31. Januar 2022

More Accelerando Than Snow Crash

The Mental Model of the Web3 Future Is Not Snow Crash. It's Accelerando. The inevitable path to a new economic model made possible by web3.


There is a lot of emphasis lately on the metaverse and virtual worlds. We believe that web3 helps to share stuff between virtual worlds, to tear down the walls between online worlds. Ready Player One shows a unified metaverse, where avatars from different virtual worlds meet. That's nice. Maybe even useful someday. There is also business and money to be made. Actually, a lot of business will be enabled or improved: entertainment, marketing, customer support, and more.

However, the real impact on the economy of the future comes from automation of business processes. In a web3 world software, scripts, and AI, can make deals. Specifically, business executing AI will have a large impact. Ultimately, AI will be able to act with tangible effect through the web3 we are currently building. It's the real economy that counts. That's why we should look to Charles Stross' Accelerando rather than Neal Stephenson's Snow Crash.

We have been learning from many examples in different fields that AI is good at finding new ways to do things. When AI optimizes a task, it often finds more efficient ways than experienced humans in the same field. For example, self-learning AI invents unconventional strategies in games. It explores strategies that the best human players would have disapproved of until they were defeated by those strategies. AlphaStar, Google's StarCraft AI once produced overwhelmingly many Oracles, a Protoss unit. A strategy no professional player tried because it has disadvantages in the later game. But still the AI beat top human players until they countered the strategy as soon as they detected it.

In another experiment self-learning AI that needed to communicate to solve a task quickly developed a more effective way of communication. They invented their own language. A protocol that was more efficient that the protocols they were given as a starting point. The language was not easily understood by humans. It was analyzed. But this took time while the AI moved forward. Understanding the AI's way is a moving target. Ultimately humans will use their optimizations without fully understanding them.

We are now at a point where software driven business processes emerge. Web3 enables software to post offers, to negotiate, to close deals, and to check fulfilment. Software is already doing significant business at stock exchanges. Software can react more quickly than humans which is important in times of high-speed trading. Some of these agents are driven by deep learning and genetic AI. While there are many details and nuances, basically trading stocks is rather simple. There are sell offers, buy offers, and real time information. The task is to optimize profit over time. A difficult task considered erratic markets, a volatile information situation, erratic market behavior, and feedback loops. But the trading model is simple: buying and selling securities.

Now, web3 promises to pull all other business into software's reach. While theoretically everything could be wrapped into a security, not everything in the real world is suited for securityfication. Partially, because it is irrelevant, like selling my own house. Selling my house is not accessible to software because nobody has made it a security, and nobody will.

In the field of patents and intellectual property rights are usually not freely tradable because there are too many barriers. IP has fundamentals that are difficult to consider automatically. Trading IP goes beyond comparing market prices. Assessing the value of IP is the domain of human experts. IP deals also need notaries, attorneys, and registers, in other words: legacy real-world mechanisms.

Car manufacturers deal with thousands of supplies, each with a detailed part specification, negotiated quality expectations, technical standards, and individual considerations. They are far from being securitized, out of reach of trading software. Until now.

Smart contracts can replace government registers like commercial and land registers. If a land register is secured by a blockchain instead of a government or an attorney, then this not only makes trading cheaper by removing the middleman. It also makes trading the goods accessible to software.

Physical properties of car parts can be measured and compared with specifications. A smart contract checks if negotiated standards are met. It decides to what extend deliveries deviate from expectations. Pricing is fixated and made transparent to all parties as a smart contract. Money flows reproducibly and reliably based on measured and negotiated parameters of contracts. In the beginning humans will create these contracts, negotiate their parameters, and set up real-world measuring equipment. Humans will also approve payments. But that is still a lot of work. After some time of waving through payments smart contracts will be made to pay without human approval on small lots. Then, when there were no major glitches for some time checking parts deliveries and payments will be automated.

Still, finding and negotiating thousands of parts is a lot of work waiting to be automated. And it will be automated. Suppliers will offer their parts through smart contracts that manage specifications and tolerances. Smart contracts also offer variations, and they will have logic, scripting, or AI to estimate production cost of variants. That makes sifting through of all these variants, specs, and tolerances for countless parts easier and humans just approve selections, confirm deals, or intervene when the AI does stupid things. And again, after some time without major glitches the industry will let software make the deals unsupervised demanding only after-the-fact reporting.

There is one more step required to completely automate the industry: planning and building factories. This will take more time. But individual manufacturing through 3D printing accelerates the process. Tesla already knows how to build Giga factories for certain products on demand. There is now so much institutional know-how that these facilities can be built in months instead of years. Factory projects are increasingly data driven and all this data will finally be used to train AI.

