Fabric Thesis 2025

Web3 x AI
First published: June 10th, 2024
A paper by Fabric Ventures

Richard Muirhead Fynn Brinkmann Thomas Crow

Fabric Ventures; Near Foundation Fabric Ventures Fabric Ventures
[email protected] [email protected] [email protected]

Abstract

Contextual, hyper-personalised AI-powered services have the potential to

fundamentally change the digital interactions of humans and enterprises. As their
influence and ubiquity grow, it is crucial that every individual take part in their
ownership and direction. AI development and success hinge on five key pillars:
compute, data, models, talent, and governance. These pillars have thrived on
open-source standards. However, rising compute costs have forced startups to
relinquish ownership and control, often to large tech companies, straying from
their open origins. Advocates of this shift argue that only major providers can
ensure AI safety, asserting that open models are too risky for the world.
Nevertheless, the issues of the centralised Web2 paradigm—including corruptible
decision-making, bias (both deliberate and accidental), lack of transparency and
privacy, prioritisation of profit over utility, and single points of failure—are
well-known. As AI advances, we must manage increasingly complex applications,
large-scale governmental uses, integrate diverse data sources, prove provenance,
respect intellectual property rights, and navigate sensitive geopolitical interests.
Therefore we need a modular, interoperable AI stack, on-chain finance and
payments, and a sustainable business model for each component to remain open
and transparent— a 'can't ever be evil' architecture. Transparent and antifragile
blockchains provide a clear, immutable rule set for AI, and token economics offer
an alternative business model for an open, modular web built with strong
ownership principles. Tokenized ownership presents an opportunity for
exceptional venture capital returns, both fostering innovation within each stack
component and encouraging the large-scale participation and capital formation
needed to compete with closed alternatives. The resulting decentralised, open,
user-owned AI ecosystem will be a more prosperous, inclusive, and aligned
alternative to the closed systems of 'big tech.'

1
2
Table of Contents

1. Introduction............................................................................................................................................... 4
2. The key pillars for the AI revolution...........................................................................................................5
3. Building on the shoulders of giants........................................................................................................... 6
3.1 Open data.........................................................................................................................................6
3.2 Open models.................................................................................................................................... 7
3.3 Open talent....................................................................................................................................... 8
4. Web2 AI: Money makes the world go round........................................................................................... 10
5. Several justifications have been expressed............................................................................................13
6. This is a worse philosophy......................................................................................................................14
7. The importance of data........................................................................................................................... 15
8. In the centralised world, innovation atrophies as platforms become extractive...................................... 16
9. The future needs a new approach.......................................................................................................... 19
10. A New Technological Fabric..................................................................................................................21
11. The incentive layer for the Open Web...................................................................................................23
12. Nothing ever changes if not driven by money (power or sex)...............................................................26
13. So what is the web3 business model?..................................................................................................27
14. What is possible with the web3 business model: the data flywheel reimagined................................... 29
15.We’ve come a long way......................................................................................................................... 31
16. We still have some way to go............................................................................................................... 31
17. Conclusion............................................................................................................................................ 33
17.1 Business model level Change...................................................................................................... 34
17.2 A Fundamental Dislocation...........................................................................................................35
17.3 Part of a Longer Arc..................................................................................................................... 35
18. Acknowledgements...............................................................................................................................37
19. Bibliography.......................................................................................................................................... 38

3
1. Introduction

We expect artificial intelligence (AI) to cause a foundational technological shift in the world.
Although chatbots and coding assistants are current trending topics, this is just the beginning of
what is possible within this new era of computing. Consider the following scenario:

‘You’re feeling tired, maybe you should stop for today’, Jarvis’s voice jolted Beth out of
her work-induced stupor. ‘You completed the urgent tasks, the rest can wait for
tomorrow’. In the background the AI had assessed her calendar and schedule for the
week ahead; scanned her emails for urgent correspondence and replied to most of them
(the ones which didn’t require her direct input), but it was concerned for her well-being
based on her heart rate, blood sugar and respiration. Beth shut the laptop and slumped
back on the couch. Jarvis was right, she was exhausted.

‘Maybe a pizza and a film tonight?’ the AI continued. Beth nodded imperceptibly. The
agent was already in action; scanning local restaurants, wait times, available dining
companions, and traffic while simultaneously comparing them to her previous orders
and preferences. The AI was optimising not for speed or platform margin but for Beth’s
culinary pleasure, all whilst taking into account a planned workout early the following
morning. Selecting a broccoli, kale, chilli and hazelnut gluten-free pizza, the agent
pulled USDC directly from her wallet and sent them directly to the restaurant, notifying
Beth of the pending transaction.

4
 A text came through to Beth. Jarvis piped up on her behalf: ‘Change of plans, I’ll
cancel the order’. The AI had processed the message before the notification appeared
on the screen, understanding the contextual information of the conversation, the content
of the message, and Beth’s likely response.

Glancing down, she nodded her approval. Behind the scenes the AI leapt into action.
Rather than returning the delivery to the restaurant, the AI submitted the transaction to
a marketplace for another agent to buy out this delivery, saving Beth most of the
purchase order. A brief bidding war later, and most of the fee was back in her account.

This scene from the near future may recently have seemed like science fiction but now seems
possible before the end of the year. We tread on a fine line between utopia and dystopia. Which
side of the line we end up on depends on how we choose to let the irrepressible forces of
capitalism play out. What if Beth's agent does not choose the best meal for her but one from a
fast food company that has signed an exclusive partnership with the AI platform? What if her
movie choice is not based on her genuine preferences but is merely the latest promotion from
Amazon? We hope to live in a world where AI can have integrity (that is, doing the right thing
without oversight). Exactly how we fund, architect, and maintain these systems will determine
on which side of the line we eventually reside. There are protagonists currently lining up on
either side: those who believe that a fully turbo-charged and closely-held centralised model is the
best guardian of humankind’s interests, opposed to those who would rather entrust this to the
emergent power of transparent and open systems. Humanity is truly at a watershed moment.

2. The key pillars for the AI revolution

The inescapable reality is that AI is increasingly controlled by a handful of corporations, the
most valuable and powerful corporations the world has ever seen1: Microsoft, Google, Meta,
Amazon, Nvidia and Apple. The interests of those corporations are profit maximisation for
shareholders, not people's health, societal justice, or the greatest benefit to all stakeholders. What
has driven this centralising force? Was it inevitable? Does it lead to better outcomes? Before we
dive into the specifics, we outline one of the major technological movements today. We describe
its history to build our case of why we think that we have been on the right path, until just
recently.

The architecture of AI stands upon five pillars: Compute, Data, Models, Talent, and Governance.

Compute: GPUs, or more recently, AI-specific data-centres for model training and inference.

1
Big Tech Market Share reaching an all-time high:
https://www.bloomberg.com/news/articles/2024-05-30/nvidia-and-five-tech-giants-now-command-30-of-the-s-p-500
-index

5
Data: As input to train models, augment generation, prompt inference, and fine tune.
Models: To be trained, perform inferences and train other models.
Talent: To build models, analyse and guide the process, and develop the data pipeline along with
other parts of machine learning operations (MLOps).
Governance: To make strategic, economic, and ethical decisions across all areas, balancing the
interests of various stakeholders, including owners, suppliers, users, regulators, and the public.

3. Building on the shoulders of giants

To understand the future, we first need to look back in time. Historically, the main pillars of AI
have been built openly and collaboratively by researchers and companies across the world. They
are based on publicly available crowdsourced resources, academic breakthroughs and open
research. In this article, we focus on the importance of open data, open models, and open talent
in shaping today's AI revolution.

3.1 Open data

Although machine learning has been around since the 1950s, the current boom traces its
immediate roots to Alexnet and the seminal 2013 paper2 from the University of Toronto that
re-introduced the world to deep learning through the ImageNet 2012 Challenge3. The key insight
from Alexnet's success was that deep learning had not previously worked because the models
were just not big enough. The lesson was that bigger models outperformed smaller, more
complex models. However, to train these big models, researchers needed orders of magnitude
more data.

Alexnet fully relied on the Imagenet database, an open database containing over 15 million
images and 22 thousand categories, which were hand-annotated with the specific goal of training
AI algorithms. This reliance on open data as not diminished; the primary source of training data
today is common crawl, an open repository that crawls the web and since 2008, it has provided
data to train AIs for free (including the training sources for gpt34).

