At the end of the day, someone needs to pay for research that requires external funding. That means it needs to provide a product or service that someone wants. If the research ecosystem is going to flourish the way we all want it to, we need to look that fact straight in the eye and be far more strategic about it.

Researchers (and generally people marinating in mainstream culture) are awash with the idea that good research “deserves” funding. I lay most of the blame on research’s role in winning WWII and its centrality to the subsequent Cold War. Vannevar Bush articulated what we could call the “central dogma of the Endless Frontier”:

“Scientific progress on a broad front results from the free play of free intellects, working on subjects of their own choice, in the manner dictated by their curiosity for exploration of the unknown.”

That dogma was then encoded in government, philanthropic, and even (to a lesser extent) corporate policies. Perhaps more importantly, it became the water that researchers ourselves swim in.

But living in the paradigm of The Endless Frontier was a historical anomaly. For the hundreds of years from Bacon to the mid 20th century, researchers knew that support for their work was contingent on delivering something that someone actually wanted. People “buy” science for many reasons — including, but not limited to: wonder, status, guilt, and existential dread on a personal or national level. But for better or worse, “because a researcher thinks an idea is really interesting” is not one of them.

Cracks in the paradigm had been appearing for years, but for reasons I won’t unpack here, events around 2020 brought the paradigmatic crisis to a head. It’s clear to more and more people that the 20th century system of doing science is no longer fit for purpose. The existence of the “independent research ecosystem” is largely downstream of that realization. But one way we are still too much like our “dependent” brethren is that we don’t think of research as a product that has customers the same as anything else that needs money. (I include myself here!) To some extent, the title “independent research ecosystem” even points at this!

Many researchers have become dodos — unaccustomed to a red-in-tooth-and-claw world where we need to sell research that addresses a need. A world of scarce resources, where research needs to compete for every dollar against a thousand other things.

So, this is the challenge I want to throw down: we need to take a cold hard look at what we’re selling to actual existing “customers” and why it’s better than, shall we call them, the “dependent” ecosystem. We need to stop fetishizing the Apollo Program, Bell Labs, the Manhattan Project, and Xerox PARC. Those organizations all sold things to customers who no longer exist. (Though they do hold lessons about how to do effective research!)

The past few years should make it clear that the question “what are you selling and to whom?” isn’t just a question for those of us trying to forge new institutional paths – it’s a question for everybody in the research ecosystem. Answering it is, I believe, a key part of figuring out what replaces the Endless Frontier paradigm and building the 21st century science ecosystem.

How the “independent research ecosystem” can help sell research (and what a research ecosystem even means)

So many of the arguments about how science will work in the 21st century bottom out around money. (There are many other things to be figured out besides money, of course!) If there was so much money floating around that everybody with a research question could all be doing exactly what they want in the way that they wanted, the conversation would look very different.

There are three ways to address money in research:

1. Get more post-economic people to do research.

2. Make research so cheap that people can do it in their spare time.

3. Create and sell something people want to buy.

The first two are worthy and underrated topics but I’m going to focus on the third for now.

Some blunt reasons I can see that different groups buy research:

• Countries buy research when it addresses an existential threat (aka legitimately enhances their military power) or satisfies some other political priority.

• Individuals also buy research when it addresses an existential threat to themselves or their loved ones (hence so much philanthropy going into health research), potentially makes them money, raises their status, or scratches some deeply personal curiosity.

• Corporations also buy research when it addresses an existential threat: they feel like if they don’t do it a competitor will get a leg up on them, or like AT&T did with Bell Labs, they use it as a decoy to save themselves from anti-trust or other regulation. They also buy it when they can see a clear path to it making them more money by cutting costs or enhancing an existing product line.

• Countries and corporations also buy research for some of its second-order effects like creating talent pipelines or marketing, but those are fickle behaviors that are quickly reversed.

There are of course many exceptions, but by and large, I fear we need to figure out how to fit our work into this framework.

Since we are here to talk about the “independent research ecosystem,” I want to suggest two ways that the ecosystem itself can help us, as research organizations and individual researchers, sell something valuable.

First, the ecosystem can help by being an actual ecosystem. Right now we’re not.

But what does it even mean to be an ecosystem? To answer that, I want to quickly introduce my preferred alternative to the “linear model” of science. (You know, the “basic science”-> “applied science” -> “development” model that is everywhere when we talk about research and even baked into law.)

A cellular metabolism is the closest thing to an accurate model of “big S” science: how things go from ideas to impactful technologies or accepted theories about how the world works.

Instead of a few discrete categories, science actually consists of millions of people doing (thousands? Hundreds of thousands?) of different ‘flavors’ of interconnected work that turn ideas, observations, and raw material into knowledge and technologies that extend our understanding and capabilities beyond our own hands, eyes, and heads. These interconnections create a complex system (in the technical sense of the word) which means that abstractions like the linear model or 2D models like Stoke’s Quadrants are unable to capture very important system dynamics.

There are far too many flavors of work to list, but to give you a sense of the granularity I’m talking about, nodes in the metabolism might look like:

• Poking at a material at a lab bench

• Building a tool so that people can do the work at a lab bench

• Doing theory to explain that experiment on a white board

• Figuring out the kinetics of that reaction at scale

• Tinkering with a scaled production system (and sometimes realizing that there is a gap in theory or a lab experiment to be done)

• Training new researchers to do more of this work

• Arguing about how resources should be allocated to do each of these things

All of this work then happens within institutions. (I use “institution” here in the sense of “a category of organization that shapes people’s behavior through a shared set of norms and incentives.”) Any flavor of work can theoretically be done in any institution, but different institutions are better or worse places for different flavors of work: universities are great places for whiteboard theorizing, for example, but bad for creating products; high-growth startups are the opposite; neither are good places for creating entirely new technology systems like the personal computer or self-improving factories.

Organizations and institutions get good at a specific subset of work. Most institutions are bad at admitting and embracing that fact. One of the competitive advantages we could have is being explicit about the flavors of work that our organizations do and do not do and take advantage of the wonders of specialization.

The thing that makes a metabolism or an ecosystem work is the interconnections between metabolic processes or ecosystem niches. Ideas rarely become reality while staying entirely within a single organization (or even usually a single group of people). These interconnections are where we currently fail. We have not really figured out how to lower the friction for ideas moving between different niches in the ecosystem — especially in a way that makes our research a more valuable product. If we could figure that out, people might bias towards supporting non-traditional organizations because they’re hooked into a more efficient metabolism.

The second way that an “independent research ecosystem” could help sell research is by becoming a value-enhancing category.

“Independent research” is itself not a product. It’s not even a product category. Perhaps it could become one — something like “organic” or “fair trade.” But nobody buys an organic apple just because it’s organic. They buy an organic apple because they want an apple and they believe an organic apple is superior to a normal apple.

People believe organic products are superior for two reasons.

The first reason is marketing. If you take a normal banana, slice it in half, and have people sample each half while telling them one is organic and the other is not, they will rate it as more delicious. That psychological flavoring is downstream of great marketing. Yes, we’re going to need to do much better marketing (which we are notoriously bad at).

This marketing needs to implicitly or explicitly answer the question “in what way is your product superior to the incumbent offering?” I don’t think “well, those guys are old-school and we’re not them” is going to cut it. One spicy possibility, though, is that “independent research” could become “fair trade” for certain ideological groups; a way of signaling to their peers.

The second reason people are willing to pay more for some categories is that they have actually experienced superior products in those categories! Produce rots more quickly without chemicals so most organic produce is fresher. Fresh produce is absolutely tastier, regardless of how it was grown. I would like to think that the new ecosystem can build a reputation for delivering superior research.

But that’s going to require a lot of hard work, time, and uncomfortably, incredibly high standards and some amount of internal policing. As a community, I’ve observed that we have a tendency to take a non-judgemental attitude towards other people’s research. There are good reasons for this! It is a reasonable reaction to the deep research conservatism and one-upmanship culture of academia. Indeed, a lot of incredible research does look crazy a-priori. But “all ideas are equally valid” is the wrong lesson to take from history and existing organizations. Most ideas and research are, frankly, garbage. If we want independent research to be a desirable category, we need to figure out how to make it become associated with quality.

There is an uncomfortable possibility that taking such an analytical and utilitarian view towards research support is, ultimately, self-defeating. That supporting research must ultimately be treated as a transcendent activity that’s incomparable to other uses of money. But that’s certainly not how it’s seen right now.

So in the meantime, as we try to build organizations and institutions that will hopefully come to define 21st century science, I want to encourage you to be in touch with reality. To ask “who are we selling to?” as individuals, organizations, and an ecosystem.

Every engineering field has its utopias: biotechnology has the immortality drug and mind uploading, computer science has the godlike AI and perfect external memory, aerospace has the flying car, warp drive, and O’Neil cylinder. The alchemists had their philosopher’s stone.

In manufacturing, we have ours too — the Star Trek replicator and the von Neumann machine. Machines that can make anything, anywhere, even more of themselves. Machines that frictionlessly turn imagination into reality.

Like any utopia, these are asymptotes, destinations so far off that they may as well be impossible to reach in a finite time. But striving for them is nevertheless valuable. Striving for the philosopher’s stone gave us modern chemistry. We haven’t won the war on cancer, but many people are alive who wouldn’t otherwise be. And while the AI labs have not (yet) created a god, they have created some pretty incredible tools.

I want to lay out a potential manufacturing paradigm that would bring us one step toward the utopian manufacturing dream. I also want to argue why building this new paradigm is both timely and important for the US in 2026.

Note: I’m using the term “paradigm” in the sense of Thomas Kuhn’s philosophy of science: self-consistent ways of doing things or thinking about the world based on a set of assumptions that occasionally change in a “paradigm shift.” The classic example of shifting paradigms in science is the transition from the geocentric model of the universe, which assumed that all celestial bodies revolved around the Earth (potentially attached to crystal spheres), to a heliocentric model that assumes objects in our solar system revolve around the sun driven by gravity in a universe with no fixed “center.” The shift from Newtonian to Einsteinian physics is another.

We can extend this idea of paradigms to technology. The cleanest example of a manufacturing paradigm is interchangeable parts, built on the assumption that you make things as assemblies of parts that are identical to within some tolerance. That was not always true! For most of history, each part was hand-shaped to fit with all the other parts of a specific device. As late as the 1960s, interchangeable parts weren’t the default; they were expensive and only worth it if production volumes were high enough.

We currently live in a manufacturing paradigm unlocked by a world where sending information and goods is cheap thanks to the internet and containerization; statistics, liquid markets, and global trade have made just-in-time production possible. We could call it the “efficient networks” paradigm. This paradigm is full of assumptions about what’s possible and how things work. We don’t think about them much — living in a paradigm is like being a fish in water. Some of the assumptions include:

• The process of prototyping a system is very different from the process of manufacturing it at scale.

• A new product requires a new production line.

• Each factory and production line is custom-built and unique, requiring untold engineering hours even if it’s highly automated

• High volume, low mix is the way to make things cost-effectively. High volume, high mix is expensive.

• Physical products have models that are updated once a year or more.

We can invalidate these assumptions.

If you can scale cost-effectively through parallelization, you can produce a system in the same way you prototyped it. If you can make things entirely with software-defined tools you don’t need to waste time and treasure building a new production line or factory for each product. These properties then drastically reduce the switching costs between different outputs, making costs agnostic to the number of things you’re making. All of this then makes creating physical products far more like software, where you can “push” updates to the next unit on the line as soon as it’s ready, drastically increasing the rate of improvement and learning.

I realize this is all incredibly abstract! It is hard to describe something that does not yet exist. So let me describe how making something within this paradigm might work.

An engineer in the not-so-distant future will start by specifying a system they want to build like a drone or a robot. This will include not just CAD files of its structural components, but all the specs — force and torque, thermal characteristics, power consumption, all of it. (If we’re being even more ambitious, the specs could be much closer to the actual things you would want a robot to do similar to “unit testing” in software.) Software – let’s call it an AI agent for both the anthropomorphism and convenient handle to refer to it — takes a given design or creates one on its own and then simulates it to see if it can indeed hit specs and pass tests. If not, the agent feeds the failures back and updates the design until it does.

The agent then breaks the target system down into components and runs a similar simulation-design feedback loop for each of them. With the system broken down into notionally-correct components, the agent then simulates how each of them will actually be made — from the tool to the exact movements it needs to go through. The agent will then need to loop between this step and the component’s design (potentially even back to the system design) because as any good engineer knows, there is often a tight coupling between a component’s design, its properties, and how it’s made. The specific path by which you lay down carbon fiber has a huge effect on its strength; the orientation of layers in a 3D-printed part determines where it is strong or weak to shear forces. After all of these loops (and loops within loops) we still have not yet moved a single atom! But abundant, powerful compute is one of the biggest answers to “why now, what’s different?”

With a manufacturing plan in place, several shipping-crate-sized boxes slide together. They’re filled with software-defined machines (or at least the business ends of them): 3D printers, CNCs, and more that have not yet been invented. Actuators — some that look like robot arms, some more exotic — pull in raw materials and in a whirl of activity, the “cell” transforms them into a finished component. More tools swing into place that test the component against the specs the agent specified. If a component is unacceptable, data feeds back to agents who update the synthesis plan and even the part’s design to address the failed test’s root cause. After the synthesis loop completes each component, the actuators assemble the components into subsystems and then the full system — autonomously testing and updating until you have a verified functional robot and a verified production process.

This way of doing things opens many possibilities. When you want to increase production, you just activate more cells. Because the entire process is encoded in software and the hardware is identical, activating new cells is like installing a new program on a computer. The price of this hardware drops because you can scale its production. This transferability means that you can easily spin up production close to where you need the product whether that’s at the ends of the earth or in space. Automated testing loops drastically lower the overhead to customizing and prototyping. The same feedback loops can go one step farther and incorporate data from operation in the field directly back into the design — there are so many problems that pop up when a robot encounters the real world that no amount of testing can surface. Manufacturing infrastructure can rapidly switch from making factory robots to masks to military drones, if need be. These properties in turn will have many incredible second order effects — drastically lowering the overhead to developing physical products, manufacturing in space, robots of all shapes and sizes, and the list goes on.

There are two skeptical reactions to the picture I’ve tried to paint that I would like to briefly address: “this is an impossible utopian pipedream” and “this is basically already happening.” Both are wrong but both have a nugget of truth.

This new paradigm won’t happen tomorrow. We are not going to get there in one giant leap — we’ll make some things this way long before others. We’re in a world of mainframes and vacuum tubes talking about graphical user interfaces and laptops. It took more than a hundred years between the invention of interchangeable parts and their widespread use. It’s not even inevitable: work to create new paradigms requires navigating potentially-impossible open research problems, many market failures, and vested interests who don’t see why the investment is worthwhile or actively oppose it.

But the pieces are already out there: AI for different parts of the engineering stack, digital twins, training robots in simulation, and an explosion of tooling-free machines. Seeing this, you might even be tempted to extrapolate and believe that this new paradigm is basically inevitable. That’s also wrong! We have precursor pieces, yes, but currently they’re being used as point changes in the existing paradigm: the simulation-trained robots are installed in big fixed production lines, AI for CAD generates 2D drawings for contract manufacturers. Building the complete system that leverages them to their full potential is still a huge hurdle that will require entirely new frameworks and reinventing existing tools.

So where do we start? The domains that will benefit the most from this approach are ones in which the design and production of a part or assembly are deeply coupled. That is, if you want the thing to work, you need to think about how you’re going to make it while you’re designing it. Everything has this property to some extent: placing two holes half a millimeter apart is easy in CAD, but unless you’re incredibly careful, real physical metal will tear when you try to drill those holes. This is why design for manufacturing is a thing. This design-manufacturing coupling is especially tight in something like carbon-fiber parts or power electronics. Two carbon-fiber objects with the exact same shape can have drastically different properties depending on how the fibers were laid down and bonded. Those process-dependent properties are especially hard to simulate, so today carbon fiber parts depend heavily on tacit knowledge and iteration, which in turn drives up price, development time, and ultimately keeps people from using it.

From there we need to simultaneously develop general purpose connective tissue — the software architectures, agents, and robotic systems — and one-by-one add “skills” and tools like software-defined forging machines.

Let me close by explaining why this matters. At the highest level, creating new manufacturing paradigms matters because materials and manufacturing underpin civilization. Ultimately, we are limited by the stuff we can make things out of and our ability to turn that stuff into useful things. New manufacturing paradigms unlock new products and let us make existing ones cheaper, better, and faster.

This matters particularly now in early 2026, not just for abstract reasons, but because of two concrete trends reshaping our world: American reindustrialization and space.

Both the supply chain shocks of the COVID-19 pandemic and political tensions with the world’s largest manufacturer have made many people want to manufacture more things in the US. If you look at history, new paradigms drive shifts of manufacturing centers of gravity. Steam power drove a shift from mainland Europe to Britain; interchangeable parts (which were colloquially known as “the American System”) drove a shift from Britain to the US; just-in-time manufacturing drove a shift from the US to Japan; and network manufacturing drove a shift to China. In each situation, a new entrant who was previously looked down on as a low-quality manufacturing “backwater” scooped the dominant incumbent using a new approach. In economics (specifically disruption theory), new entrants don’t just displace incumbents by trying really hard (although that’s necessary too) — they take a new approach to the problem that leapfrogs the old one. Building minimills for processing steel, going straight to mobile, etc. If we’re going to manufacture things cost-effectively in the US, we’ll need to do the same thing. We’re not going to out-China China. Instead, we need to leapfrog them with a new paradigm that we can become uniquely good at.

Today, space is also on the mind. If Artemis ever gets off the ground it will send astronauts around the moon for the first time in more than half a century; there’s talk of nuclear power and a permanent base on the moon; and let’s not start with space datacenters. And frankly, it’s close to my own heart. I’m a card-carrying space nerd: my PhD work focused on space robotics, I worked at NASA during the summers, and my advisor was NASA’s chief technologist. If humanity is going to become a space-faring species, we will eventually need to make things in space — both to take advantage of the unique conditions in space and because it would be insane to launch every kilogram of mass we use there up from Earth. This new paradigm is uniquely well-suited to space manufacturing: launch once and manufacture anything with designs you can test on identical hardware and then “upload” from earth.

It’s hard to give this thing a name. We can talk about what it is: software-defined, closed loop, and modular. We can talk about what it isn’t: tooling-dependent, manual, or tied to specific locations or production lines. I’m torn between two potential names for this thing because there are two core pieces that I think will really unlock this paradigm:

1. The ability to “backpropagate” differences between a spec and measured performance data back to a design and manufacturing process. This suggests Differentiable Manufacturing.

2. The ability (partially dependent on #1) to run the same process on different pieces of hardware without extensive calibration. This suggests Compiled Manufacturing.

