A Cascade of Conscientiousness

Launching FAI's new Physical Intelligence Project

The revolution in digital machine intelligence is now thoroughly underway, but the counterpart of this revolution in the physical world has barely begun. The autonomous taxis that drive in a handful of American cities are among the few concrete examples we have of the transformations that are brewing. Back in the digital world, many of us have seen demonstrations of humanoid robot prototypes doing karate, dancing, folding laundry, or sorting parts in a factory. Perhaps these videos offer us a glimpse into our future, but if they do, it is but one narrow slice. The transformations wrought by the Cambrian explosion of thinking machines that will soon roam the real world will be more surprising and counter-intuitive than we can imagine.

But therein lies the problem: the ideas for the applications of “robots” with the greatest potential are nascent, and nascent things are fragile. In today’s America, “surprise,” when it happens in the physical world, is often unlawful due to an immense thicket of regulatory and legal guardrails designed to preserve the built status quo. These very same rules are also blockers of American reindustrialization, complicating the construction of the factories where the drones, autonomous vehicles, robotic arms, quadruped robots, and humanoids of the future will be built. This institutional complex therefore imposes the principal impediment to both building and diffusing the physical-world technologies of the future. Put more simply, these laws and regulations, unchecked, will hinder us from bringing a brighter, richer future to fruition.

Environmental permitting and land use laws are one important part of this institutional complex, but only one. Just as often, the legal culprits blocking autonomous technologies are to be found in regulations far afield of these more familiar targets of accelerationist ire. Just like environmental permitting and land use, few of these regulations were put in place for no reason. Many exist for fundamentally good reasons, or at least reasons that were at one point fundamentally good. Grappling with this institutional complex, then, will not always be a simple matter of saying “afuera!” Success requires a multi-disciplinary effort that must exercise utmost technocratic and political sensitivity.

Today, the Foundation for American Innovation is pleased to launch just such an effort. The new Physical Intelligence team will sit underneath the existing AI team, but borrow expertise from other FAI policy verticals as well. It will also, of course, bring on expertise of its own, starting with our non-resident fellow, Amelia Michael. As we see it, our scope incorporates several streams:

1. Creating a favorable regulatory and legal climate for experimentation with and scaled deployment of autonomous technologies in the physical world.

2. Understanding the technical trajectory of robotics and other autonomy technologies, with an eye toward informing policymakers and the broader public about the likely contours of future technological developments.

3. Articulating the industrial strategy that will be required to ensure the core technologies of physical autonomy are designed and manufactured in the U.S. or its allies.

4. Developing policy frameworks for liability, cybersecurity, and similar areas of the emerging law of robotics.

If you are interested in working in any or multiple of these areas, please get in touch. We are actively seeking talent of all backgrounds, ages, and levels of seniority. It is our belief that the best scholars of this new field will come from a diverse range of backgrounds.

We are articulating this statement of purpose today because we hope to inspire others—entrepreneurs, researchers, policy analysts, policymakers, and interested citizens. We hope you concur with us that the vision we cast is worth pursuing, and that the hurdles we foresee are worth surmounting. Above is the short version of our story; the slightly longer version follows.

—

A central challenge with autonomy is that the low- and medium-hanging fruit has already been plucked—sometimes long ago. Agriculture, mining, commercial aviation, and much of manufacturing are already heavily automated industries. Can one automate the truck that transports raw ore from a mine pit to a processing facility? Yes, and many have. But the wage of the human truck driver pales in comparison to the fuel required to get a 400 ton, fully-burdened truck from the bottom of the mine to the top, along with the maintenance of the truck’s engine, tires, and the like.

Thus, end-to-end automation of open-pit mines may not yield the dramatic reductions in price or improvements in productivity one might expect from “end-to-end autonomy.” Similar dynamics apply in numerous other industries. The characteristic that unites these industries is that human labor is not the binding constraint on transforming the underlying economics (incidentally, very often, the binding constraint in these domains is instead energy cost or some direct derivative thereof; FAI also supports the radical abundance of energy for this reason). While we applaud such efforts at automation, these are unlikely to be the applications of physical-world AI that yield epochal benefits for the American people.

