amused tolerance as harmless little idiosyncrasies in which scholarly men must be indulged in order to keep them contented and out of mischief. In the more enlightened quarters, however it is realized that a university can not best perform any of its functions or measure up to its opportunities unless full and ungrudging support is given to attempts to advance human knowledge.

Compton goes on to defend the role of research at a university and its vital role in training students with reasons that we’d almost unanimously agree with today. Some of his points include:

  • If you want to train students to truly organize all available knowledge and use it to overcome problems and meet difficult situations, what better training than research?
  • In an age where students were excited by looking forward, Compton believed, introducing students to the concept of research progress was pivotal in fueling interest in subjects in a way that static textbook knowledge could not.
  • Compton, having worked at multiple universities, believed the difference in student interest levels between his previous university that heavily incorporated independent undergraduate research and those that did not emphasize student research to be incomparable. Students found it challenging in the best way possible.

He goes on to explain things like why he believed research makes an instructor better at their job, to point out how many of America’s glorified inventors built inventions that depended on research in one way or another, and how he thinks universities should be funded. I won’t cover all of that in this piece, but if you’re interested I encourage you to read the whole letter here.

At the end of the letter, he goes on to give a very insightful analysis of how he conceptualized the right way for universities to build up and manage research operations.

Better Administration of University Research

Being involved with both industrial research and university basic research at the time, Compton saw the clear need for both. He praised industrial R&D labs and their uncanny ability, where profits were concerned, to produce research outputs that seemed “greatly exceed the individual capacities of the research workers” involved.

But he personally enjoyed the goals of his basic university research more:

Basic research must…be a free and unfettered search for truth. It is the universities alone which can offer any considerable opportunity for such endeavor.

However, he did not think that the structure of university research departments was perfect. Compton observed that research was starting to become more specialized and cooperative. With that, he saw another area of major opportunity to improve university research as a whole: better administration.

At the time, no universities were heavily backing the research enterprise remotely to the extent that any research university would today. The rise of the research capacity at universities and its administration was often quite ad-hoc. This should be unsurprising because, as Compton noted above, many universities saw research as the thing they allowed their teaching men to do to appease them. They often viewed the enterprise with “grudging or amused tolerance as harmless little idiosyncrasies in which scholarly men must be indulged in order to keep them contented and out of mischief.”

The merits of specialization

In reflecting on what he felt universities should do to most efficiently produce knowledge, he makes several points that are quite non-obvious. The first is on the natural advantages of departmental specialization.

Another solution can advantageously be advanced by wise administration of the universities. There seems to be a widespread, but ill founded, feeling that all departments of a university should be developed together and kept closely abreast. Perhaps this relieves the administration from embarrassment, but I venture to suggest (though the suggestion is not new) that this is not sound educational policy except for an ideal institution which has unlimited resources. Such a policy dissipates effort, and if every institution followed it we should have the spectacle of a great many universities all very much alike and all with struggling, mediocre departments. Much more effective in advancing knowledge as well as in bringing distinction to the university is the policy of supporting to the available limit certain departments selected because of their already outstanding character, or because of the traditions and purposes of the university, or for any other reason. If these favored departments are chosen in a coordinated group, then the university becomes an active center for the development of that field and the promotion of cooperative effort. For example, one institution may choose to give particular facilities for advanced work in classics and languages, another in historical, economic and social sciences, another to physical and biological sciences, etc. If we were to examine the record of those universities of limited endowment which have nevertheless been preeminent in the life of the country, we should find that they attained this preeminence through concentration of effort. The words "To him that hath shall be given" apply here as well as elsewhere.

Through concentration of effort in a coordinated group of departments, a university has the opportunity not only to correct the dangers of overspecialization, but also to take a strategic position in fulfilling its obligations to society.

Compton observing that these policies would lead to “the spectacle of a great many universities all very much alike and all with struggling, mediocre departments” was remarkably prescient. In the field in which I read the most papers, economics, a dozen or so universities make up around 80% of the papers I read. The mass amount of resources spent producing under-utilized research from dozens of additional expensive-to-fund economics departments almost makes my head hurt.

And there is an empirical literature attacking questions like this as well. Just recently, a Nature paper by Wapman et. al was making the rounds covering a similar topic: that 80% of all professors have been trained in 20% of PhD programs.

“For faculty with US doctorates, we find that academia is characterized by universally extreme inequality in faculty production. Overall, 80% of all domestically trained faculty in our data were trained at just 20.4% of universities. Moreover, the five most common doctoral training universities—UC Berkeley, Harvard, University of Michigan, University of Wisconsin-Madison and Stanford—account for just over one in eight domestically trained faculty.”

Those dozen-ish top Econ departments which I usually read are almost exclusively housed in universities that are highly rated across all academic fields. These include your usual Harvards or Stanfords which seem to be winning the tournament model of academia. But, within those dozen are two specialized departments that have risen above the more middling ranks of their universities’ general research departments via specialization. Those two exceptions would be UC San Diego, with its department’s heavy focus on time series analysis, and George Mason, with its focus on areas such as public choice and Austrian economics.

