《Linux基金會：2024去中心化與人工智能報告（英文版）（54頁）.pdf》由會員分享，可在線閱讀，更多相關《Linux基金會：2024去中心化與人工智能報告（英文版）（54頁）.pdf（54頁珍藏版）》請在三個皮匠報告上搜索。

1、Decentralization and AIThe Building Blocks of a Resilient and Open Digital Future Alan Majer,Founder of GoodRForeword by Yue Chen,Head of Technology Strategy Chris Xie,Head of Open Source Strategy Futurewei Technologies,Inc.November 2024Increasingly,powerful algorithms govern online systems,putting

2、users under the thumb of those who control them.Systems of influence and fake news erode our ability to access the truth,and trust in all forms of media is at a historic low.Autonomous AI agents will be self-reflective and possess a wide scope of action,allowing decentralized systems of agents to ad

3、apt behavior and properties.Putting AI and algorithms at the edge can enhance personalized experiences yet keep data private,giving us greater control over our online identity.We must make technology choices that ensure that AI remains open and accessible to all and represents everyones valuesnot ju

4、st those of tech platforms.The attention economy means our digital experiences have fallen short of their potential and a few big companies control our online experiences.Regulations are necessary to curtail the power of private platforms and protect consumer privacy,but alone,they are insufficient

5、to catalyze the innovation required.A growing list of open source tools,platforms,and applications create alternatives to todays centralized systems,but obtaining critical mass remains difficult.Todays centralized platforms face disruptive transformation:Decentralized building blocks,AI,and algorith

6、ms can build platforms that have never been possible before.Personalized autonomous agents afford us greater control over our data and the algorithms that govern our interactions.Artificial intelligence can create easy-to-use tailored interfaces and generate code and applications on the fly,with far

7、-reaching implications for software.In order to succeed,decentralized systems mustbe agile and easy to usehaving control over ones data should be an advantage.Copyright 2024 The Linux Foundation|November 2024.This report is licensed under the Creative Commons Attribution-NoDerivatives 4.0 Internatio

8、nal Public LicenseDecentralization and AIContentsForeword 04Executive summary 05Introduction 06Where did the Internet go wrong?07Can we fix the Internet?08A closer look at the causes 10Threats to personal data and privacy 11Information overload and devalued content 12Walled gardens that erode the co

9、mmons 13Powerful algorithms that influence and control 14Gaslighting us with misinformation erodes trust and autonomy 15Regulatory currents of change 18Reimagining platforms:The Internets decentralized future 20Infrastructure for decentralized social networks 22ActivityPub:Mastodon and Threads 22AT

10、Protocol and BlueSky 23DSNP and Farcaster 24Communication and messaging 25Personal data storage and provenance 27Winning with decentralization 30AI,agents,and algorithms catalyze new platform possibilities 33Artificial intelligence in todays centralized platforms:Apple 34True decentralization via re

11、flective autonomous agents 36Open Interpreter 38Agent Zero 38Autonomous agents and our rewilded future 40Conclusion:At a crossroads 41Methodology 43About the author 43Acknowledgments 43References 44ForewordThe rise of autonomous AI agents is set to shape the decentralized Web of the future.These age

12、nts,manifesting as embodied AI,robotics,or software processes,will play a critical role in this new digital ecosystem.Acting as our ultimate digital assistants,they will permeate our living environments,handheld devices,and wearables,offering personalized services and insights.However,as these agent

13、s gain deeper access to our personal and private lives,the need for robust protections of privacy,data,and personal identity becomes paramount.Autonomous AI agents present both unprecedented opportunities and significant challenges.Their ability to operate independently at the edge of decentralized

14、networks enables a more adaptive,intelligent Web.But as they interact with sensitive data and make decisions on our behalf,a globally recognized decentralized identity system is essential to ensure security,trust,and user sovereignty.Such a system must guarantee that individuals maintain control ove

15、r their identities and information,preventing the vulnerabilities of centralized platforms from being replicated in a decentralized future.By distributing intelligence across the network,these agents can foster a more resilient and privacy-preserving Web.Yet,this also amplifies the need for open,tra

16、nsparent identity frameworks that respect user autonomy.This report outlines the path toward integrating AI and decentralized identity,creating a future where privacy,control,and trust are at the core of the digital experience.In this new era,we have the opportunity to collectively define a digital

17、frontier where AI empowers users and safeguards their rights,shaping a decentralized Web that truly serves the people.YUE CHEN,Head of Technology Strategy CHRIS XIE,Head of Open Source Strategy Futurewei Technologies,Inc.4DECENTRALIZATION AND AIExecutive summaryTodays digital platforms are not livin

18、g up to the expectations of the Internet visionaries of the past.Digital experiences have fallen short of their potential,largely because the attention economy is so deeply infused into centralized platforms that a small handful of big tech companies control.As a result,many of todays centralized pl

19、atforms pose risks to personally identifiable information and privacy,the value of content,and the benefits of our digital commonseven the Web itself.Increasingly,powerful algorithms govern these systems,putting users under the thumb of those who control them.However,the winds are beginning to chang

20、e.Regulators are noting the power of the largest tech companies and working to curtail them.Open source innovators are creating less-centralized alternatives and building decentralized infrastructures that preserve individual privacy,autonomy,and digital property rights and increase access to the gl

21、obal commons.Most of all,advances in artificial intelligence are disrupting centralized platforms.New AI algorithms are not only changing how large tech incumbents will deliver experiences on their own platforms but also creating opportunities for novel platforms and systems without precedent.Autono

22、mous AI agents optimized to the needs of users(rather than platforms or the economics of attention)afford new,more decentralized,ways of organizing.Its a chance to bring the capabilities of potent technologies back to the edgeopen technologies back into the hands of everyone.None of this will happen

23、 on its own,however.Societies and individuals must choose their future,then use these technologies and capabilities to make it so.5DECENTRALIZATION AND AIIntroduction“If it is not true,it is very well invented.”Giordano Bruno,Renaissance-era philosopher1Who or what do we trust?“The perceived truth i

24、n media,both traditional and social,is lower than at any other time in our lifetimes,”writes investor Ray Dalio.2 If that doesnt surprise you,it should.Fifty years ago,people could barely dream of the information that average people have at their fingertips.The problem isnt the availability of infor

25、mation or data;its whether we can trust the sources and platforms that deliver it to us.We recognize that platforms are vulnerable to bias.Over 80%say that outside groups influence the news that gets reported.3 We also know theyre nudging our opinions via the algorithms that mediate our experiences.

26、Its something ex-Googler James Williams calls the“largest,most standardized and most centralized form of attentional control in human history.”4 Now,the most powerful algorithm of allartificial intelligenceis nudging us.Already,bots(not all of which are intelligent)make up 49.6%of Internet traffic,a

27、ctively involved in a war over private information.5“Bad”bots that masquerade as humans and vacuum up data are replacing historic conventions on reciprocity(indexing data in exchange for search traffic).Says one publisher,“With the AI companies were getting nothing in return.”6 The Web as weve known

28、 it is becoming a tragedy of the commons.Tim Berners-Lee,the inventor of the Web,describes it as“a large-scale emergent phenomenon which is anti-human.”7 Without intervention,ever-more-powerful AI and algorithms will mediate our experiences and harvest all the data,reinforcing the dominance of today

29、s centralized platforms.Berners-Lee and many others are attempting to take the open Web back.Trust in our foundational technologies and platforms are as important to healthy functioning societies as water or air.With the right choice of technology and regulatory ingredients,we can regain online auto

30、nomy and control.Says Tim Berners-Lee,“What could happen if we give people privacy and we give people control of their data?”8 Its a process that some have referred to as the“rewilding”of the Internet.9 Doing this well isnt about mimicking todays centralized platforms and shoehorning in open technol

31、ogy.We need to replace the Internet monoculture with something new and different in kind.A daunting task?Yes.Though the timing couldnt be bettertodays dominant platforms are facing their own disruptive transformation.Decentralized building blocks,AI,and algorithms create exciting new opportunities t

32、o build platforms that have never been possible before.New breeds of personalized autonomous agents put intelligence at the edge,affording us greater control over our data and the algorithms that govern our interactions.Easy-to-use interfaces and on-the-fly code generation not only put a friendlier

33、face on powerful tools but also have far-reaching implications for platforms,software,and even open source itself.The timing is ripe for new disruptive platforms built on an open foundation that can help us reclaim what should have always been ours in the first place.6DECENTRALIZATION AND AIWhere di

34、d the Internet go wrong?We live in a world of data abundance.10 At no time in history has the worlds store of information been so vast and its potential so badly squandered.We started out with a bold vision of what the Internet could be.Take Googles mission statement,for example,“to organize the wor

35、lds information and make it universally accessible and useful.”11 But,somewhere along the way,the Internet took a wrong turn(not TCP/IP itself,of course,but rather Internet-connected experiences).Instead,we enrich large,centralized platforms that control our digital experiences.And because our world

36、 is awash with data,its easy to forget just how precious information can be.Grand ambitions for a new digital commons wasnt the problem.In fact,the Internet succeeded in ways that previous generations could scarcely imagine.Information is now at everyones fingertips,and intelligent wired and wireles

37、s networks connect billions of people on every coordinate of the globe.Instead,the challenge has been to protect and maintain such a prize.Its a grave responsibility to permanently govern and run public goods such as the Internet.Robin Berjon,Deputy Director of the IPFS Foundation,explains:“We built

38、 this system that is,you know,planetary and currently serves over 5 billion people,and we didnt approach it with a mindset that matches the scale and responsibility of the task at hand.”12 Protecting and governing a global commons is not an easy task.Says Berjon,“no one said that governing the infra

39、structure for basically all of society at the planetary scale should be easy.”13 Public decentralized infrastructures are a bit like ecosystemsthey need to be sustainable and viable.Without resources and sound governance,the economic model for such a system can become a pressure point.Left unaddress

40、ed,its an area of vulnerability where commercial systems may gain entry and co-opt the commons for their own benefit.David Clark,a former chief protocol architect in the development of the Internet and now a senior research scientist at MIT,jokes about a conversation he had with an economist at a po

41、int in time when the Internet was starting to become more commercial:14 Economist:“The Internet is about routing money,and routing packets is a side effect,and you screwed up the money routing protocols.”David Clark:“I didnt design any money routing protocols.”Economist:“Thats what I mean.”Of course