Instead, physical-world AI will transform the world when the binding constraint in the real economy is the need to have a reasonably intelligent and dexterous human being on-site, at the right time, in the right place, to perceive, decide, manipulate and act. More specifically, though noting that the following conceptual examples bleed into one another:

1. Industries where human labor drives total cost, and that cost hinders affordability. Construction is a canonical example of this, where productivity has stagnated for decades. It is dispiriting, perhaps, that we have not gotten any better at the fundamental human task of “building a home,” despite decades of technological progress. But on the other hand, fundamental, automation-driven improvements here have the potential to restructure the cost basis of the largest asset most Americans will ever own. Other examples include last-mile logistics, eldercare, housekeeping, and building maintenance. In some cases, this problem is better conceptualized as the opposite side of the coin: industries where human labor scarcity is the binding constraint on doing more of an activity we would like to do more. Here, the issue is not so much cost reduction as it is simply being able to supply services at the current levels of demand to begin with. Examples include rural healthcare delivery, environmental cleanup, disaster response and other hazardous operations, civil infrastructure inspection, and, most especially, skilled trades (electricians, plumbers, welders, etc.).

2. Novel activities that become possible only with massively scalable intelligent physical presence and/or by removing the limits of human bodies. These will be activities that benefit from continuous, near-free intelligent presence or that occur in physical locations hostile to human biology (space or extraterrestrial construction, oceans, Arctic environments, etc.). There may also be some domains where exceptionally precise autonomous systems enhance existing, currently human-dominated activities (for example, in advanced manufacturing or the budding field of robotic surgery).

In all of the domains above, and especially when considered in combination, it becomes clear how physical-world AI can deliver a radically richer and safer life for all Americans.

The most transformative deployments of physical-world AI will not be those that shave a few percentage points off the cost structure of already-automated sectors, but rather those that make intelligent physical presence abundant in sectors where it is currently scarce. The revolution of physical AI will feel like a cascade of conscientiousness has dawned throughout the built human world.

In much of the real economy, the binding constraint is not energy, raw materials, or even capital. It is the need to have a human being physically present to perceive conditions, exercise judgment, manipulate objects, and bear responsibility for the result. Where physical AI can relax that bottleneck, the gains are multiplicative.

This points our attention away from the most visually striking demos and toward the most economically consequential domains: the building of homes, factories, and infrastructure; the movement of goods through ports, roads, warehouses, and delivery networks; the inspection and repair of grids, pipelines, rail, and water systems; the execution of dangerous tasks in disaster zones and military theaters; and the provision of care and assistance in a society with too few workers to meet rising demand. These are settings in which the central scarcity is often not brute force but reliable, on-site cognition and action. Once that scarcity begins to ease, the effects will be felt not at the margins but throughout the structure of the economy. This is the industrial revolution.

—

A few points bear further elaboration. First, the conception of physical-world AI we have laid out above does not exclude the use of software-only AI to transform physical-world industries. Digital-only AI agents are likely to play key roles in automating labor-intensive administrative processes in many such fields (for example, in manufacturing), and they are also likely to serve as key orchestrators of robots, autonomous vehicles, drones, and the like.

We therefore reject the notion that innovation in “the world of atoms” is in conflict with innovation in “the world of bits.” While we concur that America is in sore need of a more rapid pace of world-of-atoms innovation, we are inclined to see digital and physical innovations as complementary to one another rather than at odds.

Second, our vision of the future does not deny, but also does not rest solely upon, the success of robotics in the human form factor. Humanoid robots carry tremendous promise as drop-in replacements for what would have been human laborers. This is especially useful for the automation of labor within environments originally built for human workers. We are enthusiastic about the innovative startups in this field for that reason.

