Tim O'Reilly 演讲:硅谷怎么了?从集体智能到攫取租金的转变
What Happened to Silicon Valley?
Tim O'Reilly 于2026年10月16日在约翰斯·霍普金斯大学发表第二届 Margaret Levi 政治经济学年度讲座,回顾其40余年硅谷经历,分析硅谷从创造价值转向攫取租金的原因。

On October 16, 2026, I delivered the second Annual Margaret Levi Lecture in Political Economy in conjunction with the AI and History conference at Johns Hopkins University in Baltimore. This is my original lecture as written, which incorporates material from many pieces I’ve written over the years. As delivered, I departed significantly from the script, as you can see from the embedded video below or the transcript that you can see at the link above.
It’s an honor to give a lecture in Margaret Levi’s name. I’ve been asked to reflect on the changes in the political economy of Silicon Valley throughout my 40-plus year career in and about it, and I can’t imagine a better framing for my remarks than Margaret’s vision.
Margaret understood that advances in our ability to work together are essential to a better future. Cooperation is the human superpower. She also asks what happens when the systems of cooperation we have constructed over time stop holding up their end of the bargain.
That is very much the question that Silicon Valley is facing today. An industry that pioneered advances in collective intelligence at scale used to engender love from its customers because of the way it solved new problems in their lives. It now too often inspires fear and loathing. How did things go so wrong?
The web I thought we were building in 1993
But let me start with my own history. I started my company more than 40 years ago with a $500 initial investment. We took in no venture capital, and we built a business that now generates hundreds of millions of dollars in profitable annual revenue, funded by customers who love our products and pay us for them. We got there the slow way, through steady organic growth. That is quite anathema to the rush to the top of whatever appears to be the latest hill that seems to characterize Silicon Valley today. Even when climbing that hill traps a company at a local maximum, the hope is to get to payday before the industry moves on to the next hill.
My original corporate goal was stated as simply “Interesting work for interesting people.” That evolved into two principles: “Change the world by spreading the knowledge of innovators,” and “Create more value than you capture.” Those principles shaped not only my company but what I expected the internet to be. When we launched GNN, the Global Network Navigator, in 1993, it was the first web portal and the first ad-supported site on the World Wide Web, and for two years it was the place people went to find other sites. What we built was a machine for pointing you at the best of what other people had made, and then getting out of your way.
Even our idea of advertising was not what you think of today. It was inspired by direct mail advertising, where you’d read articles in a print magazine or look at accompanying ads, and then you’d call an 800 number or circle a number on a “bingo card” that you’d send in to the publisher so that the advertiser could send you a brochure. I realized that the web would be a good replacement for that awkward system, that every company ought to have its own website that provided more information about its products and services. What we invented was not the “grab your attention” kind of advertising. It was all about utility, and sharing information that customers really wanted and needed. In short, it was the commercial web, which no one any longer thinks of as advertising, even though it is the most durable and useful kind.
The commercial web was only a small part of what the web became. The web was the playground of amateurs, not entrepreneurs. It was a marvelous collective work of human imagination, creativity, sharing, and ultimately (though that became toxic only later), competition for attention. Our hand-curated catalog of the best websites was soon surpassed by algorithmic solutions like those provided by Google. The amount of content produced was far larger than purely human curation could manage. It was an early example of what Richard Sutton was later to call “the bitter lesson” in his analysis of AI training.
In my 2005 essay “What Is Web 2.0?” I described how the companies that survived the dotcom bust had all, in one way or another, figured out how to help people navigate the resulting information explosion at scale. As Herbert Simon noted in 1971, an abundance of information leads to a scarcity of attention. In the face of increasing information abundance, he predicted that we would use machines to help us better allocate our time and attention. And so it has transpired: Information has become so abundant that it defies manual curation. Instead, powerful, proprietary algorithmic systems use the data they collect to match users with the answers, news, entertainment, products, applications, and services they seek. Today’s AI is in many ways a continuation of this arc.
I called this “harnessing collective intelligence.” All of it was, in economic terms, a set of remarkable advances in networked, data-enabled market coordination. It felt like a public good even when it was privately owned.
