A Scrappy Complement to FROs: Building More BBNs
Guest post by Eric Gilliam of the Good Science Project
ARPA's early decades of success have made the ARPA model iconic. Big wins from ARPA’s early history include autonomous vehicles, the internet, and stealth aircraft technology. Inspired by these successes, scientific grant funders are increasingly emulating the ARPA PM approach to R&D funding. While I wholly support the proliferation of the ARPA model, it's important to not overlook a key lesson of ARPA history: many exceptional ARPA projects resulted from exceptional contractors.
By "exceptional," I mean contractor groups who were not just staffed with elite talent, but uniquely aligned with ARPA's mission. All three aforementioned ARPA success stories had contractors — which DARPA calls performers — that shared three distinct traits, rarely found together. Each performer:
Was novelty-seeking, with a strong preference for projects that pushed the technological frontier forward substantially.
Built useful technology for actual users. This entailed professional contract management and a willingness to focus on difficult systems engineering tasks.
Used more flexible team structures than academia. When compared to academia, they more effectively hired, organized, and incentivized researchers, engineers, and other experts to collaborate on applied projects in a common-sense fashion.
I call orgs that check all three boxes BBN-model orgs — named after Bolt, Beranek & Newman (the ARPAnet contractor). In this piece, I make the case for the BBN Model and why it can be a fantastic complement to the FRO Model for the R&D community.
In the first section, I outline how BBNs compare to FROs and the bottleneck they address. In the second and third sections, I'll summarize how historically-great DARPA performers like BBN and CMU's early autonomous vehicle teams used the approach to set themselves apart from other academic departments and firms. In the fourth section, I'll list examples of problems this model is optimized to pursue and briefly discuss what is needed to get BBNs off the ground.
I've limited the scope of this piece to simply introducing the concept of BBN-model orgs. While this approach does not apply to every area of R&D, the idea is already finding traction with a small number of founders and funders. Just last week, a bootstrapped BBN, Amodo Design, was made one of nine ARIA activation partners — alongside groups like DeepMind and Convergent Research. I’ll share more details about Amodo and other aspiring BBNs in coming pieces. In the meantime, if you are an entrepreneurial scientist or philanthropist who would like to discuss this topic, please reach out (egillia3@alumni.stanford.edu).
Thanks to Thomas Milton, Corin Wagen, and Adam Marblestone for their comments and ideas on the draft.
BBNs: A Complement to FROs
In 2022, Adam Marblestone et al. published their now well-known Nature piece detailing the vision for Focused Research Organizations (FROs). The team made clear the problem they planned to address:
Our goal is to create a model to support an ecosystem of small-to-mid scale projects that fall between the cracks of what start-ups, academia and other organizations do.
The authors paint a picture of time-bound, nonprofit R&D startups wholly funded to pursue a well-defined goal — usually to build a specific thing like a dataset or tool. They point to "grand projects" like the Human Genome Project, Large Hadron Collider, Hubble Telescope, and Human Cell Atlas as having similar goals. Less than three years on from that piece, Convergent Research has brought six FROs into existence (and counting). Adam and the team have proven that you can recruit teams of the very best researchers to join these orgs, make them high-status, and get the operational scaffolding in place to begin doing great work with minimal delay. Convergent’s experiment in scientific structures is off to a great start.
But fundraising is a clear bottleneck to the creation of new FROs — often costing $20 to $100 million. There is, of course, an upside to this tradeoff. FROs raise so much money upfront to ensure they will have all the money they need to pursue their vision from day one. But until groups like universities or the NIH are convinced to set aside substantial war chests to fund and run FROs, fundraising will continue to be a bottleneck to FRO creation. In the meantime, many with the talent and technical vision to found and run FROs will not be able to do so if they are not one of the lucky few to complete a successful FRO fundraise.
BBNs offer an alternative path for these talented individuals. BBNs raise less money upfront than FROs. Instead, they heavily incorporate a contract research approach to build towards the ambitious technical vision of their founding field strategist. They can operate as firms or nonprofits, drawing on a mix of R&D grants and contracts to fund their technical agenda. Project by project, BBNs can make strides in key areas of R&D — as BBN did with real-time computing and the CMU teams did with autonomous vehicles.¹ The approach limits BBN founding areas to those in which funders are eager to pay for contracts aligned with some ambitious technical vision. But where those markets do exist, BBNs can be spun up for an order of magnitude less upfront capital than FROs in many cases — some might even be bootstrapped.
The downside is that this introduces a level of ongoing financial risk that is not present in FROs. Founding a BBN is not advisable if you can successfully fundraise for an FRO or have an offer to run a team at the Arc Institute. Those are more stable options. BBN founders may need to be a bit scrappy or creative in their revenue strategies to survive and expand. They might have grand ambitions on the scale of FROs or the Arc Institute, but they must begin with plans that, in the beginning, operate on the scale of an NIH grant, philanthropic grant, or DARPA contract. But if done well, as with FROs, the BBN model can enable a “critical mass of scientists, researchers, and managers” to assemble and pursue ambitious R&D agendas outside of the university.
BBNs offer ambitious field strategists a chance. With BBNs, the chance offered by some contracts and a bright team can be leveraged to world-changing effect. While BBNs might even compete for the same contracts as groups like Raytheon, what differentiates the two is that BBNs are steered in ways that prioritize an ambitious technical agenda over revenue-maximization. This can be tough, but it can be done. The early career of the original BBN’s first computing field strategist, J.C.R. Licklider, demonstrates this. His steering of the BBN computing portfolio provides a clear example of how the BBN approach — with the right field strategist, team, and contracts — can drive an exceptionally ambitious R&D agenda.
BBN’s Computing Agenda: A Mix of Grants, Contracts, Talent, and Vision
BBN was founded in the 1940s. It began as a psychoacoustics R&D firm spun out of MIT by two professors. The firm began as a way for the professors to take on large contracts that were cutting-edge, but too involved to pursue in their own labs — concert hall design for the UN, cockpit design R&D for the Navy, etc. Over the next decade, the firm opportunistically expanded. Hiring for their firm as if they were staffing an academic department, the rule of thumb was to only make hires that would “raise the average level of competence of the firm.” In the late 1950s, the BBN partners looked to expand into computing. They believed J.C.R. Licklider was the best person to lead this effort. Licklider had a clear vision for the near future of computing. It was a vision that might have been too engineering-heavy to be optimally pursued at MIT.
BBN offered an interesting alternative. The firm had carved out a useful niche for itself, with a comparative advantage in projects that required both:
Cutting-edge academic knowledge
Professional engineering work or contract management
Licklider, a man of impressive vision, saw the administrative scaffolding BBN provided and believed it could be used to pursue his technical vision of real-time personal computing. This technical agenda — quite different from the batch-processing paradigm then in vogue — would require substantial engineering improvements in computing interfaces and real-time computing technology. Licklider left his tenured position at MIT to join BBN. Under his leadership, BBN’s computing efforts would go on to earn BBN praise from many former MIT researchers, later being described as “the third great university of Cambridge,” “the cognac of the research business,” and the true “middle ground between academia and the commercial world.”
A field strategist with an infectious sales pitch, Lick would gradually convince many MIT, Harvard, and Lincoln Labs researchers to defect and join BBN. This included a core of rare real-time computing experts, many of whom came from Lincoln Labs. While MIT and Lincoln Labs were great research environments, many who defected to BBN felt the nature of academia forced them to leave their best ideas as under-developed prototypes or “applied” ideas in papers. Many of these individuals — more engineers than pure researchers — would have felt their minds were wasted at a large firm like Honeywell. But BBN offered an interesting alternative to them as well. They could simultaneously work on the hardest problems and build useful technology.
Licklider convinced BBN to provide funds for computers that were relatively generous by the firm’s standards, but quite small given the scope of his ambition. The work began with a $30,000 Royal McBee ($300,000 today). Licklider soon convinced the BBN partners to buy another, more powerful machine that he felt would be more useful. After this second purchase, Licklider immediately proved the partners’ confidence in him was well-placed. One BBN founder described how seamlessly contracts fell into place following the second purchase, writing:
Lick and I took off for Washington D.C. to seek research contracts that would make use of this machine, which carried a price tag of $150,000 (~$1.5 million today). Our visits to the Department of Education, National Institutes of Health, National Science Foundation, NASA, and the Department of Defense proved Lick’s convictions correct, and we soon secured several important contracts.