2
ImageNet Classification with Deep Convolutional Neural Networks, 2013, University of Toronto by Alex
Krizhevsky, Ilya Sutskever and Geoffrey E. Hinton:
https://proceedings.neurips.cc/paper_files/paper/2012/file/c399862d3b9d6b76c8436e924a68c45b-Paper.pdf
3
Wikimedia Foundation. (2024, April 25). ImageNet. Wikipedia.
https://en.wikipedia.org/wiki/ImageNet#History_of_the_ImageNet_challenge
4
Boykis, V. (2023, February 26). Everything I understand about ChatGPT.
https://gist.github.com/veekaybee/6f8885e9906aa9c5408ebe5c7e870698

6
 1 - Open data removed one of the core bottlenecks that had held back AI for
decades; sufficiently large (tagged) datasets for training models.

3.2 Open models

Arguably, just the availability of open data would not have led to the explosion of AI capabilities
we see today. Researchers the world over were also able to learn about international innovations
and incorporate them into their work. Notably, the Alexnet paper, published by researchers at the
University of Toronto, spawned a decade of innovation at the model level. AlphaGo5 came out of
DeepMind in London, and the transformers paper6 was published by Google and co-authored by
NEAR co-founder, Illia Polosukhin. The GPT-1 model was open sourced by OpenAI in
California. More recently we have seen innovation from across the world, including the USA,
UK, France (Meta AI, Mistral), and China.

5
Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M.,
Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L., van den Driessche, G., Graepel, T., & Hassabis, D. (2017,
October 19). Mastering the game of go without human knowledge. Nature News.
https://www.nature.com/articles/nature24270
6
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017,
August 2). Attention is all you need. arXiv.org. https://arxiv.org/abs/1706.03762

7
 Figure 2 - Major open and collaborative international AI/ML publications

Importantly, OpenAI were able to pivot and build ML models using a transformer-based
approach, scalable parallelisation of generalised model training based on self-attention, a move
that would not have been possible without an open, collaborative approach. In each case of major
innovation, the open publishing of models allowed the global community to understand the
innovations underpinning them, and add their own novel ideas.

3.3 Open talent

The openness around data and models allowed researchers to freely contribute to the growing
body of knowledge. Therefore, it was not the tech giants who kicked off the deep learning
revolution, but academics in Toronto and 48,000 people from over 167 countries manually
tagging ImageNet7.
The backbone of today’s AI landscape is open models built by international researchers,
academics and developers modifying open systems, using open data scraped from across the
web, and contributing from anywhere in the world. This is no accident. Open systems work, and
they work at scale to build the most innovative, resilient, battle-tested networks.

7
Li, F.-F. (2015, March). How we’re teaching computers to understand pictures. Fei-Fei Li: How we’re teaching
computers to understand pictures | TED Talk.
https://www.ted.com/talks/fei_fei_li_how_we_re_teaching_computers_to_understand_pictures?language=en

8
 Figure 3: The team from ExoLabs built an instance of Llama3 running locally on
an iPhone, or as Mo (@mo_baioumy) prefers it, right on his Apple Watch (link).

As we stated in our 2017 thesis paper8 With Satoshi Nakamoto’s 2008 release of the Bitcoin
whitepaper, we’ve entered into the fourth age of open source software development. We have
seen this play out over the past 5 years, not just in crypto but also in AI. However, a concerning
trend is pushing back against that movement.

8
Fabric Ventures Thesis by Max Mersch and Richard Muirhead, December 2018
https://medium.com/fabric-ventures/the-fabric-ventures-investment-thesis-6cd08684b467

9
 Figure 4 The fourth ‘Golden Age’ of open source software (OSS) - Fabric Ventures
2017 thesis.

4. Web2 AI: Money makes the world go round

The past couple of years have recently seen the AI industry move in a new direction. Projects
have been launched with open models, delivered against two or three key product milestones,
found market traction and then reconfigured to a closed, profit driven approach closely-coupled
with existing technology providers and ecosystem. This is in contrast not only to the
collaborative roots of AI but also to the understanding that the potential of AI was so great that it
must be developed, owned, and governed by open, equitably profitable entities (often
foundations) with open source software licensing models.

In December 2015, Sam Altman, Greg Brockman, Reid Hoffman, Jessica Livingston, Peter
Thiel, Elon Musk and AWS announced the formation of OpenAI and pledged over $1 billion to
the venture. What started as a ‘non-profit AI research company9‘, has since developed into an
overly convoluted foundation-type ownership structure10; despite (in practice) existing as a
closed shop and arguably a subsidiary of Microsoft in all but name. Stability AI led the way with
Stable Diffusion, a great product for users that was an early leader in the market for quality AI
generated images but—despite raising hundreds of millions of dollars—struggled to keep up
with the immense costs of using AWS, Google, and Coreweave. Although Mistral AI launched
with a clear statement that ‘Open source is a core part of our DNA’ and released highly
performant LLMs, it removed ‘committing to open models11‘ from its website in less than a year.

9
Introducing openai. openAI. (n.d.). https://openai.com/index/introducing-openai
10
Our structure, OpenAI (n.d.), https://openai.com/our-structure/
11
Mistral AI: Open-weight models. Mistral AI | Open-weight models. (2024, February 25).
https://web.archive.org/web/20240225001133/https://mistral.ai/

10
Inflection raised more than USD1.5 bn in venture funding, before being stripped of its CEO and
then backed/acquired by Microsoft.

Figure 5: Startups building AI models that received major investment from big tech
companies to scale and compete. Source: Crunchbase, Fabric Ventures research

What prompted the ‘bait and switch’? It seems increasingly clear that these high profile startups
did not ‘close’ for lack of acceptance of their narrative or philosophy, nor from a failure to build
more powerful technology or more compelling product, but instead on account of the inability of
their business model to provide a viable alternative to the digitally turbo-charged capitalism of
the early 21st century. It has long been known that multiagent network systems tend towards
centralisation12, especially in low or unregulated markets. Power goes to those who accrue
capital through economies of scale and vertical integration, and this tends to favour closed,
profit-seeking, patent-protected businesses because open models do not traditionally seek
sustained compounding growth.

12
Gabaix, X. (2008, September 5). Power laws in economics and finance. NBER.
https://www.nber.org/papers/w14299

11
Startups venturing into uncharted territory, even when accelerated by an unprecedented
technology, are inherently vulnerable if they are dependent on a closed / value-extractive entity
below them in the stack. They will be beholden to that entity unless they are able to find exit
velocity through an appropriate and sustainable business model. Closed business models can be
built on top of open layers (eg: android, SMTP) levels of the stack; however, open businesses
cannot sustain and thrive if built on top of closed. In such conditions, growth plateaus and the
higher levels are at constant risk of acquisition or control from monopolistic lower levels. This is
a well understood macro phenomenon: raw materials and other basic necessities generally either
operate within larger markets with several providers (such as national exports) or are regulated to
prevent monopolistic tendencies.

When confronted by the vast computing costs of training models and the compounding benefits
of having more data, many startups have been unable to resist the gravitational pull of existing
centralised business models through a combination of fear of falling behind and greed.
Specialised AI-compute resources are bought up by big tech companies, cornering supply as well
as pricing smaller participants out of the market. The numbers are staggering; with mooted
trillions of dollars13 required to spend on compute. AI startups lack the technological and
business frameworks to both remain open and scale to the degree needed for these highly
complex systems.

13
Hagey, K., & Fitch, A. (2024, February 8). Sam Altman Seeks Trillions of Dollars to Reshape Business of Chips
and AI.
https://www.wsj.com/tech/ai/sam-altman-seeks-trillions-of-dollars-to-reshape-business-of-chips-and-ai-89ab3db0

12
 Figure 6: Training costs rising over time and pricing participants out of the market14

Of course, this is the big tech business model: control the means of production and sell access to
the outputs. Well-funded startups who set the pace for AI innovation become subordinate to the
governance structures and profit maximisation imperatives of the compute providers. Even those
major companies that have made significant recent contributions towards open access to AI are
keeping their cards somewhat close to their chest. Meta’s standard bearing Llama models are not
as accessible as at first glance they appear:

Meta's licence prohibits its use to train other language models, and it requires an
additional special licence if a developer wants to deploy it in an application or
service with over 700 daily users. Why Meta’s Open source is not really open15

Taken together, these facts led us to the current paradigm. Compute, data, models and talent are
concentrated in opaque, siloed organisations and packaged under existing cloud service
offerings.

5. Several justifications have been expressed

According to the proponents of such systems, control and centralization are necessarily the
future.

Open Models are harder to control and put guardrails on… …I’m not saying we
should ban it outright but we should look carefully at their implications as they
grow. Dario Amodei (Anthropic) on The Logan Bartlett Show, October 6, 202316

We only open-source powerful AI models once we have carefully weighed the

benefits and risks, including misuse and acceleration. OpenAI spokesperson17

14
Nestor Maslej, Loredana Fattorini, Raymond Perrault, Vanessa Parli, Anka Reuel, Erik Brynjolfsson, John
Etchemendy, Katrina Ligett, Terah Lyons, James Manyika, Juan Carlos Niebles, Yoav Shoham, Russell Wald, and
Jack Clark, ‘The AI Index 2024 Annual Report,’ AI Index Steering Committee, Institute for Human-Centered AI,
Stanford University, Stanford, CA, April 2024
15
Auslender, V. (2023, August 30). Why meta’s open source is not really open. ctech.
https://www.calcalistech.com/ctechnews/article/atv6xnkya
16
Anthropic CEO on leaving openai and predictions for future of ai. YouTube. (2023, October 6).
https://youtu.be/gAaCqj6j5sQ?t=5625&si=NjmJGIoDeKqEQC-0
17
Gent, E. (2024, March 25). The tech industry can’t agree on what open-source AI means. that’s a problem. MIT
Technology Review.
https://www.technologyreview.com/2024/03/25/1090111/tech-industry-open-source-ai-definition-problem/

13
 If at some point however there's some qualitative change in what the thing is
capable of, and we feel like it's not responsible to open source it, then we won't.
It's all very difficult to predict. Mark Zuckerberg, CEO of Meta 18

Of course there is another, simpler, explanation for this situation. Big tech, now having secured
their assets, are seeking some good ‘ol' fashion regulatory capture19. After all, their business
model mandates it. Open source models are touted as uncontrollable, ungovernable, and
dangerous to society. We are regularly told that they must be controlled to stop them falling into
the wrong hands. The above leaders, who coincidentally control the largest companies in the
space, assure us that they have the correct approach, and they can govern AI on our behalf. Is
there evidence that closed source is the only way forward? We argue that it is actually quite the
opposite.

6. This is worse philosophically

In a closed and centralised model, AI-fuelled tech titans can have more power than most
countries. The consequences of such power disparity include pivotal resolutions and plebiscites
around elections, financial credit, school admissions, healthcare prioritisation, job applications,
(dis)information in the news, religious intolerance, and racism. Imagine a world with direct
brain-to-computer connections where companies or governments can regulate and restrict those
thoughts. Although Hollywood might now have numbed us to this well-trodden trope, this goes
beyond even the nightmarish imaginations of George Orwell’s Ministry of Truth in 1984. Should
the fate of humans be that of nodes in a hive mind for which the objective function might evolve
inexorably towards predominantly profit or solely subjugation?

This is clearly not the world in which we want to live. After all, in such a world, who makes the
rules? We take lessons from the past where companies like Microsoft with MSN or
AmericanOnLine’s service AOL tried to profit from a ‘walled garden’ for the Internet20. Such
approaches have, at least initially, proved incapable of matching all the world's open
experimentation and innovation within one closed platform. Are we unable to harness the power
of open systems to simultaneously carve out a better path and better business models? Can we
ultimately resist the Facebooks and TikToks of the third wave of the web and remain open?

18
Mark Zuckerberg - Llama 3, $10B models, Caesar Augustus, & 1 GW datacenters. YouTube. (2024, April 18).
https://youtu.be/bc6uFV9CJGg?t=2284&si=Ct2wai48K4rqavjJ
19
Gurley, B. (2023, September 16). All-in summit: Bill Gurley presents 2,851 miles. YouTube.
https://www.youtube.com/watch?v=F9cO3-MLHOM
20
AOL’s ‘Walled Garden’ - WSJ. WallStreetJournal. (2000, September 4).
https://www.wsj.com/articles/SB968104011203980910

14
7. The importance of data
Thirty years later, the walled gardens are back. Much of our data21 is trapped (figure 7),
unavailable to the people who produced it, who should own it, and ultimately those who should
benefit from it. In our 2019 capital formation materials Fabric noted McKinsey, IDC and EMC
research which proposed that less than 3%22 of the world's data was being leveraged to
humankind’s benefit. That trend has only accelerated under the centralised paradigm.

Figure 7: the amount of data from YouTube and TikTok surpasses all podcasts,
television archives and radio archives23

There is a second major issue with this approach. A dawning reality within AI today is that we
will run out of data24 in 2028, (Figure 8). There are two primary reasons for this emerging
situation. First, AI remains an inefficient learner, requiring an astonishing amount of data, which
is effectively the entire publicly available/scraped internet at this point. Second, AI companies
only really have access to (decreasingly) publicly available data sources. Most data is behind
paywalls, within other silos, and becoming increasingly gated off (NY times -1) (Reddit -2), as
content platforms begin to realise their worth. Furthermore, there are also concerns over

21
Jones, D. (2024, May 23). How much LLM Training Data is there, in the limit?. Educating Silicon.
https://www.educatingsilicon.com/2024/05/09/how-much-llm-training-data-is-there-in-the-limit/
22
Gantz, J., & Reinsel, D. (2012, December). THE DIGITAL UNIVERSE IN 2020: Big Data, Bigger Digital
Shadows, and Biggest Grow th in the Far East.
https://www.cs.princeton.edu/courses/archive/spring13/cos598C/idc-the-digital-universe-in-2020.pdf
23
Jones, D. (2024, May 23). How much LLM Training Data is there, in the limit?. Educating Silicon.
https://www.educatingsilicon.com/2024/05/09/how-much-llm-training-data-is-there-in-the-limit/
24
Rao, D. (2024, June 5). All-powerful, ever-pervasive AI is running out of internet. theweek.
https://theweek.com/tech/ai-running-out-of-data

15
regulation and privacy laws. It is notable that Meta, with access to huge datasets of relevant data,
trained Llama 3 only on publicly available sources.

Figure 8: Projections of the stock of public text and data usage25

8. In the centralised world, innovation atrophies as platforms become

extractive
The value of the compounding data flywheels has been well-described over the previous two
decades, creating immense enterprise value for the web2 titans. Feeding data into every step of
the AI process builds a loop that generates a greater quality of predictions, which itself increases
user stickiness who then generates more data to train the models, and thus the cycle perpetuates
itself.

25
Villalobos, P., Ho, A., Sevilla, J., Besiroglu, T., Heim, L., & Hobbhahn, M. (2024, June 6). Will we run out of data?
limits of LLM scaling based on human-generated data. Epoch AI.
https://epochai.org/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data

16
 Figure 9: The web2 data flywheel

Today’s flywheel is incredibly powerful, but it is not the only force driving product quality and
consumer satisfaction. Pushing in the other direction is the business model. Although selling ads
based on data is undeniably immensely successful, the process reduces the quality of the product,
and biases users’ data inputs in favour of monetization (figure 9). Over time, as the marginal
benefit of additional data slows, the extractive prerogative reduces product quality (figure 10),
forcing enterprises to turn to more underhand methods to protect their market26.

26
O’Laughlin, D. (2024, June 2). The internet as you know it is dying. The Internet as You Know It Is Dying - by
Doug O’Laughlin.
https://www.fabricatedknowledge.com/p/the-internet-as-you-know-it-is-dying?r=j1yw&utm_medium=ios&triedRed
irect=true

17
 Figure 10: The effect of the web2 business model on product quality

In a centralised world, firms within this paradigm are structurally bound to optimise their
Average Revenue Per User (ARPU) or marketplace rank. There is a good reason that complaints
about quality of even longstanding products are commonplace, with companies inevitably
looking to extract as much value as possible from trapped users. As tech companies control a
greater proportion of the world's media and sources of information then product experience,
politics, elections, free speech, and user wellbeing come second to maximising shareholder
value.

18
 Figure 11: Product quality diminishing as platforms become extractive.

9. The future needs a new approach

We know from history that open development is a better way to build. We now take a step
further and articulate how, longer term, in the era of AI, open development is the only way to
build in a safe and aligned manner. When AI is ubiquitous, the technological choices will be
myriad, with significant trade-offs—which we are only beginning to understand. Different use
cases will need different models (unified global models or locally run models for corporations
handling sensitive data), different data sources (private, public, IP copyrights, common use), as
well as different compute providers (factoring in geopolitics, energy consumption, local vs. cloud
for latency). How do we as humans (or even agents) choose the best approach for a particular use
case? How can we build a deterministic reputation and recommendation layer that allows

19
for these data sources and models to be leveraged on a real-time basis without human
intervention?