Every prior paradigm looked impractical until it wasn’t. Interchangeable parts were mocked as “the American System” because Europeans thought Americans too crude for real manufacturing. Toyota’s methods were “culturally specific” until they ate Detroit’s lunch. But we can make today’s assumptions as quaint as hand-fitting each musket. The pieces exist. The need is clear. What’s missing is the connective tissue that turns components into a system.

We’re actively working on this at Spectech! If you want to help support or get involved, please reach out.

Further Reading

• Austin Vernon’ Speed Can Reindustrialize America

• Brian Potter’s Is the Future “AWS for Everything”?

• David Hounshell’s From the American System to Mass Production, 1800-1932: The Development of Manufacturing Technology in the United States

• Dan Wang’s 2025 Letter

• Ben Reinhardt’s The Art of Industrial Leapfrogging

Thanks to Austin Vernon and Brian Potter for reading drafts of this doc. Claude caught a few grammar mistakes, all em-dashes are Ben’s.

For those of you who recently joined us or need a refresher: Brains is a research accelerator that helps talented scientists and technologists execute on ambitious research that is beyond the scope of individual academic labs, startups, or large companies.

If what they’re working on jumps out at you, please do get in touch with them via the internet (their profiles are linked in their names) or email brains@spec.tech and we’ll route you correctly.

Allegra Cohen

Allegra is building open-source technology for digital memory palaces. Humans have long used physical metaphor to interact with knowledge, from memory palaces to William Gibson's cyberspace. But today we increasingly access knowledge through AI agents, reducing our inclination to explore and increasing our risk of being manipulated. For example, as our digital records become ever more granular, we will quickly reach a point where agents not only remember more about us than we do but can gatekeep and gaslight us about our own thoughts. Correcting this imbalance requires representing knowledge in a medium our brains our built for. Technology to visualize and curate knowledge in the cyberphysical world could enable us to take advantage of AI agents while protecting our own agency and identities.

About Allegra

Allegra likes building technology for knowledge curation. She is currently the program director of Talk to the City, a large-scale qualitative data collection tool housed at the AI Objectives Institute. Previously, she was at DARPA where she managed a portfolio of research about how machines can help humans understand complicated geopolitical systems. Allegra holds a Ph.D. in computational modeling from the University of Florida and a B.S. in Symbolic Systems from Stanford University. She is a big fan of birds.

Andrey Poletayev

Andrey is unlocking a new type of manufacturing data by tracking the paths of atoms as they assemble into materials during syntheses, especially if they do so rapidly. To do so, Andrey builds upon methods pioneered by biologists and his work at SLAC National Lab: the x-ray camera fast enough to catch how a material comes together is much like those that capture the dynamics of proteins. Accessing the trajectories of phase transformations helps us to predict how syntheses proceed and design new ones, avoiding expensive trial and error in scale-up. This lets us make new materials for batteries, sensors, membranes, or pharmaceuticals, and improve existing ones.

About Andrey

Andrey is a battery materials chemist and x-ray scientist who has adopted simulation over the pandemic. He is an expert on how atoms move in batteries, with leading publications in Nature and Nature Materials, and now studies how we make materials: syntheses and phase transformations. Through his human-centered design work, he has started and supported social enterprises, including creating the business model for one delivering nutrition to over 10 million people daily in East Africa. Andrey is a Schmidt AI in Science Fellow at Imperial College having worked at Oxford and SLAC after a Ph.D. at Stanford.

Anubhav Sinha

Anubhav invents and deploys high-resolution imaging to systematically map how the peripheral nervous system (PNS) shapes disease processes across the body. By adapting technologies used for brain mapping in mice to the periphery, we will create “wiring diagrams” of the PNS that chart how neurons thread through the body with both cellular and anatomical precision to map how the structure and composition of the PNS changes in disease. Building such an atlas will unlock the ability to precisely dissect how the PNS impacts disease and will enable the discovery of new therapeutic targets to treat diverse conditions, including cancer, inflammatory disease, and aging.

About Anubhav

Anubhav is a neuroengineer and synthetic biologist who builds technologies to study the interplay between neurons and their spatial context. He is currently a postdoc at MIT where he is extending these technologies to the peripheral nervous system. During his PhD in the Harvard-MIT Program in Health Sciences and Technology, he co-led the development of targeted expansion sequencing, a technology for spatially precise measurements of gene expression in biological specimens, and deployed the platform to study the mouse brain, patient tumor biopsies, and mouse embryonic development.

Arila Atanassova-Barnes

Arila builds open turn-key solutions for renewable energy to reduce cost, establish local operating skills and enable peer-to-peer energy trading. We are stewards of programmable microgrids based on open standards to preserve energy choices without compromising safety. We host a network of autonomous microgrids to share data, energy, and communications management. We help communities to operate their energy assets and learn how to innovate, maintain and evolve their investment. Growing energy demands for AI alone de-prioretizes neighborhoods not only in the Global South. Energy IoT Open Source enables a model to share open microgrid technology and assure energy independence, resilience, and access for all.

About Arila

Arila is an experienced technologist and has lead engineering teams from startups to fortune 500 companies. My tenure of leadership roles in the energy industry has exposed me to various technology products: AI at the Grid Edge, Battery storage and EV Infrastructure, Asset Performance Management, Grid Analytics and Power Digital Twins and open standards like OCPP, OpenADR, and SunSpec to name a few. I founded Energy IoT Open Source (EIOT), a 501 (c)(3) non-profit with the mission to democratize microgrids for communities worldwide and have built strong partnerships with LF Energy, SunSpec and enAccess Foundation.

Ben Adler

Ben is designing next-generation antimicrobials by harnessing bacteriophages, viruses that infect bacteria. Unlike small molecule antibiotics, which depend on discovering new chemical classes, phage-based therapies are not fundamentally limited by available diversity. Instead, we are bottlenecked by our inability to systematically measure, interpret and engineer that diversity at scale. This program builds a phage design engine that integrates high-throughput phage screens, genome editing and comparative genomics to generate large-scale datasets that train AI models to predict optimized genome edits. By closing the loop between measurement, modeling and engineering, this platform enables precisely engineered antibacterial interventions that remain effective as bacterial resistance evolves.

About Ben

Ben discovers biological function encoded in the ancient evolutionary arms race between bacteria and their viruses, bacteriophages. As an integrative biologist, he bridges high-throughput screening, computational genomics and genetic design to reveal how evolution innovates, from mechanisms of anti-bacteriophage immunity to targeted elimination of human pathogens. He translates these discoveries into widely-adopted experimental platforms, including CRISPR-based systems for efficient phage genome editing. Benjamin received his Ph.D in Bioengineering at UC Berkeley & UCSF with Dr. Adam Arkin and his postdoc with gene-editing pioneer Dr. Jennifer Doudna.

Bharath Ramsundar

Bharath is working to connect AI more directly with the physical world. Today’s AI systems lack common sense due to a lack of understanding of basic physics. The scientific community has spent the last 75 years building sophisticated physical simulations for science and engineering. Unfortunately, these simulations can run for weeks or months on supercomputers, making them infeasible to integrate with modern AI. Bharath aims to transform simulations by leveraging machine learning and AI tools to increase their speed and accuracy. These accelerated simulations will directly integrate with AI training to create next generation AI systems that understand physical reality and avoid nonsensical hallucinations, thereby unlocking transformational applications in robotics and manufacturing.

About Bharath

Bharath is the founder and CEO of Deep Forest Sciences, which is building an AI-powered suite for drug and materials design and discovery. Bharath received a BA and BS from UC Berkeley in EECS and Mathematics and was valedictorian of his graduating class in mathematics. He did his PhD in computer science at Stanford University where he studied the application of deep-learning to problems in drug-discovery. At Stanford, Bharath created the DeepChem open-source project to grow the deep drug discovery open source community. Bharath's graduate education was supported by a Hertz Fellowship, the most selective graduate fellowship in the sciences.

Chris Wilmer

Chris is building a universal electronic nose (e-nose) so machines can truly smell. Today’s e-noses rely on only a few dozen sensing materials, limiting them to narrow tasks like detecting natural gas or food spoilage. As a result, they struggle with complex chemical mixtures and cannot serve as general-purpose smell systems. Our approach dramatically increases sensing resolution by integrating hundreds of chemically distinct materials onto a single chip — an order-of-magnitude leap beyond conventional designs. Our first milestone is demonstrating that performance continues improving at this scale, establishing the foundation for mass-manufacturable, general-purpose machine olfaction.

About Chris

Chris is a professor at the University of Pittsburgh specializing in the interactions between gases and surfaces. In particular his lab designs metal-organic frameworks (MOFs), which are like sponges for gases. For example, MOFs can be used to soak up methane for energy storage or to remove carbon dioxide from the atmosphere. Now Chris is using MOFs to create artificial olfactory receptors to make noses for robots. In addition to 60+ scientific publications, he has co-founded two companies (Numat & Aeronics) and was named to the Forbes 30-Under-30 in Energy. He is also a Blender enthusiast.

James Black

James is building Linnaeus, an AI-native screening platform that screens for hazardous DNA sequences using a suite of biological tools, aiming to mitigate the challenges to current biosecurity screening posed by advanced AI. Unlike other approaches, the Linnaeus agent dynamically adapts to detect engineered, novel, and AI-generated threats, providing risk scores and detailed reports based on biomolecular properties. The platform can analyse sequences & structures across DNA/protein synthesis, cloud labs, and pathogen detection. This will allow scaling of defensive capabilities with advances in AIxBio while reducing reliance on costly human review.

About James

James is an AI-biosecurity researcher studying risks from frontier AI systems applied to biology. He trained as a physician and computational biologist, and combines deep biological expertise with hands-on AI safety research. He has designed biosecurity evaluations for frontier AI developers, including work cited in the Claude 4 model card, has led work demonstrating that safeguards on genomic language models can be circumvented, and is aiming to innovate at the frontier of biosecurity screening technology.

Jiwoon Park

Jiwoon is building SAHA (Spatial Atlas of Human Anatomy), a global multi-institutional initiative to create the first comprehensive spatial maps of human tissue, charting how billions of cells are physically arranged across major organs. Her program uses AI to detect patterns in tissue architecture that predict disease outcomes, such as where a cancer will metastasize or which patients will respond to immunotherapy. By turning these atlases into clinical tools, SAHA will give physicians a fundamentally new layer of information to make better treatment decisions, one that current genetic tests and imaging cannot provide.

About Jiwoon

Jiwoon is a researcher at Harvard Medical School and Weill Cornell Medicine translating spatial omics into clinical impact through AI. She leads the Spatial Atlas of Human Anatomy (SAHA) consortium, creating the world’s largest spatial reference of human tissues by mapping billions of cells from large clinical cohorts. Her research integrates experimental and computational approaches to decode how tissues are organized and how immune systems respond to disease, with applications in cancer metastasis and precision medicine. Dr. Park holds a Ph.D. from Cornell University and contributes to major spatial biology initiatives including the Human Cell Atlas, GESTALT, and STOC.

Justin Burrell

Justin is building a scalable engineered tissue platform to restore function after neurological injury or disease. Our focus is transforming neural repair from bespoke laboratory procedures into reproducible biomedical infrastructure by automating manufacturing, implementing non-destructive imaging and electrophysiologic quality assurance, and embedding functional sensing to quantify circuit performance in real time. The underlying biology has been demonstrated in clinically relevant models. The inflection point now is industrialization. Achieving standardized, performance-validated tissue at scale requires coordinated investment in automation, quality systems, and translational infrastructure beyond traditional academic capabilities. By defining and enforcing functional benchmarks, this platform enables more predictive testing systems, accelerates therapeutic development, and establishes the foundation for restorative biohybrid neural interfaces that rebuild durable communication between the brain and body rather than compensate for loss.

About Justin

Justin is a translational neuroengineer focused on restoring damaged neural circuits through engineered neural tissue and advancing living neural technologies toward clinical deployment. Over the past decade, he has engineered and transplanted organized neural tissues at clinically relevant lengths capable of transmitting signals across lesion gaps and electrically integrating with target tissues, improving functional recovery in rodent and porcine models of neurological injury and disease. His work emphasizes reconstruction of conduction pathways, restoration of long-distance signal transmission, and development of quantitative functional benchmarks for circuit performance. He is a Research Associate in Neurosurgery at the University of Pennsylvania and holds an MS in Neuroscience and a PhD in Bioengineering.

Leo McElroy

Leo is building the core software that engineering design tools are built on: constraint solvers, geometry kernels, and the connections between design and physical simulation. Today this technology is proprietary and fragmented. Designers make shapes in one tool and test whether they work in a completely separate one. If design and simulation share a foundation, software can verify that a part works as it’s designed, and that loop can be automated. This opens the door to designs that are optimized by physics and guided by the designer’s intent, and AI that can generate engineering models a human can edit and a computer can verify.

About Leo

Leo is an independent researcher developing technology for designing objects on computers. He has built novel constraint solving algorithms and consults with companies developing CAD tools. He has created dozens of custom design programs for objects ranging from circuits to embroidery. At Hack Club he designed and built open source hardware and software for making games, music, and art used by tens of thousands of teenagers worldwide. He has published research on digital fabrication and traveled the world working with inventors and craftspeople as a Watson Fellow.

Paul Litvak

Paul is building an AI system that can weigh the balance of scientific evidence the way a careful, critical scientist would: flagging numbers that don’t add up, assessing methodology across entire bodies of research. People make health decisions based on published studies. Billions in philanthropic dollars go to whichever interventions the evidence supports. Peer review is overextended, the volume of papers keeps growing, and there’s no scalable way to know which findings hold up. Getting this right means the people and organizations making these high-stakes decisions can actually know what the evidence supports and make better ones because of it.

About Paul

Paul is the founder and Executive Director of the Robyn Dawes Institute. He has a PhD in Behavioral Decision Research from Carnegie Mellon and worked as a data scientist and product manager leading machine learning teams at Meta, Google, and Airbnb. He co-founded and led product at Rhythmic Health, a venture-backed biosensing startup. He’s currently a Visiting Scholar at UC Berkeley, creating tools for the research community to assess evidence quality.

Smrithi Sunil

Smrithi is building tools to study biomolecular structure and function at atomic resolution inside intact cells and tissues. Today’s microscopy tools capture atomic structure only in static states or function only in isolated pathways, missing how the complex molecular machinery organizes and interacts in its native cellular context. Smrithi’s program will integrate high-resolution microscopy techniques to enable comprehensive mapping of molecular structure and function inside intact cells and tissues. The result is a rich database of cellular structure and dynamics that will help researchers understand disease mechanisms, guide drug target discovery, and train predictive models of how cells actually work.

About Smrithi

Smrithi is a research scientist developing imaging techniques to study brain function. She is currently at the University of Wisconsin-Madison, advancing electron microscopy methods to visualize disease-related proteins inside intact cells and tissues. She received her PhD in Biomedical Engineering from Boston University, where she developed optical techniques to study blood flow and neural activity during stroke recovery. She was then a scientist at the Allen Institute for Neural Dynamics, developing microscopy techniques for multiplexed imaging of neural signals. She is also interested in metascience and the process of discovery, and writes about these topics on her Substack, Engineering Discovery.

Thomas Teisberg

Thomas is building a robotic observatory for Antarctica and Greenland to reduce uncertainty in future sea level rise and monitor for early warnings of catastrophic tipping points. Earth’s ice sheets are data poor and logistically complicated to access, making it expensive to scale up data collection with humans. Robotic platforms can cut the costs of data collection at scale, but achieving this scale is beyond the reach of convention research entities. Thomas plans to take new observational approaches from prototype to production and operate them at scale. The first objective is a stratospheric radar sounder, capable of delivering twice the peak annual data volume collected previously at less than a fifth the cost.

About Thomas

Thomas is an electrical engineer who uses new sensors and robotics to better understand how the Earth works. He co-developed an open-source ice-penetrating radar platform and used it to deploy a UAV-borne radar sounder in Greenland, Iceland, and Svalbard. He has also worked on making observational data freely and reproducibly available, developing tools to analyze and share hundreds of terabytes of field data ranging from the first airborne surveys of Antarctica to data collected this year. Thomas has a Ph.D. in electrical engineering from a geophysics lab at Stanford University.

Victoria Chernow

Victoria aims to transform mineral processing by developing technologies for “rock fractionation”: processes to extract and refine every compound in ore and tailings economically, slashing energy consumption and waste. Conventional mining is materially and energy inefficient, largely focused on single products that generate tens to hundreds of tons of waste per ton of metal. Left behind are valuable co-products -- from cementitious materials to battery metals to minerals for permanent CO₂ storage. She is building an integrated toolkit to unlock this full spectrum of end products, uniting electrochemistry, advanced separations, plasma processing and more. Success means cutting energy and emissions, turning waste into assets, and securing mineral supplies without excess extraction.

About Victoria

Victoria is a materials scientist who identifies the cost drivers and bottlenecks that keep high-impact climate technologies from being deployed at scale. She is a Climate & Energy Scientist at Science for America, a nonprofit solutions incubator, where she explores ways to turn captured carbon dioxide into valuable, scalable products. Previously, she was an early-stage deep tech investor and a Fellow at the U.S. Department of Energy’s Advanced Research Projects Agency-Energy (ARPA-E), shaping programs in advanced building materials and next-generation nuclear energy. She holds an M.S. and Ph.D. in Materials Science from Caltech.

Will Plishker

Will is developing a program that gives surgeons vision and skills beyond their innate human ability. Surgery is the gold standard for many treatments but has the potential for serious complications. Minimally invasive surgery reduces risk and recovery time at the cost of direct sight and touch for surgeons. Autonomous surgical robots could help, but are large, expensive, and slow. This program overcomes these limitations by replacing full autonomy with a surgeon’s existing skills and intuitions, and further improves them with AI-enhanced images and simplified handheld robots. It will enable every future surgeon to be as fast and accurate as the best surgeons today, improving patient outcomes and recovery.

About Will

Will is a medical imaging entrepreneur and technologist who bridges high-performance computing and clinical care. As co-founder of IGI Technologies and ACREW Imaging, he has led the development of multiple FDA-cleared medical imaging products and deployed them in clinical workflows. With over 15 years of experience, he has created high-performance image enhancement systems for radiology, radiation oncology, interventional procedures, and vascular and cardiac surgery. Will holds a Ph.D. from UC Berkeley and completed postdoctoral work at the University of Maryland in medical image analysis. He was a Mayfield Fellow and NIH Commercialization Acceleration Program graduate.

The problem isn’t producing proteins at scale, it’s how we turn those proteins into fibers. Spiders run silk solution through precisely tuned microfluidic channels where pH, shear forces, and flow rates change along the length, creating fibers by triggering molecular-scale assembly that we can observe but can’t yet reproduce.

“If you look at spider silk, it’s not just one thing that’s impressive,” Tim McGee explains. “It’s really strong, it’s really tough, you can spin it really fast—having multiple performance characteristics like that is rare in the material world. Our existing technology can’t do it. But spiders do it all the time.”