For anybody whose knee-jerk reaction is to think those specialized departments’ researchers are more constrained in some way, I’ve never met a member of these departments who feels like they are more constrained in their publishing than somebody at a place like UC Berkeley. In fact, a noticeable percentage of the weirder paper topics I’ve ever come across have come from the George Mason department (see this paper by George Mason’s Peter Leeson on the economic efficiency of trial by combat). These schools simply acknowledged what most businesses intuitively understand and what economists say they understand: comparative advantages are remarkably productive, particularly for up-and-comers.

Yet, specialized efforts like this that productively contribute to the diversity and progress of a field are remarkably rare.

How Compton felt a research department should be run

Compton then went on to share his ideas on how to structure research departments themselves:

Much can also be done to promote cooperation and coordination through actual methods of organization. This has been strikingly demonstrated in some of the big industrial research laboratories, from which the output has greatly exceeded the individual capacities of the research workers and has been achieved only by coordination of effort. Such organization requires a very wise and far-visioned director who can visualize the big objectives and steer through the mass of petty details which must be worked out in order to attain them.

While this research model was working quite well on an industrial scale, in which labs often held hundreds of employees and researchers, he did not think such a model was necessary desirable on the scale of a university.

In a university, where the number of workers is much smaller than in a big industrial laboratory, such army-like organization does not appear feasible or probably desirable.

The management model of a research department that he goes on to propose is remarkably insightful. It is much more dynamic than a large industrial lab, but also far more structured than a department of mostly autonomous professors doing ad-hoc research with their own separate funds.

There is another direction in which more effective organization is possible within the universities themselves! Departments of a somewhat more flexible nature than those to which we are accustomed and which could, more than now, be built around one or two outstanding men in the department, could give these men an opportunity for organization and concentration of effort which is now rarely possible. This would, of course, require careful selection of men.

His comments were brief, but they are quite interesting to consider. A physics department with $20 million and two elite scientific directors able to allocate all resources as they saw fit could do remarkably different research than a department of 50 physicists with about $400,000 each. The model is also flexible in a way that makes intuitive sense for scientific work.

If the scientific directors, the “great men” as Compton referred to them, saw it as in-line with the specific goals of their department, they could spend a large chunk of the money on equipment and technicians and maintain a smaller stable of professors. They could also hire data engineers or whoever else they needed to optimally push a specific scientific area as a whole forward. And, even in a technical department like that, they could hire, fund, and give an office to a theoretician to largely do work on their own if the scientific directors felt the theoretician was too good of a talent to pass up or they valued the individual as a sounding board for others’ work.

For many, the concept of a department as something along the lines of 50 autonomous professors with 50 separate bank accounts working more or less on their own projects feels like a natural default. But, as you consider the kind of department Compton is describing, where a few scientific directors set some general goals/a mission and can then just spend the money however makes sense, it becomes clear that this “Compton model” is also a quite reasonable default.

A (possible) modern analog employing the “Compton model”

In trying to come up with a modern analog for something like Compton described, DeepMind comes to mind. Particularly in its early days leading up to/directly following its Alphabet acquisition. People like Demis Hassabis and one or two others from the early DeepMind team could be considered the scientific directors which Compton mentions. For the three year period prior to DeepMind’s acquisition, those directors dedicated their “department” of around 50 people of varying skillsets to solving a remarkably important basic research problem in the field: beating top human Go players.

The total staff of DeepMind was around 50 at the time of the acquisition and they’d raised around $50 million in total. Just doing some back of the envelope math looking at the number of faculty in the CS Department at a place like the University of Wisconsin Madison, the rough salaries of academics at different levels, and the general size of NSF grants in the area of computer science, it would seem that DeepMind’s yearly budget was not dissimilar to a department like UW Madison. Maybe it was as something like double, but probably not much more than that prior to its acquisition.

It’s harder to tell what happened with their budget in the years after their acquisition, but the comparison to UW Madison in terms of staff size post-acquisition does not seem ridiculous. Judging from the AlphaGo documentary, it seems that the team had noticeably grown from 50 since the acquisition. But, in making the comparison, it should also be kept in mind that while university departments may only have around 50 professors, each professor often has multiple grad students and post docs contributing substantial effort into projects in addition to department administrators and others adding to the total headcount. Regardless, the small details of the comparison, whether DeepMind should be thought of as 0.5 university departments or 2.5, don’t matter very much in contributing to the general point.

What matters is that how DeepMind was able to deploy its resources and staff was remarkably different than a traditional academic department. And it seems to have worked phenomenally. Using inputs not dissimilar to the University of Wisconsin Madison, a great department that produces consistently good research, it would be hard to argue that the DeepMind/Compton model didn’t produce something quite special.

