Two Stories About Tacit Knowledge

To build a nuke or (can) not build a nuke

Here are two stories about tacit knowledge.

  1. Lawrence Radiation Laboratory hired three fresh PhDs with little nuclear physics background and asked them to go build a nuke, and they did.

  2. The DoE in the 1990s realised they didn't know how to make a particular foam crucial to nuclear warheads, and spent a decade and $70m inventing a substitute - (h/t Leopold)

They pleasingly mirror each other, which is nice for storytelling but kind of a paradox. In one, we spend untold millions in figuring out the smallest moving part to manufacture, and in the other some smart alec fresh faces postdocs made a goddamn nuke.

So which is it, tacit knowledge is impossible to easily replicate or smart folks can just figure stuff out?

The first story. How did they do it?

You just go to the library and you start looking under all the subjects, you look under plutonium and uranium and high explosives and you look under nuclear physics and you just keep going and you find articles and stumble across things and books and publications.

Selden, one of the Berkeley scientists, said that. As you'll note there's not much that's elaborate or byzantine there.

But let's not stop with nukes. When you look at the world of hobbyist biohackers, you see the same pattern. Folks who have cobbled together some knowledge from open sourced papers and YouTube, accessibility of gene-editing tools like Crispr, and go all in on self experimentation.

Whenever I was bored, I went on YouTube and watched physics and biology lectures from MIT [Massachusetts Institute of Technology],” he explains. “I tried the experiments at home, then realised I needed help and reached out to professors at MIT and Harvard. They were more than happy to do so.

They create all sorts of abominations like genetically modified beer and animals that glow like jellyfish.

The article points out that the philosophy is hardly unique. In fact some of the prize winning discoveries have been funded and conducted exactly this way!

If you really want to make important discoveries, you have to know how to work the system,” says Richard Muller, an American physicist and emeritus professor of physics at the University of California, Berkeley. Muller revealed that he had secretly redirected funding from approved projects to fund riskier rejected ones in a letter in Science magazine in 1980.

It's worthwhile noticing that this propensity to leap ahead, self experiment and generally operate outside established strictures has always been a key feature of science!

And it showed results many atime. Most recently in the data science machine learning world where a substantial proportion are self-taught. Or a generation ago when software itself came up through the eponymous founder garages.

Does this mean that through self-education and perseverance the barriers have fallen? Yes. But is the path of the autodidact all its said to be? Not quite. For one thing, back in the original story, there is still doubt cast on the postdoc's efforts to create a nuke.

It was impossible for us to detail the Nth Country predictions from the data given in the sections on Final Design and Test of the Nth Country Experiment report.

In the official conclusions it says.

According to the official chronology, the Nth Country device was tested (hypothetically) in April 1967. But its designers were not told whether their hypothetical bomb exploded or failed. They remained “knowledge virgins,”

But this I think skips over the most interesting bit. The question was whether there could be a credible attempt to create a nuke, and clearly the answer is yes. The question of whether you could design a biological organism in your garage, that's yes too. Thus have the autodidacts always driven science forward!


Now, the latter story. Here there were clearly laid out rules and steps and formulae, and the group of smart, efficient scientists had to spend about the same time recreating it as perhaps the original discovery! Is this normal or were these government nuclear scientists just especially dense? Let's look at that most rational of places, industrial espionage.

As we know from movies, these are primarily filled with morally bankrupt geniuses, which means at the very least they must be highly motivated to get results. But they too seem to struggle.

For instance, when you dig into the intricacies of corporate espionage, turns out that the biggest problem the spies have is not getting those explicit pieces of documentation, but rather the actual people with the actual knowhow. In fact, that means that the information gap is hard to fill, and China has had to resort to getting not just the information, but the bodies who contain the information.

Approximately 1,100 Chinese nationals currently work at SME firms outside China. These workers return to China at low rates, but the few who do return bring valuable engineering know-how and often secure top jobs.

They're literally relying on a glorified erasmus program to learn how to actually get things done!