The gap between possibility and reality isn’t around spider silk, but the whole class of protein-based fibers. Tim, a fellow in Speculative Technologies’ Brains Research Accelerator and now leader of the Impossible Fibers Lab at the Astera Institute, is trying to close that gap. His hypothesis is that figuring out how to translate lessons from how biology manufactures fibers into industrial practice could unlock incredible materials for fabrics, robotics, and beyond.

Twenty Years Circling the Problem

Initially motivated to study biology by the incredible things nature can create, Tim spent two decades working to apply the lessons of biology to how we make things. In all of academic biology, design, and consulting he collided with the same structural gaps.

Working with apparel and packaging companies, he hoped to unlock new technologies where they would be immediately used in products. “The apparel companies I worked with were looking at ‘what’s the future of fibers?’—but they didn’t have any ability to impact what that future actually is. They take what manufacturers produce and do small tweaks around the edges. Fundamentally changing how we make fibers was not within their capability to fund or engage.”

Even emerging protein fiber startups weren’t incorporating scientific insights published fifteen years earlier. “The things we know about how spider silk is made, how mussel byssal threads are made—that knowledge isn’t going to be translated into technology unless somebody makes an explicit effort.”

That realization crystallized the problem. The science existed but the institutions didn’t exist with the resources and incentives to operationalize it. Where do you get funding for work that’s too applied for academic grants and too blue-sky for venture capital? “I was looking for specifically how to get funding for this kind of work,” Tim recalls. “And that’s when Speculative Technologies popped into my Twitter feed.”

Finding Brains

Tim saw that Speculative Technologies was looking for ambitious ideas in materials and manufacturing so he reached out. “I said, listen—we know how nature makes fibers. It’s different from how we make fibers. I think there’s something really interesting here.” One thing led to another and Tim joined the Brains Research Accelerator.

The program delivered two things he couldn’t get elsewhere. First, a forcing function for clarity: “Brains enables you to clarify an idea to the appropriate level where you can have better conversations with others. Until you can share it concretely enough, you’re just on your own with an idea.”

Second—and Tim emphasizes this—psychological legitimacy. “You get support, encouragement, cheerleaders who are genuinely excited about you changing the world. A lot of times when you start, it’s just an idea and you’re by yourself trying to figure things out. Brains gives you agency. It gives you the ability to think that this is possible, and to realize that these things can happen.”

From Presentation to Lab

At the end of the Brains program, fellows present their ideas to an audience of connectors and decision makers. For Tim, that presentation catalyzed everything that followed.

“That led to making connections with folks in the audience. They suggested I should be talking to other companies and groups doing related work.” Conversations multiplied. One thread led to a DARPA “fast grants” style competition and a contract for Tim and his team to do a short, focused investigation into training AI models on fiber formation.

“That was a great start—it let us continue the momentum from the initial presentation, build on the work, and keep the interest moving in the community.” The DARPA project wasn’t just funding; it was credibility and trajectory maintenance while Tim developed the longer-term vision.

From there, he applied to the Astera Institute’s residency program. The fit was natural: Astera provides infrastructure for exactly this kind of translational research. Tim earned a spot in the residency and with it, the resources to build an actual lab.

“It gives us a physical place where we can work on this fiber formation technology. Now we have a platform for further growth.”

The Impossible Fibers Lab Today

The lab is in Emeryville, California where Tim leads a four-person team.

“Day to day it’s being in the lab, breaking things, pushing fluids around, and looking through the microscope to see what we did.” They dissolve proteins in solution, vary flow conditions, and image the resulting fibers with microscopes and atomic force microscopy to understand nanoscale assembly.

The goal is learning to control fiber assembly from molecular to macroscopic scales. When asked what happens if they are wildly successful, Tim responds: “We change where materials are made, how they’re made, what people think is even possible. A whole shift in manufacturing materials out of proteins—sustainable, without the toxic plastics we have now.”

The applications span domains: neural interfaces connecting body to electronics, prosthetics that feel like native tissue, artificial tendons for robots with biological-like actuation, biomedical scaffolds bridging soft tissue to bone. Protein-based fibers could dominate anywhere you need fibers with multi-objective performance—strong and tough and processable and biocompatible.

Why Ambitious Research Fails—And What Changes That

When asked why ambitious research fails, Tim’s answer centers on agency: the combination of funding, runway, and freedom from premature accountability.

“The things that will be very impactful are often things people think aren’t possible—or think are against the laws of physics because they misunderstand some piece of the puzzle. The people who see those paths have to battle against that and have enough resources, time, capability, and networks to prove it’s a different path forward.”

That’s what Brains provided: The clarity to articulate the idea, the connections to find traction, and the psychological foundation to persist. Brains took Tim’s two decades of accumulated insight and gave it somewhere to go.

“There’s a willingness to sacrifice some amount of your career toward this,” Tim says. “You have to be a little bit committed.”

Tim has been committed for twenty years—watching the gap between biological manufacturing and industrial practice, ping-ponging between disciplines, waiting for the right moment and the right structure. The bet is that evolution figured out fiber manufacturing millennia ago. Now he finally has the lab, the team, and the runway to steal its secrets.

You can learn more about Impossible Fibers here. You can find Tim on LinkedIn and X/Twitter.

Note: This piece was largely written by Claude based on a recorded conversation with Tim (that will be released separately).

The National Science Foundation’s Directorate of Technology, Innovation, and Partnerships (NSF-TIP) recently issued a request for information on a new “Tech Labs” program. From the RFI:

“Tech Labs will support full-time research, development and innovation (RDI) teams focused on overcoming persistent barriers to the commercialization of emerging technologies. These teams will benefit from operational autonomy, milestone-based funding and the ability to engage across academia, industry, national laboratories, and nonprofit sectors.”

Getting this program right is critical for the future of ambitious research in the US. If it’s successful, other agencies will hopefully copy and riff on it. Like so many new approaches to research funding and management, Tech Labs could go one of two ways: it can unlock a whole world of ideas that wouldn’t have seen the light of day, or it can experience mean reversion and capture by incumbent institutions.

We submitted the following response:

Which types of teams and organizations should be considered eligible to apply for the NSF Tech Labs program? What restrictions on team eligibility should be in place to maximize speed and ensure novel impact?

First, let’s enumerate the thing that would blunt tech labs speed and novel impact:

• Employees, especially leadership, who are simultaneously employed at other organizations after

  phase 0. This is especially true for professors and grad students. A fuzzier line is a situation where the leadership has some explicit agreement that they have a job waiting for them when the tech lab ends, like a professor on leave. This latter situation is also hard to enforce. One common important-to-avoid problem is the professor who is officially on leave, but in practice is still supervising a lab’s worth of grad students “on the side”

• If the tech labs org was a subsidiary of a large organization that has the ability to dictate things like research priorities, how money is spent, and how IP is assigned.

• If the people in the tech lab were spending a lot of their time doing outside fundraising and mixing a lot of money that has different requirements. This is another tricky one: there are definitely ways to leverage private funding, but seeking it out and aligning it all in the same direction is a huge distraction (especially during the tech labs period).

• If the lab did not own its core IP (ie. it was licensing it from a university or company.)

• If the lab has already raised significant amounts of venture capital or federal funding. Of course, what is “significant” is an open question. This one is tricky, because there are many projects that are a good fit for the tech labs initiative that have raised venture capital to get started because it was the only option. At the same time, tech labs work should not just be a way of subsidizing the venture capital industry.

In order to avoid these traps, we recommend the following restrictions:

By the end of phase 0, tech labs need to be incorporated as an independent organization with its own board of directors, and staffed by a full-time leadership team who have left their previous roles along with a majority of employees. The organization should have received less than $2M in grants and $1M in VC funding. Its IP should not create entangled relationships with external organizations.

These suggested funding limits and independence requirements are extreme. They would rule out several organizations that absolutely should get tech labs grants. For example, Convergent Research’s FROs are structured as fully-owned LLC subsidiaries. Similarly, many great organizations that would be great uses of Tech Labs money have received significantly more than $2M in grants – places like FROs, Otherlab, and Blueprint Biosecurity to name a few great orgs. Speculative Technologies ourselves would be ineligible!

The reason for these perhaps overly-restrictive requirements is that there is no other way to write the rules in order to prevent a reversion-to-the-mean/Matthew effect scenario that I worry is quite likely. If subsidiary organizations are legitimate, what’s to prevent somewhere like the Broad Institute or a university from spinning up a subsidiary subject to the same bureaucracy and incentives that (I hope) Tech Labs is meant to bypass? If they’re eligible, legacy institutions look much better on paper (track record! Institutional capabilities! brand!) and it would be hard for decision-makers in the NSF to justify overlooking them for newer, scrappier organizations. The same argument goes for funding limits. There is no rule that can distinguish between a well-funded FRO that nevertheless could do so much more and an SBIR-farm that turns government money into reports.

Some things that tech labs should be able to do:

• Hire some part-time employees (who could be grad students or postdocs) as long as they are a small fraction of the total employees.

• Use core facilities or rent space in a university, as long as they retain organizational independence and have the arrangement figured out by the end of phase 1.

• Hire contract research organizations, vendors, or subcontractors to do some of the work, up to 40% of the budget.

• Work with a fiscal sponsor.

Is the proposed timeline for Phase 0 (9 months), Phase 1 (24 months), and Phase 2 (24+ months) well-calibrated to support the program’s strategic objective of achieving high impact, accelerated outcomes? If not, what adjustments should be made and why?

The timescale seems well calibrated, assuming there is a way of calibrating the milestones at the end of Phase 0 for where the team was at the beginning of the effort. That is, there should be different expectations for a team that started off fully part time without facilities and one that was already working full time with a lab.

One potential problem with the proposed timeline is that it will be incredibly hard for phase 0 tech labs to start hiring technical and operational staff before they have phase 1 funding committed; otherwise they might all be laid off within a few months!

How should IP rights be structured to support maximum success and impact?

Any IP created during the tech lab should be owned by the tech lab entity, but the entity should be required to license that IP via a standard, non-exclusive license to any American entity. This approach balances incentivizing the creation and commercialization of IP and public good from taxpayer money. This approach to IP was one reason why Bell Labs was able to be so impactful. Anything outside their core telecom scope had to be licensed with generous terms, which allowed several different entities to do follow-on innovation on things like the transistor, cell phones, solar cells, and unix. More generally, impactful system-level innovations require the mixing of many pieces of IP, so it is important to incentivize sharing.

There are some domains where commercialization may be impossible with non-exclusive licenses. The current Tech Labs effort has explicitly said that they will not be funding therapeutics, the primary domain with this situation, but there may be others and hopefully the concept expands to other agencies that do deal with therapeutics. (It’s also worth remembering that health technology is weird.) In these situations, IP should be exclusive only within a particular field of use and the burden of proof should be strong that exclusive licenses are absolutely necessary for translation.

Tech labs should also encourage a culture of openness and trying to benefit an entire ecosystem, not secrecy and value-maximizing. The latter would be more likely if there was a potential lucrative IP prize waiting for successful tech labs. From a more hard-nosed angle, one of the reasons the Chinese hardware ecosystem has been able to flourish is because their healthy disrespect for IP enables a lot of tinkering by many parties and drives down prices through competition. The US should not loosen IP law, but tech labs should strive to replicate those productive conditions in an American context.

A tricky but likely common situation will occur when a tech lab is built around already-existing IP: a team spinning out of a university, for example. As table stakes, all licensing agreements need to be settled by the end of phase 0 (and ideally beforehand), otherwise things become incredibly messy. If at all possible, those licensing agreements should give the tech lab org ownership of any derivative IP and not put any conditions on its use (so that it can be non-exclusively licensed as mentioned above).

What degree of independence is optimal to ensure the flexibility, freedom, and speed required for the Tech Labs initiative? How should NSF define team independence?

Let’s flag two kinds of independence: team independence and organizational independence. Team independence is independence with respect to NSF TIP as specified in the funding agreement; organizational independence is independence with respect to all other entities and contracts.

Some prerequisites for the flexibility, freedom, and speed for Tech labs to maximize impact:

• The ability to spend money as quickly and flexibly as possible. Obviously there are federal rules on how money can be spent and the money needs to be audited at some point, but there should be explicit effort to make the requirements as loose as possible.

• The ability to shift technical approach as long as the concrete goal remains the same.

• A minimum amount of written reports. One way to minimize reports is to encourage and accept capability demonstrations as acceptance criteria for milestones.

• Teams should be able to set the milestones that they will be judged on (potentially after iteration with the tech labs). There should be a mechanism to renegotiate these milestones if the team heavily pivots their approach.

• Minimize committees. Committees, especially committees of experts with no skin in the game, select for consensus good things and drive results towards the mean.

There’s an open question about how much independence teams should have over the goal/output of the tech labs effort. In theory, teams should be able to pivot towards a new goal if they find a more promising one during the course of the effort; however this could also lead to a lot of wasted money and floundering if the team finds that their original idea is no good and doesn’t want to shut down the org. It’s impossible to entirely prevent this possibility but ironically, the solution might be to have more, lower-stakes touchpoints with a program officer who can get a sense of whether the team is doing good work or not.

How should funding be allotted to each proposed Tech Labs? What factors – for example: team size, team expertise, infrastructure needs, growth trajectory – should NSF consider to determine appropriate funding amounts to support successful Tech Labs teams?

The NSF should not consider team size or expertise. Teams should have the incentive to be as efficient as possible: teams should be rewarded, not punished for figuring out how to be productive with a few, relatively junior people.

Funding should actually be allocated as team-agnostically as possible. Budgeting based on team need will incentivize teams to either pad out their budget to the maximum allowed if budget isn’t part of the funding decision or skew their budgeting needs too low if it is.

However, some work is more or less expensive than others. Animal studies are far more expensive than pure software work that doesn’t involve training AI models, for example. To address this, NSF should have a simple rubric that consists of work-requirements-based multipliers that get applied to a base budget. As a made up example, say that animal studies are a 1.2x multiplier, training AI is a 1.1x multiplier, and custom hardware is a 1.1x multiplier. Then a project that is proposing to build a custom piece of hardware to take measurements on an animal model and train a foundation model would be allocated 1.2x1.1x1.1xbase budget = 1.45 x the base budget.

What opportunities do you see for synergy with research and development efforts that are or could be funded by industry or philanthropic organizations? What partnership structure would allow Tech Labs to leverage federal and private support for maximum benefit?

Synergy between private and public capital sounds great, but can often lead to traps or perverse incentives in which winners win more and teams that are good at politicking and sales are rewarded over those with the most technical capability. There are potentially good ways for private and public capital to synergize in the tech labs efforts but it needs to be carefully thought through.

Some particular things to avoid:

• Requiring private matching in phase 0 or phase 1 or giving an advantage to organizations that already have funding lined up or in the bank.

• Allowing philanthropic or industry organizations to weigh in on which teams are funded.

One way to leverage philanthropic and industry capital in a way that avoids several downsides is to require or encourage matching funding for phase two. This would give teams the freedom to hone in on a direction and show results before needing to fundraise and ‘answer to multiple masters.’ It would potentially have the advantage of providing a “forcing function” towards a specific application or transition path once the system is at a point where that is beneficial.

Teams should certainly be encouraged to coordinate closely with industrial R&D teams in order to make sure that their work is tuned for real use cases, but requiring those relationships to involve matching funding could unnecessarily weed out good teams and cause a lot of trouble around IP and transitioning in a way that maximizes impact.

An implicit synergy between tech labs and philanthropic organizations could emerge over time if tech labs became a regular, common thing across the government: philanthropic efforts could “seed” projects that become well-situated for a tech labs effort. Today, philanthropic capital is often hesitant to fund ambitious independent research projects because of sustainability questions – they don’t want to be on the hook for the entire cost of the effort but also don’t want to partially fund something that will die without their support.

What translational problems, challenges and/or bottlenecks could be addressed within 3-7 years with this program design? Answers can be broad or specific.

Some potential problems

• Building a new manufacturing paradigm that would make producing physical products much more like creating software – easily reconfigurable, rapidly scalable, and with tight feedback loops between inputs and outputs. Creating the “chatGPT” moment for this new approach would require figuring out how to integrate into a single system robotics, ai-driven feedback loops, and newly available tools like advanced magnets, lasers, or software-driven machines.

• Creating the design and scalable manufacturing tools for protein-based fibers. If we were able to design the analogues of spider silk and manufacture them at scale, it could unlock new possibilities for robotics, microelectronics, and textiles.

One thing to flag that tech labs should avoid doing is being too prescriptive about specific challenges that will be considered in the upcoming call. The ideas that will enable the most novel impact from the tech labs will not fit neatly into tight topical buckets like “new metal processing” or “advanced solar cell manufacturing.” We strongly recommend broad topical areas like “manufacturing” “biomanufacturing” “energy” and allow teams to surprise you with the challenges they think they can tackle within them.

Miscellaneous Feedback

Some things to flag that were not explicitly asked about in the RFI:

Spinning up and running a new org involves a lot of hard work beyond the core technical aspects – especially ones like tech labs will support that do not follow any standard template. Teams will need a lot of help to hit the ground running. This includes both operational pieces like setting up all the functions of an org – HR, accounting, etc – and strategic/coaching help on how to manage a fast-paced independent org, plan and roadmap multiple coordinated technical thrusts, and transition technology in order to maximize the impact of the work. In order to maximize teams’ chances of success the tech labs initiative should curate and vet some set of core services. This set of core services could include outsourced HR/accounting and legal (like that provided by Convergent Research to their FROs), and coaching+strategic planning like the Brains research accelerator.

It’s likely that this call won’t fully address the “cold-start” problem. That is, if the program selects for teams that look like they have the best chance of getting the farthest during the tech lab, it is going to heavily favor pre-existing teams that have already gotten funding in one way or another to do as much work as possible. That precondition already puts a constraint on the novel impact that the tech labs can enable because there will be a selection effect for work that already looked promising to a more traditional funding source. It’s likely out of the scope of this initial effort, but in order to maximize the value of the tech labs program, TIP should also look at funding organizations that are able to start ambitious efforts “from scratch” to get them to a place where they would be a good fit for a full tech labs proposal. Another approach to this “cold-start” problem would be to accept a few earlier-stage teams into phase 0 based on potential and then judge them on whether they can make shocking amounts of progress in a short amount of time compared to other more mature teams.

It’s important to make successful phase transitions as deterministic as possible. That is, teams should have a good sense of what “good enough” to make it to the next phase looks like and feel like the decision is not arbitrary or out of their hands. Ideally, it would be as simple as “if you hit the aggressive milestones that you set, you’ll make it to the next stage.” The proposed budgets and number of teams in phase 0 make that ideal scenario unlikely, but it can still stand as a thing to get as close to as possible. This determinism is important to incentivize teams to genuinely put it all on the table. If it feels like a team can go all out, crush their goals, and still have their funding cut then they will spend a lot of their time and effort making backup plans and going on side-quests to prepare for that contingency instead of focusing on the primary goal.