Yet we also bear in mind Henry Ford’s famous words: “if I had asked my customers what they wanted, they would have said a faster horse.” Many of the industries we have identified could be entirely reimagined using autonomous technologies, and made all the more productive because of it. Consider the now-ancient example of heating a large room: once, this could only be accomplished with great fireplaces and chimneys, which had to be maintained by human beings. Sitting from the vantage point of someone before the dawn of mechanized heating, one could imagine humanoid robots tending the fires as a great productivity improvement. Yet radically more productive—and better for human health—is abstracting away the need for indoor fire altogether through fully mechanical heating systems. It is transformations of this kind—hard to conceive of almost by definition—where we suspect the most profound gains are to be discovered.

--

Finally we turn to the regulatory challenges, which even a cursory perusal of our vision would suggest are plenty. We believe the most helpful way to conceptualize the regulatory barriers to the world we hope for is as follows. First, there are the challenges that come from when autonomous systems are acting as one-to-one replacements for what was once human labor. Self-driving taxis and trucks are the best present-day example we have of these. Second are the challenges that come when autonomous systems cause an existing industry to be reimagined in some fundamental way; a still-speculative but tangible example here would be autonomous drones replacing humans for last-mile delivery. Third are the challenges where autonomy enables a wholly novel task: the construction by robots of Lunar infrastructure, say.

This can be pictured as a spectrum of regulatory obstacles. On the one end, when autonomous systems are replacing humans, the challenges are likelier to skew political. Existing regulatory barriers to adoption (for instance, licensing or bonding requirements) will become sticking points and will likely be tweaked for greater effect; new regulatory barriers will be erected where none previously existed. Some of these will be federal laws, but a great many of the regulatory actions will probably play out at the state and local level, since that is the current central locus of regulation for many of these industries. Opposition here is likely to be the fiercest. On the positive side, however, it is here that existing regulatory standards are likeliest to give innovators a clear benchmark to strive for with respect to safety, reliability, and quality.

Toward the middle of the spectrum are instances where autonomy prompts innovators to reconceptualize how existing industrial activity is performed. Here, the challenge will likely stem from a combination of regulatory novelty (there is less likely to be an existing regulatory framework within which to operate, nor a standard to which innovators can aspire), and broader public rejection (for example, local communities finding last-mile drone delivery to be a nuisance of sight and sound). To be sure, there will be some political challenges from incumbents, but this direct pressure is likely to be lower because the innovator’s technological orthogonality also plausibly gives him legal orthogonality. This can alleviate the problem of incumbent regulatory capture. The last-mile delivery drone operator, is principally regulated by the Federal Aviation Administration (FAA), an agency that, despite its challenges, is less likely to be captured by the ground-based, human-operated freight and last-mile delivery firms the drones are seeking to upend.

Finally, at the extreme end of the spectrum lie the altogether novel activities: autonomous construction in space and at sea; persistent and distributed remediation of the natural environment; infrastructure that repairs itself; roads that reconfigure themselves on demand; perpetual and autonomous scientific fieldwork at global scale; and dare we say, flying cars. These are by definition the hardest applications of which to conceive, and the most speculative. Yet they may well be the most economically promising in the long run. Even if they are not, they are surely the most fun. The central regulatory challenge here is simple, though by no means easy to fix: the structural problem that American administrative law generally requires a pre-existing regulatory category for any activity conducted at scale in the physical world, and novel activities by definition lack one.

The positive story here is that America has faced this challenge before. At the dawn of the First Industrial Revolution, the common law system that America had inherited from England essentially prohibited novel uses of land. It was, indeed, the ultimate “NIMBY” legal regime: no land use was permitted under the common law beyond the “natural” uses, with “natural” often tautologically defined to mean “current.” When America was a young and hungry nation, common-law judges had the ambition to think beyond our inheritance from old Britain and dream of what kind of legal superstructure best suited our budding dominion. The question we face today is not so much “is it possible?” but instead “do we have the will?” We will aspire, if not aver, that the answer is a resounding yes.

This spectrum of regulatory challenges is also a spectrum of regulatory solutions, or at least of regulatory temperaments. To the extent physical AI is replacing an existing human activity, we really should aspire for dramatic improvements in at least some majority configuration of quality, safety, reliability, and cost. To the extent it is enabling new activity under the sun, regulators, legislators, and courts should be more lenient, particularly when human life is not in imminent danger. Yet we must be tolerant of errors. Make no mistake: a machine-enabled future means machine-enabled tragedies, both accidents and those intentionally caused by malicious actors. We must be steely-eyed about this, not cowed; eyes fixed icily on the prize of future prosperity.