The DeepMind office late 2015/early 2016

The “directors” chose a fundamental research question that was widely acknowledged to be a question of massive importance to the field. DeepMind had the flexibility to staff themselves the way they needed to solve it. Of course, there were talented PhD researchers from multiple related disciplines as you’d find in academic departments, but also your workhorse software engineers and data engineers whose skills make projects like this work that are sorely lacking in most university departments.

The problem was not cherry-picked either, it was considered a common opinion in the field to expect that a Go algorithm that could compete at a world-class level was ten years away. As DeepMind moves into fields like protein folding and nuclear fusion, the organization continues to show substantial promise in a variety of areas.

The success of DeepMind’s organizational model is definitely a very strong data point in favor of Compton’s 100 year old hypothesis. Compton was not an outsider who knew nothing of being an independent, unfettered researcher within a university department. He was one at Princeton. And he was not some armchair theorist who knew nothing of the nitty gritty of what made industrial research work. He was a longtime contractor with General Electric and was so valuable to them that the GE CEO, Gerard Swope, insisted on recruiting Compton to be the MIT President in 1930 even though Compton didn’t really have much initial interest in the job.

Compton, having deeply experienced both worlds, felt like this model just made sense. It allows for a certain level of freedom in choosing problems and job security that comes in academic departments, but also a more ideal allocation of resources than you’d tend to find in a university. It does seem like this model, if it were pervasive, would be much less susceptible to some noted inefficiencies present in our current academic STEM ecosystem.

Namely:

  1. The failure to compile and maintain large datasets that could be pivotal to a field
  2. The failure to build and maintain software packages that could be pivotal to a field
  3. And the underinvestment in instrumentalists who could continually improve upon the machinery that makes science possible/design new instruments altogether

The model might be susceptible to some new problems, as is the case with any model, but it does seem like it would likely be strong in areas in which our current model is weak.

Looking Ahead

If anybody with special knowledge about DeepMind disagrees with this characterization or has any comments, I’d love to speak with you! DeepMind is an organization I’m fascinated by and that I’d love to learn more about.

Also, I would love for readers to reach out to me with additional research organizations, successes or failures, that they think fit the Compton-model of an organization and are worth looking into. I’d love to build up a small “dataset” of these to better explore the model’s tradeoffs and when it seems to succeed vs. fail. For now, I’m left wondering just how many areas of research would turn out to be much less wicked than they currently seem if confronted by a Compton-model research organization.

DeepMind’s success on a problem closer to basic research, like it’s Go work, and applied problems, such as AlphaFold, should provide strong optimism that there are a wide variety of areas that we can make a major dent in with such an approach.

8

5 comments, sorted by Click to highlight new comments since: Today at 3:23 AM
New Comment

Interesting essay. I think the idea of flexibility is good in general: for example, I know of a case where a scientific programmer position on an academic project cannot be properly filled because the salary is far below the market; this is well understood by the project leader but cannot be fixed due to strong regulations of fund allocation. 

However, I'm wondering what would this model lead to for PhD students that want to stay in academia? Already in the current system, there are not enough tenure positions for everyone, in the Compton model it seems that there will be even fewer, which would discourage some amount of people from following an academic route.

So it's unclear, to me at least, whether this would grow or shrink the number of academic jobs. The number of "budgetary shot callers" would shrink by a lot since far fewer professors would have a bank account/grant to direct. But many of these teams would still require PhDs to be doing real research.

To many, that sometimes sounds like a demotion. But I think that's a knee-jerk way to look at it. You can liken many of the new roles to what it was like being one of the many many smart PhDs/professors who worked on something like the Manhattan Project or in the GE Lab. You had probably a sort of comparable amount of freedom since your mind was respected by your "bosses" and they let you explore how to find things out on your own, but you did your research while also productively working within a "team" rather than a loose collection of individuals with different goals entirely.

Of course there are certain jobs that PhDs currently do, such as writing software, that would probably be efficient to hand over to non-PhDs in many cases. But the whole operation would also work more efficiently with the unified budget and possibly have more money for PhD researchers as well. So I find it hard to tell whether jobs would be harder or easier to come by.

You mention the Manhattan Project (govt) and GE (industry), and I feel like there are still a lot of similar non-academic research groups that run on more like a Compton model. For instance I worked at the Boston Fed, the CFPB, and the World Bank (as well as in academia) and they all have that flavor of working on a team toward a common goal, with some direction from above. We were also given some time purely for our own independent research.

It would be interesting to break down R&D models into attributes, like:

  • Average team size?
  • How are problems selected?
  • How is funding allocated?
  • Average amount of funding controlled by a single manager?
  • How is talent selected?
  • How is talent developed?
  • Etc.

And then fill out a table of those attributes for all the major models (university lab, corporate lab, DARPA, etc.)

This is interesting. I guess I anecdotally have started doing this but my memory (like everyone's) can be quite faulty/hazy. Maybe I'll start keeping a running spreadsheet that I mark up as I read.