Software simulation of production processes also helps to self-train AI. A game of building factories, negotiating parts and resources to win market share against a competitor is not fundamentally different from managing resources and combat in StarCraft. AI will optimize itself with simulated competitions. Then AI will plan and build factories. As always, after some time without major glitches, some players will let AI react to market demand automatically. Even some goof-ups can be tolerated. Human decision making when estimating future demand, planning products, and executing business plans is far from perfect. If the financial impact of AI-mistakes is on the same level as the one by humans, the AI wins. Finally, the AI will win. And the first humans to adopt this way of doing business will become rich.

Then AI optimizes the business. The AIs will optimize communication by inventing new protocols. Negotiation protocols that are more efficient than the ones inherited from humans. There are many ways to optimize in a software driven world. Maybe they dispense with checking individual deliveries. Maybe they don't put up tenders anymore. Suppliers might deliver parts without prior negotiations based on information from crypto oracles. After all, the financial output of the entire operation is key. They might omit payments for supplies and just share the revenue. A smart contract takes key performance indicators and generates a pay-out scheme for all involved entities in a transparent fashion. There are hard short-term facts like revenue, time-delayed measurements like product reliability, and long-term soft information sources like polls about buyer's remorse. All this data can be used to optimize the business. At some time, there is data available from millions of products, markets, and processes over many product cycles. This data is then employed by the executing AI to find new ways.

Would humans base a car business on revenue sharing and common long-term benefits? Probably not. Human experts would reject this way of doing business for many reasons. Humans are good at coming up with reasons not to change things. Until they are outperformed.

Humans are also good at inventing possible ways for improvements based on their experience. We can imagine countless optimizations and process changes. Science fiction authors are especially good at that. But we largely fail to predict developments beyond our experience. That's where AI excels. It finds categorizations that escape us. It finds optimizations we won't think of.

AI will change the way business is done so much that humans will not understand what's going on. At first, we will. We will be surprised by AI's inventions. We will marvel at the ingenuity and frivolity of its ways. For as long as we can analyze and understand what is happening. Later we will fail to understand and just embrace the benefits.

This is what Charles Stross calls Economics 2.0. A business model more efficient than ours. Let's call it Economy3 to be in line Web3. Its economic processes that outperform the ones we know. Interactions and rules we do not understand, that can only be executed by AI. Not because of the required speed of decision making, rather because the rules will not be known. They will not be codified. They are decentralized in neural network weights or whatever AI is made of in the future. The new rules will not be programmed into AI. Rather AI will develop the rules because they work better than the inherited ones.

This sounds as if we humans have no say in the process. But we do. The key phrase is "work better". We define what "better" means. If "better" means more profit, then average people might be screwed in a way described in Accelerando. In this future the so-called Vile Offspring, basically untamed rouge AI, dominate the inner solar system and even dismantle the Earth to put its resources to "better" use. Earth's resources not meaning oil and ore, but the iron of the core, hence the dismantling.

A development that ends in the dissolution of our planet does not sound "better". And that's the key point. We will have to define the term "better" so that it serves people, and a dismantling is avoided. We need more performance indicators than profit. We need performance indicators that represent the wellbeing of people and the environment for that matter. AI optimizes along fitness functions and training data. AI designers define these fitness functions and select the training data. We decide how AI optimizes. We have a say. A society that really tries will have a deciding influence. Realistically the result will be somewhere between utopia and the planet's pulverization into smart matter. We must make sure that besides profit and wealth for as many people as possible, there is also well-being, whatever that means. Maybe the fitness function just needs as much Gross National Happiness as Gross Domestic Product.

Coming back to web3: this development path is almost inevitable because it is possible. The path is obvious. There are no unknowns, no new technologies to be developed, no new principles to be discovered. The paradigms are already in place. The rest is engineering.

There is one more thing: the smart-contractification of the real world. Paper contracts will be replaced by smart contracts. Business entities will learn that blockchains tell the truth. Companies will sue each other to honor agreements that are codified by smart contracts. Finally, courts will begin to refer to the blockchain truth in their decisions. Then, the real-world is smart-contractified. It will take some time to get there. But the path is clear.

Once the real economy (the one that builds smart phones, not just non fungible images) takes web3 serious we are bound to end up with Economy3. An automated future in which it is not necessary to work hard to pay the rent. That's where we want to go.

We are currently building the tools: web3 and AI. Then we'll get the real world to use the tools while making sure that the beast we're unleashing does not deviate too far from a good path. It is our responsibility to educate our societies about the risks and empower them to set the rules.

We must shape the future economy, not just virtual worlds. It's the real world that matters and the real economy. In this sense the mental model to guide our path is better characterized by Charles Stross' Accelerando than Neal Stephenson's Snow Crash. Read Accelerando, enjoy it, fear it, and learn from it.


Keine Kommentare: