You are sitting in a group on the steppe 2 million years ago. If someone squinted hard while driving past in a car, the least hairy of you might look vaguely human. You grab a rock, look at it for a moment, and slam it down on another rock. In doing so, the rock shears. Rock is now sharp. As far as this group knows, you just invented technology. Congrats.

Everyone around you starts doing the same. Picking rocks up, slamming them against other rocks to create sharper, sub-rocks. All but one of you. They stare at the rock to sub-rock shearing spectacle wondering: “How could this be? How could such a rock-to-sub-rock shearing spectacle possibly be?” They stare and stare. They stare for a long time, long after you and the group have started using the rocks to shave sticks into points for the next hunt and, upon returning, use them again to remove the animals’ hides.

After a long time the watcher starts to jump up and down! They know why rock became sub-rock. It comes to them in a single moment of perfect clarity. Kinetic energy from the slamming motion became elastic deformation. Stress concentrated at the tips of tiny cracks in the rock. When

the cracks propagated and the rock fractured! As far as this group knows, they are watching the first eureka.

Stunned with this new knowledge and its limitless potential applications for rock slamming and sub-rock creation, they run over to you, grab you by the shoulders, stare deep into your eyes, and say “ooga booga.”

This is the relationship between science and technology. Craft, practice, and tools create phenomena that provoke a desire for explanation, or open new ways of investigating the world. Those explanations suggest future technologies or further explanations. Advancement is entangled and contingent. Sub-rocks shape sticks and combine to form axes to chop down trees to melt rocks into metals to build engines attached to presses of sharpened melted rocks to fashion new axes. Brian Arthur calls this process combinatorial evolution. Eventually, your group on the steppe might find some use for the conceptual, rather than tacit, understanding of elastic deformation. First, you’ll have to invent language.

It’s wrong, then, to believe the linear model is an accurate story of innovation, and yet it’s wrong-er to believe the idea began with The Endless Frontier or to ignore the utility of a seemingly silly notion.

If we were asked to start from scratch and inject dollars into the entangled morass called innovation, we may again find it simply easier if every actor is assigned a role. The scientist discovers. The engineer applies. Businesses develop. Markets receive. Schoolhouse Rock’s “How an Idea Becomes an Innovation.” Where do you even exert leverage in the chain-link model?

And we may find it more satisfying to create heroic sub-plots within that story. Status can go to the idea-haver; everyone downstream is progressively more sullied. Direction is implied… the idea must be heading somewhere. If, as reality bleeds into the dream, the natural journey is interrupted, something else must be responsible: the market, the user, the realities of cost. A valley filled with death.

The first survey of the valley

It’s funny how much of Science was there from the start.

Written seven years after the creation of the Royal Society, Thomas Sprat’s History was less history and more a request for patience from a public growing unsure whether this new institution was useful, ridiculous, or heretical. It included arguments addressed to the King about the role of science funding after the English Civil War. Funding experimental science is a more glorious use of royal patronage than enlarging the empire.

Sprat listed inquiries dispatched with travelers bound for China, the East Indies, Morocco, Iceland, and elsewhere, along with observational instructions for sailors. And devoted pages to Francis Bacon’s distinction between experimenta lucifera (light-bearing experiments) and experimenta fructifera (fruit-bearing experiments). Causes and uses. Basic and applied.

For Sprat, the old philosophy, practiced by the Greeks and their admirers, produced knowledge that was “a Shipwrack in the end of the Voyage.” The preferred metaphor being a shore of death rather than a valley, which makes sense for a society that has just taken to the seas. Yet we find the same implied directionality of a voyage and the requirement for an explanation of failures to arrive.

The Society believed that experimental philosophy could find useful knowledge through observing phenomena in the world, and dreamed of building encyclopedias of such knowledge. In the early years of the Society, the mad lads actually tried it.

“The nature of things betrays itself more readily under the vexations of art than in its natural freedom.” - Francis Bacon, Novum Organum (1620)

The History of Trades was to be a comprehensive account of all the techniques, tools, and secrets employed by craftsmen throughout England. The workings of the entire economy in one convenient place. One of organized science’s first Grand Challenges.

Not everyone felt happy about visiting the workshops where this knowledge lived. John Evelyn attempted to inventory roughly three hundred subjects, from brewing and dyeing to glassmaking, engraving, and wire-drawing.

In 1659, he admitted to Robert Boyle:

“In the History of Trades, I am not advanced a step; finding (to my infinite grief) my great imperfections for the attempt, and the many subjections, which I cannot support, of conversing with mechanical capricious persons.” - John Evelyn to Robert Boyle in 1659

Translation: the scope is absurd, the company is worse.

Boyle, a bit wordily, viewed this reluctance to enter the workshop with contempt. The craftsmen might actually observe more diligently than the philosophers for the simple reason that their livelihoods depended on it.

“The phaenomena afforded by trades, are a part of the history of nature, and therefore may both challenge the naturalist’s curiosity and add to his knowledge, Nor will it suffice to justify learned men in the neglect and contempt of this part of natural history, that the men, from whom it must be learned, are illiterate mechanicks... is indeed childish, and too unworthy of a philosopher, to be worthy of an honest answer.” - Robert Boyle, The Naturalist’s Insight into Trades (1671)

Or William Petty, the son of a cloth-maker, who spent his boyhood watching smiths, watchmakers, carpenters, and joiners, and could see something the others couldn’t. He wanted techniques collected so that they could migrate between the trades. Improvement in one practice might unlock improvements in others. He proposed a college where the most ingenious practitioner of every trade could live and work alongside each other.

What had they produced that was answerable to their mighty hopes?

Histories of gunpowder and dyeing. Reports on glass, metals, brewing and other trades. Notebooks… many, many notebooks. One of Evelyn’s, Trades: Seacrets & Receipts Mechanical as they came casually to hand, is 600 largely empty pages, containing headings with nothing underneath them. I’ve been there.

These are all fragments of the totalizing vision. For all the typical reasons: the scale was too large, personnel (the virtuosi) weren’t sufficiently committed, trade secrets stood in the way, and the part of the project imagined as easy—making tacit knowledge intelligible—turned out to be the hardest thing.

Their mighty hopes gave way to what reality permits. Fragments of an idea jumping ship to arrive at shore.

A thousand valleys for every dream

We are still better at envisioning the totalizing future than the particulars of how. This will change everything.

Has 3-D printing crossed the valley? Yes and no. Judged by its mighty hopes, it’s a spectacularly successful failure.

“Automated fabrication will take its place among the major industrial developments of the 20th century.” - John P. McTague, a Vice President at Ford, in the foreword to Marshall Burns’s Automated Fabrication (1993)

Burns believed its impacts would be larger than the computer revolution. 3-D printing would put a factory in every home and collapse global supply chains. Obama partly agreed.

“3D printing has the potential to revolutionize the way we make almost everything.” -President Barack Obama, State of the Union Address, 2013

Chris Anderson thought, while writing at Wired in 2012, that kids today will grow up thinking that making stuff is as simple as publishing a blog.

That future hasn’t arrived and there are good reasons why it may never arrive. And yet, 3-D printing is remarkably successful. Want a rough and dirty prototype of an end-effector for your robot? Print one. Recently got a new dental crown? It may very well be 3-D printed. The iPhone Air and Duo use printed parts. It can have enormous advantages over milling in weight, time, and customization. In aerospace and other industries where structural integrity matters a whole lot, each new part still has to prove itself to regulators.

It’s succeeded where peculiar advantages outweigh peculiar limitations, shedding destinies until it found uses.

Maybe it’s more useful to think of the singular valley as many plural ditches. The tacit-to-formal one Evelyn drove into. The combinatorial ditch of our watcher on the steppe. The scale ditch. The diffuse-financial-capture ditch. The regulatory ditch. The society-has-a-vote ditch.

We tend to hide different causal mechanisms beneath a shared story. What is actually happening and what, if anything, can be done about it?

Many times, the idea is simply bad. Bad for its time, context, or problem at hand.

That’s okay.

If you spend weekends looking at US R&D data, one of the surprising things you’ll notice is that there’s been an explosion in businesses performing “basic research” over the past ten years. By 2023, we were at a level approximately on par with 1953, before much of the federal spending for R&D started.

Basic research is itself an odd category. Jane Calvert’s 2006 paper “What’s Special about Basic Research?”, remains a great primer on its slipperiness. She interviewed scientists and policymakers about how they defined and used the term, and found their answers were often context dependent.

