Is Software Eating the World?

Marc Andreesen's clever quip that “software is eating the world,” introduced a software-centric perspective that has shaped hundreds of billions of investment. The unstated belief lurking behind this quip is that hardware is an enemy which interferes with our ability to scale profit – the good stuff. According to this logic, business models that depend on infrastructure are likely to be quagmired in red tape and regulatory issues. A software-only axiom proved valid for enterprises like AirBnB and Uber, but what about robotics and autonomy? Is infrastructure a friend or foe? A deeper question is whether we want individual intelligence, centralized intelligence or something in the middle?

I asked Dall-E to proffer its view of an infrastructure-centric approach to self-driving and got this creepy image in response...

Today I am working to commercialize new robotics technology at DARPA which long ago created the greatest connection engine of all time -- the internet. Two decades ago, we began DARPA's early efforts to push autonomy, but today, DARPA is asking a fundamental question echoed by the Pentagon and the world at large: “Why are we struggling to reap the rewards of autonomy at scale?” Over the past decade billions have flowed into both defense and commercial markets. Technological capabilities have progressed admirably, but large-scale impact in terms of roadways, smart cities, and transportation has flagged. Despite the hype, we haven’t decreased the number of accidents on our roads or the level of congestion.

Perhaps this next image, generated by AI, offers insight into how the Silicon Valley world views reliance on infrastructure:

Within the autonomous driving community, the mindset for two decades has been that dependence on infrastructure is a crutch that can’t be guaranteed. Echoing Marc Andreesen, investors put their money into individual "on-board" AI as a scalable means to support autonomous behavior. Although these individual AI approaches purported to be self-sufficient, they ended up depending on off-board helping-hands including the comforting pulses of satellites in space, cell towers and cloud-based maps. This influence is harder to see than light posts or road-side equipment, but it still fails the test of self-sufficiency. In fact, although some self-driving companies are not forthcoming with details, recent crash investigations indicate that self-driving cars may end up stuck when they disconnect from the cloud. In fact, this is why so many cities have noticed problems with cars stopped dead, blocking the flow of traffic. Some companies use human supervisory control to help when cars have problems, but this hardly helps if the connection fails.

The central question is how to accelerate commercial adoption and limit dependence on off-board inputs that can't be guaranteed. What if GPS, cloud-based maps and off-board planning are even more vulnerable than the dreaded road-side infrastructure? Today, the smart infrastructure is not a given, but what if that changed and the necessary connectivity and positioning infrastructure was peppered into our environment the same way that ants use pheromones. If this sounds outlandish, consider that not too long ago, we went from a limited number of cell towers to countless microcells distributed through our cities.

Perhaps a history lesson can be helpful. In order to support burgeoning satellite radio, Sirius and XM Radio embedded repeater nodes throughout cities. Interestingly, when innovators at Sirius and XM Radio originally tried to tell investors the truth about the required infrastructure the investors balked. They modified their pitch to focus only on the two overhead satellites despite the fact that a proliferation of infrastructure was necessary. Fearing the travails of infrastructure rollout, the original wireless company, Wireless Telegraph & Signal Company, focused on building two giant wireless stations. Marconi's board of directors considered it less onerous to focus on these two giant transmitters than to take on the headache of installing on multiple ships. This was a mistake! As World War I began to foment, Marconi figured out that peer to peer communication between ships was a much better business model as well as a more robust technical solution. What would Marconi think today with microcells and cell phones broadcasting from every pocket and street corner?

We need to become swarm and becoming swarm means that we embed intelligence into the environment. Embedding command, control and communication into the environment helps address chaos and uncertainty. Swarm intelligence engenders fault tolerance and enables systems to fail gracefully. Within the insect world, entomologists call this principle of swarm intelligence stigmergy. Pierre Paul Grasse noticed in 1952 that ants can coordinate massive cities without explicit maps or communication. Ants use pheromones, cicadas chirp and elk create footprints in the mud. We are also swarm. Soccer players excel with lightning fast give-and-go passes and drivers in Delhi honk to stay alive. Stigmergy makes everything stronger and more robust. Our robots, despite their growing intelligence deserve a "robot-readable world."

Humans are more intelligent than any robot, but still require billions in infrastructure to drive effectively. Why would we not want to do our robotic friends the same favor? Most autonomy engineers would say we already have a robot readable world. This is correct in the same sense that the first automobile drivers had everything they needed. Imagine if Henry Ford had asserted that human drivers could already drive effectively on dirt roads? To the contrary, Ford realized that scale would come with standardized infrastructure.

The problems we care the most about are swarm problems – climate change, congestion, road fatalities – all emerge from desultory peer to peer interactions which defy our individual ability to understand, predict and control. By definition, swarm effects cannot be modulated solely by the individual. Nor do we want centralized or cloud-based control to take over. Centralized control is too slow and also engenders serious privacy and security issues. We need something in the middle and that middle ground should be a new approach based on swarm intelligence... that encompasses people, manned vehicles, robots and computational intelligence embedded into the environment.

Some might point out that self-driving teams don’t cry out for help to the infrastructure. As Brad Templeton explains: “They know how long intelligent infrastructure would take to deploy in the areas they want to service -- decades -- and have no interest in waiting on it or depending on it.” I have lived this! I agree that self-driving teams don’t blame a lack of infrastructure. Perhaps the reason they don’t want to use it is that decades of professors taught that infrastructure is cheating. I've been in the position where I'm working with large ecosystem owners including NYC, The Port Authority, and the worst of all… New Jersey Transit. I know how hard it is, but it is possible and quite possibly critical.

Innovation begins without a big outlay of infrastructure, but when it scales infrastructure becomes the means for innervating the ecosystem with intelligence. GPS required launching satellites in space. Trains required impossibly long steel rail lines. Cars required not only roads, but a staggering number of gas stations. There was a time when no one believed this change would happen. Curiously, after inventing the modern helicopter Sikorsky maintained the much smaller outlay of infrastructure was the number one reason to opt for the helicopter over the car. Ironically, the nascent autonomous air mobility world is actually happily building infrastructure to support the vision of flying cars. Even the ephemeral internet which enabled Silicon Valley required millions of routers and countless cables. Even now, fiber reigns supreme and requires thousands of miles of trenches. Replacing the undersea telegraph cables was the original reason why Marconi developed and marketed wireless, but in one of the most ironic twists of fate, these undersea cables are even more important today than they were in the 19th Century.

Still, the self-driving regime resists this notion. In the mind of many engineers, they see no need for help as their car-mounted lasers can provide 2 cm level localization on all the streets they drive. To them, the map is an adequate source of intelligence and accuracy and the idea of embedding chips into light posts seems bizarrely complicated. On the other hand, not everyone can add an expensive sensor to their car. Even if the cost came down, cars are not the only thing that needs to be localized. We cannot add spinning lasers on top of everyone’s head. Meanwhile, many autonomous vehicles still struggle today with fog, rain, dust and snow... or even just the basic absence of something to reference. My military systems twenty years ago could localize vehicles to 2cm assuming there was something to localize off of as a reference. Often there was not. Other issues include an environment where things have changed like a construction zone or a parade or even just snowbanks.

The old regime points out that even if it were true that cars could not currently do things like localization and coordination without smart infrastructure, it's a big mistake to presume we could not do those things in the future. Brad Templeton believes that we should never solve in hardware what we can do in software, not unless it's impossible under the laws of physics. Not a bad point. Let's say for Waymo that I generally agree that their system works well for city streets when there isn’t rain, dust or fog. Still, for several generations, I believe the coordination problem is all about those less proficient vehicles like my 2005 Honda Pilot (my favorite car ever). In this sense, adding a LIDAR to cars might be harder, more costly and more difficult to mass deploy than the smart roadside infrastructure.

If we only care about self-driving cars then I might agree that they can someday do everything in software, because I really do love the vision of autonomous driving. However, what about all those other cars, pedestrians and bicycles? See… it’s not just about Silicon Valley's vision. I am not against propelling that ambitious vision, but the best thing about helping others is that it solves the biggest problem of all... which isn’t individual intelligence, but FLOW. Otherwise, your self-driving car is just stuck in traffic like everyone else.

In a sense, software did eat the world, but we need to build a new one. We need to build one that is grounded in the real-world. We need smart ecosystems where position, navigation and timing are assured by an internet of moving things. Imagine what we could do with 10cm positioning, 10 millisecond latency and 10 nanosecond level synchronization. That's a world that software can't give us. Not on its own. It requires chip-scale atomic clocks, better cell networks and resilient GNSS. Once we get a taste of this new, better world, we will appreciate the value of embodied swarm intelligence -- a balanced diet of software and hardware eating each other.