Systemic decay theory of stagnation

Why hasn’t Bell Labs been replicated?

All the work at Bell Labs needed to tie back to the system

We need new institutional structures to fill the role that corporate labs once occupied

Corporate labs no longer fill the niche they did in the early-mid twentieth century

Why try to emulate DARPA and not Bell Labs

Notes on the changing structure of American Innovation

ICT Distillation

26 Oct 2020

An Institute for Creative Technology - Riffing on ARPA

ICT Outline

Startups have taken over a chunk of corporate R+Ds niche

# AI research is an exception to the decline of corporate R+D
AI research bucks the trend of the decline of corporate labs as a place where globally important, cutting edge work is done. ([[The decline of science in corporate R&D]]) If you squint, Organizations like [[Facebook AI Research - FAIR]], [[Google Brain]], [[DeepMind]], and possibly [[OpenAI]] [^1] bear strong resemblance to [[Bell Labs]], [[Xerox PARC]], and other legendary Corporate R&D orgs (I’ll just refer to these as Legendary CR&D)  in their heydays.

The parallels manifest in several ways. It’s worth speculating on why they exist in these labs but not elsewhere to understand why the dynamics may or may not be replicable in more areas.

AI research (can) require massive resources - in this case it’s just thousands of dollars of compute for training models and the datasets to train on. These resource requirements mean that there is exploratory research that people just can’t do with the resources available to most labs at universities. You saw a similar effect pulling professors away from universities and into corporate labs in the first half of the 20th century. 

At the same time, corporations running AI labs reasonably expect the research to create value commiserate to the costs of this research. Machine learning can directly improve the core product lines of all the companies listed above, in the same way that Bell Labs’ work directly improved [[“The System” of AT&T]]. [[Innovation orgs need to be aligned with their money factory]]. Additionally, AI promises massive value over long but not infinite timescales. OpenAI’s business model is implicitly based on the assumption that they will create literally infinite value in the not unforeseeable future. 

This alignment between the research and the money factory allow modern AI labs to do highly regarded work and collaborate closely with academia, *without* seeming like a waste of money. Bell Labs was the home for the work that led to nine Nobel prizes. Today, AI conferences are dominated by work from corporate AI labs, not just in quantity but quality as well. 

Like the electronics work at Bell Labs, AI research benefits from having multiple disciplines in the same place and the people who will eventually produce it. When you’re creating “AI for X”, it’s generally helpful to have someone who is an expert in X around. Similarly [[Business and finance people are utilities]]. Training and implementing ML algorithms at scale requires a different skillset than creating and prototyping the algorithms. Corporate AI labs have the budget to hire research engineers and also make it easy for researchers to talk to some of the best infrastructure and production engineers in the world. This dynamic roughly parallels the close ties between Bell Labs and Western Electric.  [[Prototyping needs manufacturing in the room]]. 

The perhaps uncomfortable parallel is between AT&T’s status as a high-margin, cash-printing monopoly and Google/Facebook’s similar situation. AI labs suggest that successful corporate labs can only exist in domains where there are large corporations with monopoly-like power. The decline of corporate labs in chemistry and physics-related domains may have been caused by commoditization of those products. GE, Dupont, Kodak, and others’ share prices suggest that they at least are no longer perceived as monopolies. However, monopoly profits -> high quality corporate research isn’t the entire story because you don’t see a lot of high-quality non-product research coming out of Boeing and Amazon[^2] ::are there other examples::?  Perhaps the missing difference is whether or not the monopoly perceives that it will benefit from significantly better technology. [[AT+T benefitted significantly from creating better technology]] because they had an effective pact with the government at as long as they kept making the system better they wouldn’t be broken up. I wonder what the counterfactual is - being left as a monopoly without the pact. While Google and Facebook have no (publicly known) agreement with the government, I believe [[Peter Thiel]]? Made the argument that at least Google’s monopoly does depend on them continuing to have the best search engine - people have no loyalty to google beyond the quality of their searches. Similarly one could make an argument that advertisers will abandon Facebook as soon as they have worse ad targeting or the kids move to another platform like TikTok or Snap. 