One biologist, when speaking with the more business-minded, would drop the word basic early on in the conversation. This signaled that the suit shouldn’t expect anything patentable or profitable. They quickly lost interest. Others would do the reverse, talking up the industrial relevance to be more competitive for funding, even if the underlying work was essentially the same.

This slipperiness makes the term useful for establishing boundaries and expectations, deferring certain types of evaluation, and preserving the image of the scientist as an independent truth seeker.

There are economic explanations for why a business would conduct exploratory, long-horizon research that we would normally classify as “basic” (see Nelson, 1959; see Rosenberg, 1990; and see Cohen and Levinthal, 1989).

These tell us why businesses might do research. But how and why do they make a big show of it?

I think it breaks down into three categories: shows of force, access tokens, and crystal balls. Scientific companies have to show that they’re capable, that they legitimately belong within science, and that they can be trusted as guides to the future. These are all claims about the type of company they are.

Shows of Force or We’re really really good at this.

In this bucket are your Deep Blues, Watsons, and AlphaGos. These take a nebulous, difficult-to-evaluate technical capability and place it inside a public spectacle that is easy to understand.

Chess is hard. Grandmasters are good at the hard thing. The computer beating a grandmaster is impressive. That’s all there is to it.

These can be contests, where the rules are understandable, or heroic triumphs over the limits of nature.

At the 1964 World’s Fair, IBM had a machine that transferred Russian text over telephone lines to a computer ninety miles away and returned it as English. People couldn’t believe it and accused IBM of using a Mechanical Turk.

The first transcontinental telephone call was a staged spectacle. Alexander Graham Bell was in New York and Thomas Watson was in San Francisco. Bell repeated the same words from the first telephone call, which covered just two miles: Mr. Watson, come here. I want you. Watson replied that it would now take a week. President Woodrow Wilson and the mayors of the respective cities were wired in on the group call. Nothing to add from my end, thanks!

Nixon called Neil and Buzz on the moon. AT&T is partly to thank for giving us this quote.

These events work because of compression. They take years of research and enormously complex technical systems and make them publicly understandable. Someone speaks; someone impossibly, unfathomably far away answers.

They invoke wonder in the capabilities of those companies. And become entwined in the stories those companies tell about themselves. Move 37 is a landmark in AI research and a naming convention for Google real estate.

Access Tokens or We belong here.

Scientific knowledge is sometimes imagined as a static object. It sits on a shelf in a library or journal somewhere, freely available to anyone capable of reading it. Like any human endeavor, though, scientific knowledge is partly social. It moves in conferences, collaborations, and informal conversations between people who are engaged in interesting work and know who else is.

To access that layer of knowledge, you have to earn your place as a genuine and trusted participant in the network.

Publishing is one way of buying an access token. New papers contribute knowledge to the community, which is a genuine public good. It signals that in this company, we believe in science. It’s also reciprocal participation in a knowledge system: I disclose some of what I know to learn some of what you know.

For the scientists, publishing papers, presenting at conferences, and collaborating with colleagues allows them to remain recognizable as a scientist. Giving them freedom tells other researchers that you can work here without ceasing to be a scientist. This pitch is surprisingly powerful. Scott Stern found that scientists were willing to trade some salary for jobs that allowed them to pursue and publish research.

Access tokens work through reciprocity. The company makes some knowledge public and, in return, gains access. Papers appear, doors open. The identity of scientists is recognized, the company gains talented people.

I explore the norms of science in businesses in much more depth in this essay:

Crystal Balls or We see the future.

The performance of science is also a way of showing strategic foresight. Research becomes evidence that the company not only understands the future but is actively creating it, confidently steering the company into its waters.

Earlier this year, Google announced that it would migrate its systems to post-quantum cryptography by 2029. The company pointed to advances in quantum hardware, error correction, and estimates of the resources required to break existing encryption. A few days later, its researchers published new estimates suggesting that a quantum computer could break widely used elliptic-curve cryptography with roughly twenty times fewer physical qubits than previously estimated. Translation: this is coming faster and cheaper than expected.

The research was a genuine contribution to the field and supported Google’s strategic story. We can see what is coming, and we know what to do about it. The company’s ability to produce scientific knowledge becomes evidence that it can be trusted to make decisions using that knowledge.

These performances often move smoothly between scientific and managerial language. You’ll find them in forward-looking statements, shareholder reports, and the like. Technical results become forecasts, which establish deadlines, which are then decomposed into near-term investments and product strategy.

Crystal balls work by translating ambiguous scientific progress into confidence in corporate judgment. We’re creating the future and plan to be here when it arrives. Trust us with your capital.

Open problem summer

As we watch open problems in mathematics begin to fall, what I find most interesting from a corporate comms standpoint is that these distinctions collapse. A new result can be all of them.

It’s a show of force. You don’t need to understand the underlying mathematics to understand that cracking a problem that has been open for fifty years is impressive. Which is good, because my prompts often look a little like this:

It’s an access token. Resolving an open problem is a genuine contribution, allowing the company to be associated with the production of scientific knowledge. It’s a signal that we take science seriously and are interested in supporting it.

And a crystal ball. The advances are often produced by internal versions of upcoming models. OpenAI reported that the token spend on cracking 10 open problems was about $2,000 at API rates. The scientific contribution becomes evidence about their upcoming products, their capabilities, and their economics. Research and advertisement merge into one thing.

Here’s what I mean.

This is from the September 8 announcement by OpenAI where they claimed a solution to a Millennium Prize Problem. This is one of seven problems established in 2000 to organize the field. It is the second to be solved after the Poincaré conjecture. There’s active discussion on the validity of all this, which I’ll leave to the mathematicians to sort out. And they are trying to sort it out! Fittingly, much of the dispute is over who gets credit for the knowledge and whether the field recognizes this contribution as legitimate.

Even stranger, perhaps, is that this isn’t always centrally staged. Models are released into the wild and users aim them at problems as weekend projects. Even if they aren’t directly a part of the company, they function as a diffuse extension of the strategy.

This diffusion is an interesting wrinkle. Stranger still is that companies are beginning to make a different kind of claim about what their systems do.

AlphaFold was revolutionary, but it was recognizable as a tool scientists could use to pursue scientific questions. Closing open problems produces scientific knowledge that did not previously exist. This renegotiates who or where we expect knowledge claims to come from. The old solution to this problem (insulate a lab, hire real scientists, and let reciprocity work) doesn’t obviously transfer when the knowledge maker isn’t socializable.

Mathematics is beginning to work out what it means to recognize machines as knowledge makers. As capabilities expand, the same question will make its way into other fields with the companies needing to pay the field’s corresponding access tokens.

Less what kind of science is this? More what kind of scientist is this?

Several years ago, at an event in the Bay, I was talking to an expert in human–robot interaction. We were watching a demo from Bear Robotics, at that point just a scrappy prototype. You can imagine it as a Roomba with a rack attached to the top, designed to move food and drinks around a restaurant.

As we watched, they remarked that they wished it went further, completely removing the interaction between the server and the served. I want to enter a restaurant, order on a screen, have the robot deliver my food, and not be bothered.

Internally, I thought: Jeez, that’s a tad odd. Externally, I mimetically repeated the pitch we’d just heard. It’s great! Servers have more time to focus on value-added tasks: making sure diners are enjoying themselves, answering questions about the menu, and upselling. Moving food is the commoditizable part of the job!

After spending much of my teens and early twenties working in restaurants, I was also thinking about Chef Mike, an archetypal, slightly imbalanced kitchen guy. If food sat in the window too long, whoever eventually came to collect it got a tongue-lashing, even if they were only helping out. Better the robot than me.

So, as we watched this robot scoot around the conference floor, I felt that I was right. The thing is, so did they. The divergence was really about something else.

What Is a Restaurant For?

If you lived in Pompeii or Herculaneum before it blew, a restaurant was a stall you brought your own plate to, were served warm slops and stews from large dolia, and then ate your meal outside. Having an oven at home was uncommon.

For most of history, this was roughly what prepared food sold in public looked like. There were places that offered whatever was available or had been prepared on that particular day, but you didn’t have a choice in what those things were. At a table d’hôte in eighteenth-century France, diners arrived at a fixed hour, sat together at a shared table, and ate the same meal.