Double Down:

• Governments run into fundamental tensions around ambitious research.

• Materials and manufacturing are an incredibly impactful place to focus for new institutional models.

• Exclusively working with external performers in the 21st century is severely limiting.

• Universities need to be unbundled.

• As an organization, we need to figure out how to use AI well.

Updated/Wrong

• Coordinated research programs are now bottom-of-the-funnel limited.

New

• A lot of technology work makes sense neither for venture funding nor philanthropy.

• There’s a specific new manufacturing paradigm we should focus on.

• We need to build a physical home for misfits and ambitious researchers.

• It is so much easier to make progress on ambitious research when you are already doing things.

• There’s a reason people rarely want to fund the high-risk part of high-risk high-reward research — successful projects need to right-side upside adjusted risk.

• Deadlines are incredibly powerful.

2025 was a weird year in the research world

There was a lot of volatility where the research world touched the government: suspended or revoked grants to universities, hiring freezes and personnel turnover at the ARPAs and other research agencies, and immigration status uncertainty for the many non-citizen scientists in America.

Many new initiatives happened as well. In the US, the Genesis Mission was announced, NSF released an RFI about a new “Tech Labs” initiative, and OSTP explicitly called for new ideas in how science is funding. Internationally, ARIA launched a “FRO founder residency” and Japan launched their Global Startup Campus.

In the private sector, “AI for science” had a big moment. While what “AI for science” means is still evolving, right now it is particular focused on drug and materials discovery. Periodic Labs, Lila Sciences, and Jeff Bezos’ Project Prometheus raised huge amounts of money among others. Basically every big AI lab announced that they had some sort of science initiative.

Double Down

Governments run into fundamental tensions around ambitious research. This year illustrated why one of our big theses is that we (and other ambitious research orgs) need to depend primarily on private funding. People regularly ask the (reasonable!) question “isn’t this the sort of work the government should fund? Can’t the government deploy far more money towards pre-commercial research than the private sector?” But two things are true simultaneously:

1. The officials in democratic governments need to justify spending to voters

2. A lot of the work that goes into ambitious research is hard to justify until long after the fact, if ever. (And most officials and voters do not viscerally understand this through no fault of their own.) We saw this dynamic play out on the national stage this year. I’m not going into the politics of the whole thing but there was a lot of research funding that a chunk of officials and voters did not think was justified and it was cut off midstream.

Materials and manufacturing are an incredibly impactful place to focus for new institutional models. The past year, materials and manufacturing were really in the spotlight both because of geopolitics and AI. The “impactful” piece is much more consensus than it was a year ago. But most of the work is being done in the form factor of old institutional models: startups, accelerators, and academic labs. Particularly, the timescales and system-level work that needs to happen raise questions whether this work will deliver on its promise. Time will tell.

Exclusively working with external performers in the 21st century is severely limiting. I’ll point at updates from 2023 and 2024 for most of the reasoning but as additional evidence, the Impossible Fibers program actually built a lab (thanks to the Astera Residency!) and both immediately started learning far more and executing far faster than when it was working to coordinate academics and startups.

Universities need to be unbundled. We first hypothesized this before all the trouble the universities ran into in 2025. Doubling down is not meant to kick universities while they’re in trouble, but the events of the past year illustrate the case for unbundling. The government used cuts to scientific grants as leverage to get concessions around issues that largely had nothing to do with technical research. The financials of many institutions are under threat because of fewer international grad students. Putting so many societal functions in one basket makes them all mutually vulnerable.

As an organization, we need to figure out how to use AI well. This is perhaps more true going into 2026 than it was going into 2025. “AI for science” was abuzz in both the private and public spheres. It’s a far bigger topic than any one organization (and frankly a lot of it is still navel-gazing) so, while we did dip our toes into the bigger discussion, in 2026 we’re going to focus on our sphere of influence and competence. Specifically, we’re asking two questions:

1. How can we use AI at Speculative Technologies to improve the work we do both creating technology and helping people start programs to do so?

2. How can people creating and running ambitious research programs leverage AI more effectively?

We suspect that models with the right context could give feedback that, while it isn’t a substitute for an experienced mentor, can give good solid feedback in a tight feedback loop. We’re starting to take advantage of coding tools to build systems and tools for gathering information and sharing it with the people we work with. I’m hoping to make building custom tools to become a knee-jerk instinct for everybody in the org, regardless of experience. In addition to helping people starting and running programs in obvious ways like finding the right people to talk to, there are surely many useful ways to leverage AI for doing technical analyses and detailed technology roadmaps.

Updated

Coordinated research programs are now bottom-of-the-funnel limited. That is, the bottleneck to more ambitious programs (especially those run by people who don’t look like traditional research leaders) is the number of people and organizations who want to support/fund them. This wasn’t necessarily the case a few years ago when the concept of an FRO was brand-new and running an ARPA-like program wasn’t on the radar of most people outside of researchers who had previously interacted with DARPA or its kindred organizations. Now that “ARPAs” and “FROs” are much more in the public consciousness, many more great people want to start programs than there are homes or support for them.

New

A lot of technology work makes sense neither for venture capital /nor/ philanthropy.

Today, most private capital for ambitious science and technology efforts falls into one of two categories: philanthropic funding, which supports work with no expectation of financial return, and venture capital, which demands high returns on relatively short timelines. However, a big chunk of transformative science and technology is a poor fit for both buckets. These efforts may ultimately generate commercial value, but the path to doing that doesn’t look like a team building a clear product for a long time; often it’s not even the original team that creates the commercial value. This puts that work in a gray area: it’s both too “commercial” for traditional philanthropy, and not commercial enough for venture capital. This misalignment creates a structural funding gap—one that leaves potentially world-changing ideas without the catalytic support they need to move from lab to real-world impact. The challenge is compounded by high uncertainty in both expected impact and financial return, making it even harder for traditional capital sources to engage.

You can see this gap if you map technoscientific work onto two axes — one for expected returns and one for expected awesomeness/impact. This gap is compounded by the fact that a lot of work has high uncertainty around even its expected returns and impact.

So, the question is — how can we change the world to support science and technology work in this “missing middle?” We’ve realized we need to focus more on that question.

There’s a new manufacturing paradigm to be created around tooling-free, rapidly reconfigurable cells. From day one, we’ve identified materials and manufacturing as one of the most impactful places to do technology research. Going into 2026, we’re especially interested in the pieces of a potential new paradigm that leverages new tools and physics. You can see the sketches of what that paradigm might look like here.

We need to build a physical home for misfits and ambitious researchers. This hypothesis is perhaps the logical extension of working with external performers being limited and where work happens being important. In the spirit of intellectual honesty, this is exactly what I called out in 2020. What I didn’t understand at the time is why there’s a strong attractor towards internalizing work, especially with modest resources in the 21st century.

“The dominant paradigm of starting an org is to do internalized R&D unless you’re a government or charitable foundation. In reality, R&D orgs lie on a spectrum between externalized and internalized, with DARPA on one end and Lockheed Skunkworks on the other. Externalized research entails working through grants and contracts while internalized research is doing everything in house. There’s nothing inherently wrong with building a Skunkworks – it just means that there are different tradeoffs and the statement “DARPA for X” is misleading.”

It is so much easier to make progress on ambitious research when you are already doing things. A pattern we’ve seen over and over now is that once people doing ambitious researchy things, there’s an inflection point where opportunities just start opening up. This makes sense: it demonstrates that you and a team can execute and showing real outputs makes an idea concrete in a way that storytelling never can. Somehow people are far better at imagining the possibilities for a technology and organizational capabilities when they can see even pieces of it in reality. You can’t start arbitrarily small: you need enough resources to get over some amount of “activation energy” and resolve some amount of uncertainty that varies on an idea-by-idea basis. But that amount is often less than people think. It’s tempting to let the perfect stand in the way of the good — “I need $50M or bust. The problem is so complicated that there are no intermediate milestones.” This year we’re going to lean hard on the question of “what is actually the smallest unit of real work you can do to make the thing real?”

There’s a reason people rarely want to fund the high-risk part of high-risk high-reward research — successful projects need to right-side upside adjusted risk.

We see several patterns over and over:

• A nonprofit or government organization announces an ambitious, very expensive research project or organization to do “high risk, high reward work”. Heavily credentialed leadership is brought in, committees of experts are given meticulous oversight, and/or every decision needs to be run by the principal. Things move slowly, maximally conservative decisions are made, and approaches are slowly nudged towards established “best practices”. “Good” work happens, but nothing great or notable.

• Coordinated research programs that want to do something that is risky and new but requires a huge chunk of capital go unfunded, while startups that are doing things that are even newer and riskier raise far more money.

• Successful efforts to do ambitious things in a new way find ways to start doing work with a relatively small amount of resources.

The synthesis here is an idea that I’ve termed the distinctly uncatchy “right-sizing upside-adjusted risk.” Basically, the more money that gets deployed in one go, the more people need to see one of two things: lower risk, or more upside to them personally. The reality is that most work without a financial upside does not have huge potential upside for the people involved in funding them. The program manager who made the DARPA grand challenge happen (Ron Kurjanowicz) is not a household name nor incredibly wealthy (I assume based on base rates and no information otherwise — I don’t know his personal finances nor is it any of my business). Funding the plant breeding research that has arguably saved millions of lives or supporting molecular biology as a field are merely footnotes in John Rockefeller’s biography, and most of the impact happened after he died.

Researchers also evaluate work based on upside-adjusted risk, though their curve is almost the inverse of a funders’. Quitting your job to work on a startup that has only raised $1M is a different proposition than quitting your job to work on an FRO that has raised $1M for two reasons: there’s both a possibility that you can become a millionaire if the startup is successful and the ecosystem for follow-on FRO funding is less developed. That is, the upside is lower and the risk is higher.

If we chart the upside-adjusted risk for both funders and researchers, they have roughly the opposite shapes. Right-sizing is the nontrivial task of crafting a point where they intersect.

This year we’re going to both try to help people figure out this right-sizing problem and try to nudge the world towards systems that make starting small more feasible, like committed tranches.

Deadlines are incredibly powerful. There’s a cultural sense that research should be free of constraints; that ideas will come when they come. Perhaps stemming from and embodied by Vannevar Bush’s quote that “Scientific progress on a broad front results from the free play of free intellects, working on subjects of their own choice, in the manner dictated by their curiosity for exploration of the unknown.” But the reality we’ve found is that external deadlines help even focused, ambitious researchers. The old adage that “work is like a gas — it expands to fill the time given to it” is doubly true in high-uncertainty situations with infinitely deep rabbit holes to go down. Many Brains fellows have told us that the most useful part of the program were the deadlines. While I think that’s true, deadlines are the tip of an iceberg: you need to take them seriously, which means that you need to believe that the goal the deadline is for is useful and trust the entity imposing the deadline. Setting up those conditions is hard. In 2026 Spectech is going to lean hard into the power of the deadline, both internally and externally.

Here’s to a 2026 full of ambition, outputs, and learning!

Manufacturing is like a sewer. We all know it’s important, but we don’t really want to think about it until something goes wrong.

Well, something is going wrong.

If you’re reading this, it means you’re not living under a rock (or at least that your rock-covered home has internet) and therefore you’ve heard some noise about manufacturing. “Reindustrialization”, “onshoring”, “advanced manufacturing”, and other terms are everywhere. Considering that materials and manufacturing underpin civilization, it’s great that such an important and often ignored area is getting the air time it deserves!

But the way that most people are thinking about manufacturing — particularly how to manufacture more things in the US, why that is important, and more generally how centers of manufacturing gravity shift — is incomplete or just plain wrong.

I’m going to unpack a different way to think about manufacturing and how it changes over time: the idea of a manufacturing paradigm.

We’ll dig into the history of manufacturing paradigms, how they shift centers of manufacturing, and explore the fuzzy contours of the new paradigm we might be able to create. Creating this new paradigm is important not just for those of you focused on the US; it’s critical for unlocking human progress both on earth and across the stars.

Manufacturing matters

First, I want to establish some table stakes: Techno-industrial civilization is downstream of manufacturing technology. Regardless of what kind of progress you want to see in the world, you should care about manufacturing.

Do you want to fight disease, improve health, and extend lifespans? You need manufacturing advances. A manufacturing innovation turned penicillin into the life-saving juggernaut that has saved hundreds of millions of lives and arguably helped win WWII: before researchers figured out how to grow huge batches of penicillium rubens in a bioreactor it was a miraculous scientific breakthrough with a worldwide supply of a few hand-made doses. Today, many drugs and medical devices are secretly bottlenecked by their manufacturability.

Do you want to prevent climate change, save the environment, and make energy too cheap to meter? You need manufacturing advances. Manufacturing innovations turned solar panels from an obscure technology used only on satellites and toys into one of the cheapest sources of energy at scale. Useful fusion technology secretly hinges on our ability to manufacture easily-replaceable, low-activation material to jacket the inside of fusion reactors and divertor components that can survive sustained 10+ MW/m² heat flux without cracking or becoming brittle.

Do you think all these problems can be solved by AI? You still need manufacturing advances! Repeated manufacturing innovations like the Planar Process are responsible for Moore’s law; we need them to continue in order to improve the physical substrate for AI.

More generally, the impact of any breakthrough is limited if you can’t make enough of it economically. Obviously “enough” and “economically” vary drastically — for LIGO, three gravitational detectors is enough and ~$500M is economical. But if you’re talking about something like a structural material, consumable, or widget, impact requires manufacturing hundreds of thousands of kilograms or units, often for cents per kg or unit.

I can’t count the times that I’ve reached out the author after reading a paper about an exciting breakthrough that can, for example, efficiently pull clean water out of thin air; I ask excitedly “what are the next steps? What would it take to get this thing out in the world so it can deliver on its promise?!” Only to have them give me a funny look and say “oh, we’ve moved on after we published – you could never make huge amounts of it because it would be too expensive/the process will never scale/etc.”

Manufacturing advances often stand between us and unlocking the future

For something so important, we don’t have great theories on how manufacturing progress happens. One framework for thinking about it is through paradigm shifts.

You’re probably familiar with the idea of paradigm shifts in science. Paradigms are the idea (originally from Thomas Kuhn) that there is a set of assumptions and ways of doing things that underpins all the work in a field, and that these foundations are occasionally overturned. The canonical example is the shift from an Earth-centric system of astronomy (where it’s assumed that all the planets and stars revolve around the Earth) to one where the solar system revolves around the sun and the sun is just one star among billions. Another example is the shift from Newtonian mechanics with absolute space and time to Einstein’s paradigm where space and time are observer-dependent.

We can turn this same paradigmatic lens to manufacturing.

Steam power is a textbook manufacturing paradigm shift. It had obvious effects like new machines, new workflows built around continuously spinning shafts and belts, and new supply chains to create and supply fuel, supplies, and machines. But it also created second-order effects like reorganizing industry and society around factories and manufacturing centers; previously manufacturing was far more distributed, happening in people’s homes and small workshops. Steam power didn’t just make existing products cheaper – it unlocked previously impossible capabilities: rolled steel at scale, precision-machined parts, and continuously processed chemicals.

Interchangeable parts were another major paradigm shift. Instead of each part in a gun or sewing machine being hand-filed to fit with each other part in that specific device, precision machinery and fixtures enabled a system where a part from one device could be used in any other device. Like steam, interchangeable parts changed manufacturing equipment, ushering in higher-precision machines and jigs. Workflows shifted: instead of one person doing several production steps and making sure each part fit with the others, now you could have workers doing a single task. Even industrial organization changed: if you can guarantee that a part will fit in an assembly, you don’t need to make all the parts in the same place. More firms can specialize in creating specific components. Interchangeable parts also unlocked products whose complexity would have made them unmaintainable otherwise: from typewriters to bicycles to automobiles.

The current manufacturing paradigm doesn’t have a catchy name like “steam” or “interchangeable parts”, but I like to call it “network manufacturing.” It’s the system enabled by a combination of global containerized shipping, just-in-time manufacturing, and the ability to send information across the world instantaneously. “Designed in California, Made in China” is emblematic of network manufacturing, which enables deep decoupling between the design and manufacturing of a product. Under this paradigm, industrial organization is both global and tiered: cobalt mined in the DRC might be refined in China, processed in Japan, integrated into a chip in Taiwan that is assembled into a phone in India, all according to a design created in the US.

Paradigm Shifts Drive Progress

Not only is manufacturing critical for progress, but so are paradigm shifts in manufacturing.

The first-order effect of paradigm shifts is to decrease prices, which gives more people access to things that were once reserved only for the rich. Mirrors, cotton clothing, and steel cutlery were all once a luxury but are now functionally free. In the same way that new scientific paradigms move us towards a “truer” understanding of the universe, manufacturing paradigms drive the cost of goods towards the raw cost of their material inputs.

Beyond increasing access, decreasing costs can cause discrete “phase-changes” thanks to thresholding effects. Once LEDs became cheaper than incandescent bulbs, the world quickly switched over, drastically reducing the amount of energy used for lighting and further driving down the cost of LEDs which then could be used for many more applications. The same happened for aluminum, electric-motor-driven tools, digital cameras, solid-state hard drives, and thousands of other technologies.

But the biggest way that paradigm shifts drive progress has nothing to do with cost. It’s how they open up design space and enable us to create things that were impossible no matter how much money you threw at them.

Steam’s consistent power output made possible uniform rails, swung forging hammers far heavier than any person could move, and powered control systems for chemical reactors that could perform the same reactions consistently. These capabilities were, in turn, key enablers for the railroads that drastically improved lives by transporting people and goods, the chemical industry that helped feed the world, and the stainless steel used in so much medical technology.

Ultimately, the precision driven by interchangeable parts unlocked things like turbines, whose blades spin at 10000 RPM mere millimeters from their casing. Interchangeable parts made complex products dependable. Without interchangeable parts, if you didn’t personally know someone who could make a new component by hand (like the handle for a shovel) you were out of luck if something broke. Even if someone could make a car by hand, it would instantly become scrap when any single component broke. Electricity generation would be impossible without the turbines unlocked by interchangeable parts and so would the appliances that turn that electricity into saved labor and more comfortable lives.

(I realize I sound like a broken record about how manufacturing technologies’ second-order-effects affect everyday lives but most people have yet to internalize it, so the yelling will continue.)

Paradigm Lifecycles

We’ve established that manufacturing is important for progress, it progresses through paradigms, and that these paradigm shifts themselves are important. But how do manufacturing paradigm shifts actually happen?

We can think about manufacturing paradigm shifts as a close parallel to “Kuhnsian” scientific paradigm shifts.