Microsoft is an interesting beast to examine through this lens. It’s unconfirmed, but arguably [[Microsoft Research]] was started as a public offering in the same way Bell Labs was. However, it both explicitly set out to do more product-focused work than PARC and Microsoft’s cash cow was Windows and the Office Suite, both of which didn’t benefit much from MSR work. MSR was subsequently pretty neglected and used for stunts like China expansion. However, Microsoft’s core business has begun to shift towards cloud services, which *do* benefit from AI research and it’s becoming clear with things like GPT-3 that AI research can augment Office Suite products that are now actually under threat. This shift coincides with a[partnership with OpenAI](https://news.microsoft.com/2019/07/22/openai-forms-exclusive-computing-partnership-with-microsoft-to-build-new-azure-ai-supercomputing-technologies/#:~:text=The%20partnership%20covers%20the%20following,promise%20of%20artificial%20general%20intelligence) that looks a bit like the relationship between AT&T and Bell Labs if you squint. Amazon remains a challenge to this just-so story because it also has a massive cloud business but is not (publicly) doing as much speculative AI research.

If it’s accurate, this narrative suggests an uncomfortable truth. Instead of declining because of  “corporate short-termism,” speculative corporate R&D might depend on companies with monopoly-level margins perceiving their fate being intimately tied to a high-potential technology.  This trifecta (monopoly + clearly high potential technology + that technology being tied to the company’s core business) can’t be created by policy or even a cultural shift.  [[Healthy corporate R+D requires a trifecta of conditions]]. The trifecta is the most straightforward answer to the question [[Why hasn’t Bell Labs been replicated?]] and suggests that attempts to “create a new Bell Labs” are doomed to failure if they try to follow the playbook too closely. At the same time, I strongly suspect there are incredibly valuable technologies that don’t fit the criteria of the trifecta (clearly high potential to solve an existential threat for a monopoly rentee) but do need the environment and resources that were once provided by healthy CR&D labs to realize that value.  This is why[[We need new institutional structures to fill the role that corporate labs once occupied]].

### Related
- [[aroraChangingStructureAmerican2020]]
- [[Bell Labs didn’t have a unique model]]
- [[Corporate R+D orgs experience a tension between Research and Development]]
- [[§Corporate R+D Constraints]]
- [[gertnerIdeaFactoryBell2012]]
- [[aroraChangingNatureCorporate2015]]
- [[Does the nature of the activities to create new technology and science change over time?]]

[^1]: At this point OpenAI resembles the Bell Labs equivalent for Microsoft
[^2]: Maybe it’s a [[Seattle]] thing?



[Web URL for this note](http://notes.benjaminreinhardt.com/AI_research_is_an_exception_to_the_decline_of_corporate_R+D)

[Comment on this note](http://via.hypothes.is/http://notes.benjaminreinhardt.com/AI_research_is_an_exception_to_the_decline_of_corporate_R+D)




Bell Labs

Bell Labs was only profitable in conjunction with AT+T

AI research is an exception to the decline of corporate R+D

Bell Labs did not give researchers complete freedom

§Bell Labs Model

Labs Document Outline

# AT+T benefitted significantly from creating better technology
AT+T benefitted from new technology in three distinct ways. First, its scale and the nature of its products and services meant that even small technological improvements could have massive effects. Second, there was a semi explicit contract with the government that AT+T would not be broken up as long as they continued to improve their services and shed excellent technology. Third, M+A blocking regulation meant that AT+T couldn’t depend on acquisitions to provide technological advances. 

![](../BearImages/5733ABA1-C990-4519-921D-35ADC34D7D5A-88049-000758A621BC58CC/IMG_0250.jpg)
[[“The System” of AT&T]] was so huge that even small efficiency gains translated to saving tons of money. Unlike in software, creating new markets (like calls to Europe, cables that can survive extreme environments, etc) required creating new, physical technology which in turn sometimes required new scientific discoveries. 