This raises the flip side of the regulatory question: how to police the activity of malicious actors using all this physical-world AI for nefarious ends. Criminals, foreign and domestic terrorists, and other evildoers will not wait for regulatory seals of approval or ISO standardization to adopt autonomous technologies for their purposes. Curiously, America’s chattering and policymaking classes seem much more eager to police the activity of lawful actors (innovative startups attempting to enrich the physical world) than they do to discuss all the perils non-lawful actors may soon pose.

Dealing with novel, technologically contingent methods of crime will almost certainly require governments to adopt novel technologies of their own. A criminal with a swarm of autonomous drones is unlikely to be thwarted by a baton or a Crown Victoria. Police will need tools of their own to fight new kinds of crime. Yet a police force equipped with, say, fleets of autonomous and intelligent drones is a qualitatively different kind of government presence than Americans are accustomed to. Bright lines in the sand restricting what governments can do using these technologies, enforced in highly transparent ways, will therefore be essential. While we do not view such policy discussions as immediately in-scope for our project, we do not rule them out for the future. We see this principle as important enough to articulate now: ordered liberty will rest neither in denying governments access to the technologies of the future nor granting governments unfettered use of those technologies.

Even more broadly, autonomous technologies of both the physical and digital worlds—“AI,” writ large—will enable qualitatively new kinds of governance institutions and regulatory mechanisms. A team of robots constructing a building can record every action they take, with precise telemetry to boot, at every instant. The need for “government inspectors,” in that world, is questionable. Instead, “auditing” and “inspection” become dramatically lighter weight; the regulator’s agents can check in at essentially no cost, and compliance becomes verifiable in seconds. The texture of “building inspections,” as an institution, the cost basis of that regulatory function, and the content of building codes themselves may therefore be enabled to change radically, if we allow it to happen. We mention this example not out of any conviction that it, in particular, is likely to come to fruition anytime soon, but as a reference class for the new institutions technology may soon make possible.

--

Here, then, is the vision. Now the problem becomes how to realize it. The United States has many strengths relative to other advanced capitalist nations. We are a frontier economy, with both frontier needs and wants, and the high labor costs that come with that status. We have a genuine need to reduce costs—a market-driven motivation—in many key areas of the real economy. We are blessed with deep capital markets and robust technical talent to invest into new technological challenges. We possess the most dynamic and innovative peoples of anywhere on Earth. With our allies, we are the beneficiaries of a mighty industrial base.

Yet our weaknesses are not to be written off. That industrial base is fraying in many areas, with outmoded approaches to production and an aging labor force in many areas of manufacturing especially. China, our principal geostrategic competitor, has built an enviable and innovative manufacturing capacity that we cannot, and almost surely will not, emulate out of whole cloth. Our only solution is to leapfrog with better tools and superior intelligence. America has done this before, but we are not used to lagging. Here, America is the underdog once more in at least some ways; we must learn to think like one. The industrial strategy we articulate will be clear-eyed about this reality. Yet it will also be optimistic about the opportunity: a leapfrog founded upon the reimagination of manufacturing and industry based on the technologies of the present rather than the past.

More profound still: America was a much less thickly institutionalized place when the last industrial revolution took place. In many ways this was bad, but it also meant that there existed far fewer institutional procedures for small interest groups to impede technological progress. At the civilizational level, we have grown older, our bones more brittle, our joints stiffer, our feet less nimble, our eyes slower to scan, our minds more hesitant to decide. How do we find our vigor once again? How do we rediscover our appetite for surprise, our tolerance of the uncertain, and our desire for the new?

Though it may not always be superficially obvious in the policy briefs, draft statutes, and analytical prose we expect to form the bulk of this project, these questions will always be on our minds. Beneath the surface of our wonkery, we will search tirelessly for that American incandescence—to brighten it wherever we find it still shining and to rekindle it where we find only its remnants, so that it may light the world for centuries more.