The modern European restaurant changed this, though Europe was a little late to the party. “Restaurant” comes from the French restaurer (to restore or refresh). Originally, they sold restorative broths to people too delicate or fussy for ordinary food. The meal was a response to an individual constitution, so it came with an individualized system of service. The restaurant was, as historian Rebecca Spang puts it, a public place people visited to be private.

The menu is so normal to us now that we don’t really see it. It’s a technology for translating the present capabilities of the kitchen into a legible set of consumer options. The waiter makes that large, shared production system responsive to the desires of a specific person. That responsiveness can be functional. Bring the food I selected to the table where I am sitting. It can also make us feel recognized, welcomed, or important.

Can a robot make me feel like Henry Hill entering the Copacabana? Probably not. It doesn’t always need to.

There isn’t one restaurant experience. Automats were fast, autonomous, and came with a futuristic pleasure in receiving lunch through a tiny glass door. Fast food sells predictability. Fine dining, theatre.

Each has their own expectations about choice, interaction, and care, even if the physical task of moving food is approximately identical. Sometimes, the food is only background to something more interesting.

Steven Shapin’s paper, “Breakfast at Buck’s,” looks at a breakfast spot in Woodside, California, famous for being a meeting place for Silicon Valley types. The legend goes that Netscape, Hotmail, PayPal, and Tesla all inked deals there. The truth isn’t all that important.

After Buck’s was known as a place where companies start, people went there hoping to start companies.

Breakfast is short, informal, and deniable. It signals more interest than an email, less commitment than dinner. Eating together allows you to see things about someone else that a pitch doesn’t. At Buck’s, you could perform informality while evaluating one another seriously.

No one really cared about the pancakes. A robot could move around the plates just fine, but moving around plates was not the only thing that restaurant was for.

From moving plates to picking them up

In an old job, I traveled around the US, visiting robotics labs and doing Gemba walks. Of all of them, BMW’s Spartanburg plant stands apart.

Automotive manufacturing has been among the easiest industries to find opportunities for automation. Motions are repeatable, material is typically non-pliable, and there’s enough production volume to prorate the cost of a robot over its lifecycle.

The best manufacturers, including BMW, chase the manual tasks left between what’s been automated. One of those is parts loading. There’s nothing exotic about using a 6-axis arm for this. It would do it faster and more reliably than a humanoid, but you’d need to redesign the cell.

That means bespoke engineering, production stoppages for install, more floor space, and a system that may need rework whenever components or vehicle models change. A humanoid can theoretically fit in this specific, human-shaped spot more easily.

Of the possible tasks one could be assigned, parts loading is perhaps the shortest straw. It’s not dirty or dangerous. It is dull, exceptionally so. You can actually learn the job right now. Stand here. Move these plates of metal from this stack onto this fixture. The fixture will rotate the piece into the robot’s cage. Then, the robot does what it does. Don’t worry too much about that, but it expects the new part to be there when it’s done. You run on its time. Repeat all day.

But from observing someone do this task, you discover qualities about them that would’ve otherwise been hidden. Do they show up on time or late? Are they dependable? Are they flaky? Do they keep pace with the cycle time or cause shutdowns? Do they understand and reliably observe safety protocols?

The visible task’s output is not necessarily its only function. Watching someone do a task can be a cheap way of identifying whether or not they’re trainable for other, higher-level ones.

In both the best and worst cases, you rotate through people. Good workers move into better positions. Others wash out. Then the job is posted and filled again, which adds time and cost. We should automate jobs like these. There are better things people can be doing and better ways of assessing their potential. Assessment will still have to happen, just somewhere else.

BMW will figure that out. They have teams devoted to understanding how tasks, people, and production systems interact. Households do not.

Kind-of useful tasks

On a good day, I shoot 10% from the 3. On bad days, I have fun trying. Having those percentages, I chase the basketball around a lot. When I saw this recent demo from Galbot, I imagined a glorious future where I didn’t have to.

The robot may not replace a specialist basketball or tennis coach, but I’m not currently participating in the market for one. Getting good isn’t all that important. A humanoid that could chase the ball and occasionally notice that my elbow is flaring is kind-of useful. I can imagine lots of kind-of useful tasks. Daisy-chain enough together, squint really hard, and a humanoid could make economic sense.

But do I want to be the person at the park playing basketball with a robot? Do enough of us? General-purpose robots will be able to do far more than we routinely ask them to do. The asking will be shaped by socially acceptable types of interaction.

Which raises the question of whether I’ll reliably ask for the whole bundle of kind-of useful tasks that underwrote the original purchase. A Sprinter van sounds fun. If weekends aren’t spent camping, it becomes a lawn ornament.

A general-purpose robot becomes economically plausible by performing many kind-of useful tasks. Performing those tasks requires it to learn your life. The better it learns your life, the less it resembles a fungible appliance, like a dishwasher.

Will there be an aftermarket for humanoids people have spent months talking to before realizing that the bundle wasn’t all that useful? Do you feel comfortable factory-resetting something that’s walked around your home?

There’s an idea that consumer robots will deploy faster because household tolerances are less demanding than those in manufacturing. A kind-of well-folded shirt is fine.

These are physical tolerances, but sentimental and moral tolerances exist too. The average home is awash with them.

The Venetian Murano vase you bought on your honeymoon is more precious than the birthday gift you politely feigned excitement over. The urn with Fido is next to a flowerpot on the mantel. One should be watered. The other should not.

Of course, the humanoid can categorize objects differently and assign them different protocols. This is a baseball. You can toss it around. This is the baseball from the first home run you scored in Little League. Treat it as you would a museum piece. What does the interface look like? Am I responsible for giving the robot the history of every object on its welcome tour of the house?

Each household will add exceptions to the general capability. What the tool can do will exceed what it’s trusted to do. And humanoids will need to make their general purpose-ness legible to the particular lives they enter.

Just like in Paris, we’ll need menus.

This post was brought to you by Erik Satie’s Gymnopédie No. 1, the first restaurant music.

Research institutions are complex systems made up of interactions between funding, governance, people, strategy, and accumulated know-how. When we write about them, we can make those interactions feel more orderly than they are.

Osteri Labs is a place for short, playable models that make these systems easier to explore. They’re made for teams or individuals who are creating research institutions or want an on-theme way to do some team building.

The first simulation is called Build a Tool That Travels. In it, you found an institution, make some important decisions, and see whether you can complete your mission on time.

The underlying game engine includes some randomness. Well designed institutions can still struggle and questionable ones can get lucky. The intention is to provoke discussion between people and surface common or orthogonal ways of thinking about the same problem.

You can access it at labs.osteri.blog. If you want to use it for a group activity, there’s also a downloadable facilitator’s guide with info on how it works, how to structure the activity, and some leading questions to get people talking.

I’m planning to build more missions in this series and some visual explorables (like you’d find here). If you have ideas or feedback, please reach out by commenting or sending me an email at kyle@osteri.blog.

The simulator

Working title for this series is “Trade-offs, Events, and Trajectories: Research Institution Simulator.” You can call it T.E.T.R.I.S. for short.

You’ll set up an initial design for an institution, making choices about how patient your funding is, how concentrated it is in one or many sources, how closely coupled you are to end-users during development, and more.

These choices impact the probabilities of project success during the run and the likelihood that certain decision points will come about. You had some early success and your founder (which is you) was noticed? Congrats! You now have to decide whether to stay with the institution or leave for the new job and let the existing governance play out. This can impact the game differently depending on how you initially configured your project governance.

Here’s what a run looks like.

Some strategies work better than others, but most can be successful

There are three main outcomes from the simulation, you can complete the mission (either by the end of the run or before schedule), reach the end of the cycle without fully completing the mission but having some transferable output or likelihood of continued existence, or run into financial failure.

Here’s the distribution of several institution archetypes across 10,000 seeded runs.

Focused delivery completed the mission most often, but other strategies were successful too. Exploratory portfolios tended to be less likely to finish, but also didn’t run into financial failure.

These aren’t meant to be empirical judgements about which institutions are best. The intent is to create plausible scenarios with plausible win percentages that allow teams to think through trade-offs and strategies.

Inspiration

There was a great piece recently in Analogue Press, which you can read below. It’s an inventory of the different characteristics an R&D organization can have.

In it, they made a provocative claim that based on the particular configurations of these traits, there are 2¹⁰ potential distinct types of organizations. Which is mind-blowing. I think the actual number is much smaller than that, but that the design space is also larger than the current types we have.