AT&T’s agreement with the government was that as long as AT&T continued to improve its offerings, stay focused on communication, and shed technology, the government wouldn’t be break it up its monopoly. As noted earlier, the way that AT&T could continue to improve its offerings was through technology. 

Finally, the government blocked large company M+A at several points in the 20th century and startups were just less of a thing before the latter half of the 20th century. At the same time, the nature of the technology increasingly seemed to be beyond the capabilities of a single inventor. [[Does the nature of the activities to create new technology and science change over time?]]  Consequently, in order to benefit from technology, AT+T needed to build it itself.

### Related
- [[Healthy corporate R+D requires a trifecta of conditions]]
- [[Bell Labs was only profitable in conjunction with AT+T]]
- [[aroraChangingStructureAmerican2020]]



[Web URL for this note](http://notes.benjaminreinhardt.com/AT+T_benefitted_significantly_from_creating_better_technology)

[Comment on this note](http://via.hypothes.is/http://notes.benjaminreinhardt.com/AT+T_benefitted_significantly_from_creating_better_technology)




Heat cycles for democratization and cutting edge experiments innovation and discovery

Heilmeier — Licklider — Braben — Elon quadrants of Research Management

Overlapping institutional constraints rule out several classes of creative work

Institutions either make things more legible or act as trust brokers

The knowledge frontier is a high-dimensional garden

The linear innovation pipeline is more harmful than useful

Grants only go so far

Research organizations have core activities that act as discrete units

Activity design indicates how power and coordination are distributed

Does the nature of the activities to create new technology and science change over time?

New kinds of ARPAs need new kinds of performers

There is a tradeoff between an organizational culture that is good at addressing market uncertainty and one that is good at addressing technological uncertainty

A taxonomy of research organizations and activities

Cross-disciplinary real research conferences are underdone

Boris Hanin conversation 30 Aug 2021

Venture Research, Bell Labs, and DARPA form non-orthogonal basis vectors over the space of innovation organizations

Precommercial scitech is a nebulous term for a certain class of frontier activities

There is no legible pathway through precommercial scitech

3 Sep 2021

2 Sep 2021

Glossary

New ways of achieving collective action enable new activities

Journalistic research is the analysis half of novel research

Spending more money on the same activity has diminishing or negative marginal returns unless it explicitly scales

Vision-goal spectrum

# Activity Space
![](../BearImages/E75C5294-7BFA-4584-AA70-A0CBFE31A1B5-2972-0005F6028EB76D67/E6EFAF86-53BA-4CC7-8A22-671DC2597375.png)

Activity space is the set of all activities that people can do that contribute to new knowledge and abilities.[^1] “Ideas” jump from activity to activity, getting added to or changed in the process. These memetic footballs merge and split, sometimes carried by one person between activities and sometimes being handed off to someone else. Obviously what entails an “idea” is obscenely [[Nebulous]] and there is no right way to draw lines between them, but I think the concept of activity space is useful, nevertheless. 

Ideas will inevitably become ill-shaped for a particular class of activity, for example when an idea that was previously well-served by curiosity-driven probing needs goal-driven systems-level work to get to the next level. There are many frameworks to describe these transitions: [[Technology readiness level - TRL]], the Basic/Applied research distinction, etc ([[Research has many orthogonal and non orthogonal classification axes]]). Of course, knowing what activities an idea “wants” at any given time is incredibly hard and there is probably   

Take an extremely ham-fisted account of modern semiconductors: the work to purify germanium and create diodes for radar (clear goal, product driven, iterative, experimental, timeline, large teams and parallel efforts) led to work to create germanium triodes (small team (3), speculative, experimental, goal but not use-case driven) which demonstrated some odd phenomenon that needed to be explained theoretically (blackboard, theoretical, white space). This theory enabled the group to hypothesize that the germanium could be used to build an amplifier (imaginative, playful, speculative). However physically realizing this idea was very hard, potentially impossible. The work to physically realize the triode involved a lot of just trying shit out (fat process, playful, but goal oriented). The team originally tried to build a FET but it proved too hard, but analyzing the failure (slack to try many shots on goal, failure analysis)

The list of all these knowledge-generating activities is probably too long to enumerate (it may just be a list of all types of human activity) but here’s an attempt [[A list of Knowledge generating activities]].