Hal Varian, formerly Google’s chief economist, once told me that only about 6% of Google search result pages carry any advertising or are otherwise monetized by Google. The other 94% were served up by search in a glorious demonstration of abundance, the surplus of value created over value captured. And while many of the pages themselves were commercial, even more were the product of the joyful exuberance of humanity, people creating and sharing for the joy of it. Google’s algorithm matched that creation to the people who wanted it, with no price involved anywhere in the transaction. Forty thousand times a second, Google’s organic search algorithms were performing the coordination that twentieth-century economists thought could only be done by self-interested actors haggling over prices. And for Google’s first decade and a half, the priced market of advertising ran alongside organic search as a sidecar, not the engine.
That’s the Silicon Valley I signed up for. Some of this is nostalgia. The early web had plenty wrong with it, its share of greedy hustlers and hypesters. But value creation for both users and suppliers was the winning strategy. Companies won by solving real problems for their customers. Investors won when the company delivered on their promises. Unfortunately, that didn’t last.
The turn to the dark side
Google, a company once known for “Don’t be evil,” now faces the same surveillance-capitalism charges as Facebook, a company that never made any such promise. On Google, advertising now pushes organic search far down the page. Especially on mobile, you often scroll through pages of ads from national chains before you reach the site that has the best answer to your question. And in today’s AI-mode search, even the content is served up directly by Google, with far less traffic sent on to the sites that provide it. Likewise, Amazon, which used to call itself the most customer-centric company on earth, now puts sixteen ads and four organic results on a typical search results page. Sometimes, it may take scrolling past three screens of results to reach the first organic search result. And by “the first organic search result,” I mean “the best guess of the company’s smartest algorithms for what the user is actually looking for.” In former times, that was the top result. No more.
The reason isn’t that good people turned bad. It’s in the overriding economic incentives of today’s shareholder capitalism. The extractive behavior the tech giants exhibit has been the norm for modern companies since Milton Friedman reset its objective in 1970, in a sentence that became doctrine: “The social responsibility of business is to increase its profits.” Harvard Business School, through the work of Meckling and Jensen, enshrined that view not only into corporate law but executive compensation. Whatever people’s original motives, the logic of the market eventually prevails, except in extraordinary cases. In his recent book Incorruptible, Eric Ries describes a set of institutions, norms, and guardrails that can help to safeguard the values of a company, but you can count on your fingers and toes the high profile companies that have successfully deployed them to resist the market. Even tax policy rewards capital appreciation by taxing it at a much lower rate than labor income.
Our corporations and the financial markets that govern them and reward them are a kind of machine, what the science fiction writer Charlie Stross calls a “slow AI.” We built it. We set its objective function to reward shareholders and ignore everyone else. And then we tell everyone that it is a marvelous self-correcting system. This is an illusion. As I put it in “We Have Already Let the Genie Out of the Bottle,” an essay I wrote for a Rockefeller Foundation workshop on regulating AI:
Attempts at governance are futile until we recognize that we have built a machine and set it on its course. Instead, we pretend that the market is a natural phenomenon best left alone, and we fail to hold its mechanism designers to account. We need to tear down and rebuild that machine, reprogramming it so that human flourishing, not corporate profit, becomes its goal.
If the problem were just villains, you could fix things by firing them or shaming them or suing them or otherwise holding them to account. But if the problem is the design of the entire system, everyone is eventually incentivized to seek the same result. The machine rewards outsized growth of revenue and profit and brutally punishes any slowdown. Not only that, stock options tie every employee’s fortune to that reward.
Criticizing a company for turning extractive isn’t really about the values of a company or the character of its executives and employees, it’s about the incentives that have been given to them.
Rising tide rents and robber baron rents
Silicon Valley is a mirror of what’s wrong with our economy and our corporate governance, not the cause of it, and not even the worst exemplar. Tobacco, oil, pharma companies, and far too much of Wall Street vie for that spot. But Silicon Valley is a great place to study what went wrong, because companies rise and fall so quickly, making them the equivalent of laboratory mice or flatworms, where patterns can be quickly identified and followed across generations.