Licklider’s computing group routinely found creative ways to offset the costs of investigations aligned with their technical agenda. Some of the group’s revenue came from relatively typical academic funding sources — federal research grants, a NASA contract to use the group’s computer to produce a textbook, etc. But BBN also took on much more applied contracts, well-suited to the group’s embrace of engineering and project management tasks. I’ll briefly describe two of these applied projects that took advantage of BBN’s comparative advantages. Both helped lay the groundwork for now-famous technical breakthroughs.
The first of these projects was BBN's Libraries of the Future Project. The project, indirectly funded by the Ford Foundation, sought to understand how computing might impact libraries moving forward. It was a contract for a philanthropy that, today, might go to a consulting firm like the Bridgespan Group. But in this unassuming contract, Lick saw an opportunity. The BBN team won the contract and used it as an excuse to explore seemingly futurist technology and make practical engineering improvements. As one example, the second half of the group’s book-length report for this unassuming contract contains technical explorations that attempt to emulate a future in which users interact with the library as a store of knowledge with a question-answering front-end, not as a way to find books. The final report also contained early work related to PC file systems, improved information retrieval methods, mathematical representations of information, and associated hardware improvements — all paid for by this unassuming contract. The opportunities to further explore and translate academic ideas via this contract proved attractive to other researchers in Cambridge. On this project alone, Marvin Minsky, John McCarthy, and Fischer Black all spent time working with the BBN group.
The second of these projects was built on a contract to develop a usable real-time computing system in a presciently chosen application area: hospital administration. The contract came about because a BBN partner made a compelling pitch to the NIH Clinical Center’s Director that hospital administration would be done using computers at some point. The director was convinced and had the partner write up a proposal for BBN to begin developing and building a computing system to pilot with a research hospital. The BBN hospital time-sharing computer project provided the firm with three years of funding, worth ~$10 million in total today. The contracting relationship would continue for several additional years, growing into an actual system deployed in the operations of Massachusetts General Hospital. With the contract, BBN was able to fund years of work a bit too far along for academia but not close to private-market-readiness. The contract required far more contract management and engineering work than would have been reasonable for a university team — they had to continually reduce machine error rates, install and maintain terminals at the hospital, coordinate with hospital staff, train doctors, provide customer service, make UX changes, etc. But with this applied work came a sizable budget. BBN’s small team of real-time computing experts used it to push the engineering frontier forward.
By the time Larry Roberts began putting out feelers for the ARPAnet contract in the late-1960s, BBN had been solving all sorts of related problems for years — similar to the ones that had stumped Roberts while he was at MIT. BBN won the contract. The first four nodes of the ARPAnet would be delivered within a year, on time and on budget. But it was BBN’s years of work on related contracts that paved the way for this overnight success.
Nonprofit BBNs: The Case of CMU's Early Autonomous Vehicle Teams
The BBN model can also be operated with a nonprofit structure. The early CMU autonomous vehicle teams — which helped pave the way for the autonomous vehicle revolution — demonstrate this. In another world, these teams probably could have operated with an FRO structure. But in practice, they operated using a model more similar to BBN’s contract research model.
The CMU Robotics Institute was primarily established to build ambitious, practical computing systems. While housed at a university, these CMU teams are more aptly compared to a group like BBN than a group like Marvin Minsky's computer science department at MIT. Contracts were a large part of the Institute’s plans from its inception. The Institute’s early autonomous vehicle group successfully deployed a BBN-style approach on DARPA’s 1980s autonomous vehicle contracts, achieving clear counterfactual impact and developing systems DARPA’s primes and academic labs under contract would not.²
The CMU group had an operational structure and incentives that enabled it to excel on these ambitious, applied contracts. As one example, the Robotics Institute created suitable tracks through which engineering and project management-focused individuals could be hired and tenured just as CMU's best professors would be. Instead of researching and teaching classes, these individuals would research and operate projects. A CMU Ph.D. and one-time project manager for the group’s early autonomous vehicle projects, Chuck Thorpe described the incentives of his role, recounting:
Raj [Reddy] took me aside. He said, “I don't care how many papers you write. I don't care how many awards you win. That vehicle has to move down the road. If you do that, you're good. I'll defend you. I'll support you, I'll promote you.”