Not all people, organizations, and institutions are equally good at enabling different activities. Different personal attributes, organizational structures, and institutional incentives lend themselves to certain activities and suppress others. [[Overlapping institutional constraints rule out several classes of creative work]]. As a result, it’s rare for an idea to achieve its full potential within a single person/org/institution and no person/org/institution should be applied to all ideas. 

### Related
- [[Research organizations have core activities that act as discrete units]]

[^1]: You could see these as all the different things that people do to cultivate the [[Knowledge frontier]]. If [[The tech tree is a scaffolding inside the knowledge frontier]], you could frame activities as the inputs to creating new heuretics. (See [[§Tech Tree Model of Heuretics]]). 

[Web URL for this note](http://notes.benjaminreinhardt.com/Activity_Space)

[Comment on this note](http://via.hypothes.is/http://notes.benjaminreinhardt.com/Activity_Space)






AT+T benefitted significantly from creating better technology

# Does the nature of the activities to create new technology and science change over time?
# Does the nature of the activities to create new technology and science change over time
Underlying many conversations about [[Stagnation]], science, technology, and innovation are implicit assumptions either that the nature of the activities 

### Related
- [[Activity Space]]

[Web URL for this note](http://notes.benjaminreinhardt.com/Does_the_nature_of_the_activities_to_create_new_technology_and_science_change_over_time)

[Comment on this note](http://via.hypothes.is/http://notes.benjaminreinhardt.com/Does_the_nature_of_the_activities_to_create_new_technology_and_science_change_over_time)






Research management matters

Research has many orthogonal and non orthogonal classification axes

Quadrant Model of Scientific Research

All funding organizations are research organizations

Programs and Program Managers exist on several different spectra

§Axes for Looking at Funding and Organizations

Paul Merolla conversation 4 Oct 2021

# Heilmeier — Licklider — Braben — Elon quadrants of Research Management
![](../BearImages/AB5D6FB5-16AE-4C04-B0E2-782142FC3387-2972-0005C1F693ED3414/640125EB-26E6-49F6-9D44-3D8133D2A963.png)

The two axes that are relevant here are the amount of “constraints” (this is a bad word — maybe precision? Shot calling?)[^1] that the people working on the projects are subject to and the “opinionation” (or maybe vision) of the manager. “Manager” here is someone who has the ability to start or kill a project. 

The constraint and opinion axes form a [[2x2s]] where each quadrant could be exemplified by a (relatively) well-known research manager.

 In the top-left (Opionated+Unconstrainted) was have [[JCR Licklider]]. He had extremely strong opinions about where he wanted the program to go and would only fund people who were on board with that vision. However, he put very few constraints on that funding — “go start a group, do awesome things.” 

In the bottom-left (Unopinionated+Unconstrainted) we have [[Donald Braben]]. He expressly tried to have very little opinion about the work itself, instead focusing completely on the quality of the person and their relationship to the idea. 

In the bottom-right (Unopinionated+Constrained) we have [[George Heilmeier]], who famously imposed the [[The Heilmeier Catechism]] on any program he approved. Answering questions like “how long will it take” “how long will it cost” “How do you check for completion” epitomizes putting constraints on a project. However, (in the platonic version of this) as long as a program manager can produce compelling answers to those questions, it will get approved. This stands in contrast to … 

In the top-right (Opinionated+Constrained) we have [[Elon Musk]], who both has strong opinions about everything from the grand vision to specific engineering decisions *and* puts a lot of time and goal constraints on the people he manages. 

### Some observations

Each of these quadrants suggests a (justifiable) management style:
- Tell people where to row
- Build trust
- Hold people accountable 
- Tell people what to do 

As you go from right to left on the constraint axis, there’s more [[Knightian Uncertainty]] and as a result the research manager needs more trust. [[The two ways to enable action is through legibility and trust]] — you could see constraints as a legibility-increasing strategy. 