1. Normal Science. Most of the time, science operates under some paradigm (say Earth-centric cosmology). People make new observations (like more precise measurements of Venus’ path through the sky) and explain them through the lens of that paradigm (“well, guess we need to add some more epicycles to explain the Venusian path, there must be some more revolving crystal spheres than we thought.”)

2. Model Drift. Over time, observations pile up that are hard to explain within the existing paradigm: how does Venus have phases if it’s always on the opposite side of the earth from the sun? How does Jupiter have moons? If stars are holes in a celestial sphere, why do they have different brightnesses and sizes? Why are there so many freaking epicycles???

3. Model Crisis. Eventually, people like Copernicus and Galileo realize that perhaps the whole thing is wrong or incomplete. (Spoiler alert: this may be where we are in manufacturing right now.)

4. Model Revolution. Going beyond just pointing out the flaws in the old paradigm, innovators propose a new paradigm, like the heliocentric model of the solar system. Massive fights break out as people defend against ideas that, if right, threaten their careers, identities, and worldviews. Notably, new paradigms often have huge gaps themselves, so the better paradigm is not a cut and dry affair. These fights can last for decades.

5. Paradigm change. Eventually, the new paradigm reaches a consensus tipping point — in the case of the heliocentric solar system, most people would point to the theory of gravity which provided a simple, consistent explanation for all observations to that time and had a lot of predictive power.

6. Normal science. The cycle begins anew. Science never gives us the ultimate truth, but merely creates ever-more-correct models of the universe. Newtonian mechanics are eventually subsumed by general relativity, which has its own paradigmatic holes waiting to be filled.

(Let’s caveat that this is an obscenely stylized description of how the world works and an extremely abbreviated description of the obscenely stylized description! A lot of science proceeds without paradigm shifts. You can have multiple overlapping paradigms at the same time. Many proposed new paradigms are just wrong.)

The analogous cycle in manufacturing paradigms looks like:

1. Normal Manufacturing. Most of the time, manufacturing progresses when process innovations improve systems, making things faster and more efficient. You can think of a manufacturing paradigm as a high-dimensional hill that we climb as we optimize more of the components of that system. If you’ve read Brian Potter’s book The Origins of Efficiency, most of it deals with this world. (And if you haven’t read it, you should!)

   

2. System Saturation. Eventually, it becomes harder and harder to eke out more efficiency gains within a given paradigm. You reach the top of the hill as you make the best steam-powered looms, the largest factories, and the most complex supply chains you can get away with.

   

3. System Building. Some people begin building components, processes, and nascent systems that are anti-paradigmatic. These things are poor fits for the current paradigm and often create products that are more expensive or poorer quality than the ones the existing paradigm creates or simply have no place in standard workflows. For years, guns made with interchangeable parts were both more expensive and lower quality than their hand-crafted counterparts. The early products of mechanized looms were lower quality than hand-woven counterparts. (And even today people prize hand-woven rugs.) Today, 3D printers are consistently criticized for their surface finish and impracticality for making huge numbers of the same thing.

3. System Revolution. If you’ve heard of “the Valley of Death,” that’s stage four for manufacturing paradigms: building them out requires a ton of resources, time, and work without any sort of inevitability. Lucky systems find a niche that values their capabilities enough to tolerate the downsides: interchangeable parts were valuable enough on the battlefield that the US military was willing to support the incremental improvements to the system.

   

5. Paradigm Change. Eventually, crawling up a new hill finally produces outputs that surpass existing paradigms on enough dimensions to create a cascading shift: as more manufacturers switch over to the new system (or new manufacturers using the new system outcompete the old ones using the old system) the trajectory accelerates and the new paradigm becomes the normal ones.

New technology and knowledge unlock new paradigms

Another key dynamic of paradigms is that they don’t just spring fully formed out of people’s heads. Henry Ford didn’t just sit down one day and say “well if we’re going to have interchangeable parts on an assembly line we’re going to need a whole system of precision gauges, machine tools, high-quality steel, and the process knowledge to use all of it; welp, better get started!” New technologies and knowledge unlock new paradigms. These prerequisites are almost always created without the full vision of the new paradigm.

Steam powered manufacturing would have been impossible without cylinder boring technologies originally meant for making canons, blast furnaces, or linkages. Interchangeable parts depended on high-quality steel from the Bessemer process, precision machine tools like mills and lathes, and a whole interconnected suite of gauges and other measuring devices that let you determine whether a part is actually within spec.

Today’s network manufacturing is downstream of containerized transport drastically lowering international shipping costs, a mastery of statistics that enables just-in-time systems to hunt down and eliminate production bottlenecks, and of course, the internet (and the infrastructure behind it) that makes information transfer practically instantaneous and free.

New manufacturing paradigms are impossible without a thousand other projects that go on below the surface. These enabling technologies are always developed for other reasons. New paradigms need people who are close enough to manufacturing to notice what’s newly possible — and who have slack to explore it.

New paradigms shift the center of manufacturing gravity

New paradigms alter the geography of the manufacturing landscape. Britain was a manufacturing backwater before steam power made it into a manufacturing powerhouse in the 18th century. Observers of the 1851 Great Exhibition in London wrote about the American products dismissively before “The American System” (which was the term for interchangeable parts at the time) catapulted the US into manufacturing dominance. The same pattern holds for Germany and chemical engineering, Japan and just-in-time manufacturing, and of course, China and network manufacturing. The places that develop the process knowledge around a new paradigm are able to race up the new slope.

Notably, new manufacturing paradigms almost always come from places that are not manufacturing leaders. It’s a straightforward innovator’s dilemma explanation: when you dominate a particular paradigm, mindsets, infrastructure, and workflows all point towards pushing further and further up the existing hill. Put a bookmark here, because it will be relevant in just a bit.

Building a new system

A final property of manufacturing paradigms (and paradigms in general) is that they look both inevitable in retrospect and unchangeable in the present. Looking back, it’s obvious that interchangeable parts were a good idea. Using cheap shipping to take advantage of specialization, trade, and scale is just economics 101. Looking forward, it’s incredibly hard to imagine things being fundamentally different: when most people say “advanced manufacturing” they mean automating what we’re already doing and figuring out how to scale up things that haven’t already been scaled.

Paradigms look unchangeable until they shift.

A new manufacturing paradigm is possible. The old system is full of epicycles and inconsistent observations are starting to pile up.

The epicycles are clear. The current paradigm relies on incredibly efficient but incredibly fragile supply chains. One shock can ripple through the system, creating shortages and cost spikes for years. It incentivizes cheap products that are meant to be thrown away instead of repaired or modified. The only things worth making are those that can be sold at massive scale. Entire industries rely on inputs from a single country, or even a single factory.

Let’s be clear: the system has delivered massive benefits to billions of people. But Newton’s physics was able to account for every observation for hundreds of years. That didn’t mean that Einstein couldn’t unlock a torrent of new discoveries.

We can put today’s piling-up observations into four buckets:

1. Leveraging new phenomena

2. Sensors and compute

3. Tooling-free processes

4. Launch and Space

Leveraging new phenomena. Near the end of the 20th century, humanity upgraded our ability to harness a handful of natural phenomena, but we still haven’t gotten anywhere near using them to their full potential. The most obvious three are lasers, magnets, and the photovoltaic effect. Lasers can (among many other things) create nano-scale structures, heat things precisely at a distance, and slice through material with exquisite precision. Powerful magnets (of both the rare-earth and superconducting sort) can move things without physical contact, weld unweldable metals, and create otherwise-impossible actuators. (Actuator is a fancy word for a generic thing that a machine uses to physically act on the world — anything from a robotic hand to a leaf blower can be an actuator.) Everybody knows about solar, but our manufacturing systems haven’t adapted at all to an energy source that is effectively free for part of the day but off the rest of it.

Sensors and compute. You may have noticed that computers have gotten way better. They can model and simulate almost anything, control systems in real time, and a host of other fancy things. And yet, by and large we’ve used them to improve existing parts of the manufacturing system rather than building new systems around their capabilities. Yes, I’m finally talking about AI (thank you to those of you who have read this far despite its absence). Sensors that let the computers know what’s actually going on have also improved dramatically.

Tooling-free processes. To make something at scale today, you first need to spend a lot of time and effort making the things to make the things. These are the custom-made casts, dies, and a plethora of specialized machines collectively known as “tooling.” There are an ever increasing number of “tooling-free” processes that can make different things without needing to swap out any physical components: 3D printers, laser cutters, EDM cutters, and more. These machines are criticized within the existing paradigm because they can’t produce identical components at scale as quickly or efficiently as tooled system. That kind of criticism is often a tell for the roots of a new paradigm.

Launch costs and space. This is the most speculative one. The cost of launching mass into orbit (and bringing it back) is decreasing while the number of things people want to do in space are increasing (I will not weigh in on orbital datacenters here). You can manufacture things in space that you cannot on Earth thanks to zero-gravity, infinite vacuum, and extremely low temperatures (if you can dump heat effectively). Obvious examples are things like fiber optics that need extremely large and pure crystals but in reality, we haven’t scratched the surface. Space lets you avoid many of the annoying things that manufacturers work very hard to deal with on Earth: material doesn’t settle so it can mix evenly, it’s easy to keep reactive material from touching the walls of a container, and you can create massive films without them tearing.

All four of these buckets are coupled, bear with me.

A quick aside on AI: You might be asking — “Ben, isn’t the next manufacturing paradigm obvious? It’s just AI. Won’t we just have lights-out factories full of humanoid robots making stuff? It’s already happening and this whole piece is just irrelevant navel-gazing."

But automation on its own does not cause paradigm shifts! “Raw” automation just makes existing paradigms more efficient (or doesn’t). We have been automating more and more pieces of the manufacturing process since the punch-card directed power loom in 1805. New paradigms require large clusters of enabling technologies and new systems built around them. The historical pattern suggests that AI may certainly be a piece of the puzzle, but it’s not the whole picture.

(Of course, there is no way to make yourself seem more foolish than to try to predict the future so I could be totally wrong here and humanoid robots could create a cascade of systemic changes that utterly transform how things are made. But I have my doubts.)

What does this new paradigm look like?

It’s incredibly hard to describe in the abstract a paradigm that does not yet exist and whose specifics depend heavily on work that has not yet been done. I’m not a great science fiction writer, but I’m going to attempt to paint some pictures of what a world with this new paradigm might look like before describing it more abstractly.

Low-effective-scale cells. Jane is an engineer at Tesla with a brilliant idea about how to make electric cars that, among other things, are powered by supercapacitors and take advantage of a new type of magnet; both of these new systems require reinventing how to make a car from the ground up. She builds a prototype using a heavily-instrumented “cell” that looks kind of like a shipping crate filled with different machines that can change what they make on the fly — think the hybrid offspring of 5-axis CNCs, 3D printers, laser cutters, and their ilk. She sells the first prototype to a millionaire with a personal racetrack and uses that money to build two more of the vehicles — a seamless process because the cell is able to replicate and improve on the exact steps she used to make the prototype using its own actuators. Selling these vehicles gives her enough evidence to get a new form of technology loan to lease several identical scales that let her start scaling production by parallelizing a continuously improving process as the cells learn from each other. This whole process is aided by new forms of regulations that use the digitized design and process data to automatically sign off on the new technology.

Rapidly reconfigurable manufacturing. COVID-37 strikes; but because the now-large car company is using this modular manufacturing technology that doesn’t require a pile of specialized machines and tooling like the current paradigm, it’s able to quickly switch production to start making masks and other gear. (The same thing is true when it needs to switch production over to microsubs during the war of ‘39.)

Closed-loop manufacturing. A different division of the company produces fire-fighting drones. Data from drones rolling out of final assembly cells feeding back into its design to continuously remove defects and improve production efficiency. More importantly, telemetry from the field shows that smoke is causing unexpected degradation in the motors (that use a new kind magnet made out of directional printed iron crystals that outperforms rare earths). That data both triggers immediate short-term design fixes and kicks off a project in the R&D department that both finds the root cause of the problem and unexpectedly discovers a new alloy that could give flying cars a 300 mile range.

Space manufacturing. Awkwardly, that alloy that can only be made at scale in zero-g, high-vacuum environments (ie. space). Luckily, cell-based modular manufacturing makes it so that Daniel can prototype the process on the ground and then instantly implement it in a twin cell in orbit instead of waiting a year for a launch slot for each iteration. His friends in the pharma industry use a similar process to discover and make miracle drugs.

The abstract pieces of the paradigm, as I see them, are the following;

• Modular, composable cells that leverage new kinds of tools and automated actuators to make it fast to iterate on designs, scale through parallelization, and quickly change what they make.

• Closed loops between the product’s design, the manufacturing process, the final products, and how they work in the field.

• The cells make it so that the minimum viable efficient scale for manufacturing drastically decreases and makes manufacturing much more location agnostic. You can drop a small “factory” outside of most major cities and propagate process improvements across all of them simultaneously.

• The remotely reconfigurable, low-minimum efficient scale, and lower-infrastructure-intensity cells and feedback loops enable space manufacturing. (Except for the hard-to-get-things-up-there-and-back-down nature of space, it’s actually a great place to manufacture things — chemicals don’t settle, you don’t have to worry about environmental damage, there are no contaminants, and much more.)

We’ve already reached this paradigms’ copernican era — there are many tools and processes that show cracks in the current paradigm and hint at the new one. 3D printers are starting to get good enough that there are a few products that it makes sense to scale up through massive printer farms. Tooling-free systems like laser cutters, electrical discharge machines, and automated metal breaks keep improving. Places like Machina Labs are creating new tooling-free systems. Sensors are getting cheaper and rudimentary “digital twin” systems are coming into their own. But right now all of these things are like new astronomical observations that we’re still trying to create paradigmatic epicycles around.

I realize that description needs far more detail! Filling out those details requires a lot of work, both back-of-the-envelope roadmapping and empirical tinkering that I am actively working to aggregate the time, resources, and people for (you’re welcome to help).

Why you should care

Even if you have no intrinsic interest in manufacturing, you should care about creating this new paradigm.

If you care about US power you should want the United States to be a place where people make things. (I’ll get to why you should want that even if you’re a globalist.) Many people claim to want this, but we’re going about it wrong. We’re not going to manufacture things cost-effectively in the US by putting up tariffs, doing top-down industrial policy, and otherwise trying to out-China China. Even ignoring their heavy government subsidies, China has gone too far down the learning curves of conventional ways of manufacturing things. History shows us that the way to shift manufacturing centers of gravity is by creating new paradigms.

If you care about humanity, could care less where things are made, and just want a flourishing future, you should still care about creating a new manufacturing paradigm in the US. The history of manufacturing paradigms (and Schumpeterian disruption theory) has shown us that radically new paradigms need to come from places that do not currently dominate manufacturing. Incumbents have the most reason to avoid the long, painful process of going down into the valley of death between the current optima and the next. So if you want the trends that have lifted billions out of poverty to continue, we need new paradigms, and those paradigms need to come from places that don’t dominate manufacturing.

I care about this for selfish reasons. I want to live in the place that unlocks the future.

There are too many stories to ignore of how important it was for great innovation orgs like Bell Labs or Skunkworks to have people with deep manufacturing experience in the room. Not only did they enable rapid feedback loops between idea and physical prototype, but they made sure that the eventual output was manufacturable (remember, inventions are ultimately useless if you can’t make enough of them economically). On top of all that, tacit manufacturing process knowledge is often critical to invention itself. One of the bottlenecks to creating the transistor was bonding a metal whisker to germanium without destroying it; the problem was solved by a machinist, John Saby, who applied a trick for manufacturing precision radio components — evaporating a thin sheet of gold onto mica, slitting it, and then pressing the gold-mica slab onto the germanium so that the stress caused two gold whisker contacts to squeeze through the mica.

I’ve belabored the point about organizations because the deep coupling between invention and manufacturing also exists at a regional level. Innovation is incredibly sensitive to friction; there is a huge difference between a world where you can run ideas by your machinist neighbor over the fence, walk through a giant components market on your way home from work, or drive over to five different specialized machine shops and one in which you cannot.

Hobbyists and tinkerers are a secret vanguard of innovation. The Homebrew Computer Club, early rocketry and aviation, drones, radios, and the list goes on. But these people thrive on discarded equipment and tools, skills they gleaned from work and their community, and borrowed infrastructure.

Furthermore, what you know how to do, and think is possible, deeply affects the kind of ideas you come up with. When everybody is just doing software, you come up with new ideas for software. When everybody is writing blogs, you come up with great essay ideas. When you make physical things and everybody around you is too, you come up with ideas for physical things. Unless you want a Ready-Player-One-WALL-E-Matrix future, the place that unlocks the future is going to be the place that comes up with brilliant ideas for physical technology and implements them.

New Paradigms Are Not Inevitable

After seeing the cracks in an old paradigm and possibilities for a new one, it’s easy to think that all we need to do is sit back, maybe invest in a few startups, and let the unstoppable march of progress do its thing. This couldn’t be farther from the truth.

New paradigms are not inevitable. Creating them requires swallowing three bitter pills that are especially at odds with our current moment.

First, new paradigms take time. Eli Whitney won the first contract to produce muskets with interchangeable parts in 1798, but the system wasn’t in widespread use until the 1870s. Steam-powered manufacturing took almost 60 years to catch on in 19th century Britain. Things move slightly faster now, but just-in-time-manufacturing still took several decades to emerge in Japan, and the same is true for the network system in China. These decades-long timescales are a poor fit both for most contemporary forms of capital and political timescales.

Second, new paradigms are not politically convenient. Increased productivity means doing more with fewer people; increasing the total amount of manufacturing will not create piles of new jobs nationally, or even where things are being built. It won’t create zero jobs, but if jobs are a political selling point, it will warp the technology to the point of uselessness. Speaking of where things are built, new paradigms like this depend on agglomeration effects in places that are hard to predict a-priori. That means the standard politically expedient moves (in the US) of top-down funding decisions and spreading money around geographically to scratch the right backs are the best ways to make sure nothing useful gets done.

Third, creating new paradigms requires work that doesn’t currently have a home. It requires uncertain work that strives to reinvent entire systems, while simultaneously being laser focused on usefulness. Much of the work that needs to be done in the next five years is distinctly pre-commercial — it does not yet make sense as a rational business investment. Startups need to focus on point changes in existing systems or the product they’re getting out the door. Most manufacturing companies don’t have the margins to stop production even to test a new machine, let alone create a research arm. And the academic system that has a near-monopoly on pre-commercial research is ill-suited to this work for more reasons than we have time for.

These bitter pills mean that without shifting trajectories, I’m pessimistic about this new paradigm’s prospects — despite excitement and money flying around reindustrialization, AI-for-science-and-manufacturing, and hardware startups. But changing those prospects is why I’m writing this and why we’re doing what we’re doing at Speculative Technologies!

Creating new paradigms always rubs people the wrong way and goes against dominant ways of doing things. That includes the government-startup-academia complex in the US.