Figure 12: A ‘pick n mix’ AI future: interoperable, fit-for-purpose elements across and
between multiple layers of the stack coordinated to produce the best outcomes for
users.

While data is arguably the most pressing concern, as described above, unlocking that is not the
only advantage of a more open, modular approach.

The open and decentralised approach of Web3 has already unlocked more computing power than
the tech giants. The bitcoin network alone produces roughly 110 exahashes per second (EH/s);
more than the combined power of the top 500 supercomputers in the world, and around 100x
more than Google, Amazon, and Facebook combined27.

Technical talent exists far more abundantly outside of these massive organisations than within
them. ‘No matter who you are, most of the smartest people work for someone else’ a principle

27
Illis, L. (2022, December 31). Tech Giants vs. the Bitcoin Network: A comparison of computing power for
Artificial Intelligence. Tech Giants vs. the Bitcoin Network: A Comparison of Computing Power for
Artificial Intelligence.
https://medium.com/@leviillis/tech-giants-vs-the-bitcoin-network-a-comparison-of-computing-power-for-art
ificial-intelligence-b029e05e08a4

20
referred to as ‘Joy’s Law’28 after Sun Microsystem’s founder Bill Joy. More open approaches
have historically been stymied by the ability to coordinate disparate providers at scale and
reward them for providing their resources—particularly when those resources are only really
valuable when combined with others. Open development has led to the majority of innovations
in the web and is the basis of today's AI. It has brought us this far, so why stop now?

10. A New Technological Fabric

The open web, encompassing decentralised data structures and distributed computing alongside
artificial intelligence, holds the potential to underpin this new and open economy—one that is
positive-sum and inclusive, benefiting everyone. By creating the necessary infrastructure and
fostering economies that support a positive-sum internet, we harness market forces to achieve
collective good, striking a balance that nurtures and protects the innate rights of the individual.

This will be powered by web3, which has as its core superpower the capability to drive the
implicit coordination of strangers without requiring the explicit trust of intermediaries
whilst shifting the power in the platform back to the individual. This capability has the real
potential to radically slash the cost of global coordination and promote more means for
innovation. We’re building for a system where corruption in the underlying infrastructure
becomes either programmatically infeasible or includes extreme financial disincentives.

For the first time, we have developed an ‘un-opinionated’ (one which does not implement
censorship at the protocol level), freely interoperable, and programmable settlement layer.
Moreover, it is not controlled by a single entity, which both reduces friction to application
builders and the cost of doing business. We can therefore empower agents and contracts with the
full suite of tools to make payments big and small, global and local, and access financial services
as required without cumbersome human intermediaries. This opens up a whole new toolset for
machines to carry out transactions and access financial services, as well as reducing switching
costs to near zero as applications can operate on open, interoperable settlement layers (such as
Solana, Near, or Ethereum) which leads to lower costs and higher competition. This would not
be possible without considerable progress in the underlying infrastructure stack:

The modular stack: a flexible, modular infrastructure that allows application builders to
seamlessly and easily prioritise and emphasise different properties of blockchains. The past 5
years have seen immense improvements in not just the core pieces of web3 development,
28
Wikimedia Foundation. (2023, November 14). Joy’s Law (Management). Wikipedia.
https://en.wikipedia.org/wiki/Joy%27s_law_(management)#:~:text=In%20management%2C%20Joy’s%20law%20is
,Microsystems%20co%2Dfounder%20Bill%20Joy.

21
including scaling, developer experience, and user experience. Moreover, the ways to package and
link them are ever increasing, allowing data, capital, and applications to move to where they are
best and most efficiently deployed. To accomplish this we have broken the blockchain into its
component pieces: sequencing (espresso, nodekit), data availability (nuffle, celestia), execution
(starknet, movement labs), and settlement (Ethereum, hyle). This approach allows innovation to
happen within these layers, and developers to pick and choose which properties are most
important to their applications. Rollups on Ethereum handle more transactions at lower costs
than the base chain. Zero Knowledge proving schemes, important for privacy and scaling
continue to improve at astonishing rates. Different execution environments expand the set of
web3 developers, and give different options to building robust applications. Finally we see
massive improvement in user experience, moving from account abstraction to chain abstraction
to (we believe) application abstraction. This is where agents, governed by intents, can execute
seamlessly on behalf of the user without requiring their constant supervision. The moments
when a human being needs to realise ‘there’s an app for that’ will dwindle until the ‘App Store’
as we know it will be no more.

The financial stack: Payments29 and finance on chains have been one of the clearest use cases
for crypto since the invention of Bitcoin, originally termed ‘a p2p electronic cash system’.
However, it has taken many years of product iterations, networks built from scratch and
regulatory landmarks to get to the point where it is seen as the possible future of digital payments
for individuals—and more importantly agents. Only recently have we established robust, secure
stablecoins backed by major institutions like BlackRock (circle USDXX reserve fund30) or
PayPal (PYUSD31) with transparent collateral enabling trust in a stable medium of exchange.
Transaction costs with blockchain rails are finally low enough to envision a world where smart
contracts and machines can transact at scale. Most importantly, legitimation of this field is

29
Hansjee, A., & Brinkmann, F. (2024, June 3). Blockchain payments: Beyond the hype.
Medium. https://fabric-vc.medium.com/blockchain-payments-beyond-the-hype-5455c56454e4
30
Circle Reserve Fund: USDXX: Institutional. BlackRock. (n.d.).
https://www.blackrock.com/cash/en-us/products/329365/circle-reserve-fund
31
PayPal stablecoin: US dollar cryptocurrency. PayPal Stablecoin | US Dollar Cryptocurrency |
PayPal US. (n.d.). https://www.paypal.com/us/digital-wallet/manage-money/crypto/pyusd

22
coming to market with the introduction of regulatory guidelines for operating in crypto economic
financial markets including MiCa in Europe32, the UK33, Singapore34 and Japan35.

In sum, with an adaptable, open infrastructure alongside a powerful financial network we open
up the door to fully transact value at scale and move from the internet of information to the
internet of value. A core mechanism that ties these together are an incentive layer and a
mechanism to distribute ownership powering both the cold start problem for ambitious global
networks but also solve for more dynamic reward mechanisms to enable fluid development in
this field. Tokens are still an entirely new field of economic research today but experimentation
over the past decade has come a long way.

11. The incentive layer for the Open Web

As laid out in our 2017 thesis36, tokens are an incredible coordination mechanism37 for
distributed yet aligned actors in a network. Their ability to define finite resources in a flexible
and composable way allows protocols to build scalable networks on top of assets both physical
and digital. Tokens have already been used to digitise dollars, create new stores of value,
coordinate hardware compute worldwide, and govern protocols. We first published our taxonomy
of tokens in 2017 and it is updated here for the AI era.

32
European Parliament. (2023, May). Markets in crypto-assets regulation (MICA).
https://www.esma.europa.eu/esmas-activities/digital-finance-and-innovation/markets-crypto-asse
ts-regulation-mica
33
UK Government. (2024, March 1). Economic crime and corporate transparency act 2023:
Factsheets. GOV.UK.
https://www.gov.uk/government/publications/economic-crime-and-corporate-transparency-act-20
23-factsheets
34
Monetary Authority of Singapore. (2023, August). MAS Finalises Stablecoin Regulatory
Framework.
https://www.mas.gov.sg/news/media-releases/2023/mas-finalises-stablecoin-regulatory-framewor
k
35
Financial Services Agency, the Japanese Government. (2022, September). Regulatory
framework for crypto-assets and Stablecoins. https://www.fsa.go.jp/inter/etc/20220914-2/02.pdf
36
Mersch, M, Muihead, R. (2018, December 20). The Fabric Ventures Investment Thesis. Medium.
https://medium.com/fabric-ventures/the-fabric-ventures-investment-thesis-6cd08684b467
37
Thevenard, J. (2020, June 24). Tokenised ownership is the best coordination tool since equity. Medium.
https://medium.com/fabric-ventures/tokenised-ownership-is-the-best-coordination-tool-since-equity-7c957a00d818

23
 Figure 13: Taxonomy of Token characteristics

Table 1: Table of token characteristics

Store of Value tokens Such as Bitcoin and Ethereum. These are understood as the native
base currency of their respective platforms and are sufficiently
mature, widely held, and possessing of a clear community to have
made the leap to represent a store of value. These tokens are
necessarily few in number, representing their own network’s store of
value.