Navigating this space is a challenge and I thought simulations could be a fun way for people to do it.

Thought experiment. A medical-components manufacturer is chasing hairline cracks in an injection-molded housing. It’s cycled through resin suppliers, adjusted cooling cycles, and increased injection pressure.

The manufacturer has tried everything it knows to try, so it calls in two outside advisers. They arrive with access to the same AI.

The first asks the company what it needs. Writes down “assistance with injection-molding defects.” Queries the literature. Identifies several plausible specialists, including one at a university fifty miles away. Knows someone there, it turns out. Mentions they’ve just gotten new microscopy equipment. Offers to make an intro. Finds some outside funding that might be available for an assessment.

This is useful, but the search space is enormous.

The second adviser has worked with the company on other projects for the last fifteen years, including on one previous attempt to fix this process. They know its machines, its people, and its customer’s requirements.

They start probing, and, in their probing, compress the search space.

Didn’t you recently change your material handling process? Huh, that’s weird, these cracks are mostly appearing near one weld line. Whoa, who do you have running the dryer on the night shift? Defects rise 20% then.

They ask the AI what could cause delayed cracking near a weld line in this particular polymer after sterilization, given inconsistent pre-molding moisture levels, the pressure limits of the company’s machine, and the results of previous experiments.

Maybe neither adviser solves the problem. But which one is more likely to?

Coming abundance and bottlenecks

The recent White House OSTP report, A New Golden Age, sets an ambitious goal of doubling the productivity and impact of American science within a decade. The report correctly describes discovery as an iterative loop in which academia, industry, engineering, and manufacturing all play a role in shaping research questions and discoveries. Adding AI will, the report argues, accelerate that loop.

This raises a question: what must exist inside, or immediately around, a company for that loop to work?

There are already good summaries of AI takeoff and diffusion constraints, see Cowen’s. As it relates to institutions becoming the bottleneck, there are at least three different problems.

The first is a throughput bottleneck. Drug candidates are generated faster than regulators, clinical trials, or factories can process them.

The second is a jagged frontier bottleneck. Some sciences accelerate faster than others, but the slow movers are on the critical path for overall progress.

The third, and subject of this post, is an absorption bottleneck. Useful knowledge exists, but firms cannot recognize what matters, relate it to their particular circumstances, and make use of it.

This concept is borrowed from Cohen and Levinthal. Each organization has a variable ability to see the value of external knowledge, assimilate it, and apply it commercially. This variable is called absorptive capacity. A firm needs internal expertise to understand external expertise, so internal R&D has value because it both produces new firm-specific knowledge and allows the firm to benefit from knowledge produced elsewhere.

A manufacturer can be located inside a research cluster. A scientist can rotate into industry for six months. Neither means that the company will recognize which discoveries matter or connect them to its own problems.

This is already true without AI. But dramatically expanding the search space compounds the challenge of navigating it. A small manufacturer will not read ten times as much science simply because ten times as much is being generated. Its attention and knowledge of its own systems have not increased at the same rate.

This is especially true if the manufacturer is reading no science at all.

Variety and where firms once kept their attenuators

Let’s borrow Ashby’s Law of Requisite Variety here. A system, in our case a manufacturer, can only respond effectively to its environment if it possesses as much variety as the environment. In this push and pull, amplifiers increase the variety available to a system, while attenuators reduce it to something that can be managed.

Our wicked smart AI might become a fantastic variety amplifier, but where are the attenuators? Maybe it’s also AI, but it still needs firm-specific knowledge to navigate the search space.

There’s a tendency to treat the heyday of corporate labs as solely a story of a time when firms created small internal universities and produced new science (see my article below).

But these internal universities were also sensory organs and variety attenuators. Their scientists followed external research, translated it into company problems, and created technically capable counterparts for the broader scientific community to communicate with. This expanded the range of scientific signals the company could perceive and use. Beginning in the 1980s, large American firms have steadily withdrawn from science as corporate R&D has shifted toward developing and commercializing existing knowledge.

The machinery for producing science travels with the machinery for absorbing it. Most small companies never built the machine.

Back at the farm

All of this is a bit too ones and zeros. How about a pastoral palate cleanser?

At the beginning of the twentieth century, the boll weevil took a road trip across the American South, destroying cotton crops as it went. The federal government had developed control methods, but a leaflet from Washington was unlikely to get farmers to actually use it. Why trust someone who has never worked your land?

So Seaman Knapp, a USDA official, tried something different. He arranged for farmers to test new cotton varieties and cultivation methods on a Texas farm. He let them see how these new methods performed under conditions resembling their own. Knapp’s been quoted as saying, “what a man hears, he may doubt; what he sees, he may possibly doubt; but what he does himself, he cannot doubt.”

The demonstrations spread throughout the 1900s and 10s through county agents. Agricultural journals existed. Experiment stations produced research. Colleges trained experts. But the Cooperative Extensions founded through the Smith-Lever Act of 1914 built a chain long enough for all of this to reach the farm.

A farmer could not be expected to personally follow developments in all the sciences relevant to their farm. Through the county agent, they could externalize absorptive capacity. The agent lived near the farmers, understood local conditions, and could reach back into a larger network of specialists and experiment stations when necessary. The problems farmers encountered could travel back toward researchers, too. This connected two bodies of knowledge: agricultural science and farm-specific.

On the shop floor

Land-grant universities were originally created for both agriculture and the mechanic arts. Why did only one side acquire an extension system of comparable reach and permanence?

Part of the answer is that the need to externalize absorptive capacity is easier to see on the farm. They’re numerous, dispersed, and it’s a bit absurd to expect they’ll employ their own soil chemists, plant pathologists, and entomologists. Local similarity of soils, pests, and weather makes a territorial system more plausible.

Manufacturing is jagged. The same region might contain a foundry, a machine shop, a food processor, a chemical plant, and an electronics manufacturer. Each working with different equipment and materials, drawing on different bodies of technical knowledge. An industrial county agent would either have to remain a generalist or have access to a much larger network of specialists.

At the time these things were debated, it wasn’t clear that American manufacturers couldn’t look after themselves. This was the era of GE, DuPont, Bell, Kodak…, all maintaining their own labs. Smaller firms could just obtain whatever knowledge they needed from a network of equipment suppliers, customers, and consultants.

This type of government support also looks like a subsidy to private companies or public competition with consultants who sell the same service. But if we’re renegotiating science and industrial policy, maybe that debate needs to be reopened?

In 1965, America briefly attempted something resembling an industrial Smith–Lever Act. The program began developing field services through universities and state institutions. It lasted five years.

One of the early experiments that preceded the legislation was done in 1964. NASA funded a Technology Use Studies Center at a small college in southeastern Oklahoma. It connected NASA’s vast scientific and technical information system to manufacturers across seventeen mostly rural counties.

Reading the report, it becomes clear that the challenge for manufacturers was more formulating technical questions and less access to technical information. Firms with limited engineering capabilities required an “interpretive step” between research and use. A specialist had to define sophisticated knowledge in terms of the client’s particular problem.

The U.S. tried to create an extension service again in 1988, partly due to rising anxiety about American manufacturing’s decline relative to Japan. This produced the Manufacturing Extension Partnership (MEP), with centers still operating in every state plus Puerto Rico. They help small and medium-sized manufacturers improve quality, introduce automation, train workers, adopt lean, comply with new standards, and locate outside expertise.

That can be very valuable in the same way our first adviser was valuable. But MEP was not built like agricultural extension. Its centers are heterogeneous public-private organizations rather than the local arms of a common research system. Half of their funding comes from nonfederal sources, and they’re evaluated through the attributable results of client projects.

These incentives favor a particular kind of intermediary and a particular kind of problem solving. A manufacturer arrives with an issue. The center scopes a service, provides it directly or locates someone who can, records the resulting cost savings or new sales, and moves on to the next project.

Some MEPs have in-house engineering capabilities, others operate as consultancies and network brokers. The system, though, does not require every center to maintain laboratories, conduct research, or accumulate technical knowledge of local firms.

This may be changing. A new NIST competition for fourteen MEP centers placed greater emphasis on helping manufacturers demonstrate, deploy, and adopt advanced technologies. This could pull them toward deeper technical capabilities.