It may sound odd, but I see creating new manufacturing paradigms as a humanistic endeavor on par with curing cancer, ending poverty, or discovering the secrets of the universe. Without new manufacturing paradigms, progress will grind to a halt; whether you care about progress in health, wealth, or going to the stars. The importance is just under the surface.

Remember, manufacturing is like sewer systems.

Sewers are a miracle, arguably saving hundreds of millions of lives and making life so much more pleasant to boot. But they didn’t just happen by default. It took hard, thankless work by many individuals who captured little of the value they created.

If we have that opportunity to create that kind of progress, shouldn’t we take it?

But giving is the wrong way to think about it. Money is a tool for enabling more of what you want to see in the world. We’re selling a world with more wonderful technology, more ambitious research, and more pathways for talented scientists than just academia or startups.

Supporting Spectech is also how to keep your finger on the pulse of what science and technology is coming after the next big thing. Take the Brains research accelerator: because it’s domain agnostic, it’s basically a cross-section of what the most ambitious scientists want to work on, and supporters get an inside look.

Supporting a fellow through the program takes $60k. That covers staff time (the program is operationally heavy and we work closely with the fellows), mentor honoraria, and travel/ events (we’ve found that it’s incredibly important that the fellows meet each other and present live to potential allies).

$60k is not a small amount of money. But it’s also $6000 from 10 people or $600 from 100 people. Coordination is powerful. And when a fellow successfully starts a FRO or ARPA program, that $60k has unlocked $10M or more. That’s more than 150x leverage.

Since the first cohort in 2024, our fellows have gone on to raise $40M of philanthropic and government money to run ambitious programs in health, climate, materials, and space.

If we’re able to get more than the bare minimum to run the program, there are so many things we could do:

• Create tools to help the fellows iterate faster and help them roadmap their program

• Hire a director who could spend all their time making the program better

• Give the fellows seed money to start working on their programs immediately

We already have several anchor donors who have committed 20% of the program’s budget for the next several years. If you want to join their ranks, please reach out. You can support us in smaller ways here or sign up for a paid Substack subscription.

Today is also a great day to become a supporter because we’re hosting a supporters call on Wednesday December 3rd at 8p EST, where we’re sharing what we’ve been up to, a peek at what’s coming next, and giving supporters a chance to ask us about anything.

However you choose to be involved — we’re grateful for you coming on this journey with us.

—Ben

That upbringing shaped how Brachman, a fellow in the first cohort of Spectech’s Brains Research Accelerator, approaches problems. Today she’s leveraging that approach to the problem of decoding the human immune system at Imprint, a Focused Research Organization (FRO) that is building the Rosetta Stone for the human immune system.

The Problem

Your immune system keeps a diary. Every infection you’ve fought, every allergen you’ve encountered, every cancer cell your body has quietly destroyed—it’s all recorded in the sequences of your B- and T-cell receptors. Accessing this biological ledger could revolutionize how we understand and treat disease, revealing hidden causes of autoimmunity, predicting who’s at risk for chronic conditions, identifying the therapeutic antibodies that nature has already perfected and more.

There’s just one problem: we can’t read it.

Unlike your genome, which sits neatly packaged in nearly every cell and is 99.9% identical between people, your immune repertoire is scattered across billions of constantly evolving immune cells. Each person carries roughly 10^18 to 10^20 unique receptor sequences that aren’t encoded in the genome. They’re generated on the fly as you move through the world, resulting in immune repertoires that are 99% unique to each individual.

Current technologies can’t cut it. “It’s like trying to understand an entire ecosystem by tagging a few hundred animals,” Beck explains. “You’re sampling something incredibly diverse and dynamic from a single snapshot, then trying to infer the whole.” But with today’s massive data sets and powerful machine learning models, it is possible to decipher meaningful patterns from inherently sparse data. She gives the example of Epstein-Barr virus and multiple sclerosis: it originally took a massive 20-year study of 10 million people to reveal that a “mono” infection dramatically increases MS risk, but that result was replicated by a group at Stanford with just a handful of people and a single timepoint. “You need machine learning to extract the patterns—what antibodies recognize what proteins, how infections lead to autoimmunity years later, which signatures predict disease risk.”

By building the tools and models needed to read the immune system, Beck and her team at IMPRINT are tackling one of the most audacious problems in human health.

The Discovery

In 2008, while working on her PhD in a neuroimmunology lab, Brachman transferred peripheral immune cells from chronically stressed, depressed mice to healthy recipients, expecting to transfer depressive behavior along with the cells.

The opposite happened. The immune cells protected against, and even countered, depression. The result also pointed to something bigger: the immune system wasn’t just involved in psychiatric disorders—it was a component of nearly every disease state. “The same tools needed to understand immune mechanisms in psychiatric disorders are the same tools for infectious disease, autoimmunity, and cancer. These are fundamental immunomics tools for chronic disease.” And these tools didn’t exist.

Beck tried tackling this tooling problem through traditional routes – first in academia, then as a startup. Neither fit. In each case, she found that the “default templates” of those paths did not match the incentives and potential output of this type of project. “I didn’t have a straightforward path, and there have been a lot of iterations of this and related ideas,” she says. “We had a phenomenon but not a mechanism. We needed fundamental tooling before we could commercialize anything.”

The standard routes—academic lab or venture-backed startup—couldn’t provide the right structure for the fundamental tool-building required.

The FRO Solution

Beck and her co-founder Victor Greiff, a computational immunologist, started IMPRINT as one of the six inaugural Schmidt-funded FROs. The structure provided what neither academia nor startups could: patient capital for tool-building, technical milestones without publication pressure, and crucially, the ability to pursue fundamental questions while maintaining a focus on clear deliverables.

Because they are so new, FROs have no standard wisdom or communities of people who have gone down the same path. “FROs have all the standard problems of startups, but the solutions don’t work” Beck says.

This is where the Brains accelerator proved invaluable. Beyond providing community for isolated leaders, it paired Brachman with Geoff Ling, founding director of DARPA’s Biological Technologies Office, as a mentor, and he became an Imprint advisor. “He gave us the advice early on to ‘let the science lead,’ which has been our North Star.”

The program also helped navigate challenges unique to the FRO model—from rapid scaling to creating useful feedback loops in absence of market signals.

The Inflection Point

Today, IMPRINT has hit an inflection point. “We’ve built the fundamental models and tools for being able to read the patient record encoded within the immune system,” Beck says. The team can now infer immune repertoires from samples, extract patterns about antigen recognition, and predict which antibodies bind which proteins—essentially solving the “AlphaFold problem” for the immune system.

The next phase involves shifting the technology towards specific applications like antibody-antigen binding prediction for therapeutic design. The possibilities remain vast: diagnostics that detect disease before symptoms appear, therapeutics pulled from nature’s own library of tried-and-true antibodies, insights into the hidden causes of chronic disease.

Brachman’s journey—from finding antidepressant immune cells to building industrial-scale tools— is why alternate research structures, like FROs, matter. The immune repertoire holds a record of every pathogen encountered, every cancer cell detected. Reading that record at scale could transform our understanding of chronic disease. But first, someone had to build the tools.

You can learn more about IMPRINT here.

Thank you to Lauren Richardson for conducting the interview and writing the vast majority of this piece.

This is a repost of a piece we published last year about the idea behind Brains and how it works, with some updates based on a year’s worth of lessons and progress.

Why an accelerator for coordinated research programs?

The idea behind Brains is simple: pave a “third path” for talented scientists to start potentially game-changing research programs that are poor fits for both startups and academic labs.

This kind of big-if-true but still pre-commercial work unlocked much of the modern world — from the Internet to mRNA vaccines. Riding in Waymo cars still feels like the future; autonomous vehicles got a huge boost from the DARPA Grand and Urban Challenges — both were ambitious research programs beyond the scope of a single company or lab.

Over time, these programs have proliferated — in government ARPAs (Advanced Research Projects Agencies) like ARPA-H, ARIA, ARPA-E, IARPA on top of the OG DARPA, at nonprofit organizations like Convergent Research and Spectech, and through programs that are either affiliated with foundations or are entirely independent.

However, many scientists still only see two possible paths to turn an idea that needs coordinated group effort into reality: starting an academic lab or a startup. Neither of those is wrong, but many ideas are poor fits for either of those institutional boxes. Even for people who know that coordinated research programs exist as a third path (like many of you reading this!) the pathway to starting one and running it successfully is opaque.

Starting and running both startups and academic labs is hard but there are straightforward ways to learn how to do it: as a grad student or postdoc, you can watch what your PI does; you can work for a startup and learn from the CEO. Libraries’ worth of writing has gone into the experience of being a founder or a professor. There are hundreds of universities, venture capital firms, and startup accelerators with open applications.

By contrast, starting an ARPA program or an FRO is an opaque process demanding a skillset that is hard to pick up through other professional experience: a combination of hustle, vision, pragmatism, sales, and organization. This skillset is hard to compress into an interview or test, so it’s hard to know if someone will be good at leading a program until they’ve led a program. In high-uncertainty situations like this, funders and hiring managers fall back on heuristics like credentials and networks instead of evaluating people and ideas on their merits.

It’s a market failure: on one side scientists don’t know that running a program is an option and don’t know how to get started even if they do; on the other side, it’s hard for ARPAs and other organizations to know whether people will be good at running programs without pre-existing relationships.

Today, a lot of potentially awesome program leads cram their ideas into poorly shaped institutional structures or abandon them all together; ARPAs and other organizations find themselves “fishing from the same small pond” of people they’ve worked with before — missing out on great talent and ideas.

So where does Brains come in?

You can think Brains as performing several roles:

1. Putting up a signal flare for people who might be great program leads. (Like this post!)

2. Primary filtering (that’s the application and interview process)

3. A crash course in the skills people need to successfully run programs

4. 1:1 mentoring to help fellows get ideas as close to “shovel ready” as possible

5. Building a supportive community of peers and advisors

6. Connecting the two sides of the marketplace.

Over time, we’re hoping to do for ARPA programs, FROs, and other coordinated research programs what YC and other accelerators did for startups. Starting and working at startups went from an obscure pursuit that was hard to get into unless you knew the right people to a normalized (if still high-variance) path for people all over the world.

We believe that the market for coordinated research programs is elastic: as more people start them and get results, more people and organizations will want to fund and support them. And we’re starting to see this hypothesis play out!

Since the first cohort of fellows graduated a year a half ago, they have gone on to start coordinated research programs that in total have more than $40 million committed from philanthropists and governments. The 2025 cohort has a ton of exciting work in the pipeline and the AI fellows will be unveiling their programs at showcase soon. And we’re just getting started.

Onwards!

The 2026 Brains cohort will run from February through June 2026. If you have a deep technical background and an ambitious idea, we encourage you to apply!

An ideal fellow:

• Has a deep technical background, has done the equivalent of PhD-level research, and has several years of work experience.

• Has one or more ambitious but precise ideas that don’t fit into existing institutions.

• Has worked both inside and outside of academia.

• Has a strong bias towards action.

• Is comfortable with acting under uncertainty.

Of course, there are exceptions to every rule, especially on the frontiers of human capabilities, so people who don’t exactly fit this profile should still apply if they think they can crush it!

So:

1. If you or anybody you know has what it takes — you can find the application here.

2. If you have more questions, you have several options:

   • You can check out the FAQ here.

   • We’re going to be doing online info sessions on October 16 at 12p ET and November 5 at 8p ET.

   • You can email questions to brains@spec.tech.

3. If you want to get involved outside of being a fellow — please do reach out at brains@spec.tech.
   
   We’re excited to help make your awesome ideas a reality.

Funding preferences and situations can change quickly, so if any of this is incorrect or incomplete, please leave a note in the comments!

Unsurprisingly, there is no default way to fund misfit research: support can range from a group of philanthropists starting a new institute, to DARPA running robot competitions, to DAOs funding longevity projects, to VCs funding research projects gussied up as a company.

To get our brains around the funding landscape, it’s useful to divide this funding into non-dilutive (funding that comes without ownership or expectation of financial return) and dilutive (funding that comes with an expectation of financial return and often involves some ownership of an organization). This division is useful because, in broad strokes, non-dilutive and dilutive funding come with very different expectations, evaluation criteria, and “sales” processes.

Be aware that these categories have a lot of fuzziness (like many things in non-traditional research). Several entities, like family offices, do both dilutive and non-dilutive funding; they have their own section. Furthermore, non-dilutive and dilutive funding are not always mutually exclusive: some technology projects get off the ground with a mix of non-dilutive grants from foundations or governments and investments from angel or impact investors. (There are still a lot of gaps in this “messy middle” between pure-public-goods work and profit-maximizing company.)

Below are the major groups in each category and what they're actually funding. The end of this section touches on what to actually do with this information when you’re trying to fund research.

Non-Dilutive Funders

Non-dilutive funding doesn’t come with any expectations of repayment or organizational ownership. This sort of funding is important for research that is a poor financial investment, whether because it will never create capturable value, or has long timescales and high uncertainty. However, non-dilutive funding isn’t just “free” money. Raising non-dilutive funding usually takes significantly more time and effort than the equivalent amount of investment dollars; most funders impose much more process up front and restrictions on how money can be spent.

Foundations: Foundations are organizations with full-time professional staff deploying money that has been set aside explicitly for philanthropic purposes. Foundations can range in size from a tiny org with one or two staff to hundreds or thousands of employees at the largest foundations like the Rockefeller or Gates Foundations.

In aggregate, foundations gave $30B towards research in 2019, which is more than the NSF and comparable to NIH.1 While these numbers are large, the median grant is significantly less than $1M. Even large foundations are usually spending money that comes from the interest on an endowment, so they may have much smaller budgets than you might expect based on the wealth of their founder or size of their endowment.

As of 2025, most research funding from foundations goes towards traditional research. The bureaucracy and processes of most foundations make it hard for them to support non-traditional work. Foundations typically deploy money through program officers who work within tight bounds set by the board of trustees on a yearly basis. Programs often have explicit mandates to work within traditional institutions through graduate fellowships or awards to professors.

There are, of course, exceptions. The now-defunct Schmidt Futures helped a number of ambitious research organizations get off the ground.

Philanthropic Aggregators: Philanthropic aggregators are organizations that use their brand and connections to fundraise for specific projects from wealthy individuals or foundations. Some examples include Renaissance Philanthropy, Founders Pledge, and XPrize. Each philanthropic aggregator has their own process and funding “form factor”: Renaissance Philanthropy creates “philanthropic funds” that they use to deploy grants, while the X-Prize creates prize competitions.

Philanthropic aggregators have funded a lot of non-traditional research. The process of recruiting on a case-by-case basis makes aggregators more flexible than foundations with board-specified programs.

Government Organizations: Government is, of course, a major research funder. The vast majority of government research funding is little-c conservative and intended for traditional PI-driven academic work. However, certain agencies and programs have supported some misfit work work.

DARPA pioneered using Other Transaction Authority (OTA) to run prize competitions like the DARPA Grand Challenge, Urban Challenge, and Robotics Challenge. SBIR (Small Business Innovation Research) grants provide non-dilutive funding towards research-heavy startups. Recently, two British government organizations (ARIA and the Department for Science, Innovation, and Technology) announced that they were funding Focused Research Organizations.

Crowdfunding Platforms: Platforms like Experiment.com enable researchers to raise small amounts of money from a large number of people. Crowdfunding can work for non-traditional projects that have public appeal like a citizen science endeavor to catalog ocean plastics or a team building a field microscope. However, crowdfunding rarely raises amounts more than tens of thousands of dollars, so it’s suited for modest-scale projects or early seed funding.

Dilutive Funders

Dilutive funding can be double-edged: it injects significant resources, but it may steer the work towards shorter-term commercial goals and away from research work or even longer-term commercial goals. Professional investors need to show their investors portfolio growth and exits, which means they need companies in their portfolio to show year-over-year growth, which can be at odds with the uncertainty baked into research.

Angel Investors: Angels are individuals who invest their own money in early-stage startups. Technically, angel investment is driven by prospects of eventual financial return, but some angels think less about whether a startup is a good investment but instead about whether the work would be cool or impactful.

Venture Capitalists (VCs): The main context in which professional venture capitalists fund non-traditional work is (almost by definition) bubbles. When an area is “hot” enough, even professional VC funds put money towards work that has no sense of how it becomes a product or a business. There are also a few VC firms that have different structures, like longer fund lifetimes, that enable them to invest differently from other firms. Some examples of VC funding into research includes most quantum computing companies, Colossal Biosciences, and Physical Intelligence.

Corporate Research: While corporate research has drastically contracted, some large companies with large margins still support exploratory research. As obvious 2025, AI research is an obvious example. Corporations rarely fund nontraditional research – most external research funding goes towards universities primarily as a hiring pipeline. Teams can sometimes carve out a niche within corporate research to develop something ambitious – the team that started the Lean FRO worked at Microsoft Research for a long time.

Corporate Venture: Corporate venture capital arms invest in startups based on the company’s “strategic interest” in addition to pure returns. For example, a car company may invest in a research-heavy battery startup or a chip company may invest in a photonics company long before they have a product. Large companies sometimes acquire and continue to fund organizations that focus more on research than products: Hyundai’s acquisition of Boston Dynamics, for example.

Impact Investors: Some investors are willing to accept lower financial returns in exchange for high social or environmental impact. These investors fund for-profit ventures within some impact area (like climate or health) that don’t fit the profile for normal VC funding because of factors like time scales or capital requirements. This kind of investment is also sometimes called “patient capital” or “concessionary funding.” For example, Breakthrough Energy Ventures (BEV) explicitly operates on a 20-year timeline and invests in risky clean energy companies with the understanding that some may only yield societal benefit without huge profits. In the medical world impact investors sometimes invest in exchange for royalties or revenue sharing rather than explosive startup growth.

Program-related Investments (PRIs): Foundations and Donor Advised Funds (which we will talk more about in the next section) can make dilutive investments out of their endowments that count towards their legal deployment quotas as long as they are mission aligned. Unpacking that jargon: Foundations (but not DAFs) legally must spend 5% of their assets annually; normally this money is deployed as grants, but dilutive investments that are aligned with the foundation’s mission can also count towards that 5%.

Investments with the possibility of a return enable Foundations and DAFs to fulfill their charitable missions without eating as much into their bank accounts. This upside means that PRIs that do happen are often larger than normal grants and theoretically people with DAFs and Foundations should be excited to do them. However, the potential for IRS scrutiny, divisions between investment and granting teams, and DAF sponsors that don’t support PRIs means that they are fairly rare.

Two words of caution about funding research-heavy work with dilutive alternatives to venture capital:

1. While PRIs and impact investors explicitly fund work that struggles to raise VC funding, they often focus on work that can’t raise VC funding because of characteristics like timescales, capital requirements, and market sizes, not because they are too research-heavy.