Stablecoins These have significantly matured in the last 7 years, and are
increasingly seen as fit-for-purpose as an alternative mechanism to
traditional digital payment rails. Especially within the context of AI,
they are likely a superior mechanism for machine-based payments
enabling transfer of value in a digitally native way just as machines
communicate via API’s today. Stablecoin payments are:
1. Unencumbered by legacy overheads based around validating
human bank accounts
2. Programmable and interoperable by design. Crypto has spent
years wrestling with questions around security and
permissionless algorithmic payment execution; questions
which are naturally relevant to agent-based systems. Existing
user flows such as session keys (currently focused on
improving UX for on-chain games) could clearly be co-opted
to provide defined guardrails for agents to execute restricted
actions without requiring constant approval.

Payment tokens Expand the set of methods of payment for contributors and agents
beyond the previously discussed stable/pegged existing currencies by
allowing the invisible hand to dictate an appropriate price for the
exchange of services. This provides flexibility to agent-based or

24
 algorithmic payments, particularly in cases where the payment token
is related to the value of the underlying network and/or there is an
asymmetry of information between buyers and sellers in the value of
that network.

Governance tokens Giving holders a vote in how a network is run will be crucial in
making sure all facets of the software value chain are run in the
interest of all relevant stakeholders. Although crypto governance is by
no means in its final state, it continues to iterate, and represents a
more egalitarian and communitarian approach vs. the top down
opaque structures of existing AI firms.

Work tokens Allow contribution to be measured, validated, and ultimately

rewarded for desirable actors in a network, where access is granted
based on their appropriate provisioning of sufficient resources. From
decentralised compute networks rewarding the provision (and
ultimately utilisation) of computational resources, where these
resources are appropriate for that use case, to data-networks
rewarding the collection, tagging, and refining of data, to model
networks rewarding developers for providing better appropriate
models, to applications it rewarding user feedback and power users to
drive the development flywheel; there are opportunities to use work
tokens to incentivise better outcomes across the entire open economy.

This class of tokens not only helps to overcome the cold start problem
but also aligns the incentives of contributors by valuing their access as
a function of the total success of the network.

Identity tokens Allow network participants to represent facets of their identity in an

open/public sphere. This allows us to solve some of the main
problems we face as we enter an era of abundant AI and machine
made transactions: When are we interacting with a human or an
algorithm? What is real and what is synthetic?
Identity tokens can represent absolute personhood (as in the case of
WorldCoin) and orthogonal (but related) facets of identity, such as
nationality, reputation in a network, age, or location.

25
12. Nothing ever changes if not driven by money (power or sex)

We are not so naive to believe that doing the right thing, or even a superior technological
solution are enough, on their own, to change behaviours and recalibrate organisational structures.
Rather, we believe that the combination of a new philosophy and new technology with the most
powerful coordination force, a superior business model, will help a new solution to cross the
chasm to mass adoption.

Each technological revolution is received as a shock, and its diffusion

encounters powerful resistance both in the established institutions and in people
themselves. Hence, the full unfolding of its wealth-creating potential at first has
rather chaotic and contradictory social effects, it later will demand a significant
institutional recomposition. This will include changes in the regulatory
framework affecting all markets and economic activities as well as the
redesign of a whole range of institutions, from government, through
financial regulation, to education, as well as modifications in social
behaviours and ideas. It is thanks to that restructuring of the context to fit the
potential of the revolution that ‘golden ages’ can occur. Carlotta Perez38
(emphasis added)

38
Perez, C. (2014). Technological Revolutions and financial capital: The Dynamics of Bubbles and golden ages.
Edward Elgar.

26
13. So what is the web3 business model?

Figure 14: The closed versus open economy stack.

Historically, for argument's sake,there are four types of stakeholders that businesses endeavour to
keep happy via specific reward mechanisms, for example:

Customer: net product utility

Employee: salary
Supplier: cash payments
Backer: equity ownership

Each stakeholder has pretty distinct primary reward mechanisms which incentivise a particular
type of behaviour. Suppliers receive their payments, and have no further interest in the success of
the customers beyond their survival. Employees receive their salary, and are incentivised to fulfil
their requirements to maintain payment but little more. Customers understand they are being
manipulated, pushed towards platform lock-in, where retention becomes more important than
acquisition. Backers think in quarterly terms, rewarding short-term price fluctuations without
ever having to engage with the underlying products. On the fringes, enterprises play to some
extent with these mechanisms (supplier financing, equity options, loyalty schemes) but they are
costly to implement and struggle to fit into the framework of the modern corporation.

With tokens, we are able to denote ownership, track and reward behaviour in software and
instantiate granular mechanisms that can be triggered automatically by their context and

27
life cycle in a manner never before possible. Customers can be employees, incentivised to go
above and beyond to supply useful work. Suppliers can be backers, encouraged to think about
the long term needs of the project. Demand can be generated by consumers, incentivised for
growth in the network.

We can focus resources and attention where they are most needed, and engage all players where
they can be most useful, producing a more Pareto efficient outcome. Many products today
already represent this morphing of roles as users become influencers and distributors. This will
only become more flexible with digital agents free to participate flexibly and simultaneously in
any side of the market, allowing them to enjoy the product's utility or earning rewards as a
supplier. Managing all these different stakeholders and rewarding those performing useful
actions requires massive coordination at scale: a scale we believe is only possible when
underpinned by crypto economic incentives and the flexibly open business model of web3.

Figure 15: Fabric ventures OpenWebForum Keynote 2021

The beauty of this approach is that it allows for much more granular alignment of interest and
greatly reduces the acquisition cost of new stakeholders in the business. Businesses can create a
vehicle specific to a subset of the business, product feature, or asset to give ownership in and
coordinate stakeholders in a targeted manner via transferable licence bearing tokens. This allows
us to consider a more modular, flexible, and composable approach to an enterprise. With
crypto-economic incentives, it is not necessary for one player to own the entire stack from top to
bottom: each area can coordinate on its own, and distribute value across whatever combination
of networks makes sense. Participants in a network can earn a stake in the system that delivers a

28
reward that is commensurate to the value created.

Every new network will require community bootstrapping prior to a product or protocol launch.
With distribution of ownership that grants the supply and demand side of a network ‘skin in the
game’ then the customer acquisition cost or ‘contribution acquisition cost’ to incorporate an
aligned participant significantly reduces and can even turn negative as it introduces incentives to
maximise value. Users change their behaviour and are more likely to put their own resources to
work to improve the network, even without receiving commensurate compensation.

14. What is possible with the web3 business model: the data flywheel
reimagined
There is a performance advantage to the open approach; it incentivises contribution from all the
stakeholders in a system with a promise of collective upside, not just maximising profit for
disconnected shareholders.

Rather than extracting as much value as possible from users to juice shareholder returns, the
open approach incentivises users to continue to contribute towards improving the flywheel at
step on its journey (figure 16).

Figure 16: The user-centric value flywheel

Not only does purposeful incentivisation of data throughout the flywheel increase the quality of
the flywheel and the feedback loop (shifting the product quality curve up) but users get rewarded
for the quality of data they provide, getting value back from the product, further increasing the

29
value for a user (both from earnings, as well as from a better product experience – figure 17).
Novel approaches to intellectual property and licensing allows proprietary value to flow more
easily throughout a network, in an appropriate and customisable manner. This allows for
intellectual property to confidently contribute to the open collective, secure in the knowledge that
the benefit will flow back to them and their community.

As with any system, there is of course a trade-off to be considered. Here the trade-off is that the
company (or more precisely the founding, scaling or backer contingents of the community) does
not get to extract all of the value from the product, being required to provide benefits back to the
stakeholders. However, the product experience ultimately benefits from active stakeholder
participation, delivering a greater economic outcome.

Figure 17: The effect of the web3 business model on product quality

This approach also allows the balance of contribution to shift over time. Protocols are built by
the founding team in the early stages, then progressively decentralise and receive more
contributions from external builders or even the consumers of the product. Moreover, even if an
actor does not make code contributions, they are still able to contribute to the design of policies
within it through governance, an opportunity open to all stakeholders.