Each end of the constraint axis also suggests a flavor of operation strategy: when you’re operating in a low-constraint regime, it’s much more important to *filter* people a-priori, either based on trust or their alignment with your vision. In the high-constraint regime, you don’t need to filter as much but do end up needing to be more “bossy.”

### Organizational positions on the 2x2 
It’s interesting to think where different organizations sit on the constraint/opinion axis. 

As suggested by the fact that two different quadrants are both named after DARPA people, [[DARPA]] itself is all over the place on an office, PM, and program basis. Some TTO or STO programs are effectively high-risk weapons engineering programs and fall into the Musk quadrant. Some DSO programs tend towards Licklider or even Braben depending on how opinionated the program manager is. Over time though, my sense is that the organization has  moved on average towards the right. DARPA programs can only be so far towards unopinionated because at the end of the day they need to justify their military applicability. 

The [[National Science Foundation - NSF]] is incredibly unopinionated vision-wise, but does constrain how PIs spend money and the scope of the work (at least nominally — there’s of course the strategy to propose work that you’ve already done <[[People proposing work that they have already done is a sign that feedback loops are broken]]>) . The [[National Institute of Health - NIH]] is like the NSF but with more opinion (things need to be related to health). 

[[Howard Hughes Medical Institute (HHMI)]] operates pretty far towards the unconstrained end of the spectrum but is opinionated insofar as the work needs to nominally be related to health/medicine. 

### Orthogonal axes
These are of course, not the only two relevant axes. Some others include:
- Observation frequency 
- Feedback loop with reality frequency 

What others? The dream is to create some set of “basis vectors” for researcher management and be able to say which contexts are best served by which locations in management space.[^2] 

### Background
The 2x2 started off as a spectrum between [[George Heilmeier]] and [[JCR Licklider]] — who are both some of the most storied leaders of [[DARPA]] and also seem to have very different management styles. Heilmeier is the originator of the eponymous [[The Heilmeier Catechism]] which demands very concrete answers and planning from PMs (and presumably the researchers they’re working with). Licklider basically gave people free rein to do whatever they wanted with the money he gave them.   [[Two contradicting things that are both true is a strong signal that you need more theory]]. The seemingly contradictory styles suggest that there isn’t actually one good approach but a continuum of control that can be effective (contingent on what though?) 

But then thinking about where people like [[Donald Braben]] or [[Elon Musk]] would fall on this spectrum didn’t feel quite right. While he didn’t demand specific plans, Licklider was extremely opinionated about the vision that the projects were pursuing and would only give money to people who shared that vision. <see [[Program design is a coordination tool]]>. Contrast this to [[Donald Braben]], who worked hard to have *zero* opinion about what people he funded worked on as long as he built up trust with them. While he *did* demand specific plans, Heilmeier was (as far as I can tell) fairly unopinionated about the vision of the plans as long as it was well justified. Contrast this to Elon, who both demands precise planning and has his own strong vision of where a thing needs to go. These dichotomies suggest something like a 2x2 more than a spectrum. 

[[Theory gives you the language to make important subtle distinctions between things]]

### Related
- [[DARPA does multiple levels of top-down problem generation and bottom-up solution generation]]
- [[Programs and Program Managers exist on several different spectra]]
- Contrast this 2x2 about research *management* to [[stokesPasteurQuadrantBasic1997]] which is about the research itself. I don’t think there’s a 1-1 mapping between the quadrants (or even a linear mapping?)
- [[The more trust an institution has, the less it needs formal process]]
- [[Research requires more trust than other disciplines]]
- [[2x2s]]
- [[Research management matters]]

[^1]:Note that constraints are not necessarily the same thing as control. For example, in its platonic form, once you pass through the Heilmeier Catechism you actually have very few reporting requirements. “Control” is some combination of constraints, yes, but with short *observation periods*/high observation frequencies. (This suggests that there are many other axes that we can be looking at here — ideally we could get to a periodic table type ontology. [[Periodic tables are a useful way of categorizing]].)  
[^2]: See [[Does the nature of the activities to create new technology and science change over time?]], [[Activity Space]]