2. Most dilutive funding puts an organization on a trajectory where they do need to raise venture capital eventually. Many misfit research projects will never be a good fit for venture capital, so raising dilutive funding can be a trap.

Both

These entities can (but do not always) do both dilutive and non-dilutive funding.

Decentralized Autonomous Organizations (DAOs): DAOs are communities that pool funds for a common purpose by selling blockchain-based “tokens” that give their owners a say in the organization’s governance. Some DAOs, like VitaDAO or CerebrumDAO focus on research. DAOs fund work through many different mechanisms — both traditional grants and dilutive funding as well as newer blockchain-based mechanisms for capturing some of the value that the work could create.

High-Net-Worth Individuals (HNWIs): Wealthy individuals sometimes bankroll research personally. Individuals have the most flexibility of any funders to do unconventional things. As a result, a lot of non-traditional research has been funded directly by individuals. HNWIs often fund things quietly and make themselves hard to contact for obvious reasons: everybody would be asking them for money and funding strange things can open people up to reputational risk.

Family Offices: Family offices are professional organizations that handle the money of a wealthy individual or family. The big thing that differentiates family offices from foundations is that they don’t have a pile of money that has been explicitly set aside for philanthropy. Instead, they have multiple mandates – increase wealth, hedge against risk, maintain liquidity, and do philanthropy.

Keep in mind that people and organizations can have completely different focuses, mindsets, and processes whether they’re doing dilutive or non-dilutive funding: a wealthy individual who will write a million dollar check to a startup the same day might only donate $10k at a time only to projects focusing on a specific disease.

As a reminder, the Brains is a research accelerator that helps talented scientists and technologists execute on ambitious ideas that are beyond the scope of individual academic labs, startups, or large companies. This is a special “interstitial” cohort focused on AI security and governance capabilities.

If you find any of their ideas particularly exciting or intriguing, please get in touch with them via LinkedIn (linked to in their names) or email brains@spec.tech and we’ll route you correctly.

Daniel Kang

Daniel is building the first red-line benchmark for AI agents, focused on the ability of AI agents to perform complex, cybersecurity attacks on real-world systems. Assessing their products’ offensive cybersecurity capabilities is highly complex and contrary to the commercial incentives of frontier labs so no benchmark yet exists. As a result, policymakers and AI labs are flying blind on measuring the ability of AI agents’ ability to exploit complex cyber “kill-chains,” making it difficult to generate policy recommendations or determine when responsible scaling policies should be enacted.

About Daniel

Daniel is a professor of computer science at UIUC, where he studies the progress of AI. His work includes award-winning benchmarks (CVE-Bench), widely cited standards for AI benchmarks (ABC), and work on understanding reinforcement learning for LLMs. Daniel's lab has been awarded grants from the Open Philanthropy Project, Schmidt Sciences, and Google.

Lisa Thiergart

About Lisa

Lisa is an AI Security researcher who has worked on projects spanning high security datacenters, AI interpretability and AI technical governance. Lisa previously led an 18-person research team at MIRI and pioneered the AI interpretability technique of Activation Steering together with Dr. Alex Turner. Lisa holds a BSc in Computer Science from TU Munich, completed graduate research in reinforcement learning for robotics at Georgia Institute of Technology, and earned an honors degree in Technology Management and Entrepreneurship. Lisa is a Fullbright alumnus, fellow at the Foresight Institute, a Manifund regrantor and mentor at the MATS AI safety fellowship.

Mayank Kejriwal

Mayank is developing methods to forecast and prevent rare, high-impact failures in AI agents operating over long time horizons in real-world environments. Today’s AI testing methods focus on the short tail (common, easily reproducible bugs) while long-tailed risks remain largely invisible until it’s too late. This project seeks to shift those rare failure modes into the observable range, creating an “existential risk calculator” for advanced AI systems. The work will help surface edge-case errors in autonomous vehicles, military decision aids, AI scientists, financial trading agents, and long-lived systems managing critical infrastructure.

About Mayank

Mayank is a research scientist and professor at the University of Southern California, where he directs the Artificial Intelligence and Complex Systems group at USC’s Information Sciences Institute. His work bridges AI and human-computer interaction to tackle real-world challenges, from mapping health disparities and combating human trafficking, to building research copilots that help scientists write better grants. Mayank's research, funded by DARPA and NIH, has been published across 100+ peer-reviewed papers, and featured in Fast Company, The Conversation, and the San Francisco Chronicle. He is also the lead author of an MIT Press textbook on knowledge graphs.

Melissa Carraway

Melissa is developing tools and methodologies to build better AI-enabled sociotechnical systems for high-stakes expert decision-making. By developing an understanding of the kinds of decisions that domain experts must make in high-information and high-stakes environments, we can refine and generalize our understanding of the dynamics of human-machine teaming in these kinds of scenarios. By focusing on the needs of expert users and teams, we will be able to create systems that are more readily adoptable, safer, and provide the right levels of insight and accountability while enabling AI to appropriately support work in high-stakes contexts.

About Melissa

Melissa Carraway is a human factors researcher, UX strategist, and systems thinker who designs mission-critical tools. With a background spanning visual design to cognitive modeling, she brings clarity to complex systems, crafting intuitive and trustworthy interfaces for high-stakes domains like the DoD and intelligence communities. Her work has influenced AI adoption strategies across government programs, reducing cognitive load and enhancing decision-making. Currently a Research Scientist at ARLIS, Melissa integrates technical insight with user advocacy, shaping the future of human-centered AI through research, design, and cross-functional collaboration.

Michael Hsu

Michael is giving human-like memory to AI by building computing architectures inspired by neuroscience. Existing memory solutions for AI are more like notebooks than like brains, leading to severe cost and capability shortcomings. A machine that possesses biomimetic memory, on the other hand, would be able to shift the substrate of learning from fixed patterns in the model to memory structures that can be continuously updated. The result is AI that can maintain context indefinitely, learn in real-time, and dynamically leverage its internal representation of the world to optimize its performance. Together, these capabilities enable us to better understand what the machine is thinking, while supporting more robust and efficient forms of reasoning across a diverse range of tasks.

About Michael

Michael is a software engineer exploring how complex systems science and computational neuroscience can help AI think more like humans. He brings a systems perspective from designing software architectures and leading engineering teams, combined with deep curiosity about nonlinearity, emergence, and the relationship between time and space. In the past, he was Director of Engineering at Voltus, which sought to bring balance to the North American energy grid. He also worked briefly on Google Assistant. Michael studied computer science, literature, and mathematics at McGill University, and has interdisciplinary experiences across bioinformatics, biomimicry, ecology, geography, and philosophy.

Rick Goldstein

Rick is developing techniques to reverse engineer neural networks into precise, human-readable code – not just approximate explanations, but exact algorithmic reconstructions that reveal the computational principles behind AI capabilities. This enables designing more efficient AI architectures and training recipes, verifying safety through rigorous analysis of system behavior, and extracting scientific knowledge from models trained on complex real-world data. For instance, we could train a model on biological data and then extract the cellular mechanisms it discovered. This transforms AI models from mysterious black boxes into transparent and reliable scientific tools.

About Rick

Rick is the CTO of Freestyle Research, a robotics startup developing AI models for dexterous grasping. For the past three years, he has conducted AI Safety research into understanding how AI models reason, and he previously worked as a software engineer at Waymo developing routing algorithms for autonomous vehicles. Additionally, he mentors early-career AI Safety researchers. He holds a B.A. in Applied Mathematics from Harvard University and M.S. and Ph.D. in Robotics from Carnegie Mellon University.

Sonia Joseph

Sonia is reverse-engineering video and robotics models to drive unprecedented advances in both capability and security. Despite the rise of "world models" like Genie 3, Veo, Cosmos, and Sora, we still lack a foundational science of world models– we don’t yet know if they truly capture causality or physical laws. This knowledge gap carries real risks: as video models are increasingly integrated into robotics, hidden failure modes or even malicious “sleeper programs” within model weights could lead to dangerous autonomous behavior. Sonia’s research develops methods to detect and neutralize these threats while rigorously probing how these models represent the world, including their understanding of fundamental physics. Her work seeks to establish a scientific framework for interpreting, securing, and improving world models.

About Sonia

Sonia is a PhD candidate at McGill and Visiting Researcher at Meta on the JEPA video understanding team, where she is writing her thesis on the interpretability of multimodal models and physical world models. Previously, she was a researcher at Janelia Research Institute and the Princeton Neuroscience Institute, and worked for various AI startups in San Francisco.

Before you even undertake the work to create a thing, there’s some assessment of whether creating it is possible and how valuable it will be.1 Some ideas take a large chunk of resources to assess — let’s call these ‘fat’ ideas.2 In other cases, a spreadsheet and well-founded assumptions can make a strong case that “it will be hard, but if we pull off all these things that don’t violate physics along this critical path, it will be incredible” — let’s call these ‘lean’ ideas.

SpaceX’s reusable rocket is a canonical lean idea: “If we can land a rocket (which is very hard but has a clear critical path) then we will increase launch cadence and reduce $/kg to orbit by an order of magnitude, which is a complete game changer.” Ideas can also be lean even without a clear value proposition or critical path if the cost of experimenting is sufficiently low. Many bit-based ideas are lean despite a lot of uncertainty about where they’ll end up thanks to software’s low cost and fast iteration speeds.

‘Basic Science’3 that winds its way to a useful technology a classic example of fat ideas: what started with curiosity-driven study of gila monster venom ended up as GLP-1 inhibitors; what started with poking at yogurt cultures ended up at CRISPR. However, fat ideas include more than just curiosity-based discovery. Most systems research — of the sort that were the precursors of a good chunk of modern technology — is also a fat idea.4

The dichotomy between lean and fat ideas warps which ideas get support, which ones get killed too early, and which ideas people choose to pursue in the first place.

If you want to maximize efficiency — returns per dollar in either financial terms or ‘impact’ — there is a clear strategy: quickly kill ideas and bias towards a thousand false negatives (prematurely killing an idea that would eventually work out) over one expensive false positive (an idea that fails to work out after pouring resources into it).

Imagine that lean ideas take 1 unit of work assess and will create 1000 units of value, while fat ideas take 100 units of work to assess but produce 1100 units of value. Each idea creates the same net value but if you care about returns on investment you should only assess lean ideas, ignoring every single one of the fat ideas and their potential results. In order to make more in-depth explorations worthwhile, the fat ideas would need to, on average, be two orders of magnitude more valuable than the lean ones.

Another framing is through the lens of scarce resources. If you have 100 units of work to deploy, your choice is whether to spend that 100 units of work on 100 lean ideas or 1 fat idea. If each project has the same expected value, you create 100 times the expected value by focusing 100% on lean ideas.

This assessment assumes that you even know up front whether an idea is lean or fat (but not whether it it is worthwhile or not). The fact that it’s hard to distinguish between many fat and lean ideas compounds the argument for a “lean assessment only” strategy.

In reality, ideas don’t fall neatly into a lean bucket or a fat bucket. Instead, there is a continuum of how much effort an idea takes to evaluate, . However, because we’re talking about a scalar evaluation effort, it’s reasonable to draw a line that is the maximum effort you’re willing to spend per idea. Everything to the left of the line is lean and everything to the right is fat. This bucketing system means that you lose out on the marginal idea immediately to the right of the line.

The advantage of the lean strategy has nothing to do with whether you’re trying to capture the value you create or not — as long as you prioritize efficiency, it makes sense for both financial investments and philanthropic spending.

The efficiency of the lean strategy drives many research organizations to either implicitly or explicitly pursue the lean strategy — opting for many false negatives over an expensive false positive. Indeed, many organizations play up the efficiency of their process: “we kill ideas fast!” And given the assumptions and constraints they’re operating under, they’re not wrong to use that strategy.

Let’s examine those assumptions:

1. Lean and fat ideas create roughly the same amount of value on average.

2. The difference in cost between evaluating lean and fat ideas is several orders of magnitude.

3. There are plenty of lean ideas.

4. It’s impossible to know a-priori how much effort it will take to evaluate an idea.

5. The results of lean and fat ideas are interchangeable.

6. Efficiency should be prioritized.

1. Lean and fat ideas create, on average, roughly the same amount of value.

This assumption seems reasonable — to my knowledge there is no comprehensive study of innovations, the value they created, and whether it was easy for people to evaluate whether they were worth pursuing a priori,5 but there are plenty of valuable examples of both lean and fat ideas. Reusable rockets, electric cars, transistors, airplanes, AI, artificial fertilizers, the structure of DNA, and atomic power (after a certain point) are all lean ideas — generally they’re characterized by either a gold rush or assertions that they are impossible to pull off. Lasers, cars, Unix, The Internet, germ theory, and solar panels are all fat ideas — generally they’re characterized by people calling them niche or useless and are preceded by a lot of “piddling around.”

Implicitly, this assumption relies on there being some one-dimensional “value” number that you can use to compare ideas apples-to-apples. Regardless of whether this assumption is correct, I’m going to push back against the underlying framework in point #5.

2. The difference in cost between evaluating lean and fat ideas is several orders of magnitude.

This assumption is also reasonable. Lean ideas are characterized by being able to put a lot of assumptions into a spreadsheet and possibly identifying one or two key experiments to validate some assumptions. Fat ideas often involve years of piddling around (and thus years of salaries and expensive equipment) before even having a clear idea of what the valuable output even could be.

3. There are plenty of lean ideas.

The frequency of fat ideas is irrelevant unless the frequency of lean ideas is so low compared to the frequency of fat ideas that you hit a tipping point where pursuing a lean strategy forces you to spend more resources throwing away ideas quickly than you would spend evaluating fat ones. Thus, the lean strategy is rational if you believe there are plenty of lean ideas.

It’s (almost? actually?) impossible to evaluate this assumption. My personal hunch is that a component of the Low Hanging Fruit theory of stagnation is that so many institutions have (rationally) pursued the lean strategy that lean ideas have become depleted. You could call lean ideas “low hanging fruit” themselves, but generally the “low hanging fruit” is understood to refer to the total effort to realize the idea, not to evaluate whether it’s worthwhile.

My other hunch is that while the distribution of evaluation costs is continuous, it is not flat. It likely has at least two peaks — one that’s lean-esque and one that’s fat-esque. If that’s the case, institutions pursuing the lean evaluation strategy can’t just incrementally increase their threshold for leanness as lean ideas are depleted.

The first three sets of assumptions all fit within a nice numerical framework. You can play with those numbers here.

4. It’s impossible to know a-priori how much effort it will take to evaluate an idea.

The lean strategy implicitly assumes that it’s impossible to know how much effort it will take to evaluate an idea (meta assessment!). If it were possible to know up-front how much work it would take to fully assess an idea, you could know how fat each idea actually is — whether it takes two units of effort or the full 100, and go after the lean-er ideas. The dream of program design as a discipline is to be able to cheaply determine how much work it will take to assess an idea.

5. The results of lean and fat ideas are interchangeable.

Collapsing value into a single number assumes that the outputs of pursuing different ideas are the same in kind. That is, the types of innovations that will come out of pursuing only lean ideas look the same as if you pursued both lean and fat ideas. Or at least the types of innovations created by lean and fat ideas are equally desirable.

Here is where thinking in terms of monetary outcomes violently diverges from a more nebulous concept of ‘value.’ If all you care about is cash in and cash out, the results of all ideas are interchangeable. However, I suspect that pursuing primarily lean ideas systemically biases towards certain types of innovations. As a result, whole classes of ideas are systematically neglected. It will take a lot more work to unpack the systemic differences between the outputs of lean and fat ideas, but my hunch is that a lot of software and single-node-in-a-system changes are lean ideas, while many general purpose technologies and new systems are fat ideas. Even without knowing which specific classes of innovations we’re missing out on by pursuing primarily lean ideas, I suspect that pursuing an “unbalanced” distribution of lean/fat ideas is unhealthy for long-term innovation because ideas feed on each other — fat ideas generate lean ideas and vice versa. The bias towards lean ideas may be a factor in why some technology domains feel like they have stagnated.

6. Efficiency should be prioritized.

The final assumption underlying lean strategies is that efficiency (value created per unit of input) is the most important thing. When you have a fiduciary duty to maximize shareholder value, this assumption is correct. There are many important innovations that are just bad monetary investments. Speaking to several high-ranking people who run corporate research labs, they constantly need to justify not the absolute cost of the work they’re doing, but the return on investment they generate.

You would expect the unbreakable coupling between profit and efficiency to suggest a clear opportunity for philanthropic money to step into the gap and enable the fat, inefficient assessments that can unlock entirely different classes of innovation. Disappointingly, I’ve found this to not be the case. Instead, the majority of philanthropic funding decisions I’ve run into are still implicitly or explicitly dominated by efficiency. Questions like: “What is the most impactful way to use this dollar?” “How do we minimize risk of failure?” “How do you know it will work?” “What will the impact of this be?” are all trying to maximize efficiency. Prioritizing efficiency is not inherently bad, but it is creating bad outcomes when it is killing fat ideas at a societal level. (Assuming, of course, that fat and lean ideas are not actually interchangeable).

Some final thoughts

Where does this all leave us? If the outputs of lean and fat ideas do indeed generate different types of innovations, a relentless prioritization of return on investment (whether in dollars or impact) has created a world of systemically skewed innovations. Many incredible discoveries and inventions have been sacrificed on the altar of efficiency. The lean-only approach is completely rational for profit-seeking businesses. However, institutions with other mandates, like philanthropy and governments, have also been captured by the same mindset. Prioritizing something besides efficiency is a heretical notion in the current zeitgeist (and there are many ways it can go wrong), but it would be powerful for institutions that can get away with it. Perhaps efficiency is overrated.

Related Reading

• ’s Studies on Slack

• ’s One Process

• ’s GUTS: The Grand Unified Theory of Striving (or Slacking)

• Rob Pike’s Systems Software Research is Irrelevant

Updates

wrote a thoughtful response here flagging that “thorough” and “efficient” are much better terms for what I’m describing than “fat” and “lean” and also digging into the nature of actually-fat ideas.

Thanks to Matt Clancy and Luke Constable for reading drafts of this piece and amazing ideas and Omar Rizwan for pointing me towards the piece on Systems Research.

Detailed work on how money was spent enables unique comparisons to the past. Historians have dug into the archives and poured over old account books to pull out the amounts spent on dowries, salaries, and household items in Renaissance Italy, comparing that spending to the yearly wage of a day laborer, and then used that wage to convert all of those amounts to contemporary dollars. From this work we can learn that a well-paid Renaissance professor made potentially more than a tenured professor today (up to $800k/year vs ~$200k/year) but what you could buy with that money was very different – a nice set of clothes could cost as much as a Lamborghini today. (See Ada Palmer’s excellent new book for more.)

As far as I know, nobody has taken a similar look at finances in the history of science and technology. Some historian friends and LLMs have corroborated this, so I’m pretty sure these are legitimately open research questions.