30
Short, medium and long-term, both individuals and collectives can benefit economically with an
open approach. Taken together, this should ultimately lead to a higher optimum than any closed
system could ever achieve.

15.The first decade's experiments are promising.

Since its inception the space has proven it can successfully coordinate and reward global
networks of contributors and break down subsets of economic value into separate tokens.
Bitcoin, with a market cap close to USD1.5 tn, never had to ‘own’ its miners or create on- or
off-ramps (even though all of those players extract value symbiotically). Ethereum, with a
market cap north of USD400 bn, does not try to ‘own the stack’ either. From security (provided
by third-party validators who earn rewards) to the application layer (another set of disparate
actors) and to stablecoin payment networks, there exists a symbiotic relationship where value is
able to flow across different networks and stakeholders are either rewarded for good behaviour
or slashed for breaking the rules demonstrating true ‘skin in the game’ incentives. Furthermore,
additional tokens within an ecosystem are not a zero sum game in terms of resources (either
labour or capital). Instead, subsets of economic value are granularized into different vehicles in a
manner that avoids cannibalising the existing pool of value; a 'positive-sum' game. For example,
Optimism’s token coordinates contribution around the Optimism ecosystem, which ultimately
brings more value to Ethereum, and its token holders. We believe the same should exist for the
AI stack. Each layer of the stack (compute, data, models, governance and talent) can exist on its
own terms, rewarding contribution to all the myriad stakeholders who provide their time, talent,
or resources to help develop it.

16. We still have some way to go

The shift to a new business model takes time. We are conditioned to the web2 moat of corporate
command and conquer (charging for access, controlling ‘proprietary’ data, vertical integration,
vendor lock-in) and existing in a limited sum game. In the same way that we skeuomorphically
use notes on our phones to represent a yellow legal pad, we still treat web3 protocols with an
antiquated web2 SaaS point of view. To sustain a truly open web, we need to leverage the unique
composability characteristics and “skin in the game” community that can themselves be moats
when built on open source rails; and value can aggregate at many layers and between layers,
much more flexibly than ever before in truly open ecosystems.

🍣
Rightful criticism comes from those worrying about what a sustainable moat within an open
system looks like. Crypto protocols and apps like Uniswap are open, and have been forked ( ),
yet still remain on top—their code base copied and re-deployed in their entirety. However, it is
well understood in this space that truly open projects have defined new moats driven by integrity
(trust), billions of dollars in capital, open development standards (such as token standards), and a
loyal developer community.

31
 Figure 18: Commentary on moats in open source crypto networks by David Phelps39

There also remain technical challenges to be solved before we are truly ready to fully unlock the
world’s data bottom ups. Most notably, we need mechanisms for preserving user privacy to
ensure transparency and control over their data. Cryptography provides the answer, with multiple
different approaches for solving slightly different problems. We believe that we will all be
required to have truly private data solutions at scale:
● Homomorphic encryption for running computation over publicly available, yet
encrypted data (useful for public blockchains as well as outsourcing compute in a
privacy preserving manner). Teams like Zama and Lattica are pushing
performance with novel encryption schemes as well as hardware-based
optimizations, and the likes of Fhenix and Inco are putting FHE into production in
the blockchain arena.
● Multi-party computation (MPC) for allowing the joint computation of separate
entities private inputs, without needing to share data. Because of the importance
of communication between the parties, the primary use case to date has been
implementing custody solutions. More recent developments40 have shown
significant speed improvements.

39
Phelps, D. (2024, May 9). I need to rant... Twitter.
https://twitter.com/divine_economy/status/1788679122631156139
40
Tay, V. (2024, March 14). Introducing MPC-cmp: Pushing MPC wallet signing speeds 8x. Fireblocks.
https://www.fireblocks.com/blog/pushing-mpc-wallet-signing-speeds-8x-with-mpc-cmp-9/

32
 ● Zero knowledge (ZK) proofs for validating that a statement made by an actor is
true without them having to reveal information beyond showing the validity of the
statement itself. We have seen a Cambrian-type explosion41 in zk proving
schemes, with orders of magnitude42 improvements43 seemingly occurring every
few months. Practical zk-based solutions for blockchain privacy, requiring client
side proofs, move closer and closer to reality44.
● Compute-to-data45 where we bring computation to the data, allowing it to stay on
premise. This allows for outputs to be aggregated or anonymized after the
computation is completed, maintaining privacy for the individuals.

We have seen the empirical success of the web3 business model first-hand from over a decade of
total immersion in this space: from the permissionless development of Ethereum to the
conviction and intensity of the ownership-powered Sorare collector community. However, the
work is not complete. As the number of experiments playing out in the world of web3
accelerates, Fabric continues to assemble more data-rich case-studies of not just why this is
better, but by how much this is better. Beyond that we believe it should be possible to formulate
some axioms and identities of this new token-economics design space that will have both
superior predictive capability and higher predictions of prosperity that their dismal cousins from
both macro and micro economics to-date. We continue to flesh out ways to measure this new
economy and explore research into appropriate token-economic incentives and distribution
mechanisms for ownership and invite others to join us.

17. Conclusion
The past few decades of open source experimentation in cryptography, game theory and artificial
intelligence have led us to today's watershed moment where we must choose between
supercharged dystopian surveillance (state) capitalism or human-centric software that acts in our
best interest.

41
StarkWare. (2023, June 9). Cambrian explosion of cryptographic proofs. Medium.
https://medium.com/starkware/cambrian-explosion-of-cryptographic-proofs-5740a41cdbd2
42
Zhu, M., & Ragsdale, S. (2024, April 9). Building jolt: A fast, easy-to-use ZKVM. a16z crypto.
https://a16zcrypto.com/posts/article/building-jolt/
43
Sasson, E. B. (2024, May 29). Stark @ Home: Starkware’s lightning fast next-gen prover. StarkWare.
https://starkware.co/starkwares-lightning-fast-next-gen-prover/
44
Wu, J. (2023, October 13). Announcing the aztec sandbox: The endgame for smart contract privacy - aztec. Aztec
Protocol. https://aztec.network/blog/announcing-aztec-sandbox-the-endgame-for-smart-contract-privacy/
45
McConaghy, T. (2022, November 9). How does ocean compute-to-data relate to other privacy-preserving
approaches?. Medium.
https://blog.oceanprotocol.com/how-ocean-compute-to-data-relates-to-other-privacy-preserving-technology-b4e1c33
0483?gi=725ceaa381a0

33
17.1 Business model level Change
The most important technological innovations do not just spur innovation in product or
distribution, they also lead to a core reimagining of business models. From internet marketplaces
to advert-supported free-to-use networks, and to SaaS: finding new ways to construct business
propositions has the biggest impact on the economy as a whole. However, such situations are
also the hardest to imagine in a nascent sector.

Figure 19: ‘The real winners on the ‘Net will be firms selling directly to consumers'. Barron's
magazine Profile on Amazon in 199946

Web3 allows us to imagine a more flexible approach to the concept of ‘ownership’. No longer
are stakeholders forced into restrictive situations with limited upside, they are now able to
participate across the economic activities of the underlying organisation. The lines dividing
stakeholders in software products have already blurred as technology eliminates the difficulties
inherent in juggling multiple roles. Most apps today attempt to utilise the power of incentives to
make customers their influencers, users their sales people, and advertisers their content suppliers.
Programmable incentives and open ownership primitives brought to networks and protocols take

46
Amazon.bomb - Barron’s. Barron’s. (1999, May). https://www.barrons.com/articles/SB927932262753284707

34
this a step further: supercharging true stakeholder capitalism, taking zero sum dynamics to
positive sum—a modern version of mutualism47.

17.2 A Fundamental Dislocation

To understand the level of disruption occurring one must be aware that the software primitives
were simply not possible without hard silicon, network architecture, and system level
innovations to support them. The capabilities of distributed, specialised, accelerated computing
addressing the specific requirements of cloud storage, low latency cloud compute, and
end-to-end encryption required at this scale were propelled by the economic cycle of
infrastructure and applications. The scale of the change prompted not only impacts how we
work, play and shop but sends fundamental shockwaves in every direction. From investment in
silicon and robotics hardware manufacturing to the mining of raw materials and geographically
decentralised and sustainable forms of energy generation48, all the way up to the organisation
principles of enterprises and even nation states and whatever will come after them.