[Web URL for this note](http://notes.benjaminreinhardt.com/Heilmeier_—_Licklider_—_Braben_—_Elon_quadrants_of_Research_Management)

[Comment on this note](http://via.hypothes.is/http://notes.benjaminreinhardt.com/Heilmeier_—_Licklider_—_Braben_—_Elon_quadrants_of_Research_Management)






Phenomena-based cycles

Low Hanging fruit theory of stagnation

Graduate students are the labor in academic science and engineering

Regulations slow down technological advancement

Historical narrative is not scientific inquiry or engineering design but it is valuable

Regulations create paradigm lock-in

The peer review and citation system incentivizes people to work on things that other people think is interesting

Shareholder revolution caused public companies to focus on short-term profits

Conquest state of nature

Phenomena-based cycles are stuck

Competition and lack of slack  creates specialization

Exponential hardware does not lead to exponential software

Human capital theory of stagnation

Progress getting exponentially harder should be the null hypothesis

Disambiguating stagnation arguments

Technology has enabled progress to be a net positive

Regulations have increased by an order of magnitude

Are we actually living in a world where humanity is retreating into a world of digital imagination?

Contemporary research has multiple failure modes

Why is there so little slack in mission-focused research?

The possible upsides of a healthy program design discipline seem like they could be huge

Missing productivity thread - BalajiS

The term moonshot is romantic bullshit

Is Science Hitting a Wall?

Of Flying Cars and the Declining Rate of Profit

Research success in the first half of the 20th century led to many more people becoming researchers, which changed the nature of research in the second half of the 20th century

There are private companies that do planning on the timescale of decades

Vague language only works when dealing with competent people

What if the thoughts we think are nudged by our mediums of communication

# Stagnation
### List of as atomic as possible stagnation theories
1. [[Low Hanging fruit theory of stagnation]]
	- Specifically: [[Progress getting exponentially harder should be the null hypothesis]]
2. [[Regulations slow down technological advancement]]
3. [[Universities have been pushed more towards applied research since the end of WWII]] which then means that academics have been pushed away from more long-term and genera work 
	- This could be attributed to the peer review and citation system
 4. [[The peer review and citation system incentivizes people to work on things that other people think is interesting]] 
	 - This could be attributed to more government funding
	 - This could be attributed to 
 5. [[Embedded growth obligations have shifted the cultural narrative to the idea that everything can and needs to be a high-return investment]]
 6. [[Exponential hardware does not lead to exponential software]]
 7. [[Knowledge of how markets and innovation work has made us jaded]]
 8. [[Research has become a commodity]]
 9. [[The number of researchers has exploded]]
 10. [[Successful professors are not necessarily the best researchers]]
 11. [[Specialization in research and careers in general has cut off a lot of possible innovations]]
	

* What is the rate of discovering new S-curves? How has that changed over time?
* Are today’s S-curves, individually, developing faster or slower than before?
* Are the steep parts as steep as they used to be?
* Are they plateauing at a higher or lower level?
* What is the distribution of these characteristics, over the set of all S-curves?
From 
https://rootsofprogress.org/teasing-apart-the-s-curves
- [[Stagnation and Scientific Incentives]]
- [[Stagnation and Scientific Incentives argues against the low hanging fruit hypothesis by historical analogy]]
- [[Low Hanging fruit theory of stagnation]]
- [[There are three main theories of stagnation]]
- [[Are Ideas Getting Harder to Find?]]
- [[Notes on Are Ideas Getting Harder to Find?]]



[Web URL for this note](http://notes.benjaminreinhardt.com/Stagnation)

[Comment on this note](http://via.hypothes.is/http://notes.benjaminreinhardt.com/Stagnation)




Benʼs working notes

Does the nature of the activities to create new technology and science change over time?

Does the nature of the activities to create new technology and science change over time

Related

Links to this note