42、,one solution to this problem is baking economics into the system from the beginning.It has worked for many blockchain-based systems,where arrangements of incentives drive entire ecosystems.Indeed,its an interesting thought experiment:Would the“attention economy”(treating human attention and private

43、 data as a commodity that can be bought and sold)have arisen at all if the Web had a payment infrastructure built in from the start?David Clark,for one,suggests that the economic models of open systems are essential to their success.“The really critical issue in making a sophisticated,distributed ec

44、osystem work is going to be the model of how money flows through the system.So,I am a complete believer in open source as a facilitator of this space,but the design of the ecosystem has to be right,or open source alone wont be enough to make it happen.”15“No one said that governing the infrastructur

45、e for basically all of society at the planetary scale should be easy.”Robin BerjonIPFS Foundation7DECENTRALIZATION AND AIOthers say its not the economic models that are missingsimply better decisions and governance of complex systems.According to author and futurist Karl Schroeder,“the fundamental p

46、roblem facing us is our inability to govern ourselves well,particularly when it comes to common goods.”16 Robin Berjon of the IPFS Foundation suggests that it boils down to an issue of epistemic democracythe capacity of the“many”to make the correct collective decisions and achieve better results.17

47、Without the right tools and decisions,our systems fail to escape“enclosure,”a well-established pattern that Schroeder says has repeated itself over hundreds of years:“Theres a 250-year-old,roughly,process that has repeated itself a number of times,and its called the enclosure of the commons,and it s

48、tarted with the literal fencing in of common land used for grazing.”18In 15th-and 16th-century Britain,for example,as wool became more profitable,landowners started fencing in arable land and displacing commons.Adds Schroeder,“So,this is the classic example of the enclosure of the commons.Its an enc

49、irclement where,technically there is a commons,but it gets harder and harder to access,and walls get put up that necessitate gatekeepers.”19 Today,the same phenomenon is unfolding in digital form.Its visible in the ownership of the airwaves,vast collections of data,and now even in computing and netw

50、orking itself.“Its taken a little while for it to arrive in the Internet,but here we are,”says Schroeder.20 Robin Berjon agrees:“Its been captured.And yes,the means of capture have often been to take one piece and then leverage that into controlling other parts.”21 Recognizing the problem is the fir

51、st step.These systems are pervasive,and so is their influence,so its easy for our values and control to slip away without us being aware of it.James Williams,a former architect of Googles search ad business,asks,“Will we be able to recognise it,if and when it happens?And if we cant,then how do we kn

52、ow it hasnt happened already?”22 Its not an idle question.For example,much of the thought-provoking discussion about the influence of social media on society occurred from 2017 to 2020,a time when established journalism and media were under threat and when we began to understand the challenges of th

53、e“attention economy.”Now,as social media platforms emerge as winners,is it reasonable to expect open,self-reflective dialogue on these platforms?Especially if it explores the nature of their dominance?In 2017,Justin Rosenstein,the inventor of the Facebook“like”button,prophetically suggested,“One rea

54、son I think it is particularly important for us to talk about this now is that we may be the last generation that can remember life before.”23 Can we fix the Internet?“The world doesnt have to be the way it is right now,”says Wendy Wong,Professor of Political Science at UBC.24 While it might feel li

55、ke a daunting task,“We need to stop thinking of Internet infrastructure as too hard to fix.Its the underlying system we use for nearly everything we do,”suggest Maria Farrell and Robin Berjon in their article,“We Need to Rewild the Internet.”25 The Web belongs to everyone,says Tim Berners-Lee.“Colle

56、ctively we hold the power to change it.It wont be easy.But if we dream a little and work a lot,we can get the Web we want.”26 David Clark certainly hasnt given up on the Internet he helped create.“I absolutely believe,with a considerable passion,that we should try to build a decentralized user exper

57、ience.”27 Shroeder suggests that the recovery of a commons is easier than one might think.“The thing about the enclosed commons is that its highly artificial.It rests on mercantilist and capitalist construction of reality.Basically,if the owners go away,the locals just come by and knock down the wal

58、ls,and then you have a commons again.”28 Its about the return to a natural,less-8DECENTRALIZATION AND AIcentralized,state.“The real world is decentralized,”says Munindar P.Singh,SAS Institute Distinguished Professor of Computer Science at North Carolina State University.We need to shed our closed-lo

59、op assumptions and allow sufficient autonomy at the edge of a network.29 Dominance also means accountability.Daniel Goldscheider,founder of the OpenWallet Foundation,says that platform dominance should also come with responsibilities:“At the very least,you should not use or abuse that position of po

60、wer just to further your own business goals.”30 What if,rather than furthering the interests of todays dominant players,our platforms benefited us by amplifying our capabilities instead?To appreciate the opportunities to steer future platforms,lets look at the root of the problem.9DECENTRALIZATION A

61、ND AIA closer look at the causes“Data is not the new oil its the new plutonium.”Jim Balsillie31The Internet didnt begin with a built-in business model.Those in the packet routing business soon found oneconsumers proved more than willing to pay subscriptions to Internet service providers.But the busi

62、ness models of burgeoning content and application providers was more uncertain.In the beginning,secure payment on the Web wasnt even possible,and even today,wallets remain fragmented.Says Daniel Goldscheider,“If browsers behaved like wallets,then you would basically need to download a new browser fo

63、r every website you visit.”32 This friction,and the lack of ubiquitous forms of payment,breathed life into advertising-based business models.We began paying for content and servicesnot with our money,but with our attention.The miracle of the Internet is that everywhere we go,anywhere we spend our ti

64、me,were traversing across platforms in which billions have been invested.Many of them are freein that we dont pay any money to use them.Instead,we cede control over our private data and attention.This“attention economy”puts a vast range of services and experiences at our fingertips,but it comes at a

65、 steep cost.Says James Williams,“The dynamics of the attention economy are structurally set up to undermine the human will.”33 Jim Balsillie agrees:“The online advertisement-driven business model subverts choice and represents a foundational threat to markets,election integrity and democracy itself.

66、”34 They“capture”and enclose the commons.“Capture”of the commons relies on one piece that lends to control of the others.It often involves platforms that extend their footprint into key pieces of infrastructure,overlapping with domains that public goods,institutions,and national interests occupy.Tha

67、ts a tricky role to occupy,suggests Daniel Goldscheider:“Whoever is trying to do something that is fundamental infrastructureespecially when you are entering an area where you are recreating functions that typically governments heldyou should probably do so in the interest of the public and check yo

68、ur own business goals at the door.”35 Centralized platforms and the ad-based models that power them are the thin edge of the wedge that now threatens the integrity of the Internet we set out to build.Specifically,we are increasingly participating in systems that:Threaten our personal data and privac

69、y Subject us to information overload and devalue content Create walled gardens that erode the commons Use powerful algorithms to influence and control Gaslight us with misinformation to erode our trust and autonomy10DECENTRALIZATION AND AIThreats to personal data and privacyThe power that comes from

70、 the sensitive personal data that big tech has accumulated has not gone unnoticed.Data mining is often personal and invasivefor example,allowing Target to learn of a teens pregnancy before her family did.36 The White House itself has warned of the dangers of“Big Tech platforms gathering too much per

71、sonal information.”37 Others,including Mozilla,have a take that is more blunt and to the point:“Ubiquitous surveillance.harms individuals and society.”38 The value of information collected from“free”services is staggering,evident in Googles paying Apple$20 billion in 2022 for the privilege of being

72、Safaris default search engine.39 But how much do these platforms really know about us?Quite a lot,it turns out:“Research has shown that up to 52 companies can theoretically observe up to 91%of the average users Web browsing history.”40 “Large players know on average 47%of each users browsing history

73、 Google alone,which is the biggest player in the tracking ecosystem,covers 64%of the average users history logs.”41 “Large trackers know,on average,nearly half of the browsing history of almost all users.”42 “Online trackers can capture up to 80%of users browsing histories.”43 “We discover that 81.3

74、%of visited websites send search terms to third parties in some form,representing a potential privacy threat to users.”44 However,individuals themselves are also complicit in this tracking and surveillance.Its not unusual to give up sensitive personal data in exchange for a few loyalty points(and 80

75、%are willing to sell it for$100 or less).45 Yet its not just giving data;sometimes people surrender control of valuable data that contains their history of experience.Rodrigo Mendoza-Smith,Founder of a start-up,Quira,which he describes as specializing in experience quantification of software develop

76、ers through data,explains the difference between the two.Says Mendoza-Smith,for developers,“When you write code for a company,theres a big difference between your work output(and your code)and your own experience The company a developer works for owns the code and the IP they create.But the develope

77、r owns the reputation and experience associated with writing that code and creating that IP.”46 While it might sound like a minor difference,owning your work experience forms part of your reputation,allowing you to access new opportunities.Its what allows Quira to glean reputation information from p

78、ublic data.“My work is yours,but my experience is not,”suggests Mendoza-Smith.47 Thats why access to and ownership of personal data are so important.Histories and experience help define who someone is.Nor do people approach data sharing rationally.Research suggests that we share our data not because

79、 we believe its safe or because we trust the platforms we share it with but because everyone else is doing it.48 11DECENTRALIZATION AND AIInformation overload and devalued contentOnline platforms also subject us to information overload.Research suggests that additional information aids decision-maki

80、ng,but only up to a point.Too much information impairs our decision-making performance,49 leading to analysis paralysis.Social media amplifies the flow of low-quality content,50 and AI-generated content will only make things worse.According to Sam Altman at OpenAI,in 2024,ChatGPT generated approxima

81、tely 100 billion words daily.51 In a year,thats about 1%of the worlds total estimated store of 4.1 quadrillion total written words.52 Its a flood of chat sessions,articles,blogs,social media posts,and even vast amounts of synthetic training data used to generate future AI models.Even entire applicat

82、ions will be coded up in an instant.53 At a projected growth rate of 70%annually,54 by 2032,AI agents would annually author 61%of todays total store of written words.55“Now were in a position in which knowledge is free,and its available anywhere and everywhere,”says Rodrigo Mendoza-Smith.56 The prob

83、lem?The explosion of AI-generated content will overwhelm humans capacity to absorb it.A decade from now,the average written word is unlikely to be read by a human at all.What will it mean if AI can produce anything on the fly?Books,articles,software(even open source software)the value of all of thes