17.3 Part of a Longer Arc

Two hundred and fifty years ago, European empires used economic, religious and monarchic
powers to subjugate both colonists and colonised alike across the globe. As the scale and scope
extended, This despotic model began to strain in a number of location aroound the worlds,
notably France and America. In the United States the founding fathers are revered for their
success in leading a country to the establishment of a new constitution. This instantiated freedom
of religion and separation of such from state, the heart of which was documented in the 1791 Bill
of Rights and specifically the 1st and 4th amendments relating to the rights to bring grievances
against governments and to maintain personal liberty. The intention was that the yoke of the
European empires was not simply swapped for a new master and the full potential of the human
spirit might be fulfilled.

As Thomas Paine, one of the founding fathers of the United States stated, ‘Of more worth is one
honest man to society...than all the crowned ruffians that ever lived’ and James Madison said in
1785 that ‘Religion is wholly exempt from its [the government’s] cognizance. A mutual
independence between Religion & Government is essential to the purity of both, and as
guaranteed by the Constitution of the United States’. In this way the founding fathers stopped
despotic monarchs using religion to assert power and re-established the right of the individual to
have a direct relationship with their personal God. There is a parallel here with Satoshi
Nakamoto’s ‘constitutional’ bitcoin white paper that established the principles and prototypes for
‘one honest man’s’ or sovereign individual’s direct relationship with their personal money,
47
Wikimedia Foundation. (2024b, June 4). Mutualism (economic theory). Wikipedia.
https://en.wikipedia.org/wiki/Mutualism_(economic_theory)#:~:text=Mutualists%20seek%20to%20construct%20an
,of%20creating%20a%20federal%20society.
48
Dwarkesh Podcast https://youtu.be/bc6uFV9CJGg?t=1567&si=-MtV7znTWOR_X-HV

35
thereby removing the interference of banks and governments. This empowerment of the
individual must now continue through the work of a new peloton of ‘founding fathers’ who are
architecting of an open, decentralised and mutually-owned approach to AI that guarantees the
‘honest man’ a direct relationship with the accelerating use of our own personal data, such that
it can both safe and generally and abundantly fruitful.

We have sketched out our vision for the myriad AI-driven application experiences that will
deliver upon this new Open Economy. We have discussed how with the correct starting
conditions and building blocks we can architect a ‘can’t ever be evil’ approach to unlocking
these services. Although we have seen that this approach is better philosophically, better
technically, sadly this will come to naught without a better business model. Thankfully these
same technologies also make possible a new more equitable, sustainable, and prosperous form of
capitalism through token-economics. This is the frontier where we must push ahead. With this
endeavour in mind, let us see how Beth might now be doing:

‘Food’s on its way’. Beth never looked up from her conversation on the couch, not
breaking flow at understanding this news. She didn’t need to, for no one had ‘told
her’ about the food as understood in the early 21st century, instead she had become
aware of it, somewhere in her brain. The question of whether the USDC23.16 it cost
was the correct amount or how it would be paid never crossed that same mind.
‘Jarvis’ was a thing of the past, antiquated tech long rendered obsolete by modern
telekinesis based agent communication methods.

Her date would be impressed, she was sure. The simple act of desiring food was
enough for action to be taken; her personal preferences and the correct choices
refined not only on her personal thoughts, but those of billions of others. Why had
she been comfortable doing this? Thought control had long been considered a
dystopian nightmare, the potential for abuse too powerful to ignore. Technology had
provided the answer, the ability to use your most personal innermost data while
controlling its use in an absolute manner, without trusting intermediaries or
counterparties.

That wasn’t the only perk of the night. Beth had bought into various aspects of the
food supply chain, in some instances supplying her money, in other instances her
time, her personal network, her feedback, even her excess home energy when she had
been out of town. Each of those had given her a ‘piece’ of the underlying networks;
not just understanding and comfort about the quality of the produce that would soon
be on her plate - from farm to fork - but an upside in their success.

36
 And not just ‘success’ in the short term economic sense, but also to drive zero
inventory supply chain management that is so critical to minimise food wastage and
energy consumption, and also to further optimise her diet and the diet of millions of
celiac sufferers like her round the world. Ultimately this system thrives on the health
of all of its actors, not just on growth in shareholder profits. So whilst the economics
of the whole system are larger than ever it has indeed grown there slower but much
more predictably and sustainably.

As we wrote in our thesis paper, published in 2017 :

“We’re convinced that the coming decades of technical innovation will be driven
by the interplay between Edge Computing, Machine Learning, and
Decentralised Data: the edge computing on swarms of devices capturing
millions of data points, the advancements in machine learning algorithms
ingesting this wealth of data and a substrate of decentralised data enabling
secure and scalable communication, coordination and fair incentivisation.”

We believe this today—even more strongly than we did then. In particular we must fulfil the
potential of token-powered ‘modern mutualism’. If you are currently addressing this specific
point, or any of the myriad other opportunities, or even something we have not considered,
please get in touch.

37
18. Acknowledgements
With great thanks to Trent McConaghy, Shruti Appiah, Steven Waterhouse, Max Mersch, Julien
Thevenard, Ian Emerson, Camilla McFarland and Lata Persson for their review and commentary
during the writing of this paper.

Notice: We may hold positions in the above mentioned companies, therefore no content should
be taken as financial advice.

19. Bibliography
➢ Amazon.bomb - Barron’s. Barron’s. (1999, May).
https://www.barrons.com/articles/SB927932262753284707
➢ Amber Group. (2023, April 20). Rebuilding connections. Medium.
https://ambergroup.medium.com/rebuilding-connections-faf0fe9e8bd7
➢ Anthropic CEO on leaving openai and predictions for future of ai. YouTube. (2023,
October 6). https://youtu.be/gAaCqj6j5sQ?t=5625&si=NjmJGIoDeKqEQC-0

38
➢ AOL’s ‘Walled Garden’ - WSJ. WallStreetJournal. (2000, September 4).
https://www.wsj.com/articles/SB968104011203980910
➢ Auslender, V. (2023, August 30). Why meta’s open source is not really open. ctech.
https://www.calcalistech.com/ctechnews/article/atv6xnkya
➢ Boykis, V. (2023, February 26). Everything I understand about chatgpt.
https://gist.github.com/veekaybee/6f8885e9906aa9c5408ebe5c7e870698
➢ Circle Reserve Fund: USDXX: Institutional. BlackRock. (n.d.).
https://www.blackrock.com/cash/en-us/products/329365/circle-reserve-fund
➢ Earn a stake in the AI Revolution. Grass. (n.d.). https://www.getgrass.io/
➢ European Parliament. (2023, May). Markets in crypto-assets regulation (MICA).
https://www.esma.europa.eu/esmas-activities/digital-finance-and-innovation/markets-cry
pto-assets-regulation-mica
➢ Financial Services Agency, the Japanese Government. (2022, September). Regulatory
framework for crypto-assets and Stablecoins.
https://www.fsa.go.jp/inter/etc/20220914-2/02.pdf
➢ Gabaix, X. (2008, September 5). Power laws in economics and finance. NBER.
https://www.nber.org/papers/w14299
➢ Gantz, J., & Reinsel, D. (2012, December). THE DIGITAL UNIVERSE IN 2020: Big
Data, Bigger Digital Shadows, and Biggest Grow th in the Far East.
https://www.cs.princeton.edu/courses/archive/spring13/cos598C/idc-the-digital-universe-
in-2020.pdf
➢ Gent, E. (2024, March 25). The tech industry can’t agree on what open-source AI means.
that’s a problem. MIT Technology Review.
https://www.technologyreview.com/2024/03/25/1090111/tech-industry-open-source-ai-de
finition-problem/
➢ Gurley, B. (2023, September 16). All-in summit: Bill Gurley presents 2,851 miles.
YouTube. https://www.youtube.com/watch?v=F9cO3-MLHOM
➢ Hagey, K., & Fitch, A. (2024, February 8). Sam Altman Seeks Trillions of Dollars to
Reshape Business of Chips and AI.
https://www.wsj.com/tech/ai/sam-altman-seeks-trillions-of-dollars-to-reshape-business-of
-chips-and-ai-89ab3db0
➢ Hansjee, A., & Brinkmann, F. (2024, June 3). Blockchain payments: Beyond the hype.
Medium.
https://fabric-vc.medium.com/blockchain-payments-beyond-the-hype-5455c56454e4
➢ Introducing openai. openAI. (n.d.). https://openai.com/index/introducing-openai
➢ Jones, D. (2024, May 23). How much LLM Training Data is there, in the limit?.
Educating Silicon.
https://www.educatingsilicon.com/2024/05/09/how-much-llm-training-data-is-there-in-th
e-limit/