First, I’m going to describe these questions (each of which is a hardcore, dig-into-the-undigitized-archives-and-try-to-piece-things-together-from-account-books, history research project) and then I’ll explain why answering them is important (and just interesting and cool). Ideally, someone would find out the following for a wide range of technologies, not just the most charismatic ones:

• How long did the technology take from conception to being sold? Taking into account both when the person or group who started selling it successfully started working on it and when other people who maybe went bankrupt or never brought it to market. (Many people who invented important things went bankrupt!) And the corollary, how long did it take for the technology to go beyond being a niche thing once it was being sold?

• How much did a technology cost to develop in current $$? Again, taking into account both when the person or group who started selling it successfully started working on it and when other people who maybe went bankrupt or never brought it to market did work.

• What was the return on investment for the people who put money into creating the technology? This question goes for both the inventor and other backers. Ideally these returns would be recorded over time so you could calculate time-discounted returns (a dollar ten years from now is worth less than a dollar a year from now).

• What were the returns on comparable investments when a technology was created? If someone took a chunk of money they spent inventing a technology and bought bonds, real-estate, or stocks (if they existed) — what returns could they expect with what levels of risk?

• What fraction of people’s wealth did they put into research and invention? How much of their savings did the Wright Brothers put into the airplane? How much of his fortune did Alfred Loomis put into Tuxedo Park?

I’m particularly interested in the answers to these questions in the 19th and early 20th centuries. Before that period, the idea of deliberate invention for the sake of making a profit was more rare; after that period we enter the era of modern venture capital.

Why do these questions matter?

The answers to these questions add color to the assertion that “we have picked all the low-hanging fruit.” It might be that indeed, it was once faster and cheaper to create valuable technologies. But I don’t think that’s obviously true. It may also be that our expectations around timelines, risk, or returns have shifted so that comparable work that yielded “low hanging fruit” in the past are no longer rational today (or were never actually rational).

We’ll learn things about how the enterprise of research has changed over time. It might turn out that building technology has become a worse (or better!) investment. Likely the answer depends on the type of technology. More generally, we might see patterns that hint at constants in the “nature” of technology.

We would also learn things about ourselves. The questions about comparable investments and fractions of wealth would give a sense of how risk tolerance around doing science and technology has changed. (I have a personal hypothesis that collectively we’ve all become much more risk averse.) We’ll also see how the personal value that people put on tinkering around or doing science has changed over time.

Finally, it would just be cool to know! Having more concrete information about the nitty-gritties of technology development in the past is another lens on history and the human enterprise.

If you’re a historian or funder interested in collaborating on this, get in touch!

Frontier AI Tamperproofing (FAIT)

Taylor Kulp-McDowell

Read more

• Discovering new materials with high-throughput automated labs

• Building nano-manipulators for atomic-scale manufacturing

• Creating datasets that could unlock precision medicine for neurology

• An automated atlas of high-biorisk research

• Fixing copper shortages with e-waste

• Open-source AI-manipulable CAD software

• Electric ammonia production

• Diagnosing and preventing age-induced cognitive decline

• Preventing the misuse of open-source AI models

• High precision atmospheric aerosol and climate models

• Automated manufacturing design

• Biosurveillance systems to nip outbreaks in the bud

• Precision forecasting animal movement

• Transmitting electric power directly through the ground

• Carbon conductors that could replace copper

You can see the fellows’ profiles and brief descriptions of their program ideas here.

If you or anybody you know would like to connect to any of the fellows or learn more about what they’re working on, feel free to reach out to them directly or contact us at brains@spec.tech.)

These results were a team effort. Thank you to so many people who were critical to the program’s success: the mentors who worked with the fellows, the research leaders who shared their stories, the team who helped run the events, our awesome advisors and ops team, and more.

There are compounding returns to talent programs like Brains: repeated iterations enable us to refine our selection criteria and how we help the fellows; if you can maintain quality, you will build a reputation that both attracts more of the right people and makes it easier for them to be successful; and a growing community of people who have gone through a program create powerful network effects. As such, we plan to run regular cohorts as long as we can, with the next one tentatively kicking off next January.

So:

• If you have an ambitious science or technology research idea that is a poor fit for a single academic lab and doesn’t make sense as a startup (or are already working on one), keep your eyes peeled for when we open applications for cohort three!

• If you know someone who might be a good fit for the program, please`` route them our way!

• If you work at an organization that might want to hire or fund current or future fellows, or support the Brains program itself – please get in touch!

Some tactical meta-lessons

I will do a longer lessons-learned post about the entire cohort soon, but in the meantime I wanted to leave some Tactical Meta-Lessons:

• We changed the format from last year and did not do Q&A for each talk. I think this was the right call — it maintained more momentum and let us compress the event into a single day. There was a tradeoff in that it made the audience more passive but I think it was worth it because the topics were so diverse that only a fraction of the audience would have questions for any given speaker and we left copious breaks

• Speaking of breaks, copious breaks where people can mingle is critical for any event like this. The reality is that talks are basically teasers for “talk to me during the break.”

• In retrospect, the correct videography move is to focus entirely on the speaker, ignore the live slides, and expect to add them back in post. It’s never possible to capture both the speaker and the slides perfectly and video editing has gotten so easy that it makes more sense to put digital slides directly into a digital video instead of needing to go digital → analog → digital.

• There is one way to win at any presentation to a non-expert audience about a technical subject: shifting someone’s mindset from “I am not interested in <general topic>” to “wow, that’s really cool.” The way to do this is not to present a lot of evidence for why they should care, but instead to try to teach them something that gives them an “aha!” moment.

Don’t forget — Applications for an AI-focused Brains cohort close June 16th!

Apply Here

We are doing two Q+A sessions about AI Brains in the coming weeks:

1. June 3 at 8p ET — Register here

2. June 9 at 1p ET — Register here

Applications close June 16th

On Monday, June 24 at 2-3pm ET, we’re hosting our very first “Shareholder Meeting” on Zoom! It’s a chance for anyone who’s ever funded Speculative Technologies to hear what we’ve been up to, get a peek at what’s coming next, and ask us anything.

While nonprofits technically have no sharehold…

Read more

You might have noticed that the world of AI research is moving blisteringly fast. The world needs ambitious, high-leverage AI programs whose scope is bigger than a single company or lab. That’s where AI Brains comes in.

This 4-month, part time program is meant to help talented researchers with experience in AI hardware and software build skills, refine ideas, and make the connections to spin up coordinated research programs in governments and nonprofits.

If you or someone you know:
1. Has hands-on experience building the modern AI stack (software or hardware)

2. Has ideas for one or more ambitious AI research programs that are beyond the scope of a single company or lab.

3. Wants to start a nonprofit or government program to execute on that idea (or has already started and wants to take it to the next level).

You should apply (or encourage them to!) Applications close on June 16th.

I want to caveat up front that this analysis is only about humanoid robots in manufacturing. There are many situations where you’re neither optimizing for efficiency nor able to reconfigure environments easily where humanoid robots might have a big impact: these are your janitors, housekeepers, and gardeners. Those environments are incredibly unstructured, so people often point to manufacturing, which involves more repetitive work and structured environments, as the first useful application for humanoid robots.

When people imagine humanoid robots, they’re usually assuming that they can do basically anything a person can do. While I do want to flag how hard that will be to achieve, I think the more interesting question is “in a world where AI is good enough to enable human-parity humanoid robots, what other manufacturing paradigms would be unlocked? How do they compare to just dropping humanoid robots where we have people right now?”

We could break these big questions down into three, slightly more tractable, ones:

1. What does it take for a humanoid robot to be at cost parity with a person?

2. What role do people actually do in the manufacturing process? What will automating that do to the speed and price of manufactured goods?

3. How do humanoid robots compare to other, more specialized forms of automation?

This piece does a rough numerical analysis on cost and then some more qualitative analysis on the latter two questions.

Spoiler alert: My hypothesis is that in a world where there is good enough software and hardware to create humanoid robots that are as good and flexible as a human at manufacturing tasks, we will also be able to quickly create more task-specific hardware or use less complex hardware that can do those roles cheaper, faster, and better.

Cost

One can evaluate a manufacturing process on a few axes:

• How cheap can you make a thing?

• How high-quality is a thing?

• How quickly can you spin up the ability to make a new thing/adjust how you make an existing thing?

The costs that go into a manufacturing system can be broken down into upfront costs and ongoing costs. Upfront costs include hardware/capital, system design, and programming robots or training people. Ongoing costs include maintenance on machines, salary for people, associated operational costs like light and heating for humans, electricity for robots, etc.

In the limit, you would be able to instantly spin up a manufacturing process that made high-quality widgets for the price of the raw inputs. In reality, there are tradeoffs: we all know that you can often make a thing cheaper by reducing the quality; making an automated system to rapidly spit out widgets can drive down the cost per widget but it often requires a ton of setup time and cost to get the tooling right.

Let’s look at (extreme back of the envelope) costs for three different systems:

1. Humans

2. Humanoid Robots

3. Specialized Manufacturing Equipment

I’m going to talk about top-level costs, what factors drive them, and the assumptions that go into them. You can see a much more detailed spreadsheet of the numbers I used here. If you think I am massively wrong on any of these, please let me know!

Costs — Humans, Humanoids, Specialized Machines

Human labor costs anywhere between $20k/year (this would be your developing world sweatshop) and $200k/year for skilled American labor. While there are costs associated with tooling and system design, the vast majority of human labor costs are salaries.

Humanoid robot labor costs somewhere between ~$60k/year and ~$13k/year. These numbers involve a metric fuckton of assumptions, considering there are no existent general purpose humanoid robots in use in factories.

The first assumption is around the cost of hardware. The high number is $150k, which is the cost of Unitree’s top-of-the line unit without hands or grippers. Even that price assumes costs come down because grippers or hands can often be half the cost of a traditional robot arm installation. A hand that has all the dexterity of a human hand costs ~$100k on its own right now. While you don’t need a hand with all the joints and dexterity of a human hand to do most manufacturing tasks, there is a big difference between being able to be a drop-in replacement for literally all human tasks and a finite set of them that you can do with a simplified gripper. Another way to get this higher humanoid price is to note that an actually-existant commercial robot arm costs ~$30k and at bare minimum a humanoid robot is approximately four robot arms and a ton of sensors in a trenchcoat.

The low number for hardware is $16k. This is the sticker cost of Unitree’s cheapest humanoid (also without hands). Currently, the $16k model is basically a toy, but there is potentially a world where economies of scale and learning curves could drive the cost of a functional humanoid this low. (Personally, I’m skeptical.)

I’m also assuming that it is more expensive to program a humanoid robot than it is to train a person, but not absurdly so. This assumption does mean we’re not yet living in a world where robotic ~intelligence is completely at parity with humans. However, if you think that’s going to happen soon, the analysis doesn’t hinge too heavily on training cost.

All mechanical systems require maintenance and break down. The numbers for humanoid robots around maintenance costs and lifetimes are also based on the approximation of humanoid robots as four robot arms in a trench coat. An interesting cost driver here is that the standard accounting lifetime of robot arms is five years, which is closer to electronics than the 15-year lifetime of heavy machinery.

Specialized machines cost ~$70k/year. Most of this cost is driven by the tooling/hardware cost, which is extremely variable based on the actual machinery. What’s notable is that the cost of hardware is roughly order-of-magnitude the same as a humanoid robot.

Labor Costs Per Widget

The cost of labor only matters insofar as it affects the cost of making a widget. One of the advantages of robots and machines over humans is that they don’t need breaks or sleep, so while they might be more expensive, they can work 24 hours a day... ish.

All mechanical systems need regular maintenance and also break down unexpectedly. Planned downtime is often ~10% of total potential working time, with failure rates accounting for another 2.5% of theoretical potential uptime. Maintenance and mechanical failures is one way in which robots are very different from computers! Of course, even with downtime, machines can work many more hours per day than a human.

Another big factor in the labor cost per widget is just how fast a system can produce a widget. Here is where humanoid robots fall behind specialized machines: humanoids produce things at the same speed as a human, while a specialized system can move many times faster on almost any given task. Many videos of automated assembly lines need to be slowed down to make it not just a blur. The analysis assumed a modest 5x speed increase for a specialized system over a human or humanoid.

Given working hours, downtime, failure rates, and speed assumptions, the labor going into human-made widgets costs $0.48 per widget on the low end and $4.31 on the high end, humanoid labor cost $0.54 per widget on the high end and $0.12 on the low end, specialized-machine labor cost $0.13 per widgets. (Obviously, all these numbers are far more precise than reality.) So, humanoid-made widgets are somewhere between the cheapest human-made widgets and specialized-machine-made widgets.

These numbers are all based on the assumption that specialized machines effectively stay the same. We’re comparing an actually-existing technology to a hypothetical one. In reality, I would bet that the same capability increases that enable functional humanoids will also improve specialized machines. If you only assume that the costs of system design and programming for specialized machines drop as much as they do between expensive humanoids and cheap ones, the cost of specialized-machine-made widgets drop to $0.07 per widget, almost half of those produced by cheap humanoids.

Labor as a fraction of total cost

So far we’ve been ignoring the fact that labor is only one input to the total cost to manufacture something. Usually the other biggest cost driver is the cost of the raw materials, but yield rates (how often there’s a defect), depreciation on the factory, energy, quality assurance, and a long tail of other factors also contribute to the final cost. Here are the labor fractions of the cost of some common items, just to get a sense of the labor fraction of the cost of manufacturing:

• Car: 7%

• Airplane: 30%

• Toaster: 50%

• Electric Motor: 50%

So there is actually a lot of room to decrease costs by automating labor — generally these are products that require a lot of fiddly final assembly.

Qualitative Arguments

At this point, you probably have a lot of objections to this extremely-assumption-laden analysis.

If a process could have been automated by a specialized system, someone would have done it by now, therefore we need humanoids to automate the rest. This is true, assuming a fixed level of technology in specialized systems. However, in a world where humanoid robotics are good enough to tackle most currently-not-automated tasks, we will have big advances in control, perception, and planning that will not be humanoid-robot specific. Those advances will make many more situations automatable by specialized systems that will be faster and more reliable than humanoid robots.

The real advantage of humanoid robots is not on cost, but setup speed: they make it so you don’t have to make or buy specialized equipment so you can retool very quickly. There are many processes that one could automate today with existing technology, but it’s not worth it because there’s not enough demand for a huge volume over many years. (We could shoehorn that into the price calculation above as a much shorter depreciation time, which would indeed drastically increase the price of the specialized system.) However, a world where humanoid robots are both smart and capable enough to be easily “reassigned” to a new task is one where we’ve had massive advances in computers’ ability to understand manufacturing processes, manipulation, planning, etc. These advances would open a whole host of harder-to-imagine, but likely better, options than humanoid robots.

One possibility is that these advanced capabilities could make designing, building, and modifying specialized systems dramatically faster and cheaper.

Another possibility is that advanced capabilities could enable simpler robots to do the job just as well. The vast majority of tasks don’t need a full-on humanoid. Instead, many tasks could be done by something like a single robot arm on a wheeled base. Or two hands on sticks. Or itty bitty spider robots. In a world with the robotic capabilities for a humanoid to be as good as a person, it’s extremely likely we would have the capabilities to have all of these form factors and dispatch them appropriately, which would be cheaper, faster, and less prone to failure. It’s just harder to imagine than drop-in replacements for people.

The reason simpler robots are cheaper and less prone to failure has a lot to do with the sheer complexity of humanoid robots. Humanoid robots with dextrous hands have roughly 70 “degrees of freedom.” Degrees of freedom (DoF in the lingo) are “ways a robot can move” — so one joint that can move in only one way (like an elbow) has one DoF, a joint that can move in three dimensions (roll, pitch, yaw, like a shoulder)1 has three DoF, etc. Each DoF adds additional complexity, cost, and ways for the robot to fail; in many robots, each DoF has a motor controlling it and each of those motors add costs and potential independent failure modes. The probability that a system composed of multiple independent failure points is

which means that each additional failure point has a pretty significant effect on the entire system’s failure probability. The upshot of all of this is that you really want to minimize the number of DoFs you’re using.

Humanoid robots will be like microprocessors: a general purpose tool that is so much cheaper and easier to use than its specialized counterpart that most people use it, despite not being optimal. In the world of microprocessors, you can create a custom chip that can do almost any specific computing task better than a general-purpose chip. However, there are only a few applications where the time and cost to create a custom chip is worth it. One could imagine a world where the same is true for robotics. While I can’t rule this out, I think there are some key differences between robots and microprocessors that make this scenario unlikely. The vast majority of the cost of a chip is in the design and initial tape-out, so there are huge economies of scale: you need to make many of that specific chip to make the cost per chip reasonable. On the other hand, a lot of the cost of a robot is in the components – the motors, sensors, and raw material. These components are general purpose – as you drive down the cost of the motors for a humanoid robot, you also drive down the cost of other robots that also use motors.

Humanoid robots being general-purpose and easy-to-train at all assumes that we have much better robotic software. (I realize I sound like a broken record at this point.) It’s likely that the same software that makes humanoid robots work would also make non-humanoid robots work. It would also be a strange world where we have software that enables humanoid robots to easily do most work but we don’t have software that can help design better specialized systems, or teams of specialized robots.

If it’s just as cheap and easy to use the specialized tool as a general-purpose tool, you’ll use the specialized tool. My hard-to-verify assertion is that the same capabilities that will make general-purpose humanoids cheap and easy will also make more specialized systems cheap and easy, unlike microprocessors and custom ASICs.

The point of humanoids are as drop-in replacements for people in an existing manufacturing line. This is a fair point. So far we’ve been talking about creating new processes, not just replacing humans in existing processes. There are several awkward things there: it will take some time for humanoid robots to get good enough to be drop-in replacement, and by then many systems will have been overhauled. Overhauled systems will increasingly be built around actually-existing automation, which will not be humanoids. It will be a sad state of affairs if we keep everything the same.

Conclusion

General-purpose humanoid robots will be useful for many things, especially in environments that are primarily populated by people. However, manufacturing is not one of those environments. Most people think about the difference between humanoid robots and other robots along two axes: (1) morphology and (2) intelligence. People tend to assume that the two axes are coupled, but that isn't true. Assuming that AI breakthroughs spill over to robotics, we can make larger gains applying vastly improved automation algorithms to “classical” robot arms, specialized robots, or designing and building custom robots themselves.

In high-volume situations, it already makes sense to build specialized tools that can produce widgets far faster than humanoids. The same technology that we need to make useful humanoids will drive down the volume that makes sense for specialized machines. As the gap between human and robotic capabilities closes in low volume situations, the same software and hardware that will enable useful humanoid robots will also enable cheaper, more robust, and faster automated systems that are just harder to imagine right now.

Thanks to Brian Potter and Matt K for reading through drafts of this piece.