84、e things depends upon assumptions of scarcity,which may soon cease to exist.Without that scarcity,the value of apps,or even the Web itself,may fall off a cliff.12DECENTRALIZATION AND AIWalled gardens that erode the commonsEven if users dont treat personal data as precious,the companies that they div

85、ulge it to certainly do.Theres a quiet arms race taking place for content and personal data among the platforms that control it.Content-harvesting bots have become pervasive,and so have the efforts to wall up data in order to keep them out.Bots today make up 49.6%of Internet traffic,and 32%of those

86、are labeled“bad”because they attempt to disguise their behavior and mimic users,often for tasks such as extracting data without permission.57 The open Web has become a walled garden.According to the Data Provenance Initiative,between 2023 and 2024,Web data sources have restricted themselves from aut

87、omated accessapproximately 5%of all data,as well as 28%of what it labels as the most active critical sources.58 Its not just the public Web thats being harvested.Increasingly,dominant platforms are hiding data from each other.We are held hostage by applications that own our data,logins,and even iden

88、titiesand every morsel of personal information is invisible to other platforms in an effort to make switching more difficult.For example,Amazon order confirmations no longer include emailed purchase descriptions for fear of information scraping.59 The ubiquity of Secure Sockets Layer ensures that fa

89、cts,such as what we search for,remain tracked but not eavesdropped upon.60 There are many other measures,too:banning VPNs,removing APIs(or enforcing limits),or requiring CAPTCHAs(short for“completely automated public Turing test to tell computers and humans apart”).Five hundred human-years are waste

90、d on CAPTCHAs every single day.61 Companies such as Cloudflare have built a business on warding off bots,62 while others such as CapSolver offer tools to help foil them.63 A recent study suggests that chatbots displace public discourse,too,which“threatens the future of the open Web,as interactions w

91、ith AI models are not added to the shared pool of online knowledge.”64 65 66 13DECENTRALIZATION AND AISome of these activities occur under the banner of consumer privacy protection.However,for many platforms,data protections also raise users switching costs.Bans on bots mean consumers cant use them

92、for legitimate purposes on these platforms,either.If theres a bot,its generally one with strict limitations,and is not a users own bot.Users are often stuck with whatever algorithms,automation,or bots the platform owner provides.The market for attention has created an unhealthy arms race as walled g

93、ardens attempt to corral and lock in humans and keep out the bots.Platforms can detect bots,but the technologies for skipping ads get better,too.“I just want to ask my computer for something,and then it should give me a window into just what I want,”says Killian Lucas,and let the computer sift throu

94、gh the ads and“bear the brunt of the terrible Internet experience that the attention economy has created.”67 Lucas adds,“But its going to piss off a lot of advertisers.I have no idea if theyre going to accept that,or if there are going to be more sophisticated ways of detecting that kind of computin

95、g activity.Ill say,though,that CAPTCHAs are totally easily solvable now by AI.”68 Powerful algorithms that influence and control“Social medias toxicity is not a bugits a feature.Technology works exactly as designed.”Jim Balsillie69Algorithms are at the heart of many social media experiences.The goal

96、 is to hold our attention for as long as possible and then leverage that influence to generate revenue.Its the science of building and maintaining addiction in which companies are investing billions.Meta,for example,spends more than 25%of its revenues on R&D,a significant portion of which goes to AI

97、 and algorithms.70 Its not about whats good for users;its the capacity to keep them engaged.As Guillaume Chaslot told the New York Times,“Watch time is key.The algorithm tries to get people addicted rather than giving them what they really want.”71 Some algorithms are particularly adept at steering

98、our preferences.Take Netflix,for example:“More than 80 per cent of the TV shows people watch on Netflix are discovered through the platforms recommendation system.”72 Some of the best algorithms consistently explore the frontier of our interests to get to know us better.TikTok,for example,recommends

99、 30%50%of videos based on past interactions,but it also mixes in a wider range of“exploratory”videos to glean deeper insights into users tastes.73 74 75 Social media is an experiment,and we are the guinea pigs.Optimizing for attention doesnt maximize individual or societal outcomes,however.Its easy

100、to find oneself in a filter bubble.Touching our phones 2,617 times a day hasnt made us smarter.76 Instead,says Justin Rosenstein,formerly of Facebook,“Everyone is distracted All of the time.”77 Eric Schmidt describes how he views the end-to-end process:“The CEOs,in general,are maximizing revenue.To

101、maximize revenue,you maximize engagement.To maximize engagement,you maximize outrage.The algorithms choose outrage because that generates more revenue.”78 The problem is what these algorithms optimize for;they leave the values of users and society out of the loop,or even harm us directly.79 At a soc

102、ietal level,the result can be an unhealthy level of addiction with questionable benefits for individuals.14DECENTRALIZATION AND AIIf algorithms are the problem,then should we do away with them?David Clark suggests theyre a critical part of what makes social platforms successful:“At heart,you have to

103、 have a recommendation engine.There has to be a feed.This is what makes it exciting.”80 Theyre essential to the platform,but do they really benefit users?“I dont think we should get rid of those algorithms completely,I do see the benefits of them,”says Rui Zhao,Research Associate at the University o

104、f Oxford,especially when it comes to exploring new or unfamiliar areas.81 Yet,developers can design these algorithms to exploit human weaknesses,82 and so problems can arise when users arent in control of those algorithms.We should be able to turn them on or offperhaps the“intention”economy rather t

105、han the“attention”economy,suggests Zhao.83 Algorithms should reflect users values,needs,and interests.Gaslighting us with misinformation erodes trust and autonomy“We have a trust problem in our society,democracies can fail,and the greatest threat to democracy is misinformation because were going to

106、get really good at it.”Eric Schmidt84The following charts(Figures 1 and 2)depict the growth of social media(left)and the rise in the use of the term“fake news”(right).Although correlation is not causation,it is interesting to note that the term“fake news”(which first appeared in 1860)really came int

107、o vogue in 2016,just as many social media platforms rose to global dominance.15DECENTRALIZATION AND AIFigures 1 and 2:Growth of social media(left)and rise in the use of the term“fake news”(right).FIGURE 2:USE OF“FAKE NEWS”OVER TIMESource:Google Trends,https:/ 1:GROWTH OF SOCIAL MEDIASource:Statista

108、and TNW,2019,OurWorldinData.org/Internet,CC BY,https:/ourworkdindata.org/rise-of-social-mediaIn 2018,just two short years later,57%of consumers sourcing news on social media expected it to be“largely inaccurate.”85 Among users who got their news via social media,the level of inaccuracy is also what

109、they dislike about it.In 2018,18%cited inaccuracy as the thing they“disliked most”about getting news on social media;by 2023,that number had risen to 40%.86 And yet,convenience means social media remains an important source of news,one thats viewed as particularly valuable when it comes to following

110、 events in real time or when seeking opinions that help form our own opinions on current events.87 Original fact and evidence(even for this report)may be difficult to source in cases where authors dont include references,let alone cases where news is outright fake(something 59%profess concern about)

111、.88 Increasingly,were wondering about the facts,whats real,and what isnt.Theres a name for thatgaslighting,which Merriam-Webster defines as“manipulating someone into questioning their own perception of reality.”89 Artificial intelligence only makes things worse,whether its creating misleading conten

112、t by accident90 16DECENTRALIZATION AND AIor amplifying our ability to generate fakes.We can fork reality itself,suggests Karl Schroeder,thereby inventing entire worlds that dont exist:“Any amount of information,about anything,can be created on the fly.Entire libraries full of books,manuals,histories

113、,cross referenced with the autobiographies of authors,maps,and so on for entire historical epochs that did not happen.Scientific developments,political theories,anything you can think of,social movements,religious systems with histories going back thousands of years.And then hundreds or thousands of

114、 different documents,sacred textsall of it whipped up in seconds Well be able to do that.”91How can anyone know what to believe?If social media puts individuals in a position where they begin to mistrust reality or question their perceptions,they become more receptive to outside influenceimperiling

115、both individuals and society at large.Martin Baron,former Executive Editor at the Washington Post,is one of many who have warned of these risks:“If you have a society where people cant agree on the basic facts,how do you have a functioning democracy?”92 Grounding points of reference and retaining a

116、grip on the truth are difficult on platforms designed for purveying influence.Living in a free society means having the ability to make educated choices and being responsible for their consequences.“Autonomy is very important,I think,for humans as a species,”offers Wong.93“We are not being treated a

117、s though we have dignity,because data just gets collected about people in ways that we dont fully understand.We certainly dont have a good grasp of where the data goes once its collected.”94 The values embedded in these algorithms dont belong to us,yet theyre changing the very fabric of our society.

118、“Were nudged toward certain types of social engineering outcomes that maybe we dont want,”and we havent consented to those changes.95“We were never asked,”suggests Wong,“and it pushes against what it means to be a human being.”96 17DECENTRALIZATION AND AIRegulatory currents of changeAs the influence

119、 of major tech platforms has grown,regulators have started to put big tech under the microscope:scrutinizing data and consumer privacy,reinforcing consumer protection and liability,and ensuring fair market access and competition.Europe in particular has been a hotbed of legislation.Mike Milinkovich,

120、Executive Director of the Eclipse Foundation,which recently moved from the United States to Europe,describes it as“a cargo ship of regulatory activities that have all been passed in the last couple of years.”97 Whether these actions can breathe life into decentralized players or mitigate some of the

121、 consequences of centralized platform power remains to be seen.Some changes,such as consumer data protection,have already resonated around the world,while others are still being tested.98 Some tech pundits,such as Benedict Evans,suggest that the E.U.isnt a large-enough market to force these platform

122、s to change the way they do business,with the result that“EU consumers get the downsides of the regulations without the upside.”99 For example,additional E.U.regulatory requirements might delay European access to new offerings or,in some cases,lead companies to avoid regional-specific releases at al

123、l.Others see it as a human rights issue where regulation is critical to fundamental protections.100 Some examples of the most prominent regulations are:Human rights:The GDPR offers protection and privacy for consumer data,and the E.U.AI Act aims to improve AI safety and mitigate risks.Fair competiti

124、on:The Digital Markets Act(DMA)emphasizes fair competition in digital markets,and the DSA defines rules for large online platforms and search engines.Property rights:The Data Act covers greater control over generated data(including data spaces).101 Liability:The Cyber Resilience Act introduces liabi

