AI Swarms on the Blockchain
The future won’t just have robots, it will have huge swarms of them. They’ll be small, probably non-specialized, and they’ll work together to accomplish larger goals. Think about ants. Ants individually are so much simpler than what they accomplish together. There is a logic to their behavior that transcends an individual’s programming. We should build machines to behave the same way.
Robots could move in huge, mutually-reinforcing swarms. Sophisticated capabilities that would require a large, expensive robot may be distributed among many, replaceable lesser ones. This lends the swarm far more redundancy. As individual units fail, new ones take their place. Since there are so many of them and they are built to be non-specialized, they can be mass-produced with staggering efficiency.
A robot swarm could be more than the sum of its parts. For example, by sharing sensor data,they could achieve an accuracy far superior to that which could be acquired individually. Consider how your ear works. Individual hairs of different lengths resonate at different frequencies. Only with many hairs, at different lengths, can you integrate their numerous signals to achieve sophisticated hearing. The individual units, the hairs, don’t have to be effective at all frequencies, merely one. Indeed, the point goes further; certain capabilities such as clever inteferometric techniques are basically only possible with large networks of discrete elements. Masses of simple robots can accomplish what highly-engineered individuals cannot.
Consider the parallel with multicellularity in nature. If you stop to think about it, multicellularity is deeply bizarre. If it had not yet happened, I would judge it phenomenally unlikely. What began as loosely collaborating colonies of plankton became tightly bound and highly specialized organizations, such as mammals, where only a small subset of cells were destined to reproduce at all. It’s staggering to consider that the overwhelming majority of cells in your body are meant to die, are ready to die, in the service of the germ cells (sperm and eggs). Our bodies are fantastically fine-tuned collections of celibate specialists engaged in monkish harmony for the purpose of highly orchestrated reproduction. It’s as if skyscrapers conspired to build more skyscrapers.
Robotics may take yet another page from biology. The future may contain fleets of drones, building our structures, excavating, performing grand projects and heavy lifting, finding people lost in the wilderness, helping to form mesh networks, deliveries, and much more. If this comes to pass, we should consider first and foremost the network architecture underlying so many AI elements. It may not make sense for numerous little robots to obey a central server somewhere else. What should the command and control hierarchy look like? Already the US military chafes against the limitations inherent to controlling drones with human operators from thousands of miles away. Delegating responsibility to local AI will become essential as the number and scope of drones grow. But how to delegate without focusing too much responsibility on trusted ‘master’ nodes? What happens when master nodes fail? What happens when malicious actors try to spoof commands? What is true and what is untrue in the world of peer-to-peer networks? Who has access? What happens when the drones start lying to each other?
Here I believe the Bitcoin Blockchain (or another blockchain built from its principles) may offer some potential. It is a universal consensus ledger: a place to provably and irrevocably write things. Using the power of strong encryption, Truth is as it is written in the Blockchain. Drones may use that. Since it may be read algorithmically, computers may verify the legitimacy of the Blockchain and its immutable history. Bitcoin ownership is totally self-evident to a Bitcoin node. So too can other data be mathematically self-evident on the Blockchain.
The Blockchain resolves a crucial failing in multicellularity. What happens when cells go rogue? In us, when cells stop heeding the interests of the organism, when the celibate specialists cast aside their virginal raiments and pursue their own aggrandizement, we call it cancer. It is a problem that is basically intrinsic to multicellularity. Any cell is only a finite number of mutations away from total rebellion at any given time. When this occurs, as statistics dictates it eventually must, the interests of the individual and the collection diverge. Havoc ensues. This is a fundamental design flaw with, to my knowledge, no ultimate cure.
However if robots were designed to obey intructions according to rules posted on the Blockchain, the system of obedience keeping them in check would be vastly stronger than any biological equivalent. Cells slide down a slippery slope; as mutations accumulate, a cell’s risk for cancer grows steadily. However in deciphering strong encryption, the results are always binary, total success or total failure. In NP problems, there is no steady progress, no slow, upward path to success. Robots could not mutate their way outside of constraints imposed from the Blockchain along a gentle and gradual evolutionary fitness landscape. They would fall off a fitness cliff, from 1 to 0. It would be immediately obvious whenever a robot had gone rogue.
Just as alterations to Bitcoin constitute hard forks that are immediately identifiable, rogue or aberrant drones could be identified from the very first instant. Perhaps honest robots could mathematically prove their own ‘loyalty’ to the common governing protocol. A Blockchain ledger could ensure honesty between drones, even as they consist of large assemblies of peers with no trusted center.
Discrete math arms us with powers our multicellular bodies will never have. The enforceability of strong encryption, and the power of the Blockchain, mean that blockchain-truths are mathematically self-evident. This could be used to curtail the divergence of AI robot swarms in a way our bodies can never do chemically to cancerous cells.
Follow me on Twitter @abarisser