A few years ago, I found a useful language for describing what I was seeing in the theory of economic rents.
The essential idea is that when the reward accruing to an actor is larger than their contribution to value creation, the difference is rent. To the classical economists, to Smith and Ricardo and Mill, living in a world of inherited privilege, rents were not a temporary aberration on the way to equilibrium. They were both a persistent consequence and a persistent source of inequality. Joseph Schumpeter had a different idea, that disproportionate profits can be a reward for innovation. An innovator can earn outsized profits during the expansive phase of a new technology, when it’s solving new problems and creating value not just for consumers but for a whole ecosystem of suppliers and intermediaries and even competitors.
When economists talk about rent-seeking today, though, many of them mean what George Stigler called “regulatory capture” and what the public choice school built into its doctrine: rents extracted through government intervention at the behest of those who lobby officials and bureaucrats to get their way in despite the free market.
The rents I want to talk about are a different animal. They are collected in the market, not around it, by private firms with enormous power, whose power is unchecked precisely because the government has stepped back.
For my purposes here, speaking about the rents that accrue to Silicon Valley companies, I divide them into two rough categories.
The first is what I call rising tide rents. This fits with the classical idea of rent as a reward accruing to an actor that is larger than its contribution to value creation, but it is a virtuous kind, a kind of communal value creation, often the result of positive government interventions. Consider the rise in home prices that accrues to a homeowner who lives in a neighborhood that is improved by better schools or improved roads, which then may lead to new shops and restaurants or the home improvements made by neighbors. This is also a kind of unearned income. Everyone shares the improvements in quality of life. But homeowners get something bigger, the unearned income from the appreciation in home value.
In a similar way, leading companies make massive improvements to the neighborhood, so to speak, and we all benefit. I should point out that government plays a significant role in these improvements. It’s easy to forget the NSF grant that funded the research that became Google, the loan guarantees that made Tesla possible, and government as the founding customer for SpaceX’s commercial launch services. It’s a rising tide that lifts all boats, and the fact that the market leader takes the biggest share of it is not, to my mind, a scandal. But Schumpeter thought those outsized profits would eventually be competed away as knowledge diffuses through society. What is a scandal is the way that Business Strategy 101 and corporate governance today teach companies to build moats and barriers to keep profits high even when the company stops earning rewards by innovating and better serving customers.
Companies do this by turning to the second category of rents, which I call robber baron rents, the income above a competitive rate that gets collected because of asymmetries in power, the way the medieval baron’s troops collected their share of the harvest whether or not the baron had contributed anything to growing it.
The same company can collect both kinds of rent at different stages in its lifecycle. As far as I can tell, the shift from one to the other is close to a law of the technology business cycle. The rising tide doesn’t last. Once the growth of a new market slows, the now-powerful incumbents can no longer rely on new users and a vibrant ecosystem to sustain their extraordinary profits. So they turn to extraction. They use their market power to hold on to their now-customary level of profit in the face of competition that ought to be eating it away. Cory Doctorow came up with a catchy name for this. He calls it enshittification.
Two economies, two sets of rules
There’s another problem with Silicon Valley and the economy writ large.
At one point in early 2021, when Elon Musk first became the richest person in the world, Bernie Sanders tweeted about it, contrasting Musk’s fortune with the frozen federal minimum wage. Bernie was right that the comparison was shocking. But he was wrong to imply that Musk’s wealth came from profits extracted at the expense of Tesla’s workers. The median Tesla worker makes considerably more than the median American worker, and up to that time, Tesla had few profits to extract. Musk didn’t hoard Tesla’s profits like a robber baron of old. He won a different kind of game.
Musk was paid in what the financial writer Jerry Goodman, writing under the pseudonym Adam Smith, called “supermoney.” The essence of Goodman’s insight is that any company that has been financialized through stock issuance is valued today as if it were already delivering the decades of future earnings reflected in its stock price. This is another of those ideas that help you to see the world in a new way. Once you realize that owners of capital are literally living in a different economy from those who earn income from their labor, you start seeing two different economies operating under two completely different sets of rules.