125、lity for software providers.With so much at stake,striking the right balance isnt easy,but its important to find common ground.Daniel Goldscheider,Founder of the OpenWallet Foundation,suggests that“there are some areas where governments and corporations may want to meet each other at eye level.”Forc

126、e is rarely the answer when it comes to the adoption of open technology;instead,its better to focus on the benefits.102 Many regulatory efforts seek to restore the balance of power between dominant platforms and the people who use them every day.Data ownership and portability,the right to be forgott

127、en,and privacy rights are all worthy causes.But some rights arent clear cut.In vehicles,for example,OEMs often believe they own the data,while others say it belongs to a vehicles driver.Ted Guild,Connectivity Standards Lead at GEOTAB,has been involved in a number of automotive standards bodies.Says

128、Guild,“You cant make everything public if you dont know who it belongs to.”103 Other regulations aim to curb the dominance of major platforms and the market power they wield.“We should be using regulation more to break up monopoly power,”suggests Robin Berjon,yet regulation isnt good at exploring ne

129、w possibilities.“You cant innovate by regulation.”104 Major players have been subject to a steady stream of antitrust scrutiny(Google is the most recent).105 The DMA,for example,seeks to further open competition and fair markets,leveling the playing field in areas that a handful of firms dominate.Wh

130、ile some question the effectiveness of such legislation,even the most complex and fast-moving technologies18DECENTRALIZATION AND AIstill offer important points of leverage.“AI has three fundamental componentsdata,compute,and algorithms,”all of which are amenable to regulation,suggests Wendy Wong.106

131、 Other regulations,such as Canadas Online News Act,appear more protectionist.The Act obliges large players such as Google and Meta(Facebook)to compensate news publishers directly.107 Facebook simply refused,dropping publishers from Canadian links.Other regulatory skirmishes are more personal.France

132、arrested Telegrams founder Pavel Durov,who now faces criminal charges from a failure to moderate illegal or extremist content.108 In Brazil,a battle over censorship boiled over into a ban of X throughout the country and a seizure of Starlinks funds.109 Elon Musk called Brazils Alexandre de Moraes“an

133、 evil dictator,”110 while Moras maintains that Musk has“total disrespect for Brazilian sovereignty.”111 These personal and political battles are difficult to decipher.Are these vendettas against free speech or worthy attempts to foster competition and protect national sovereignty?Its not only regula

134、tors who are catalyzing change.Online advocates,open source communities,and their foundations play important roles in maintaining the open Internet as well as many of the public goods we often take for granted.Sometimes the work is simple,ensuring that regulations dont have unintended consequencesor

135、,worse,that“what they have legislated is an impossibility.”112 113 Thats where advocacy work comes in,says Joshua Simmons,aligning regulation with practical considerations.114 Open source is also about fostering community,says Mirko Boehm,Senior Director of Community Development for Linux Foundation

136、 Europe.“They want a neutral home for the project where its not owned by one entity.”115 Thats where a foundation can help.“We basically encourage support and catalyzeif you willthe way different people work together.”116 Daniela Barbosa,General Manager of Decentralized Technologies and Executive Di

137、rector of LF Decentralized Trust at the Linux Foundation,echoes these sentiments:Its“a place where people can actually collaborate on these technologies in order to build something better together.”117 Rebecca Rumbul,of the Rust Foundation,says its important for foundations to fill in the gaps,takin

138、g care of all the things unlikely to happen on their own and engaging on a grassroots basis to help get things done.118 Supporting the momentum of the community is key,allowing it to keep making a difference even when progress itself is difficult to ascertain.Says Rebecca Rumbul:“How do you define s

139、uccess and impact in a field where everyone has a different idea of success and impact can be good or bad(or nothing)depending on which particular vantage point you have?”119 Building consensus on what matters is perhaps the most important contribution of all.How do we build that consensus?Says Dani

140、el Goldsheider,“I think it is really important to come up with fundamental principles that are independent of specific technological manifestations of human ingenuity.”120 That means open collaboration where differences are put aside.Says Goldsheider,“My hope is that governments and private sector c

141、ompanies will increasingly work together to ensure that critical technologies act in the interest of the public.”121 19DECENTRALIZATION AND AIReimagining platforms:The Internets decentralized futureIf centralization is the problem,is decentralization the fix?Its tempting to think so.After all,the In

142、ternet itself is a global decentralized network.Surely we can recast the digital platforms in a more effective decentralized form.But why decentralization,and what are its benefits?The answer depends on who we ask.Vitalek Buterin,the founder of Ethereum,has distinguished among:Physical decentralizat

143、ion,that is,the number of computers comprising a system Political decentralization,that is,who controls those computers Logical decentralization,that is,whether the system is unified and monolithic or can be chopped into viable pieces122Those distinctions help describe system differences.For example

144、,blockchain is architecturally and politically decentralized but logically centralized because it runs like a single computer.123 The benefits of decentralization,according to Buterin,are:Fault tolerance,because of redundancy and diversity Resistance to attack,because theres no single point of weakn

145、ess Collusion resistance,because its harder to collude in ways that undermine the system124 Ultimately,decentralization is about fostering greater autonomy at the edge to enable trusted systems that wont run counter to our interests.Moxie Marlinspike,Co-Founder of the open source messaging applicati

146、on Signal,has a different take on the subject.125 In his view,when many people think about decentralization,they have specific benefits in mind.He believes well-structured centralized systems can offer many of those benefits“increased privacy,censorship resistance,availability and control.”126 Signa

147、l,with a focus on open source,privacy,governance,and user experience,is the realization of Marlinspikes ambitions:“What we want is for technology to better serve us.”127 While not fully decentralized in an architectural sense,Signal appears to fit Buterins description of“political”decentralization.M

148、arlinspike elaborates on these choices:“To the extent that people are manifesting the control they would like to see,its interesting to me that it doesnt seem to have much to do with the decentralized nature of these protocols,that it has more to do with the open source nature of these projects,that

149、 because these projects are open source,its very easy for people to take whats there and just change it and,you know,redeploy it as something else.So,in a sense,I feel like that open source is sort of the best tool that we have in terms of manifesting control.”128 Instead,the decentralized piece of

150、Signals infrastructureMarlinspike doesnt call it thatis the address book already found on phones.129 Social apps need a social network,and one of the best ways to access that network is via an address bookits personal,user owned,and very powerfulbecause users maintain that network even as they trave

151、rse different platforms and services.130 20DECENTRALIZATION AND AIMust everyone,as Marlinspike suggests,accept centralized structures as a necessary evil in order to assert control over technology and enjoy its benefits?Must it be“under the control of a benign dictator?”131 Isnt there a way to have

152、the best of both worlds instead?David Clark suggests its a worthy goal:“Im willing to put some energy into it,because I believe in that outcome.”132 Can we have a structure thats sufficiently agile but also decentralized?There are no guarantees;Clark calls it“a grand experiment that we dont know the

153、 results of yet.”133 Lets take a closer look at some of the decentralized systems and technologies that are making headway on these challenges,as visualized in Figure 3.FIGURE 3:VISUALIZING COMPONENTS OF FUTURE DECENTRALIZED PLATFORMS.Source:Author;Chris Xie and Lalana Kagal also contributed to this

154、 diagram.In addition,the author consulted with Anthropics Claude.ai,which also made helpful suggestions about the visualization and its components.21DECENTRALIZATION AND AIInfrastructure for decentralized social networks“The Times 03/Jan/2009 Chancellor on brink of second bailout for banks.”Satoshi

155、Nakamoto134Using decentralized systems for communications isnt new.In the 1970s,citizens band radios were popular,and centuries ago,seafarers could receive letters at sea.With the advent of the Internet,some might argue that Usenet and Internet Relay Chat systems were decentralized as well.Satoshis

156、message embedded in the genesis block(see Figure 4)may quote a headline from a British newspaper,but it almost certainly qualifies as a social use of blockchain.There were numerous other early systems as well,such as Steemit,Diaspora,and Eth-tweet.135 FIGURE 4:THE SCRIPTSIG FOUND IN THE GENESIS BLOC

157、K ALONG WITH ITS ASCII CONVERSION.Source:AuthorWhile no list is exhaustive,the following is a selection of key protocols comprising a few of the most popular decentralized social media systems worth a closer look:ActivityPub AT Protocol Decentralized Social Networking Protocol(DSNP)136 FarcasterActi

158、vityPub:Mastodon and ThreadsActivityPub is one of the most well-established protocols for decentralized social media.Its what many would describe as a federated system,with endorsement from W3C and running on open protocols.Mastodon and Threads are some of the Twitter-like social networks using the

159、protocol.137 Together,these networks illustrate some of the successes,but also the challenges,of less centralized systems.Freedom and resistance to censorship have been value propositions of these alternative social networks,Renaud Chaput,Mastodons CTO,says its hard to solve the problem of large soc

160、ial networks where youre“really a prisoner in their platforms If your friends are using another platform,or if the platform owner is starting to do things you disagree with,then you are stuck.”138 Thats made these platforms home to marginalized groups,strong supporters of digital freedom,and those w

161、ho have fled the constraints of centralized systems.Its a strange mix of strange bedfellows.Some of these networks received large influxes of users from events such as Elon Musks purchase(and changes to)Twitter,as well as Donald Trumps ban from social media.Not surprisingly,not everyone gets along.F

162、or example,Truth Social(which runs a non-federated social network)was initially accused of using source code without giving proper credit.And Metas interest in embracing open22DECENTRALIZATION AND AIprotocols is tempered by a hesitance to adopt technologies that could put it under regulatory scrutin

163、y(European content sharing requires user approvalsa difficult challenge for a federated protocol).Others are wary of Metas motives,questioning its use of open protocols or shared AI models,for example.Says Nathan Lambert,“The intent is clearMeta wants the Llama brand to touch as much of the open eco

164、system that it can.”139“Altruism is not a business model,”suggests Mike Milinkovich,so its important to consider practical considerations:“Every one of these companies is going to take the actions that they think are best for their business.”140 Implementation differences can also mar user experienc

165、es and bottleneck growth.Making the platform more accessible(not just for those with technical skills)is a challenge Mastodon is still working on.“Compared with many other open source projects,we take a lot of care about user interface,onboarding,and accessibility,”says Chaput.141 The choices are so

166、metimes controversial.For example,Mastodons initial signup system was very open,allowing users to pick from hundreds of servers.It was a daunting process,and users simply walked away:“They wont do it because its so complex and they dont understand it.”142 The solution was a new default button to cre