Broadening the aperture

Ok, we’ve tried some experiments. Maybe this particular externalized absorptive capacity institution is just not possible for manufacturing. Let’s broaden the aperture a bit.

Since the early twentieth century, Japanese prefectures have had public industrial research and testing centers known collectively as kohsetsushi. Initially, they served agriculture and other regional industries, but gradually extended to manufacturing.

The centers collectively employ thousands of engineers and researchers. They maintain testing equipment and laboratories, provide technical consultation and training, conduct applied research, and sometimes develop inventions that local firms can license. A small manufacturer that cannot employ a specialist can temporarily reach one via a center nearby. And because they’re a steady local presence, trust accumulates between the actors.

They’re proof that absorptive capacity can sit outside the firm, while still accumulating firm-specific knowledge.

The English-language literature on kohsetsushi is a bit thin. Philip Shapira produced an account in the 90s and, more recently, Nobuya Fukugawa has assembled empirical evidence on how the centers solve problems and learn about local technical needs.

Kohsetsushi sometimes make referrals and MEPs sometimes contribute to deep technical work. But the center of gravity is different. The kohsetsushi can occupy the overlap between two bodies of knowledge: the evolving technical frontier and the accumulated problem histories of local firms.

Back to our two advisers. AI should make both of them better. The first will search a larger literature, identify specialists more quickly, and make better referrals.

But the second adviser can relate what science knows generally to what has happened inside this factory specifically.

Scientific abundance amplifies that difference. The more possible explanations, the more important it is to know which possibilities can be discarded. At least in the near-term, the value of situated knowledge may increase disproportionately.

This doesn’t mean America should reproduce kohsetsushi or that MEPs couldn’t take on this work. But absorptive capacity has to exist somewhere. Large firms may maintain it internally. Small firms may still depend on a network of institutions that already exist or ones we have yet to build.

Someone still has to know the factory.

What happens when old questions survive into a new age? Hans Blumenberg called this reoccupation. Old questions don’t disappear, but old answers lose their authority, leaving an opening for a new answer. They are assembled from whatever is lying around at the time.

Let’s approach 1945 from the opposite direction, looking at the shape of ideas leading up to it. How far back should we start? Twenty-five hundred years? After all, our words for science (epistēmē) and technology (technē) come from the Greeks.

Why pursue questions?

If you asked Plato, he would simply respond: because knowledge of the Forms is an ascent towards the Good. It perfects the soul.

Aristotle would take it further. In Book X of the Nicomachean Ethics, theoria is the highest human activity because nous, the human capacity for grasping first principles, is the divine element in us. What it observes through contemplation is eternal, and these objects are the best of knowable things.

In other forms of activity, of the practical variety, we gain something beyond the acting. The desire to build a table results in a table.

Virtues need the world to provide an occasion. Justice needs people to be just towards. Temperance needs temptation to decline. Bravery needs something to be afraid of. Everything we do, we do for the sake of or in relation to something else. This chain has to end somewhere, and contemplation is unique because it reaches its end simply by being exercised. For Aristotle, the highest activity has the least need to justify itself by what it produces.

These values were even embedded in the language. The Greek word scholē, from which we get “school,” meant leisure. Its opposite was ascholia (busyness), the absence of leisure. Work is defined as the absence of the good. It’s what remains when the good thing disappears.

Now you may be thinking to yourself, “wow, sounds like this Aristotle guy wasn’t a big fan of work.” And you’d be right.

Banausic labor, the manual kind you’re paid for, is an obstacle to the good life. The body gets bent. The leisure required for cultivating virtue becomes exhausted. You become tied to necessity and utility. In the Politics, Aristotle even argues that those performing such labor should not be full citizens.

But this was Athens, and some of the men doing that labor were citizens. They voted. The poorest among them rowed the fleet. Potters signed their pots.

And these citizens went to the theater, occasionally poking fun at the ideas from these philosophers. Two generations before Aristotle, the citizens watched Aristophanes hoist Socrates above the stage in a basket, explaining that he had to suspend his mind in the air to think about lofty things. These lofty things included how many of its own feet a flea could jump and whether a gnat hums through its mouth or its rear.

In 423 BC, Aristophanes’ The Clouds came home empty-handed from the City Dionysia festival, losing to both Cratinus's Pytine and Ameipsias’s Konnos, which also lampooned Socrates. He couldn’t catch a break in those years.

So, we should understand this preeminence of contemplation as two things. An honest attempt to say what a human life is for. And a bid, from one competing vision of what Athens ought to admire, to place certain activity above that of everyone else’s.

This wasn’t a defense of scientific research in the modern sense, yet familiar strands are starting to emerge. Distinctions drawn between kinds of knowledge and kinds of knowers. That usefulness is the wrong standard for judging certain activity. And an unease among the populace, in jest or persecution, in this idea.

Why do you want to pursue questions?

Warnings about transgressive knowledge appear throughout myth and scripture. Eve reaches for forbidden knowledge, Lot’s wife looks back, Actaeon sees what he ought not see, and Psyche raises the lamp to look at her husband. Apuleius said what drove her was curiositas.

Beyond the social embarrassment of being the subject of satire or cosmic punishment, is there any limit to what we ought to seek to know? Or is the only meaningful demarcation between kinds of activity? Is it polite to ask “why do you want to know?”

Let’s move forward in history a bit. The year 400 or so should do. Augustine installs a moral test on the Platonic hierarchy focused on motive rather than object. Eternal truths, arrived at through contemplation, can be a way of approaching the divine, but they must be ordered.

To Augustine, curiositas is the candidate for the lust of the eyes slot of the three temptations in the First Epistle of John. Curiositas was not simply the desire to know, but the desire to know divorced from the proper end of knowing.

Gawking at a car or, in this case, a wagon crash; fascination with strange or marvelous things; learning pursued from pride or vanity; the magical arts. These are all knowledge that distract from God.

The Confessions is a self-audit, addressed to God. Augustine is directing this criticism inward and finding himself wanting: “In this so vast a wilderness, replete with snares and dangers, lo, many of them have I lopped off, and expelled from my heart.”

But if curiosity is so morally perilous, how do we navigate the territory? Aquinas arrives about a millennium later with an answer. Aquinas is writing as both a continuation of Augustine and during a time when the Latin West is recovering Aristotle’s work. He accepts Aristotle’s premise that “all men by nature desire to know,” but also Augustine’s that this desire must be ordered. For Aquinas, the appetite is natural but we must temper it, the same way we do food and drink.

So he named two forms: curiositas, the appetite let off its leash, and studiositas, the same appetite disciplined and ordered toward wisdom, and ultimately God. Accompanying the curiositas are four ways it can be disordered, only one deals with the source of knowledge, like divination or demons.

The others concern its ordering: what we choose to know, why we seek it, and whether the inquiry is proportionate to our capacities.

Two people can study the same thing, ask the same question, and reach the same answer. One of them is exercising a virtue while the other commits a sin.

Three centuries after Aquinas, Francis Bacon wanted the systematic study of nature taken up as serious work, but realized he could not get there without answering for Aquinas, Augustine, and Adam.

What ruined Adam was not natural knowledge, he named every beast of the field, after all, but “the proud knowledge of good and evil.” The Fall was due to man’s attempt at judging like God rather than knowing like a creature.

Bacon reoccupied an idea that was already fifteen centuries old. Paul had said that the invisible things of God are understood through the things that are made. Augustine turned that into a book, or rather two: “It is the divine page that you must listen to; it is the book of the universe that you must observe.” Augustine’s view of this second book was remarkably egalitarian in a way the eventual Royal Society would not be. Anyone, even the illiterate, could observe the universe as a testament to God’s creation.

Bacon wanted nature read for its causal structure, and he wanted the two books distinguished. Scripture could tell us about God; nature could tell us how the created world worked. No one can search too deeply in either book, provided they do not confuse them. In both, aim at endless progress.

Inquiry, the patient and disciplined kind, can then be a form of repair, recovering what the Fall took from man. The relief of man’s estate, and all that.

“Curious” as a compliment is surprisingly new. It sat for a millennium closer to sin than salvation. Einstein’s quip that “I have no special talents, I am only passionately curious,” would be concerning to Augustine, but ordinary to the early modern gentleman scientist.

During this period (roughly 1500–1800), it began to transform into an increasingly legitimate way of engaging with the world, even if they remained anxious about what counted as proper curiosity.

Curiosity was a cultivated virtue, provided it was disciplined. The ideal gentleman was expected to observe carefully, travel widely, collect thoughtfully, and converse intelligently. Increasing global exploration, collecting, natural history, and print culture created new social spaces where this disciplined curiosity could be displayed.

This introduces an element of social performance and etiquette to inquiry. Collections, including Curiosity Cabinets, became a way to represent the world in miniature and periodicals extolled the proper way of pursuing knowledge well.

Science did not only have to establish a process for producing reliable knowledge. It had to establish a recognizable type of person who could be trusted to produce it.

“A Man of a Polite Imagination is let into a great many Pleasures, that the Vulgar are not capable of receiving.”

No. 411 The Spectator by Joseph Addison

But, as Socrates discovered, recognizable traits can make one the target of caricature. In 1676, Thomas Shadwell put the curious figure on a stage in The Virtuoso. His Sir Nicholas Gimcrack is a man consumed by experiments and discoveries. In the play, he lies on a table mimicking his observations of a frog swimming in a cup, having no intention of ever entering the water himself.

Robert Hooke, one of the leading experimentalists of the Royal Society, recognized himself.

“With Godfrey and Tompion at Play. Met Oliver there. Damned Doggs. Vindica me deus (God avenge me). People almost pointed.”

Robert Hooke’s diary entry on June 2, 1676, after seeing The Virtuoso

The knower had become a particular type of person, and particular types of people can be judged not only by their discoveries, but by their character.

This moral tradition and its anxieties around curiosity haven’t vanished entirely. We share Bacon’s concerns with our moral preparedness for the enormous, godlike powers science gives us. We inherit Augustine’s idea that the internal state of the knower is meaningful. Some motivations are purer than others, even if the underlying object is the same. The qualities of the person asking still matter.

What’s going on inside the mind of the knowledge-seeker can even categorize funding lines, though auditing those motivations remains imperfect.

Why should a person support those who pursue questions?

So far, our concern has been on whether contemplation perfects the person doing it, but why should anyone else support that pursuit? Who pays the contemplator? Curiosity alone doesn’t buy food, even bread and lentils. Maybe I’m sick of bread, lentils, and sleeping in a barrel. Let’s try the flattering the king route.

If you were wealthy in Latin Christendom you might endow a chantry so priests would say masses for your soul. You built a chapel, funded a hospital, left land to a monastery.

For a long time, nobody endowed “research” because research was not yet a thing you could endow. It was something clerics did inside institutions that existed for other reasons. Inquiry happened in their margins.

In 1264, Walter de Merton, Chancellor of England, endowed a house of scholars at Oxford for the good of his soul and his family’s. Over the following century, four men worked out how to describe motion mathematically.

Nobody in their time was asking for this. It rode along on money given for something else. Galileo found their work useful three hundred years later.

At some point in between, the rewards patrons expected from inquiry began to move from the next life toward this one.

The first known scientific publication for research using a telescope was done by Galileo in 1610. It announced the discovery of four celestial bodies we know of as the Moons of Jupiter.

Cool find, Galileo. Is there any way to leverage this into something tangible?

Why not name it after Cosimo Medici? That guy’s pockets are deep. Or even better, name it after the whole family. Cosimo is one of four brothers, right? Four brothers, four stars. It works. And for about a century, they were known as the Medicean Stars. Like any good fundraising manager, Galileo sent along a telescope with the book so the family could see what they now owned.

“God graced me with being able, through such a singular sign, to reveal to my Lord my devotion and the desire I have that his glorious name live as equal among the stars.”

Written by Galileo to the secretary of Cosimo Medici

Galileo figured out how to make his research appropriable to his funder. But because of this value he gave, he expected some to be returned.

He requested a new appointment at the University of Pisa; a doubling of the rate of scudi flowing to his pocket; and a new title in the royal court, that of Philosopher rather than mere Mathematician. Mathematicians only compute after all. They don’t explain the nature of things.

A bit self-dealing, sure, but the underlying innovation is rather startling. The justification for supporting inquiry need not be contained within the activity itself. It can come from somewhere else: salvation, prestige, power, or the prospect of useful knowledge. And these justifications can underwrite the involvement of others.

Justifications are also promises, and promises can be audited.

Enter our final satirist, Jonathan Swift. Swift, an Irishman, looked around at man’s estate and questioned whether the Royal Society was making good.

In the Academy of Lagado, projects are justified by promised improvement to human life. These promises underwrote projects like extracting sunlight from cucumbers, building houses from the roof downward, and finding ways to make food grow where it cannot.

In the fictional land of Balnibarbi, farming has been reorganized on the Academy's advice and its fields are barren. The only estate still producing anything belongs to Lord Munodi, who kept the old methods. Munodi is thought a fool for this.

The knowledge that works—the knowledge of the humble farmer derived from banausic labor—is the lower kind. The kind of knowledge that doesn’t work is higher. It’s ingenious, novel, theoretically impressive, and accompanied by an account of what it might someday accomplish. Munodi can only point to his harvest.

Once inquiry is supported by promises of future benefit, the promise itself can become part of the thing being rewarded. Not does this work? But does this sound like the sort of thing that might work?

Why should a society support those who pursue questions?

There’s a problem in involving other people in your curiosity. The person doing the inquiry wants to know something. The person paying for it wants something different, chiefly a return.

Demand too much justification, and you fund only what already looks useful. Demand too little, you get Lagado.

If we grant that sometimes these scientists stumble upon something useful, then who has the authority to decide what is worth knowing?

When Wilhelm von Humboldt set about organizing the new University of Berlin in 1810, he proposed that the state fund Wissenschaft, or knowledge pursued for its own sake.

Then, having funded it, you can please leave us alone. Thank you very much. Einsamkeit, Freiheit, and Bildung: solitude, freedom, and self-cultivation.

A nation cultivates itself through disinterested inquiry, and that only works if no one is standing over your shoulder asking what it’s all for. The state ought to pay; the state ought not decide.

Humboldt was pushing against the useful university: the school that exists to train officials and engineers, knowledge measured by the trades it feeds. Brotstudium, or study for bread.

In the late 1800s, curious Americans went to Germany by the boat-load and came home wanting the same thing. Johns Hopkins opened in 1876 as the first attempt, which is where we find our next figure, slightly appalled by his own country.

Henry Rowland was a professor of physics at Johns Hopkins who had studied in Germany. In August 1883, he stood at a lectern in Minneapolis and told the nation that it had, as yet, produced no science at all. He declined to dignify telegraphs and electric lights with the name of science.

“Some obscure American who steals the ideas of some great mind of the past, and enriches himself by the application of the same to domestic uses, is often lauded above the great originator of the idea, who might have worked out hundreds of such applications had his mind possessed the necessary element of vulgarity”

Henry Rowland’s “A Plea for Pure Science”

The kinds of people American society celebrated were all wrong. Edison was a folk hero, but he was self-taught and of obscure intellectual origins. Application is beneath the genius. The potters are not the best of us. Rowland eventually got over himself.

About a year after Rowland’s speech, the engineer Robert Thurston presented a counterargument. Where Rowland imagined a hierarchy of pure science first with application following, Thurston insisted that the two could not so neatly be separated. The attempt to solve practical problems could itself be a mode of scientific inquiry and need not wait for science to produce usable knowledge. This Gilded Age quarrel, as nearly as these things can be dated, is the moment “pure” and “applied” harden into a matched pair.

During the Gilded Age, American inquiry was being institutionalized by private fortunes. Rockefeller, Carnegie, and other industrialists began endowing foundations to consecrate private wealth into permanent institutions. A great many Americans looked at this and saw menace.

In 1915, in the wake of the Ludlow massacre, the United States Commission on Industrial Relations, chaired by Frank Walsh, turned its attention to the foundations. For Walsh, a perpetual, tax-exempt institution could convert one family’s fortune into a claim on the public agenda. What should be studied, which social problems deserved attention, and what counted as knowledge worth producing are political questions.

But is the public any good at answering these questions? The whole point of inquiry is that the answer is not yet known. If the public decides which questions will be useful, will they exclude the discoveries whose usefulness they can’t yet imagine?

Abraham Flexner thought this an impossible ask. To him, scientists were closer to poets and artists than engineers and accountants. If they thought about practical ends, they would be paralyzed. The life of the mind and of the spirit satisfy something in us beyond the material. The activity undertaken therein may appear useless to you, but in the scope of history can result in fortunes as yet unseen.

“To be sure, we shall thus free some harmless cranks. To be sure, we shall thus waste some precious dollars. But what is infinitely more important is that we shall be striking the shackles off the human mind and setting it free for the adventures which in our own day have, on the one hand, taken Hale and Rutherford and Einstein and their peers millions upon millions of miles into the utter-most realms of space and, on the other, loosed the bound-less energy imprisoned in the atom”

Abraham Flexner’s Usefulness of Useless Knowledge (1939)

The possibility of future usefulness is unknowable at the time of inquiry. Therefore usefulness cannot serve as the prior criterion for deciding what to fund.

The mathematician does not know what the engineer might need twenty years from now. Nor does the physicist know what problem the state will find itself desperate to solve.

Abstract work on nuclear physics, differential equations, probability, etc. can, suddenly, become bomb math.

Vannevar Bush was out there in the desert when it did.

This post was brought to you by Gavin Bryars’s The Sinking of the Titanic.

There’s a difference. For Shockley, corporate science is water. Everyone’s swimming in it and everyone knows everyone’s swimming in it. For Hassabis, or more accurately his interviewer, corporate science is an exception.

If corporate science were judged solely by its fruits during those seventy years, it would seem strange to ask the question. Corporate science gave us the math of communication, the relic radiation of the Big Bang, the microscope that sees single atoms, coherent light, the math of rough shapes, the first synthetic fiber...

Besides discoveries, the question of how to organize corporate science and its workers has diffused into unexpected places. Lab directors were on the frontier of how to manage knowledge workers, though they didn’t call them that yet. The dual technical-and-managerial career track that let researchers advance without becoming people-managers, what we now call the Individual Contributor, was formalized at IBM in 1962. The campus-style corporate headquarters with their famous commissioned architects. The practice of giving employees non-accountable time for side projects, whether formalized or informally tolerated as bootlegging.

So where does the suspicion come from?

Maybe it comes from broken promises between what corporate employers offered those idealistic scientists and what they received?

In 1927, Du Pont tried to steal the chemist Louis Fieser away from Bryn Mawr. Fieser went kicking and screaming: “I never expected to go into industrial work but the thing which makes a decision so difficult in this case is that I don’t have to sell my soul at all; they even said I could bring my quinones along and continue my present work” (quoted in Shapin, 2008).

John Hopfield sweated over where he would fit in the 1950s job market: “Academia or industry? The answer was clear from the Solid State Seminar speakers I had heard. AT&T’s Bell Labs (Murray Hill) and General Electric (Schenectady) had broader, more vigorous research programs in solid state physics than any university did.”

Some managed to keep a foot in both camps. Like Boaz Barak, who splits time between Harvard and OpenAI. “I’m now also very interested in having an actual impact on AI as it will affect people’s lives, and being at OpenAI is a good place to do it.” He added that Harvard’s compute resources, though extensive, weren’t comparable to what industry could provide.

For Venkatesh Narayanamurti, a 20-year veteran of Bell Labs, if a prospective hire shared that they wanted to continue their thesis work once they got hired it was actually a countersignal, as it meant the researcher wasn’t thinking big enough. His researchers never had to write reports for him, either, because he was always there and knew what they were working on. Sounds intolerable.

This sentiment is similar to the view of others at Bell through its history: the problems of the firm are interesting scientific problems and researchers naturally gravitate to interesting problems.

I now invite you to waste an hour exploring the work of Benoit Mandelbrot, who felt that his research could only have been done outside the university: “Some today argue that my kind of science belongs in a public university, not a corporate laboratory. In fact, none of my discoveries could have been done in the formal atmosphere of a university.”

Ok, in some cases they did have academic freedom on par with, possibly surpassing, their academic peers. Not every corporate lab was trying, though. A Monsanto training film from the mid-century depicted laboratory workers in a corridor while a narrator remarked, “No geniuses here; just a bunch of average Americans working together” (quoted in Shapin, 2008). I wonder if those average Americans got a look at this script on filming day.

But surely financial pressures were ever-present, the crushing logic of the market weighing down on the caught-in-the-clouds heads of these researchers?

In 1912, George Eastman asked Kenneth Mees to build him a photographic research laboratory at Kodak. Mees told his new boss that commercially useful results were not to be expected for ten years. Eastman wrote the check.

Striking a similarly conciliatory tone in his demands was Charles “Boss” Kettering, who ran General Motors Research from 1920 to 1947: “I would never be accountable for the money I spent... You can’t keep books on research, because you don’t know when you are going to get anything out of it or what it is going to be worth when you get it” (quoted in Shapin, 2008).

This may have been the view from the corner office, but was it felt by the researchers? Henry Pollak, a researcher at Bell, remarked: “there was the philosophy: look, what you’re doing might not be important for 10 years or 20 years, but that’s fine, we’ll be there then.”

Today, Alphabet has the same arrangement. The Other Bets segment, including Waymo, Verily, Wing, Calico, Intrinsic, and others, has collectively burned billions on long-horizon work that the search-advertising business covers each quarter, and the shareholders are told so in plain language.

Is this frivolous spending or portfolio logic applied to corporate exploration? Place many bets anticipating that one will pay off enormously. We associate this logic with VCs, but “Boss” got to it first: “We take a chance on spending so much for research in the hope that the boys may stumble onto something and we may make a little money out of it” (quoted in Shapin, 2008).

“Management is challenged to put aside its own technical judgments and assessments and prejudices and just bet on the people” could have come from A16Z or Peter Lee, former head of Microsoft Research.

Another complication: it turns out, maintaining academic norms against financial pressures is a two-way street. That same Lee once feared that “all of our researchers are perhaps too devoted to helping Microsoft win in the market today.” Mees observed the same nearly a century before about researchers in Kodak’s industrial laboratory: “[they] naturally want to engage in work which will result in direct and visible financial gain, and hesitate to carry on fundamental work for which no commercial application can be seen” (quoted in Shapin, 2008). Researchers may still internalize the logic of the firm at the organization-level, even when the promise at the unit-level is freedom and curiosity.

Let’s grant then that some pockets of corporate science had academic freedom and financial cushion. But they didn’t share science’s ideals of communalism, right?

Here comes “Boss” once again. “When you lock the laboratory door, you lock out more than you lock in” (quoted in Shapin, 2008).

Ralph Bown, Bell’s vice-president through the transistor era, was desperate to keep a lid on the lab’s secrets: “Prompt scientific publication is a sound policy for any research enterprise to follow as much as it can… No one laboratory has more than a fraction of the talent that can operate usefully on a new scientific idea” (quoted in Shapin, 2008).

Also playing close to the vest, was Kodak’s Mees: “[Mees] felt that publication of much of the work would enlarge the knowledge basis of the subject, and even though it might help his competitors in business the resulting advance in the field as a whole could not fail to benefit his Company and science.”

Despite valid objections to field narrowing and transparency on methods, almost any modern AI conference is dominated by submissions from corporate labs. In 2022, DeepMind was responsible for 12% of the most-cited AI papers globally, with only Meta and Google proper producing higher impact. 67% of the top 100 most cited AI research papers were from corporate sources that year.

On being one of the good eggs, Yann LeCun said “[Meta FAIR] caused other labs like Google to become more open, and other labs to also kind of publish much more systematically than before.”

Economists have names for this. Nelson: a diversified monopolist can be an exception to firms underinvesting in basic research, because it can internalize its own spillovers. Rosenberg: publishing is a sort of admission ticket to the information network. Stern: scientists pay to be scientists, trading wages for the right to publish and freedom to pursue interesting questions. Which explains the cascade LeCun describes. One actor offering wages plus freedom reprices all the other researchers in the system.

So why do we still call it corporate science?

Part of the answer is that we’re suspicious of knowledge claims made by people with financial interests in them. Tobacco companies funding cancer research. Oil companies funding climate science. Pharmaceutical companies funding trials of their own drugs. Startups talking their own book. We’re all familiar with these examples.

But these are heuristics and “corrupted science” isn’t an exclusive pathology of corporates. It’s institutional and rears its head wherever the incentives are aligned. Academic careers reward publications, citations, and novelty, giving us p-hacking, salami slicing, and replication crises. Firms reward commercial success, giving us secrecy, hype, and conflicts of interest.

We tend to judge academic science by its ideals and corporate science by its perversions.

Those incentives may be more visible in industry, but the voices of these scientists show that norms can hold quite radically. Norms that were received through their training, acculturation into the world of science, and upheld when management reproduced the conditions where scientists could recognize one another as scientists.

As Hassabis said, doing science here is only a question of doing it in the right way.

Then what does doing it in the right way mean?

Let’s end where we began, with Shockley. To understate, Shockley was a difficult person. One of the milder anecdotes about his management style was submitting his staff to a lie detector test after allegations of sabotage, and those familiar with the history of Silicon Valley will be aware of the “Traitorous Eight.” His name, not theirs.

This paranoid style was not newfound at Shockley Semiconductor; it was unconstrained there. His competitiveness, jealousy over credit, and certainty that he was right at the expense of other leads were all present at Bell. But his faults could be absorbed so long as his authority was bounded, and Bell built institutional scaffolding around him.

Bell intervened relentlessly on the institution and surprisingly little on ideas.

Take the patent: Shockley wanted the transistor patent written around his own field-effect concept, which Bell’s attorneys disputed due to prior art, filing on Bardeen and Brattain’s device instead. When he absconded to a hotel room to work out the junction transistor and froze his colleagues out of the follow-on work, their appeals to their boss’s boss worked, and Brattain was moved out from under him. And he was locked out of the executive track throughout his time at Bell, because they judged scientific and managerial competence separately.

Not paradise, but workable.

All of this was run back at his own firm sans scaffolding. Senior staff appealed to his backer Arnold Beckman to move Shockley out of management; Beckman backed his founder, and the eight walked. He bet the company on the four-layer diode at the expense of anything else. Even though he was right on the science (it anticipated the thyristor), he was wrong on the market and the manufacturing.

Nothing at his shop could unleash what made Shockley a genius, and constrain what made him difficult.

That’s the rub.

The freedom to work on anything happened within a context designed so that “anything” collided with the problems of the firm. Financial logic wasn’t suspended, it was a bet, and the bet was extended on trust to “the boys” that they might stumble onto something. Publications were free, so long as the patent attorneys got a first look.

And Shockley has the freedom to be Shockley, so long as being Shockley doesn’t get in the way of the other geniuses.

This post was brought to you by Laurie Spiegel’s The Expanding Universe, composed at Bell Labs.

I submitted an application portfolio of one-off DIY furniture builds, put together with basic tools and basic skill. In the weeks leading up to the retreat, I sweated over whether I’d be able to roll with a group of aspiring and professional designers. To self-soothe, I told myself that I could make up for the gap with enthusiasm and a blue, vintage work shirt with paint stains that weren’t mine.

Nevertheless, the day came. Our work started by walking the property, finding interesting spots. The underside of a tree with a population of firebugs, a patch of ground by the sauna and stream, a neglected well. The idea was to choose from among those locations one that would be the focus of our designs for the remainder of the week. The well and a talented, professional designer would be my companions (you can find her work here).

As someone enrolled in a business program, my justification for being there was clunky. By the second day, I landed on this: I’m researching organizational design for creativity in support of a thesis on corporate research labs. This was partly true. I’m unsure anyone was convinced.

On the third morning, during a 5Rhythms dance class, somewhere between the ecstatic gyrating of chaos and stillness, I realized I might be onto something. While staring down at the patterns in the grass, two things became clear.

One: the organizers had accomplished something remarkable. The social cost of looking stupid was run down to near zero, fast, giving people (including, maybe especially, myself) license to actually try things.

The other: the grass was suddenly inspiring. The looping patterns, the small spontaneous architecture of it, the texture under my hands and feet. These were inputs I was suddenly open to noticing.

It’s cliché to say we should pull inspiration from nature, in part because it glides over underlying mechanisms. It only works if there’s an individual openness to these weird inputs. For some, rocks are rocks, chickens are chickens. For Thorleif Schjelderup-Ebbe, a bored teenager spending the Norwegian summer at his grandma’s farm, they were the source of his insight on the pecking order. I would’ve just stared at clouds.

At that point, I realized the design of the retreat was as important as the design of the objects the retreat produced. The morning movement, the communal meals, the design meetings sitting in circles on the grass, the pastoral setting. These were deliberate choices, and therefore inputs to the creative process.

And this isn’t unique to arty retreats for designers. The same impulse to design the environment shows up wherever people are serious about creativity, though the stakes may vary. Mervin Kelly, of Bell Labs fame, wrote extensively about the physical design of the lab and how it was instrumental to creativity. Or Black Mountain College in the 40s, where Cage, Cunningham, and Rauschenberg, alongside their art, farmed their own food, built their own buildings, including Fuller’s first geodesic dome. These interventions don’t always work, and they don’t need to be elaborate. The coffeehouses of 17th-century London ran on penny entry and shared newspapers. Xerox PARC held meetings in a conference room filled with bean bags.

These are designed spaces in which openness becomes available, and maintaining that openness is as integral to the work as the work.

Back at the well, we wrote about what the space evoked. Storytelling, childhood, surrealism, and the crude shapes of the structure itself, a large cylinder with a comically oversized triangular roof.

We conceptualized and sketched. The first idea we presented to our lab directors was a set of large pieces with a point that could be driven into the dirt, a long back that arched towards the top, and a head. Basically bowing stick figures. They’d be arranged in a circle around a mirror to look like they were staring down a well, but could be plucked out of the ground and flipped over to make a rocker. Très conceptual. Très arty.

Misled by the beginner’s mind, I thought physically demonstrating what this looked like would help. So, first standing with my arms outstretched, back and head arched downwards, I flopped onto the ground and made the shape of the rocker. I don’t have access to their internal monologue, but from their expression it was something like this: “oh, we understood what you meant, it’s just a bad idea.” The dirt was too dry to reliably drive the point in. The pieces were reliant on a staged artificial presentation state to make sense.

The lab directors did something else, too: they pointed the groups towards reference designers worth drawing from. Isamu Noguchi’s playscapes, Donald Judd’s boxes, and, for the well project, the Shakers.

The Shakers emerged from a group of English dissenters led by former Quakers, and were sometimes derisively called “Shaking Quakers.” Their furniture was simple, humble, and built for use, premised on the idea that the act of making something well was itself a form of worship. Their chairs and tables now sit in design museums. They may have found this embarrassing.

Given this, we settled on a humbler ambition: a stool. One of the gems in the idea pile was to use the curve of the well in the dimensions of the seat. But what would be the base? We tested lots of options. Stark boxes with solid squared bases. One, an upright cylinder. The design didn’t land until we put the cylinder on its side and started rocking on the cardboard model. It made us laugh, which at that point in the week was what we needed.

On build day, we scrambled to get everything fabricated and stained. Cutting and sanding thirty circles from scrap plywood, gluing in sets of ten to form the bases. The seat was cut, sanded, and secured to the bases with wooden pegs. At the workshop, craftsmen were on-hand to give advice on the builds and ensure that fingers stayed attached. Other teams buzzed around the workshop passing tools back and forth, competing for time with the CNC.

Creative environments tend to be surprisingly crowded. They’re filled with builders, designers, and researchers working in close proximity. Venkatesh Narayanamurti, a veteran of Bell, Sandia, and the founding dean of Harvard’s SEAS, calls this body-contact sport. The productive friction across expertise levels that generates new, better ideas. The lab director who points the researcher down an interesting alley. The craftsman who brings the designer back to reality.

After the stain dried, the last hurdle was presenting to the group. Everyone sat in a large circle near the well. The stools were placed against it. I went to the well, grabbed a stool, brought it back to my spot, told a short story, and then returned the stool. My design partner followed the same pattern, sharing a brief anecdote instead. Then, our pre-conscripted lab director did the same, but put his legs in half-lotus and meditated in silence on the stool for a beat.

At that point, the pattern was clear, and the rest of the retreat-goers followed suit: they grabbed the stool, sat with it, and shared whatever they felt compelled to share.

This is actually a broad Quaker practice, where groups of worshippers sit in a circle in silence, sometimes for hours, until one feels compelled to speak. The Shakers added ecstatic dancing. Our presentation didn’t last as long, though someone did jump on the stool and ride it like a surfboard.

In the end, iteration was less about trying to stumble upon a good idea than about learning which instinct in a bad one was worth keeping.

From the flopping on the ground, we kept the seesaw. From the pile of ideas, the curve of the well. And from my attempt to blend in with designers, some stains on the work shirt of my own.