Spectech Lessons and Updated Hypotheses from 2023

Things we got right, things we got wrong, and things we didn't know we didn't know

In the spirit of being an institutional experiment, we wanted to share some of the key takeaways from 2023. In addition to our actual outputs, we hope that through the feedback loop between meta scientific ideas and executing on those ideas, we can pave the way for other institutional experiments.  I realize that each of these points wants its own memo unpacking it. I hope to create those over the coming year (which will hit the donors-only section of the substack first!) 

Some of the things that happened over the past year re-enforced the hypotheses we went into the year with; we realized that some of our hypotheses were wrong or needed updating; and there were a number of important lessons that were basically unknown unknowns going into the year. 

Double Down

• There are far more ideas that don’t fit into existing institutions than a single organization can handle. Between launching an organization that explicitly enables work that doesn’t have a home in other institutions and starting an accelerator for that sort of idea has made us a Schelling Point for people with homeless ideas. Of course, some of these are just bad ideas. But many are not! There are far more of these than a single organization can support: some are out of our focus areas, some want “differently-shaped” institutional structures, and there are simply too many among those that are within scope.

• Governments run into fundamental tensions around ambitious research. Research is expensive and creates a lot of public goods. Governments have a lot of money and provide many public goods. As a result, many of us expect governments to  be the main supporters of research. However, government organizations answer to many different stakeholders, impose uniform rules, need to extensively justify their actions, and in most situations move slowly. In many situations these are good things, but create problems for ambitious research. We ran into these impedance mismatches several times over the past year, making us more confident that nimble, low-bureaucracy, (slightly less democratic!) private organizations are essential to enabling

• Materials and manufacturing are an incredibly impactful place to focus for new institutional models. We consistently ran into ideas in materials and manufacturing that don’t have institutional homes: they don’t make good startups, challenge the paradigms of large organizations, and are very non-academic. We found that it’s consistently hard to find funding for ambitious materials and manufacturing projects which is frustrating, on the one hand, but also means that fighting to support them is incredibly impactful on the margin, especially if you buy the importance of materials and manufacturing for other types of life-improving technology.

• Being a nonprofit is important. There are so many pressures towards being a for-profit organization: it’s far easier to raise money, less paperwork, the ability to use equity as an incentive and bargaining chip, the perception that all nonprofits are inefficient and naive, clearer feedback loops… . But throughout 2023, we just kept running into places where there is technology work that could be extremely valuable for the world but don’t make sense for a profit-maximizing business to pursue. That being said, the “nonprofit” designation is primarily an artifact of the American legal system and us being honest that our goal is not to create a return that beats an index fund. It does not mean that we get a free pass on inefficiency, complacency, or being a bad business (see below for more).

Updated/Wrong

• Exclusively working with external performers in the 21st century is severely limiting. The “traditional” ARPA model exclusively depends on coordinating other organizations to do the hands-on research work. One of Spectech’s original hypotheses was that exclusively externalized research was an aspect of the ARPA model that was worth copying (at least early in programs). Experiences over the past year have suggested that exclusively externalized research is more a path-dependent and government-rules-based artifact than an optimal point in institutional design space.

• Remote research organizations take a lot of work to make work. This is less something that we were wrong about but more naive about. On its surface, our model lends itself to a remote organization: program leads work independently and travel frequently. We went into the year with only two full-time employees, now we’re up to five. As we’ve grown, we’ve realized that there’s a lot of intangible things you get from being in the same place that we need to actively work to replace. We still believe we can make it work and there are many advantages, but it takes active effort to replace what you get “for free” when everything is colocated.

• Starting an organization as a hybrid non/for-profit comes along with a lot of overhead.  When we started out, I thought that it would be possible to raise money simultaneously for a for-profit and non-profit entity. Maybe it’s a skill issue on my part, but the work and type of pitch you need to do in order to raise nonprofit money and for-profit money is completely different and both are full-time jobs. You can’t just go in and give someone the option of putting money into a for- or non-profit because people think about them with entirely different parts of their brains and often it’s entirely different people in an organization. Furthermore, for-profit fundraising generally needs to be done in chunks that are sufficient to get you to the next stage of funding. Because of the importance of being a nonprofit, we’ve put the for-profit piece on hold for now.

• There are a lot of non-obvious non-ideal things about philanthropic funding. Going into 2023, I thought the main downside of philanthropic funding was the difficulty of convincing individuals and foundations with many options that your particular effort was a good use of funds without a clear metric like profit. Getting deeper into the system has highlighted more frustrating aspects that are hard to know about a priori. (Note that we’re still incredibly grateful to the people and organizations who have supported us!)

• We need to build out business models that aren’t just donation or investment-based. The hypothesis going into this year was that we could construct a business model that got its short-term cash flows from donations, grants, and investment. That may still be true, but I’m increasingly convinced that there are other business models we should explore. Not just to have additional revenue streams, but because some business models (which I will dig into in more detail on the donor feed) could potentially make us better at executing on our mission to unlock technology that can enable a more abundant, wonderful future (which in turn requires people to be using that technology!)

• The core thing that makes the ARPA model work is that it gives program leaders the agency to do different kinds of leadership.  I went into this year believing that a core aspect behind why the ARPA model worked is that its Program Managers, well, managed research. (See Bottling Lightning). That idea was consistently at odds with stories of how the legendary PM JCR Licklider and other wildly successful PMs operated: it could only be described as “lay out a vision, find good people who bought into it, and give them money.” That was certainly leadership, but it wasn’t management.  At the same time, there are many successful PMs who attribute DARPA’s success to their ability to tightly manage performers. The synthesis of those two contradictory ideas is that it’s not that there is a particular kind of leadership the ARPA model enables but that instead, DARPA’s success stems from giving program leaders the agency to lead in the way that works best for them and for the technology that they are tackling.

New

• Universities have developed a monopoly on pre- and non-commercial research. Today, the vast majority of people who do pre- or non-commercial research are affiliated with a University in one form or another. This wasn’t always the case, but a number of historical trees have converged to make it so unless a research organization is going to do all work in-house, it inevitably needs to interface with a university. Among many other effects, university’s near-monopolies on pre- and non-commercial research means that academic constraints tend to leak into all pre- and non-commercial research projects.

• The skillset to run coordinated research programs is hard to find and teach for everybody, not just us. This year I realized that even talented, experienced people rarely have the complete skillset you need to run a coordinated research program. It’s a skillset that it’s hard to learn in the wild: you learn to be an academic by doing grad school; you learn how startups work by working at one; where do you learn the weird combination of vision, hustle, non-commercial results-focused technical rigor, and coalition-forging? Furthermore, it wasn’t that Spectech had a uniquely hard time finding these people, but so did all the ARPAs and other organizations that needed it. This was the genesis of the Brains program.

• Working with Bureaucracies is incredibly hard for a new organization. There’s been a trend towards increased bureaucracy in the research ecosystem over time. Interfacing with bureaucracies requires a lot of time and effort to make sure you dot all the I’s and cross all the T’s. Established organizations have bureaucracies to interface with other bureaucracies, but in new organizations the bureaucratic interfacing needs to be done by people who would otherwise be running the organization or doing the actual work. On top of the time professors spend writing grants, Universities have whole departments of people whose job it is to manage those grants. The result is that the more organizations need to interface with bureaucracies, the more advantage incumbents have.

• SETAs are an underrated component of the ARPA model. Most analyses of how DARPA works (including my own!) tend to ignore Systems Engineering and Technical Assistance (SETA) contractors. These folks are contractors who help Program Managers with everything from contracting to technical due diligence to how to spin up a program in the first place. Program managers’ short tenures are a load-bearing feature of the ARPA model but its success depends on SETAs who, despite being contractors, sometimes work with the organization for more than a decade and often coach first-time Program Managers.

• Technology research needs tight connections to the people who will nominally use it.  We’ve always known that getting technology out into the world is important and can be one of the weaknesses of the ARPA model outside of a military context. Over the past year we’ve learned that an important component of successful transitions is hanging out with and trying to understand the people who you expect to use the technology from day one, even though you’re doing research instead of building a product for them.

• Contracts (contracting authority, creating them, what they say) really matter. From the outside, it’s easy to ignore contracts as “implementation details” of agreements, but what is in a contract, how an organization goes about contracting, and what sorts of contracts an organization can write or agree to has a huge effect on research. Some examples: PIs at the vast majority of universities can’t contract directly — instead everything needs to go through one or more of several university departments; by law there is a limited set of contracts that government agencies can write, meaning that (without a lot of the work to get exceptions) everything needs to be squeezed into potentially awkward contractual structures.

• It’s less about public vs. private and more about levels of bureaucratization. We went into the year believing there was a particular tension between government funding and weird ambitious research. That’s not wrong, but the causal link runs through bureaucracy. The same forces that hamstring government funded research (consensus-seeking, paperwork, diffuse authority, lethargy, etc.) are characteristic of any heavily bureaucratic organization — whether they are governmental, foundations, universities, or even private companies.

• Nonprofit does not need to mean ‘bad business’ Before 1913, the idea of a 501(c)(3) did not exist, nor did the codified concept of a corporation’s job being to maximize shareholder value. There were just businesses with different governance structures and goals. Speculative Technologies is a nonprofit for two reasons: as a signal and a fundraising mechanism. We’re trying to signal that we really are striving to create public-goods technologies that don’t have a home in other institutions, including companies whose ostensible job is to make their investors as much money as possible. We can’t in good faith tell an investor that by pursuing our goal, we’ll be able to beat even an index fund. However, being a nonprofit does not mean that we need to be a bad business. Many nonprofits are bad businesses — they’re inefficient, have broken feedback loops, survive solely by telling a good story or appealing to vanity, and are frankly a waste of resources. Correlation is not causation. Nonprofits can still make products and services that people love. They can be self-sustaining without depending on handouts. They can create tight feedback loops. It’s certainly hard — to some extent it’s playing on Hard Mode — but nothing worthwhile is easy.

Here’s to a 2024 full of more lessons and surprises!