167、ate an account on mastodon.social.“We had a lot of backlash”about that,says Chaput,but“we think its better for them to create an account than to leave.”143 In the future,offering hosting services may be another way to keep things simple.For the European Commission,for example,a hosted solution may b

168、e easier than trying to do it themselves.144 Infrastructure constraints and fragmentation have also held up key features.Allowing users social graphs to follow them from one service to another has always been part of the attraction for less centralized systems.But the feature has proved difficult to

169、 implement.Today,users can move their accounts,but their posted content doesnt always follow them.Rather,users“broadcast”their move to a new address so that followers can follow them there.145 But putting additional burdens on users isnt the answer.“Content creators are burned out by the amount of s

170、ocial networks they need to manage,”said Chaput.146 Another issue is that the federated system of ActivityPub is itself unevenly adopted.Chaput says that despite the fact that Truth Social is based on Mastodon,“one of the first things they did was to disable Federation They are a closed social netwo

171、rk.”147 Nor is it clear that a server should automatically accept content from other instances,either.Some servers lack moderation,with the result that they may host undesirable,banned,or even illegal content or images.148 While algorithms are often key to such moderation,the opacity of such algorit

172、hms(especially when they dictate whats in users feeds)means their use is something that“people in the Fediverse are very strongly against.”149 As a result,the software will lack algorithmsmaking moderation,automated spam fighting,trust and safety,detection of illegal content,and even content discove

173、ry and recommendations a little more challenging.150 The resolution is decentralized search and discovery that is implemented externally instead.151 AT Protocol and BlueSkyAnother popular destination for those leaving centralized systems is BlueSky.For example,in August 2024,the Brazilian government

174、s disagreement with Musks X led to a country-wide ban in Brazil.152 Many switched to BlueSky,which claimed a 2.6 million increase in users in a single week,85%of whom were Brazilian.153“Altruism is not a business model,”Mike Milinkovich Eclipse Foundation23DECENTRALIZATION AND AIThe founders of Blue

175、Sky viewed it as a proof of concept for the AT Protocol,which is based on a microservices architecture(rather than large monolithic server instances).154 Initially spun out of Twitter,the application for BlueSky is open source(see GitHub).The AT Protocol is currently governed by a public benefit cor

176、poration,whose stated mission is“to develop and drive large-scale adoption of technologies for open and decentralized public conversation.”155 Its CEO,Jay Graber,claims the company is“moving toward standardizing pieces of the protocol.”156 BlueSkys microservices architecture is a deliberate choice,s

177、omething Graber says makes it easier for features such as“composable moderation,”a decentralized solution for tailored moderation.157 Joshua Simmons of Matrix suggests that these kinds of innovative algorithms are valuable features when it comes to accommodating diverse values and points of view:“Yo

178、u can layer on different frameworks of moderation to create the experience that you need.”158“Theres about 40,000 custom algorithms out there and custom feeds,and some of them are very simple,some of them are very complex,and users can install them and switch between them,”offers Graber.Yet other fe

179、atures are global,such as search and discovery,account portability,or the ability to consolidate all feedsthese key features were planned from the outset.159 DSNP and Farcaster DSNP was created by Project Liberty,160 a nonprofit founded in 2021.161 The goal was to create a public protocol that would

180、“free the social graph from central control.”162 Like ActivityPub,social application functions such as moderation and algorithms were not designed as a part of the protocol but remain possible at higher levels of the technology stack.The project distinguishes itself by its use of blockchain as part

181、of the protocol.The attention to scalability(the stated aim is to support billions of users economically)means that running significant parts of the protocol over blockchain is infeasible.So blockchain technology was considered for narrow,yet critical,parts of the protocol.As David Clark puts it,“th

182、e root of the tree is always,always,magic.”163 For Clark,the“magic”part is identity.164 Identity is tricky because so much rides on its implementation.“Key management is the ugly duckling of all of this stuff,”and there are a lot of edge cases to handle.165 People lose passwords and devices,and so t

183、hey need a recovery strategy.Even with a mobile device,says Clark,“itll be stolen,itll fall down a manhole.Its got to be replaceable.”166 That means having either a trusted agent(or system)that can vouch for ones identity or a publicly visible“identity assertion on a blockchain.”167 Project Liberty

184、opted for the latter:using blockchain to help minimize the number of points of“trust”in the system.Today,a social network called MeWe is one of the largest adopters of DSNP.In August 2024,MeWe announced 1 million on-chain users,500,000 of whom control their complete social graph on the blockchain.16

185、8 The foundations main contributor is also said to be considering other acquisitions,the most ambitious of which could shorten the path to a critical mass of adopters.169 Farcaster is another open protocol with on-chain components(Warpcast is a client)designed to aid the creation of social media app

186、s.170 The protocols stated goal is to address concerns of“privacy,monopolization,and censorship.”171 The project has a number of on-chain components;however,it uses this“friction”to its advantage.A$5 fee helps to deter spam accounts,for example.While adoption levels appear to have leveled off,one of

187、 its users,Vitalek Buterin,praises it as a“Twitter alternative.”172 When somebody hacked Vitaliks X account,Vitalik used Farcaster to explain what happeneda socially engineered SIM swapand plugged Farcasters strong form of identity management:“Anyway,glad to be on farcaster,where my account recovery

188、 can be controlled by a good wholesome ethereum address:)”173 24DECENTRALIZATION AND AICommunication and messagingWhile many decentralized applications have focused on the social graph or microblogging,others,such as Matrix,have focused on messaging itself.David Clark scrutinized Matrix as part of h

189、is own research for Project Liberty.He describes the elegant simplicity of its decentralized ecosystem:“Its about as simple as you can get.Each user is represented by some software that runs on a client,and you can have multiple clients,and then those are synchronized by a package that they call the

190、 home server.when you want to send a post,it looks at all of the people its going to,which are defined as being part of a room.Thats their vocabulary.”174 Joshua Simmons,Managing Director of the Matrix.org Foundation,describes Matrix as“a federated decentralized communications protocol.”175 Its suit

191、ed to many different kinds of communication,real time chat,voice,or video callsvirtually any form of communication,even virtual reality clients,or AI-enhanced chat.176 While Simmons is well versed in the technologies behind Matrix,hes most excited about the communities that Matrix enables.What is Ma

192、trix really about?“Free,open,secure,private decentralized communication for all humans,”177 says Simmons,“Our view is that that is a human right.”178 Simmons is passionate about open source and the alternative these communities offer.In the early days,platforms such as Twitter were very developer fr

193、iendly,and it was easy to build applications on top of them.Over time,these platforms began to close themselves off and became increasingly centralized.Their closure created an opportunity to be seized.Says Simmons,“That drove home for me the power we give away when we buy in to closed social platfo

194、rms.”179 Less-centralized systems mean no single platform or entity can pull out the rug.180“Ultimately,we need to empower people first to have private secure messaging that they own.”181 That passion led Simmons to various open source and community roles before he joined the Matrix.org Foundation.B

195、ut how does an open platform maintain and steward the resources needed to sustain itself?For open source ecosystems,sustainability can be a constant struggle,one that requires balancing tensions with commercial interests.With Matrix,for example,the for-profit company Element has been a primary contr

196、ibutor.Yet,to help fund its own development,Element has switched to an AGPL v3 license.182 It forked the original Apache-licensed repositories of Matrix.org to exploit dual-use licensing opportunities for Element.183 184“We didnt love that.And our policy is still to stick to permissive licenses that

197、 maximize adoption and experimentation,”says Simmons.“Our role as an open source foundation is to be a steward,not an organization thats turning a profit For Matrix,our mission is much broader.”185 Sound stewardship,protocol simplicity,and an emphasis on values have contributed to the adoption of th

198、e Matrix open communications software.Those values resonate strongly;in some cases,they have outweighed other drawbacks,such as ease of use.Simmons tactfully suggests that Matrix sometimes requires a little bit of patience:“There might be some sharp edges here and there,and,truth be told,building fe

199、derated,end-to-end encrypted systems is a thorny problem.And building a polished user experience on top of that is not easy.”186 But,ultimately,the importance of security,privacy,and data sovereignty helps overshadow some of the drawbacks.Public sector adoption has been strong,particularly in German

200、y and France.187 Matrix has done an excellent job of matching its value and mission with adoption by parties that are seeking those requirements(and,often,highly discerning in their technology choices):188 25DECENTRALIZATION AND AI Educationespecially at a number of German universities.Healthcarethe

201、 national agency of Germany,for example,will use Matrix for messaging interoperability.189 190 Defenseadoption in military contexts has included Germany,the United Kingdom,the United States,and Ukraine.Open sourceLinux distributions,as well as foundations such as Mozilla,and a number of Linux Founda

202、tion projects.Marginalized groups or the oppressedexamples include minorities,journalists,dissidents,queer/trans communities,and communities that harbor concerns about safety on other platforms.As adoption becomes more mainstream,identity is an area that demands immediate attention.“I see people fru

203、strated with the state of identity who say,I would really like to be able to use one account for all of these things,”says Simmons.191 The challenge is to offer consistent experiences with identity without relying on centralization.Matrix will soon work with other identity systems,such as OIDC,and s

204、upport a new native alternative.Says Simmons,“We have our own in-house matrix authentication service that weve built.”192 Users will benefit from the convenience of logging in to Matrix with other identity providers,but in the long run,they may also use their Matrix identity to authenticate with oth

205、er services,too.Whats in store for Matrix users in the future?Joshua Simmons believes its about creating a more polished user experience and delivering the values its users hold dear:193“Were at over 100 million users today,which we feel very good about.But to get to a billion and beyond,the two thi

206、ngs we have to solve is we have to deliver a polished user experience that feels much more like a drop-in replacement to existing communication tools.The other thing that we have to solve is delivering on trust and safety.”26DECENTRALIZATION AND AIPersonal data storage and provenance“If you want to

207、keep a secret,you must also hide it from yourself.”George Orwell,1984In the attention economy,it is not possible to retrieve“spent”private data.In the digital world,the ability to make unlimited perfect copies is both a blessing and a curse.In the same way that absence of scarcity creates the“double

208、 spend”problem for digital currencies,it also creates privacy issues.194 Call it the double spend problem of private data.Those who receive private data may store,share,or make copies of itoften in ways that are difficult to anticipate(take vast libraries of AI training data,for example).Today,the p