The CMU teams also found ways to effectively use the grad students as both project staff and researchers. Requiring dissertation problems of their own they could “sole author,” grad students were assigned components of the system for their dissertation to improve upon. Each project had to easily plug into the larger system. High-risk, high-reward projects still had a place in this structure. If projects were high-risk, project managers like Chuck Thorpe might simply assign more grad students to improve a component to ensure at least one succeeded. A good project might reduce the processing time for data from some sensor 10X. The best grad student project would prove to be world-changing.
In 1988, a sharp grad student named Dean Pommerleau combined his knowledge of CMU’s systolic array machine and neural nets to build a neural net-based steering system for the vehicle — ALVINN. In the following years, the team built this neural net steering system into a new vehicle system. In 1995, with this new system, the CMU vehicle would drive cross-country 98.5% autonomously on its “No Hands Across America” tour.
Getting BBNs Off The Ground
Attempting to found and operate a BBN is not easy. You might fail to raise a required seed grant or win contracts to get off the ground. You might sell a contract funder on your idea but need a partnership with an existing org or philanthropy to get the deal over the finish line. You might win an initial batch of revenue and still have to shut down after a year due to a lack of funds. You might even succeed in raising funds, but raise them from misaligned funders and become a myopic R&D contractor. This is a failure mode all the same; becoming Battelle or Raytheon is not the goal of a BBN. But great field strategists can use the BBN model to turn revenue, people, and vision into what Patrick Collison recently called a “great scene of discovery.” This is what BBN made itself into for real-time computing and CMU for autonomous vehicles.
Building a BBN is a viable, alternative path for those field strategists and scientific entrepreneurs eager to do work in the cracks of the R&D ecosystem. Done well, BBNs can be a force multiplier to investments made by scientific funders, enabling them to undertake projects that are cheaper, better, or different than they would otherwise. This “create the performer” approach still has its uses in the context of modern philanthropy.³ As Jacob Trefethen recently discussed on his blog, the Gates Foundation funds “half or more” of many global health product development partnerships (PDP) budgets because they usefully fill cracks vital to Gate’s global health mission.
But founding BBNs can require overcoming a somewhat delicate coordination problem. While some BBNs can be bootstrapped, many BBNs will require some combination of a low-seven-figure seed grant to get started, contracts, and (maybe) NSF/NIH grants to optimally pursue their research agenda. It takes a bit of talent to overcome the associated chicken-and-egg problems that might arise. Philanthropies might want to provide seed funds to spin up a BBN, but only if they can be assured contracts for R&D projects will materialize. A contract funder like DARPA might require the org actually exist before engaging in some negotiations. OpenPhil might happily provide 60% of the budget for ~2 years for some BBN useful to its mission but require the group to source additional funds and contracts to cover the rest. The list goes on. It’s not an easy path. But if the crack is big enough and a field strategist is proactive enough, these challenges can be overcome.⁴
But none of this can happen without an ambitious, open-minded contract funder. Funders with the level of vision and ambition of ARIA are ideal. I would encourage funders reading this to ask themselves the question, “Could I take my portfolio to another level if I was able to create a single performer to work in a targeted area?” If the answer is “yes” or “maybe,” you might benefit from a BBN.
BBNs might apply to many areas. Some areas in which BBNs might have natural comparative advantages include:
Development of new instruments
Producing bespoke hardware
Engineering new pipelines to serve a group of researchers
Building and maintaining research software or datasets
Pilot-plant style work with the intention to:
Scale promising lab bench ideas in manufacturable processes
Produce a key research material for labs
Translation of academic ideas to non-venture markets.
One example could be a CivTech BBN that wins city government contracts to both translate the best social science insights into practice and conducts research projects on the org’s internal data.