There’s the operating economy, where companies make and sell products and services and earn a profit by, as Nick Hanauer and Eric Beinhocker put it, solving human problems. And there’s the betting economy, where wealthy people bet on which companies will win the beauty contest that the stock market has become. In the operating economy, a dollar is a dollar. The neighborhood dry cleaner earns one dollar of profit and has that one dollar to spend. In the betting economy, a dollar of profit becomes many dollars of market value. At the time when Musk first became the world’s richest man, a dollar of Tesla profit turned into roughly $600 of stock market value. A dollar of Amazon profit at that time was worth about $67. A dollar of Google profit, about $34. The local business that isn’t financialized gets only $1.
That gap is one of the biggest unacknowledged drivers of inequality in this country, and it’s why one slice of society has gotten so much richer in recent decades while everyone else languished. Most rich people—and really, all holders of capital assets—live in the supermoney economy, where our assets are valued as if the future had already arrived and can be borrowed against, tax-free, today. This is the ultimate unearned income, or rent. Most everyone else lives in the dollar economy. If you’re a public-company executive paid in stock, or a venture capitalist, or an entrepreneur lucky enough to place in the financial horse race, you live in the other one, and you get a large interest-free loan from the future.
The betting economy, within reason, is a good thing. Speculative bets on the future are how we fund innovation and build infrastructure ahead of the prosperity it will bring. Tesla is a real company building real factories and real charging networks, and it kick-started a genuine transition to renewable energy. SpaceX and Starlink likewise required massive amounts of capital. And now AI has dwarfed even those bets.
Carlota Perez has shown that bubbles are a normal part of how revolutionary technologies get adopted. I’m not against betting. I’m worried about what we bet on, and what will be left over when the bubble pops.
Hype is a big problem in the casino
If your biggest lever for wealth creation is not in the market for real goods and services, but in the market for future expectations, it’s easy to continue down the slippery slope to making money by manipulating those expectations. And that’s the road Silicon Valley has charged down.
Sam Bankman-Fried captured the zeitgeist with his line “Number go up.” Silicon Valley has gotten addicted to value capture without value creation. Think about it: 75 to 80% of IPOs are for companies without any profits. Add in acquisitions of pre-IPO companies and the numbers are likely even worse. It’s like a horse race where bettors can collect their winnings when the horse has rounded the first or second turn.
There are still venture capitalists who look for companies that succeed by delivering real value to customers, but the prevalence of the profitless IPO has turned the venture-backed company into a kind of financial instrument. Much as the bankers who invented the derivatives that nearly sank the global economy in the 2009 financial crisis eventually ran out of real mortgages to collateralize and turned to “liar loans” so they could keep the game going a little longer, the measure is no longer whether a company has a chance of living up to its promises but whether it can convince the investor in the next round that it can. Most startups are built to be refinanced all the way to the exit rather than to turn into sustainable businesses.
This dynamic is essential to the current AI infrastructure buildout. Data centers are built to meet expectations of massive future demand. But much of that demand is from companies that have not yet, as Silicon Valley used to describe it, found “product-market fit.” That is, they have not yet found customers willing to pay more than the product costs to deliver. In short, the future cash flows with which data center builders are expecting to pay back the trillions they have borrowed or invested are dependent on those customers’ ability to raise their next round at an ever-increasing valuation.
AI is the real deal, a transformative technology like the internet. And its leading company, Anthropic, looks to be turning into an actual business with cash flow and profits. But others, even those like OpenAI with enormous revenues, are still dependent on massive additional capital infusions to survive. But it’s OK, as long as “number go up.”
A structure like this does not need demand to collapse. It only needs the hype to recede enough that it is no longer possible to raise more capital at ever more stratospheric valuations. That’s why we see hype in increasingly desperate amounts.
I first started thinking about this turn to hype in Silicon Valley in the days of Uber and Lyft. The innovation of using GPS and smartphones to create crowdsourced taxis on demand using people’s personal cars was real and important. But building the matching marketplace of riders and drivers took enormous subsidies, and it wasn’t terribly profitable. So Travis Kalanick, the CEO of Uber, told investors that Uber wasn’t out to replace taxis, but the private automobile. Then in the run-up to the IPO, he started talking about self-driving cars, presumably to suggest that Uber would be much more profitable once it was able to get rid of the cost of drivers. Meanwhile, Waymo kept its head down and did the hard work of actually building out the technology, which is only now coming to fruition.