209、rotection of our identity and personal data relies on promises,often with the backing of legislation about how the owners of that data must store and share it.Are these promises trustworthy?History has proven that once the genie is out of the bottle,it is very hard to stuff it back in.Bad actors may

210、 break these promises,but more often,users forego these rights,clicking them away in complex user agreements.Even with the very best of intentions,personal data is vulnerable to hacking or theft.Increasingly,users face zero-sum choices between disclosing personal information and benefiting from the

211、systems and services that feed on them.Today,shielding ones data means cutting oneself off from services and platformsa steep price to pay.We need something bettergreater control and transparency over our personal data.Says author Karl Schroeder,“We should own our own information.”195 Thats the idea

212、 behind the regulation underpinning Data Spaces,says Mike Milinkovich;systems of federated data in which data can aggregatebut only under specific,user-defined terms and conditions.196 An open protocol called Solid takes this a step further,with a technology option that literally puts users in contr

213、ol of their data via“decentralized data stores called Pods.”197 Its the right place to start,suggests Jesse Wright,a doctoral student at the University of Oxford:“This creates a much better online experience.By decoupling applications from user data,the Web can become user centric rather than compan

214、y centric.”198“As a user,I dont want to have to create a new login and re-enter the same personal details across every website I visit,”so the principle of owning your own data is a natural way to enhance interoperability,says Wright.The“Solid and Semantic Web technologies provide solutions for many

215、 of these issues”.199 Equally important is the ability to transparently specify permissions related to our private data.“Some kind of additional information needs to come with the data,”says Rui Zhao,formal descriptions of how it should or shouldnt be used,and what the obligations are.200 The sharin

216、g of data gives users explicit assurances about what will happen to it.Says Zhao,“We are generating new data,and we are sending data to others,and we are processing data.And its no longer about just access control,its more than access control,because were talking about the rights or permissions of t

217、he new data.”201“True self-sovereign identity and data portability”are becoming increasingly important,suggests Daniela Barbosa.202 Its the idea that“Im going to take the stuff I own and actually create better experiences for myself.”203 Alone,this doesnt exactly solve the double spend problem for p

218、rivate data;we still rely on promises to protect it.But users gain formal explicit control over what those promises are,making data obligations far more transparent.Regulations give these obligations legal teeth to ensure compliance,but technology helps here,too.When aggregating data,for example,it

219、can be difficult to ensure that individual data remains private.27DECENTRALIZATION AND AI“Differential privacy is one definition that is typically very strong,”suggests Hart Montgomery,CTO of LF Decentralized Trust,and its application in the U.S.census is a well-known example.204“Differential privac

220、y is essentially a mathematical statement that says its hard to tell whether you are included in a database or not;it effectively means that the database doesnt leak any of your personal information.”205 Federated learning is another key enabler,allowing models to train on data even when its not sto

221、red centrally.Instead,federated learning helps preserve data rights,allowing training data to be held by separate decentralized entities.In addition,we also have techniques such as zero-knowledge proofs that allow us to verify information about private data without having to share it.For example,a u

222、ser may have a credential from a government agency,and it is demonstrable with zero knowledge that they possess properties related to that credential without having to directly disclose all its details.“As you might suspect,”suggests Hart Montgomery,“this is extremely powerful for privacy.”206 At th

223、e opposite end of the scale are techniques that help make data publicdecentralized systems that,instead of focusing on privacy,can help ground content or data,creating a common point of reference.The provenance of shared information is critical,and its important to access and verify sources of groun

224、d truth.Blockchain does this rather well,of course,and IPFS does,too.But Robin Berjon doesnt think of IPFS as merely a file system:“Its called the interplanetary file system,and its neither interplanetary nor a file system.”207 Says Berjon.“Whats quite interesting about IPFS is that its basically an

225、 architecture of protocols built around content addressing.”208 That“shifts architectural power on the Web,”says Berjon,because the server and its address is no longer a point of control.209 With IPFS,when you request content,the pointer to that content is the hash of the content itself.It bestows m

226、ore power to clients at the edge;users dont need to“connect to the world to function.”210 What is a word worth?The word purple is worth$40,000.At least,it was to Luke Farritor.Farritor decoded the word“purple”from a 2,000-year-old lump of charcoal using X-ray tomography,winning the Vesuvius Challeng

227、es grand prize in the process.211 Its one of 1,800 carbonized lumps just like it,called the Herculaneum scrolls.All of these scrolls underwent instant carbonization in 79 AD when Mount Vesuvius eruptedand the buried scrolls werent found until 1752.Because their significance was not known,many are th

228、ought to have been thrown out or lost.Others were damaged or destroyed in the attempts to read them:Some were cut in half,some disintegrated as they were unrolled,and the contents of some simply vanished when exposure to air faded any trace of the writing.Today,AI and 3D imaging are finally unlockin

229、g the centuries-old secrets hidden within each scroll.Ironically,carbonization itself helped preserve thema regular scroll would be unlikely to survive the test of time.Every word from such ancient documents is precious because the further we go back in time,the more scant our information becomes.Go

230、 further back than 2,000 years and provenance itself begins to break down.All that remains are gaps,contradictions,mysticism,broken clay tablets,and rubble.Even the existence of entire ancient libraries can remain a mystery.The library at Sarough(modern-day Iran),for example,may date as far back as

231、500 BCE.212 By one account,it once inspired awe similar to the Egyptian pyramids.213 Yet today,the ruins at Sarough are scarcely more than a small hill.We only know of the existence of the library from passing mentions by 10th-century scholars who,in turn,reiterated others reports of books discovere

232、d among the ruins in 350 AD.214 215 28DECENTRALIZATION AND AIAlgorithms,too,can shift behavior to the edge.GitHub,for example,is a treasure trove of public behavioral data.Rodrigo Mendoza-Smith says Quira uses that data to“quantify developer reputation from the data they generate.”216 The goal,Mendo

233、za-Smith says,is to create a market with perfect information:“We are building a market oracle that enables this market and gives people back ownership and control of their experience data.”217 Todays job market for software developers is far from that ideal;Mendoza-Smith calls interviews a“costly an

234、d lengthy process of talent validation.”218 Without this friction,everything changes;developers“can just jump into the market and then get the best economic opportunity for them right away.”219 None of this is possible if activity records are stuck in proprietary databases.Thats why ownership of exp

235、erience is so important.Artificial intelligence adds another element to the mix.While artificial intelligence is normally thought of as orthogonal to privacy,220 it also affords new avenues for processing and sharing personal information.Its a possibility explored in a recent research paper,where ge

236、nerative AI language models use retrieval-augmented generation to access Solid Pods storing private documents.221 The result combines the ease of natural language retrieval with structured rules for information management.“New LLM technology is going to open up a lot of opportunities because more an

237、d more of our systems will be able to interact with us in natural language,”says Tim Finin,Willard and Lillian Hackerman Chair in Engineering and Professor of Computer Science and Electrical Engineering at the University of Maryland,Baltimore County.222 This offers a valuable complement to ontologie

238、s or structured knowledge graphs.223 Apple is another company making use of AI in the context of privacy(see the section starting on page 34).Suggests Killian Lucas,“The direction that Apple is going in with privacy is a great one.They are thinking about trying to have local models.”224“Theres no re

239、ason why we cant have assistants that know us extremely well and act as liaisons.They act as mediators between ourselves and other entities,which could be companies,but it could also be other individuals,”suggests Schroeder.225 Think of it as“an assistant that you own that can isolate your informati

240、on.”226 Its a compelling idea.For as long as humanity has held secrets,compartmentalization has been useful to manage them.Large organizations may break information up among silos,none of which has a full picture.Anthropic,for example,practices internal compartmentalization,describing how its platfo

241、rm provides access to sensitive or proprietary information and code on a need-to-know basis.227 Who makes the decisions about these compartmentsand whether they are trustworthyis an important question.Unfortunately,conventional forms of compartmentalization are difficult in practice.On one hand,comp

242、artmentalization itself limits the benefits and insights possible with a birds-eye view.On the other hand,managing compartments is onerous and difficult to scale because of labor-intensive processes such as cleaning data passed among compartments.Artificial intelligence affords new decentralized app

243、roaches for tackling both these problemstackling the double spend problem for Why is history so important?Because it tells us who we are and how we got here.Its the same for personal information.In a world rife with misinformation and distorted facts,the provenance of information matters.We need acc

244、ess to ground truth:sources of information and facts we can trust.It informs who we are and how we interact with others,like looking at ourselves in a mirror.Gauging the clarity and distortions of that mirror is important for an accurate picture.So it is with information.29DECENTRALIZATION AND AIdat

245、a head-on.AI agents represent a new,and potentially disruptive,opportunity in their own right,something Section 5 examines in more detail.Winning with decentralizationInnovators have made tremendous progress in these domains,in many cases approaching parity with centralized platforms.Decentralized p

246、latforms for private and secure messaging,social uses,and even payment applications are all within reach.But competition is stiff,and feature parity alone may not be enough.An installed base of billions of users creates significant network effects,and switching costs are high.Despite the rapid growt

247、h trajectories of new and emerging platforms,the bulk of them barely register against the traffic levels of incumbents(see Figure 5).FIGURE 5:TRAFFIC DATA(WEB)BY PLATFORM FOR JUNE,JULY,AND AUGUST 2024.Source:Author,compiled with data sourced from SimilarWeb228 30DECENTRALIZATION AND AIPlatform lock-

248、in is no accident.One of the most desirable features of decentralized platforms is the portability of personal data and social graphs,the very thing these centralized platforms attempt to control at any cost.Users remain loyal to these platforms,not necessarily because of exceptional features or exp

249、eriences(though many have these too)but because its where others they know hang out;its a community.Thats a powerful barrier to entry that wont be lightly relinquished.New platforms need critical mass.There are many approaches to obtaining critical mass.Here well consider just two,which we might lab

250、el(1)building a better mousetrap229 and(2)disruptive innovation.For many areas of application,blockchain and other key decentralized infrastructures are the“better mousetrap.”Slowly and inevitably,assets and infrastructure are starting to migrate to a Web3 stack.Over time,the balance of power will s

251、hift as key use cases and applications make their way into the ecosystem.Blockchain and other decentralized technologies are making headway in some of the areas that matter most.Security,for example,is an area where the underlying blockchain technology is extremely sound,says Hart Montgomery:“People