This also presents itself as a challenge in poaching talent. Occasionally it helps to bring someone over and have them rebuild a talent or a division, but oftentimes it results in mediocre results if not outright failure. And it rarely is as easy or straightforward as it appears to be. The story of Lewandowski is but one example. Despite having taken a ton of Waymo's IP (allegedly), Uber eventually ended up shutting down its research for not bearing enough fruit. Sure, this could have something to do with the need for making money instead of burning it, but at least its clear that getting explicit information and getting a top employee still wasn't enough.

Its a curious thing though. Companies are highly motivated to get things right. If they go to the trouble of paying millions to poach someone shouldn't it work?

Turns out it's really bloody hard to replicate an organisation. You can call this irreplicable part culture, because that's what it is. Which is a way of saying sure you can hire some folks with some data but that's not gonna help you build a self driving car. That's why the loss of a person hurts so much, because they take knowledge with them that's not only not written down anywhere but can't be written down. And yet, just because it hurts the poachee doesn't mean it will help the poacher. The exciting stuff is all culture, which is as numinous and emergent and intangible as it is real.


My hunch is that the difference amongst the stories is in the forms of tacit knowledge they imply. The first story had its tacit implicit knowledge regarding a problem. Re these we're pretty good at experimenting, thinking and figuring out a solution. The second story had its regarding a complex coordination issue where its tacit process knowledge, we find ourselves floundering. It's a version of the butterfly effect but for compounding errors.

Even the types of tacit knowledge often discussed, like embodied knowledge, are aspects that are similar to the implicit knowledge above. Learning to ride a bike? Same, the embodied aspects of it have to be experimented with to be understood. But once experimented it can be understood, by almost everyone. Even an AI being trained on a set of images to learn what a "marina" is, its identifying the implicit features that make a marina what it is.

And all the while, the culture of Stripe or Google or the FDA is also what makes them stand apart. The cadence of meetings, and information processing, and particular rituals, and decision making criteria and the hundred exceptions they all have, and innovation encouragement, and talent recruitment, and retention. These and a thousand other things small and big build the framework that creates just that particular flow of information in just that pattern which seems incredible or incredibly dull as the case may be.

Which leaves us with these two conclusions, which are either uplifting or disquieting depending on your mood.

  1. Almost any tool or technique eventually will be, and can be, figured out by smart people working on the problem

  2. Almost all scaling up that requires manufacturing and organisational knowhow will remain hard to easily copy

Its worth noting that we have blocked both of these in different ways. The former by gatekeeping and calling the discovery process of science DIY or biohacking, and the latter by trapping the manufacturing knowhow in such incredible bureaucratisation that even we can't replicate it, let alone the spies. That's why we love looking at institutions like DARPA and marvel at its efficacy while not being able to replicate it ... It's not just about what they did, but culture. And that's hard.

It's the difference between knowledge nobody can explain and knowledge nobody knows to explain.

Share

Appendix 1

Also, another way to think about it is as knowledge production at a “node” vs knowledge produced on the “edges”. The individuals, the teams are nodes. The interactions amongst the nodes are the edges. When the knowledge is within a node, it’s not easy to explain but its easier to reproduce. When the knowledge however is embedded within the edges, its much more diffuse, and harder to reproduce.

Appendix 2

A key concern about pocket nukes or biological materials is of course terrorism. This thesis would indicate that yes, it's theoretically highly feasible for small scale creations to truly disrupt our lives. But what it also says is that this is difficult and requires highly trained, sophisticated folks who are willing to nerd out for extended periods of time. And honestly, if you were gonna do that anyway, wouldn't you rather spend that time joining FAANG and complaining about the nap pods?

Appendix 3

  1. I made a Google colab sheet learn how to draw a marina from looking at a whole lot of different pictures of it

  2. A set of employees left Company X to join Company Y and found themselves unable to replicate the speed and innovativeness of X, the reason they were hired

  3. I taught my son how to ride his bike by explaining how it works broadly, letting his play with it, and encouraging when he did the right things

In all these instances, the commonality is that there is a black box that sits between the input and the output. You have to somehow wriggle your way through the problem.

You could call it tacit knowledge, on which Cedric has written a fantastic series of articles. Or the Eureka effect for that matter.