That kind of hype continued through the next one of Silicon Valley’s dry spells. There wasn’t much real innovation to talk about, so the masters of capital sold “the Metaverse” and eventually “Web3.” As the person who defined “Web 2.0,” I was asked frequently to opine on Web3. I tried to avoid answering. Most prognostications about the future turn out to be wrong. What we can do is ask questions that help us see the present more clearly. So finally, I laid out my reservations in a piece called “Why It’s Too Early to Get Excited About Web3.”
I explained how Carlota Perez and Bill Janeway distinguish productive bubbles from unproductive ones. In a productive bubble, speculative excess builds lasting infrastructure the future can capitalize on. The dotcom bubble was, in that sense, productive. When it popped, it left behind billions of miles of high-speed fiber and the data centers and the know-how that the next decade’s boom was built on. The 2009 mortgage bubble, on the other hand, left behind mostly wreckage.
The issue with Web3 wasn’t that the prices were crazy, since prices are always crazy in a bubble. It was whether anything was being built that would outlast the crash. What Silicon Valley delivered in that period mostly turned out to be betting with only the barest nod towards actual product. You probably all remember the craze for “bored apes.”
Even when there is real substance to an investment bubble, Silicon Valley has learned to lean into hype as an instrument for drawing investment. For example, when Tesla’s stock began descending from the stratosphere, Elon Musk started touting humanoid robots as the future of Tesla and pitched SpaceX not on the basis of its actual business of commercial space launches and Starlink’s huge network of communications satellites, but on AI, and on promises of a self-sustaining city on Mars housing at least one million people, travel to any point on earth in 30 minutes, and a possible enterprise value of $100 trillion!
The quest for monopoly
Not everyone can be as bold and imaginative as Elon Musk. But there’s one pretty sure bet that gets the investors to pile on, and that’s the promise of a monopoly.
Look no further than that race between Uber and Lyft. When they went toward their IPOs, neither was profitable, and their enormous valuations rested entirely on the premise that if you grow big enough and fast enough, profits will follow. Most monopolies and duopolies used to develop slowly and were understood as dangerous to competitive markets. Now Silicon Valley decided they are something to seek from the start, as the holy grail for investors. Reid Hoffman and Chris Yeh called this strategy Blitzscaling, a term they cheerfully derived from the Nazi blitzkrieg. Their argument is that you need to move fast enough to seize a market before anyone else can, even when you don’t yet know how you’ll make money. If you can promise a monopoly, you can raise a great deal of money.
I don’t have a blanket objection to moving fast. But do I think blitzscaling is a morally bankrupt idea, a game that defines winning as extracting value rather than creating it.
The saddest part of blitzscaling is that it has turned the original genius of the competitive market on its head. It’s now a kind of oligopoly of capital. Instead of customers choosing the winners, the VCs do it up front. Sunil Paul, the inventor of the technology and business model that Uber and Lyft soon dominated, raised only $35 million for his company Sidecar, which was soon crushed when Uber and Lyft raised and spent billions on marketing and subsidized fares to quickly gain users.
If I had to name a villain in this transition, it would be Andreessen Horowitz. The firm was founded in 2009, at the start of the era of cheap capital I’ve been describing, and it built a media and narrative operation from the start. Benedict Evans, who worked there, later described the firm as “a media company that monetizes through VC.” The story that a company will win because of how much money it has raised becomes part of how it actually wins. But I don’t think it allocates capital anything like as effectively as a well functioning market where customers pick the winners.
The next rising tide
What connects the enshittification of search, the supermoney economy, the blitzscaling of Uber, the narrative machine of a16z, and the crypto casino is not a run of bad actors. It’s shareholder capitalism reaching the late stage of its own logic. In its youth, that system was oriented, however imperfectly, toward creating value. In its endgame it has reoriented toward extracting it.
We are standing at the beginning of the next turn of the cycle. This time, AI is the technology driving the boom. And once again, it represents the possibility of huge advances in human collective intelligence. Like the web was 25 years ago, it’s a new creative frontier enabled by technology.
The pattern I described earlier usually has a rising-tide youth that comes before the extractive turn. However, in some ways, AI has been enshittified from the very beginning. It skipped the youth. The AI labs have taken in hundreds of billions of dollars, with trillions committed to data center buildout, and capital on that scale doesn’t wait for a business to grow into itself. It demands a return, and the pressure to repay it will inevitably push these companies toward value capture before they’ve finished creating value.
AI was also built on an extractive foundation. Last year, when The New York Times reported that OpenAI and Google were training on transcribed YouTube videos and other copyrighted work, the host of The Daily, Michael Barbaro, called copyright violation “AI’s original sin.” Margaret Levi might have called it an act of enclosure, the privatization of what had been a commons, the accumulated writing and code and conversation and images that billions of people made and shared, much of it in the joyful exuberance I started this talk by describing.
AI also breaks the old bargain in a way even the enshittified web did not. Google, at its best, sent you somewhere else. It read the commons and pointed you back into it, 94% of the time without even trying to monetize it. AI doesn’t send you elsewhere at anything like the same rate. It offers a replacement for what you would once have gone out and found, an answer in place of the referral.
We’ve been told, again and again, that the secret of future prosperity is to create enormous, disproportionate rewards for innovators and their backers. The kind of people that tech companies used to serve as their customers don’t really matter. They are necessary casualties on the road to an unimaginably better future.
This is profoundly short sighted. An economy is a circulatory system. In The Marriage of Heaven and Hell, the visionary poet and painter William Blake (writing during the first Industrial Revolution) put the circulatory logic perfectly: “The Prolific would cease to be prolific unless the Devourer as a sea received the excess of his delights.” In other words: Output has to be consumed. The system has to flow.
Today, many AGI narratives assume that the “prolific” can keep producing and the broad mass of customers (“the devourer”) somehow continue to buy, even as more and more human labor is displaced and labor income and bargaining power collapses. That’s not a future of abundance. It’s a recipe for a kind of congestive heart failure for the economy.
If the AI labs wish to be architects of a prosperous future, they must work as hard on inventing the new economy’s circulatory system as they do on improving model capabilities. They need to measure success by diffusion, not just capability. They have to ask always how their customers are going to make money, not just how they themselves will. They have to treat the labor transition as a core problem to be solved, not just studied. They have to be willing to win in the marketplace, not through artificial moats. That means committing to open interfaces, portability, and interoperability. General-purpose capabilities should not become a private toll road.
It’s an economy crying out for mechanism design. Mechanism design is a term that I’ve shamelessly stolen from the economists and redefined to describe what I think is missing from the AI economy. It’s a branch of microeconomics that studies how to design the rules of a market in such a way that even self-interested actors who have private information are more likely to produce outcomes that are better for everyone. Leonid Hurwicz, the founder of the field, called it “reverse game theory” because it works backward from a desired outcome to develop the rules that will produce it.
They give Nobel prizes in economics for advances in the theory of mechanism design. But it seems to me that many of the most successful mechanism designs happen in the real world. They evolve. They are discovered. Ken Thompson and Dennis Ritchie weren’t trying to invent the marvelous open-ended game that we now call open source software when they designed Unix. They were just trying to build a portable operating system able to run on multiple hardware architectures. That also happened to mean that you could only distribute software via its source code rather than just the binaries. And their breakthrough design of a protocol-centric operating system with a small kernel surrounded by cooperating, communicating utilities created what I call “an architecture of participation.” Anyone could add to the system, or easily replace one of its standard parts with an alternative.
That was a very different marketplace design than the proprietary operating systems of the day, which were created to tie software developers to specific hardware platforms. It led to a vibrant software sharing economy that is still going on today and that transformed all of software development.