It’s hard to predict how many areas might benefit from a BBN, but the model offers opportunities for the ideas of special individuals to surprise you. Progress studies funder and Stripe CEO Patrick Collison recently spoke of a similar phenomenon regarding a new Stripe product, tweeting:
We launched Stripe Issuing a few years ago. I wasn't sure how many use-cases there would be — like, issuing a card is surely a pretty niche thing to do. As usual, developer platforms surprise, and customers have now issued a cumulative 200 million cards.
BBN’s idea to bid on the Libraries of the Future project is one example of this cleverness at work. The BBN model has the potential to not just plug obvious cracks in the R&D ecosystem, but create great scenes of discovery in their place. In which cracks these scenes will take hold is hard to predict.
The fact that FROs now exist and are a fixture of the new science lexicon makes the task of coming up with ideas for BBNs easier for field strategists. A good FRO problem is, in many cases, rather similar to a problem for a BBN. Adam Marblestone recently shared one example, explaining that an FRO might arise from a problem like neuroscientists requiring a different kind of microchip. One complex chip solving many life sciences problems is a great example of an FRO problem. But what if many neuroscience applications required many, simpler microchips? That’s a problem well-suited to a BBN.⁵
Tyler Cowen points to the history of Hollywood as proof that exciting young people can find ways to create exceptional things if given an opportunity to take charge and figure things out. FROs, with their time-bound nature, remind him of Hollywood movies. BBNs, with the scrappy nature of their origins, have the chance to become a vehicle for unproven R&D entrepreneurs to prove they are masters of their craft — just as Tyler believes Paul McCartney did with the Beatles.
It is an exciting proposition. But founders and funders with vision are required to make it a reality.
Thanks for reading.
Please reach out if…
I am already working with a few field strategists who believe they might have ideas for BBNs in areas like automated material discovery, bespoke semiconductor production as test bed hardware for certain areas of semiconductor design, and modifying instruments to enable a wider range of life science experiments. I will release those pieces in the coming weeks. Please reach out if you are a funder or interested field strategist to discuss ideas (egillia3@alumni.stanford.edu or on Twitter). I’d love to help!
I am setting aside as much time as necessary to assist would-be BBN founders and funders. I believe we have J.C.R. Lickliders sitting on the shelf with vision, philanthropists with exciting ideas that are best served by a new contractor, and outside funders who might happily provide the seed funding to catalyze this change. BBNs can be a vehicle to turn this potential energy into action.
Various ideas and details included in this piece have been explored in FreakTakes pieces on BBN, DARPA’s Autonomous Land Vehicle and NavLab projects, an interview with CMU’s Chuck Thorpe, and how BBNs might apply to new era semiconductor research.
I’d also like to mention that Adam Marblestone listed OtherLab as an org that does great work with a BBN-type model.
1 The model offers other advantages in relation to FROs. BBNs are well-suited to projects/areas where you need to work closely with users to define your problem over time, gradually build up your team in response to user needs, or find it advantageous to not be overly influenced by a single funder’s wishes.
2 In the early years of DARPA's first autonomous recon vehicle project, while many academic DARPA fundees plead for data from DARPA’s vehicle to train academic vision algorithms, the CMU team plead for $1.2 million ($3.5 million today) to build a pair of their own test bed vehicles. In 1986, DARPA agreed. This filled a key crack in DARPA’s performer ecosystem. To that point, the prime R&D contractor, Martin Marietta (the “Martin” in Lockheed Martin), had shown little enthusiasm for testing cutting-edge academic ideas. This test bed work entailed piles of engineering and administrative work academic departments did not want to touch. This niche of “test bed contractor for cutting-edge ideas” was largely unwanted. But CMU, with its systems engineering focus, wanted it.
3 I credit Adam Marblestone with this phrasing. He used it in an email exchange.
4 In many areas, the market for BBNs might be growing. The more difficult it becomes to pursue applied contracts at universities, the less common it becomes to do ambitious research at R&D firms like Raytheon, and the more field strategists take positions of authority as scientific funders, the more cracks form which BBNs might reasonably fill.
5 In fact, in the 1970s DARPA created its own performer to make bespoke chips for its MOSIS program — supplying DARPA researchers, professors, and EE grad students with cheap, fabricated chips of their own design.