252、 should consider it mature.”230 The ongoing safety of attractive targets(in this case,billions of dollars worth of assets)is a really good way to judge security,suggests Hart Montgomery:“Thats pretty good evidence that the technology is in good shape.”231 Blockchain is also becoming home to many oth

253、er pieces of robust essential infrastructure,and even nation states are starting to take notice.Take identity for example:“We see a lot of people,including governments,use blockchain as the root of trust for distributed and decentralized identity.”232 Governments around the worldin Brazil,Hong Kong,

254、and Singapore,for exampleare exploring central bank digital currencies(CBDCs).Suggests Daniela Barbosa,CBDCs are something fundamental“that every sovereign nation needs to figure out.”233 Infrastructure changes can happen quickly.Take the India Stack,for example,“the moniker for a set of open APIs a

255、nd digital public goods that aim to unlock the economic primitives of identity,data,and payments at population scale.”234 Billions of people have already adopted the initiative,and it is now being offered to other countries around the world.Even areas where gaps exist are receiving balanced attentio

256、n by developers,communities,and early adoptersa must for grand technical undertakings:Ease of use.Blockchain technologies may not be easy to use,but“the early Internet was pretty awful,too.”235 Imperfection creates plenty of room for improvement.Infrastructure thats robust.Mature infrastructure does

257、nt happen overnight.A robust infrastructure is very different from a proof of concept,says the Linux Foundations Daniela Barbosa:“It takes a long time for these technologies to be adopted and for the technologies to be hardened and proven right.”236 Secure.“Security,even among devs,is a very niche p

258、ursuit,”suggests Rebecca Rumbul of the Rust Foundation:“Historically,it was always kind of a bolt-on thing.”237 Security gaps are finally gaining recognition,both within languages themselves and at every level of application development.Enterprise adoption is also gathering steam.Theres a lot of sup

259、ply chain use cases in areas such as track-and-trace,for example,says Hart Montgomery:“Were seeing a big wave of adoption in finance.”238 The outlines of an emerging finternet are beginning to manifest:“multiple financial ecosystems interconnected with each other a user-centric approach that lowers

260、barriers between financial services and systems,thus promoting access for all.”239 31DECENTRALIZATION AND AITom SerresInvestor,Strategic Advisor,and Co-Founder of Warburg Serressuggests that mainstream adoption is just a matter of time,because“network effect is stronger over here than it is over the

261、re.”240“The lock-in to the network will be even more difficult to disrupt than the original Internet.“Its not just sending and receiving data,”asserts Serres,“Im actually tying my assets to this thing.”241“The lock-in of that network effect is just going to be that much more robustand more difficult

262、 to displaceonce it hits that escape velocity.”242 Serres predicts that we can expect adoption to occur roughly twice as fast as the Internet.By 2030,we will“have full proliferation and saturation of Web3 around the globe.”243 But what if we dont want to wait till the end of the decade?Can disruptiv

263、e innovation catalyze more immediate change?What extraordinary areas of opportunity promise to destabilize incumbent industries and platforms?Artificial intelligence.32DECENTRALIZATION AND AIAI,agents,and algorithms catalyze new platform possibilities“The hottest new programming language is English.

264、”Andrej Karpathy244Moshi is one of the lowest-latency conversational agents in the world(reacting in tens of milliseconds),capable of listening and speaking at the same timeso fast that it exhibits a tendency to interrupt and take over conversations.245 CTO Laurent Mazar laughs,explaining that the l

265、atency(at a recent demo)was set as low as possible to showcase the speed and fluidity of conversations.“Its probably not what you want in real life,because an assistant should probably not interrupt you every two sentences.”246 Mazar is part of a team of eight at Kyutai,a French nonprofit lab dedica

266、ted to open research in AI.247 How did this small team develop a world-class disruptive technology?“We reused lots of open source technology,”suggests Mazar.248 That“allowed us to go very quickly.”249 Artificial intelligence itself also played a central role in accelerating development,both through

267、the use of synthetically generated training data as well as the use of AI as a development productivity tool.“I would be a far worse coder if I didnt have access to the different AI models,”said Mazar.“I use them a lot,and I find that they boost my productivity by quite some margin.”250 Thats a stro

268、ng statement coming from a world-class developer who,in his“spare time,”developed Candle,a minimalist ML framework for Rust(a repository that has 15.1k GitHub stars).251 Artificial intelligence is not only a source of new features and applications;its becoming a transformative platform in its own ri

269、ght.252 Centralized players swim against that tide,reshaping artificial intelligence to conventional business models and platforms in an effort to cement centralization and dominance.Yet decentralized building blocks,AI,and algorithms also create fresh possibilities.New breeds of personalized autono

270、mous agents put intelligence at the edge,affording us greater control over our data and the algorithms that govern interactions.Easy-to-use interfaces and on-the-fly code generation not only democratize powerful tools but also have far-reaching implications for platforms,software,and even open sourc

271、e itself.The following sections take a look at both how conventional players are using AI and how AI catalyzes new and disruptive platform opportunities.33DECENTRALIZATION AND AIArtificial intelligence in todays centralized platforms:Apple“It is not inherently evil to bring people back to your produ

272、ct Its capitalism.”Chris Marcellino,former developer at Apple253Tech giants and other large companies are expected to invest$1 trillion in capital expenditures in the next few years,with Meta alone spending$40 billion on AI infrastructure in 2024.254 255 Incumbents find themselves between a rock and

273、 a hard place.On one hand,disruptive changes threaten core business models.On the other,Wall Street is questioning whether the trillions invested in AI will ever pay off.256 Tech giants are responding in real time,adjusting offerings,altering business models,and making eye-watering investments in an

274、 effort to outspend and outperform peers.It shows few signs of abating.There are rumors of OpenAI attempting to secure another$6.5 billion round of investment at a valuation of nearly$150 billion(up$50 billion from a week prior).257 While a closer look at the strategies of any number of tech giants

275、would prove instructive,well instead focus on just one:Apple.The company has commandeered the very term“AI”in service of its own business model,purveying it to customers as Apple Intelligence.258 Its a field where Apple has been forced to play catchup.259 Siri,for example,has yet to match the Q&A pe

276、rformance of other smart assistantsone of the things that makes Apples approach a little more interesting.260Privacy and user experience are core to Apples strategy.Its goal is a personalized contextual experience that keeps user data private.That privacy comes through a mix of local LLMs,private da

277、ta storage,and a secure auditable cloud,which Apple stated that even it cant access.261 Queries undergo triage:simple requests can stay local while more complex queries are relayed to advanced LLMs in the cloud.Those with premium devices(such as Apples latest 35 TOPS processor)will run more advanced

278、 models locally.262 Apple intends to make the most local LLMs through what it calls adapters,“small neural network modules that can be plugged into various layers of the pre-trained model,to fine-tune our models for specific tasks.”263 These adapters are tailored to experiential contexts:summarizati

279、on,proofreading,mail replies,tone adjustment,refining,query handling,friendly,and urgency.The idea is that AI(with contextual knowledge)will streamline experiences.264 Contextual knowledge means that the AI“knows”whats on a users screen and the actions available to it,including via other apps,so tha

280、t it can gauge the users intent and respond intelligently to queries.265 Attention to user experiences and safety has resulted in a carefully curated migration process that provides Apple Intelligence with deeper access to an apps capabilities.266 Christopher Nebel,an Apple engineer,describes how Ap

281、p Intents work:“They take the core features from inside your app,actions and content that are meaningful to someone using your app,and present them outside your app.To do that,you need to expose your apps core features in a way that the system can understand.App Intents is the framework to do that.”

282、267 Figure 6 below offers a simplified view of Apple Intelligence;a more complete diagram is available in an explanatory video,268 and Apple has also published a detailed discussion of its foundational models.269 34DECENTRALIZATION AND AIFIGURE 6:A SIMPLIFIED VIEW OF APPLE INTELLIGENCE.Source:Author

283、Detailed personal data plus an intelligent agent(Siri)that can act upon it is a potent combination.Siri will receive privileged access to users,their apps,and their devices.OpenAIs Sam Altman professes wishing for a tailored AI with nearly unlimited access to personal data that“knows absolutely ever

284、ything about my whole life,every email,every conversation Ive ever had.”270 Apple hasnt shared its road map,but such an assistant,were it to exist,would possess an unparalleled level of influence over its users.Today,user queries map to specific intents and actions that can summon and work with apps

285、(or even third-party AIs)to help minimize“friction”and put AI in the drivers seat.In the long term,these advanced AI capabilities may deliver more of what users need on the flyraising interesting questions about whether Apple itself might one day“forward integrate”into the domains that apps occupy o

286、n its platform today.User interfaces are changing,too.Apples Math Notes application(link)offers a sneak peek at the UI experiences that AI in the future might afford us.271 Today,these features allow Apple to champion clear wins for user experiences and private data protection.The challenge,of cours

287、e,is that user experience and privacy will improve,but only for those within Apples ecosystem.That will cement Apples control over users and their data(and perhaps even the App Store itself)even further.User data may remain private,but without direct access and portability to other platforms,the dat

288、a really belongs to Apple.This caters to privacy and regulations,says Robin Berjon,“but it doesnt change the power dynamic,which is the thing that matters.”272 Apple Intelligence strategically locks private data into its platform,then uses privileged contextual knowledge and advanced intelligence to

289、 contain the user experience.Platforms with these characteristics become digital gatekeepers,disintermediating basic experience.35DECENTRALIZATION AND AITrue decentralization via reflective autonomous agents“If we want to advance AI beyond its current capabilitieswe want more than AI that can see an

290、d talk.We want AI that can do.”Fei-Fei Li273If artificial intelligence can further centralize todays dominant platforms,can it also do the opposite?How might decentralized AI systems create something fundamentally new and disruptive instead?Decentralized AI agents are an exciting possibility.The ide

291、a of autonomous decentralized agents sounds well suited to blockchainits easy to imagine networks of AI agents orchestrated through smart contracts.Yet deterministic computing systemsblockchain or otherwiseare rigidly constrained.Overcoming those constraints requires a broader definition of“decentra

292、lization”something beyond what blockchains alone offer.“That blockchain with the central ledger is trying to impose a unified perspective on the system,”says Munindar Singh.“This puts a limit on what decentralization we can have because your perspective may be overwritten unless you are in control o