You can see the same kind of evolution in AI. I don’t know that David Soria Parra and his colleagues at Anthropic meant to change the market structure of AI when they developed MCP, but that’s what happened. Rather than a winner-takes-most market in which developers call the APIs of the dominant, centralized platforms (as had happened previously with Windows for the PC, and iOs and Android for mobile), the agentic web began to take on the shape of the participatory economy of open source software and the web. People can bring their own models and infrastructure to the table as an alternative to the big centralized players. And that story is far from over. In fact, it is just beginning.
Every major technology wave arrives in two stages. The first stage is the development of the technology itself: the personal computer, the World Wide Web, the smartphone, and now, generative and agentic and physical AI. The second stage involves the development of new economic mechanisms that enable transformative business models to take advantage of the technology. The earliest business models often imitate what went before and encode the business ambitions of the old technologies the creators grew up with. As the technology matures, the problems with the original business models become apparent, and eventually, through a process of trial, error, and invention, new mechanisms enable new business models.
The web was invented in 1991. But the web advertising economy didn’t really take off until Google figured out PageRank, and then perfected pay-per-click advertising and the ad auction. Ecommerce didn’t take off until companies like PayPal made it safe to type your credit card number into a stranger’s website, and until Amazon built the warehousing and delivery networks that made rapid delivery possible. The smartphone put GPS in everyone’s pocket, but it wasn’t until Sunil Paul envisioned and built the first real-time location matching marketplace for ride sharing that Uber and Lyft and DoorDash were able to turn that capability into a huge on-demand economy. YouTube didn’t become the engine of the creator economy until it answered the takedown notices and lawsuits of the music industry not with stonewalling or settlements, but with Content ID.
It’s easy to look at the winners, and think the path to get there was obvious. It wasn’t. There were all kinds of failed approaches to solving the problems that only became apparent during the technology diffusion and adoption period. I don’t have time to chronicle the graveyard of failed search, advertising, and ecommerce experiments and standards. But you don’t need to look back. Just look around. How many proposals there are today for managing rights, blocking or allowing bots, enabling agents to cooperate, and payment. There are at least six competing agent payment protocols alone. We’re in the exploratory phase of mechanism design for AI.
That’s where I come back to the idea of mechanism design as reverse game theory. Rather than asking ourselves the rules of the game we are now playing, it calls us to ask “What game ought we to be playing?” And that question leads me imperatively to my answer: “The game of human flourishing.”
We have been told that the game is the quest for superintelligence. We have been told it is a winner-takes-all game that will bring untold riches to the winners and impoverish everyone else. And we have seen mechanisms designed to support that game: massive investments in data center capacity such that only the well-capitalized can play, colonialist appropriation of existing intellectual property, closed model weights with terms of service that forbid others from learning from them, memory that locks our context inside one provider’s walls, and PR stunts like dumping millions of dollars of compute to solve high-profile math problems, or hyping the cybersecurity risks of their offerings, for which only their latest, most powerful models are offered as the only remedy.
If the game we want to play instead is human flourishing, we need to design mechanisms that lead us to that outcome. We need mechanisms that provide a circulatory economy of value exchange, for and by humans in partnership with AI, rather than replaced by it. And in my experience, those mechanisms look like open source, open protocols, open memory, because these are things that give the freedom for users to shape and control this most wonderful new technology wave for ourselves.
Henry Farrell, Alison Gopnik, Cosma Shalizi, and James Evans argued in Science last year that large models are best understood not as independent intelligences but as a new kind of cultural and social technology, one that lets us take advantage of the information other humans have accumulated, in the lineage of pictures, writing, print, and internet search. AI may be the most powerful means of harnessing collective intelligence yet. The commons isn’t incidental to AI, it’s the material AI is made from.
In the world I see coming, AI is a medium, like painting or writing or music. If we get this right, everyone will get the same paints and brushes, the same words, the same notes, but some people draw more out of them than others, or do it better. We won’t decide up front that those who raise the biggest pool of capital are due all the rewards.
Our job is to draw something more, something better, out of the ocean of possibilities in the collected knowledge inside the model. AI was built from our collective intelligence. It should not be designed to replace us. It depends on us to create its value.
Thank you very much.
来源:O'Reilly Radar · oreilly.com