293、f the entire universe,i.e.,the blockchain.”274“We have to go beyond our closed-loop thinking that we are accustomed to in computer science and technology generally,”suggests Singh.275 True decentralization occurs in the real world where people use multiple systems,each with its own rules and structu

294、re.People choose the systems they use and the rules to abide by.Thats what truly autonomous decentralized agents look like;picking the systems they wish to use and the rules they choose to follow.There are always norms that give us acceptable paths,says Munindar Singh,but deviations are possible.Dep

295、ending on the situation,some deviations may be desirable but not others.276 Those extra degrees of freedom are quite valuable,and in more rigid systems,its why a human is kept in the loop.Says Amit Chopra,Associate Professor at Lancaster University,“Autonomy requires modeling systems at the social l

296、evel,and thats where norms come in.”277“A socialtechnical system should be able to improve itself,”says Singh.278 To do that,it must be possible to break the rules.“Thats the heart of autonomy,”suggests Singh,an agent thats“able to violate a contract.”279 Its a significant departure from blockchain,

297、but its how self-organizing systems improve and get better.“People do that all the time.”280 Changing a system and“showing ones work”is an essential practice in many self-organizing systems,whether its a fact-checker making edits to Wikipedia or a developer making a commit after updating a line of c

298、ode.However,it can apply to altering a systems rules as well.Choosing to violate an agreement might be as simple as missing an appointment when conflicting priorities arise,or it might have more serious consequences,such as if a disagreement leads to the forking of a project.Rigidity constrains our

299、ability to innovate,adapt,and reach consensus,and that lack of flexibility has been a big part of whats been missing from conventional automation or from decentralized computing infrastructures.Rigidly designed systems running on structured code cant do this;they simply follow the rules.To understan

300、d just how valuable rule-breaking can be,its useful to revisit the Cold War,when tensions among superpowers ran high.In 1983,an early warning system detected a missile.Stanislav Petrov failed to follow orders to launch a retaliatory nuclear attack after the enemy launch was“detected.”36DECENTRALIZAT

301、ION AND AIPetrovs judgment(that the system was malfunctioning)was credited with preventing a full-scale nuclear war.281 Systems without a human in the loop,operating in pure deterministic fashion,can put themselves at risk from these kinds of unanticipated consequences.282 This inability to adapt is

302、 why some,including Marlinspike(see Section 4),say that centralized infrastructures are necessary.Greater autonomy is the solution.While systems may have rules and norms,an autonomous agent“can do more than what the norm says.So the agent scope is larger than the norm scope,”suggests Singh.283 Samue

303、l Christie,an academic researcher from North Carolina State University,says,“An important aspect of the way that we look at agents and decentralization is that all interaction,state,and reasoning is local,so its all held by the agents instead of the system,and that allows you to decentralize it and

304、give those agents the autonomy that they need.”284 This allows for new automation of systems without the constraints and baggage that automation tends to bring with it.Its the freedom from constraint that many centralized systems have been missinga redistribution of power to the edge.Stafford Beer,a

305、n early proponent of such systems,suggests that“it is a prerequisite of viability that a system should develop maximum autonomy in its parts,where maximum is defined to mean short of threatening the integrity of the whole.”285 Its easy to implement,too;some might call it a simple rewilding.Weve neve

306、r had an opportunity to build automated agents like this beforebecause,until recently,those capabilities didnt exist.Today,were seeing the advent of large action models(LAMs)AI models that are also capable of actions(such as browsing the Web or writing code).So far,proprietary hardware-based version

307、s of the concept(Rabbit and AI Pin)have proved somewhat disappointing.It remains to be seen whether AI Friend will experience the same fate.LAMs can interact with digital systems or even act directly upon the world itself,almost as a quasi-assistant or-advisor.The idea of AI systems interacting dire

308、ctly with the world isnt new to the likes of Tesla(automobiles)or 1x(robots),which are already using neural networks for end-to-end control.But its an interesting idea for other kinds of platformsthe billions of edge devices we use every day.Autonomous agentsagents that are open,have access to platf

309、orms and resources(not necessarily“stranded”on one),reflect upon their own actions,286 and can organize with one anotherare a potent and disruptive recipe.It opens new architecture possibilities(see Figure 7)that are a sharp departure from AI on centralized proprietary platforms.As these agents get

310、smarter,theyre increasingly able to plan ahead and even think through their actions via hidden chain-of-thought reasoning.Greater agency resolves the timeless dilemma facing designers of decentralized systemshow to create something open and decentralized yet flexible enough to adapt to a world thats

311、 constantly changing.Autonomy at the edge allows us to finally cut this Gordian knot.37DECENTRALIZATION AND AIFIGURE 7:ONE POSSIBLE MODEL FOR USER-CENTRIC AI AGENTS AT THE EDGE.Source:AuthorTwo very interesting open source examples highlight what emerging edge-based autonomy might look like:Open Int

312、erpreter and Agent Zero.Open InterpreterOpen Interpreter is an AI that has direct access to your laptop and uses code(via an LLM interpreter)in order to operate it.Its more open than Rabbit287 and allows you to specify an interpreter and,of course,operate locally with local compute resources.Think o

313、f it as a language model computer interface,Killian Lucas,the founder of Open Interpreter,suggests.Its“an intermediary layer between you and the operating system.”288 In the same way that Apple and Microsoft imbue their systems with AI,an LLM interpreter is a very powerful way to do the same thing f

314、or Linux.And it has two key advantages:First,a voice AI offers an easy and incredibly intuitive user interface.Its a fantastic way to control a desktop or laptop(or any edge device)and a natural way to bring the power of open source tools to mainstream(nontechnical)audiences.It can lift the constrai

315、nts of todays user interfaces,too,so“you dont have to go through any layers of abstraction.”289 In the future,perhaps well modify things the way we imagine them in our head,suggests Lucas,“much more like drawing on a piece of paper.”290 Second,unlike Apple(or Microsoft),which might place limits or c

316、ontrols on operations within a“walled garden”OS,those using an AI interpreter on their machine face no restrictions.That means fewer rails of protection when things go wrong,but it also means rapid,unconstrained innovation.These are “tools that let you,as an individual,have power and control.Softwar

317、e has not been that.”291“We are building incredible tools that amplify the natural abilities of humans this is another chapter in computing history,”suggests Lucas.Agent-based tools and interfaces hold the promise to transform what an OS is and the purpose it ought to serve.A lot of effort goes into

318、 making code and applications accessible by humans,a process thats historically been difficult.“AI is the advent of personal software,”allowing us direct and tailored access,not only to an entire universe of open source software but to any digital experience we desire.292 Intelligent agents turbocha

319、rge the world of open source,putting a friendlier face on operating systems and providing a simple gateway to the power of an entire universe of open source applications.Says Lucas,“You can totally rethink what an operating system is,and thats absolutely what we want to do.”293 Agent ZeroUnlike Open

320、 Interpreter,Agent Zero is slightly less wedded to a particular platform.Its not designed to mediate the experience of ones desktop computer or any other system.Agent Zero is an autonomous agent,a simple AI that performs arbitrary tasks.According to its GitHub description,“Agent Zero is not a 38DECE

321、NTRALIZATION AND AIpredefined agentic framework.It is designed to be dynamic,organically growing,and learning as you use it.”294 Jan Tomek came up with the idea for Agent Zero as an alternative to setting up systems of specialized static agents.Instead,wondered Jan Tomek,why not allow agents to brea

322、k up projects into smaller specialized tasks(something AIs are particularly good at)?Tomek prototyped a system that allowed agents to summon more agents(recursively as needed):“I started experimenting with this,and it worked.”295 Each agent acts as a quasi-atomic primitive,breaking larger tasks into

323、 smaller chunks and then summoning other sub-agents to perform them.“Thats the moment where I saw that it had great potential,because AI is very good at writing code,and its also very good at analyzing errors,”suggests Tomek.If you execute code then feed the result back to the AI(along with self-ref

324、lections),“it can fix like 90%of the errors.”296 The priority is to make Agent Zero as flexible and useful as possible,allowing it to improve itself over time yet still benefit from more-capable versions of AI in the future.“I wanted it all to be customizable,”suggests Tomek,making it easy for peopl

325、e to modify it as needed and explore its full potential.This flexibility means people are testing all kinds of use cases with it,from running an RPG to rendering geometric shapes in 3D.Agent Zero represents a highly autonomous agent,one that can operate across different centralized and decentralized

326、 systems and applications.Instead of specifying how to do a given task,these agents are autonomous,self-organizing their activities in the manner of their choosing.Are these the droids were looking for?Like many AI agents,Agent Zero has its moments,sometimes acting the role of a“genius engineer”and

327、at other times a“foolish child.”However,with the right amount of coaxing,some of its capabilities are astonishing.Yes,it can install an Ethereum wallet and access the private key;even installing a Matrix client proves to be within its capabilities.297 It has a vast set of capabilities at its disposa

328、l,says Tomek:“Python and Node.js and Debian Linuxthose are probably the three biggest repositories of open source software on the Web.So,if it can use all of these three,it can do almost anything with a virtual computer.”298 With further improvement,Agent Zeros capabilities will only grow.Areas Tome

329、k would like to see improvements in include:Usability:Today,novices still find Agent Zero difficult to install and use.A graphical application and installer will make it more accessible to non-programmers.Memory:Improvements to the memory and knowledge tools that require frequent manual intervention

330、 are necessary.In future,relevant information pertaining to a task should be summoned automatically,and agents should learn from past successes and failures.New research suggests that “Agent Workflow Memory”may allow agents to adapt and improve their performance over time.299 In the future,perhaps A

331、gent Zero will also rewrite its own code(nearly 80%of which is AI generated anyway).300 Prompts:Prompts were designed to be transparent and easily modified.Says Tomek,“more than half of the potential of the framework lies in the prompt and not in the code itself.”301 To date,much of this potential r

332、emains unexplored,as most users stick with the default prompts.Networking:While Agent Zero is currently confined to the resources of its Docker container(and the sub-agents spawned there),it lends itself easily to stitching with others of its kind in a network.In fact,Tomek has already announced pla

333、ns for an API endpoint,meaning“the framework can run in the background and be called via API from other applications.”302 39DECENTRALIZATION AND AIWhats next for Agent Zero?The project started as open,and its important to keep it that way,suggests Tomek.“Agent Zero itself should remain free for everyone.”303 He sees the project as an open source counterweight to companies with large proprietary LL