39
➢ Li, F.-F. (2015, March). How we’re teaching computers to understand pictures. Fei-Fei
Li: How we’re teaching computers to understand pictures | TED Talk.
https://www.ted.com/talks/fei_fei_li_how_we_re_teaching_computers_to_understand_pi
ctures?language=en
➢ Mark Zuckerberg - Llama 3, $10B models, Caesar Augustus, & 1 GW datacenters.
YouTube. (2024, April 18).
https://youtu.be/bc6uFV9CJGg?t=2284&si=Ct2wai48K4rqavjJ
➢ McConaghy, T. (2022a, November 9). How does ocean compute-to-data relate to other
privacy-preserving approaches?. Medium.
https://blog.oceanprotocol.com/how-ocean-compute-to-data-relates-to-other-privacy-pres
erving-technology-b4e1c330483
➢ McConaghy, T. (2022b, November 9). How does ocean compute-to-data relate to other
privacy-preserving approaches?. Medium.
https://blog.oceanprotocol.com/how-ocean-compute-to-data-relates-to-other-privacy-pres
erving-technology-b4e1c330483?gi=725ceaa381a0
➢ Mersch, M. (2018, December 20). The Fabric Ventures Investment Thesis. Medium.
https://medium.com/fabric-ventures/the-fabric-ventures-investment-thesis-6cd08684b467
➢ Mistral AI: Open-weight models. Mistral AI | Open-weight models. (2024, February 25).
https://web.archive.org/web/20240225001133/https://mistral.ai/
➢ Monetary Authority of Singapore. (2023, August). MAS Finalises Stablecoin Regulatory
Framework.
https://www.mas.gov.sg/news/media-releases/2023/mas-finalises-stablecoin-regulatory-fr
amework
➢ Optimism. (2024, April 18). Retro funding 4: Onchain Builders - round details. Optimism
Collective. https://gov.optimism.io/t/retro-funding-4-onchain-builders-round-details/7988
➢ O’Laughlin, D. (2024, June 2). The internet as you know it is dying. The Internet as You
Know It Is Dying - by Doug O’Laughlin.
https://www.fabricatedknowledge.com/p/the-internet-as-you-know-it-is-dying?r=j1yw&u
tm_medium=ios&triedRedirect=true
➢ PayPal stablecoin: US dollar cryptocurrency. PayPal Stablecoin | US Dollar
Cryptocurrency | PayPal US. (n.d.).
https://www.paypal.com/us/digital-wallet/manage-money/crypto/pyusd
➢ Perez, C. (2014). Technological Revolutions and financial capital: The Dynamics of
Bubbles and golden ages. Edward Elgar.
➢ Peters, J. (2023, June 8). Apollo for Reddit is shutting down. The Verge.
https://www.theverge.com/2023/6/8/23754183/apollo-reddit-app-shutting-down-api
➢ Phelps, D. (2024, May 9). I need to rant... Twitter.
https://twitter.com/divine_economy/status/1788679122631156139
➢ Rao, D. (2024, June 5). All-powerful, ever-pervasive AI is running out of internet.
theweek. https://theweek.com/tech/ai-running-out-of-data

40
➢ Sasson, E. B. (2024, May 29). Stark @ Home: Starkware’s lightning fast next-gen prover.
StarkWare. https://starkware.co/starkwares-lightning-fast-next-gen-prover/
➢ Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert,
T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L., van den
Driessche, G., Graepel, T., & Hassabis, D. (2017, October 19). Mastering the game of go
without human knowledge. Nature News. https://www.nature.com/articles/nature24270
➢ StarkWare. (2023, June 9). Cambrian explosion of cryptographic proofs. Medium.
https://medium.com/starkware/cambrian-explosion-of-cryptographic-proofs-5740a41cdb
d2
➢ Tao bittensor: Decentralized machine learning 2024. (2024, March 27).
https://bittensor.org/
➢ Tay, V. (2024, March 14). Introducing MPC-cmp: Pushing MPC wallet signing speeds
8x. Fireblocks.
https://www.fireblocks.com/blog/pushing-mpc-wallet-signing-speeds-8x-with-mpc-cmp-
9/
➢ Thevenard, J. (2020, June 24). Tokenised ownership is the best coordination tool since
equity. Medium.
https://medium.com/fabric-ventures/tokenised-ownership-is-the-best-coordination-tool-si
nce-equity-7c957a00d818
➢ uf.eek.eth. (2024, March). Activate Uniswap Protocol Governance. Snapshot.
https://snapshot.org/#/uniswapgovernance.eth/proposal/0xad7e3612d11d56b21f0b2274e4
ce825163bc1873d0e2ef809a3a98733df992a7
➢ UK Government. (2024, March 1). Economic crime and corporate transparency act 2023:
Factsheets. GOV.UK.
https://www.gov.uk/government/publications/economic-crime-and-corporate-transparenc
y-act-2023-factsheets
➢ Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L.,
& Polosukhin, I. (2017, August 2). Attention is all you need. arXiv.org.
https://arxiv.org/abs/1706.03762
➢ We connect wi-fi and mobile to improve coverage and enable mobile data traffic
offloading to our customers. Telefónica. (2024, January 26).
https://www.telefonica.com/en/communication-room/blog/wi-fi-and-mobile-to-improve-c
overage-enable-mobile-data-traffic-offloading/
➢ Wikimedia Foundation. (2024a, April 25). ImageNet. Wikipedia.
https://en.wikipedia.org/wiki/ImageNet#History_of_the_ImageNet_challenge
➢ Wikimedia Foundation. (2024b, June 4). Mutualism (economic theory). Wikipedia.
https://en.wikipedia.org/wiki/Mutualism_(economic_theory)#:~:text=Mutualists%20seek
%20to%20construct%20an,of%20creating%20a%20federal%20society.
➢ Wu, J. (2023, October 13). Announcing the aztec sandbox: The endgame for smart
contract privacy - aztec. Aztec Protocol.

41
 https://aztec.network/blog/announcing-aztec-sandbox-the-endgame-for-smart-contract-pri
vacy/
➢ Zhu, M., & Ragsdale, S. (2024, April 9). Building jolt: A fast, easy-to-use ZKVM. a16z
crypto. https://a16zcrypto.com/posts/article/building-jolt/

Further readings:
1. Vitalik - https://vitalik.eth.limo/general/2024/01/30/cryptoai.html
2. Trent McConaghy:
a. https://blog.oceanprotocol.com/blockchains-for-artificial-intelligence-ec63b02849
84
b. https://medium.com/@trentmc0/ai-daos-and-three-paths-to-get-there-cfa0a4cc37b
8
c. https://blog.oceanprotocol.com/from-ai-to-blockchain-to-data-meet-ocean-f210ff4
60465
d. https://medium.com/@trentmc0/bci-acc-a-path-to-balance-ai-superintelligence-80
bb6f32e39c
e. https://www.youtube.com/watch?v=JVr0a4_ttIg
3. Joel Monegro - AI Belongs Onchain -
https://www.placeholder.vc/blog/2023/10/23/artificial-intelligence-belongs-onchain
4. Ai X Crypto primer - by Alex cheema -
https://alexcheema.github.io/AIxCryptoPrimer.pdf
5. https://switchonleadership.com/magazine/regenerative-tech-slow-down-mend-things/
6. Need for Energy:
a. Data Centers Now Need a Reactor’s Worth of Power, Dominion Says
https://finance.yahoo.com/news/data-centers-now-reactor-worth-171044009.html
7. Naveen Rao on Open Source AI -
https://twitter.com/NaveenGRao/status/1781491370114633816
8. Ada Ventures Framework for Investing in AI-first Companies -
https://medium.com/ada-ventures/ada-ventures-framework-for-investing-in-ai-first-comp
anies-df066049f096
9. The rise of DePin (R.Muirhead Token2049 Panel with Raullen Chai (IoTeX), Trevor
Harries-Jones (Render Network), Ahmad Shadid (io.net), Ariel Seidman (Hivemapper)
https://www.youtube.com/watch?v=Ta1-ecd3Pto
10. Rishi Sunak testing AI safety institute:
https://www.politico.eu/article/rishi-sunak-ai-testing-tech-ai-safety-institute/
11. Considerations for Governing Open Foundation Models -
https://hai.stanford.edu/sites/default/files/2023-12/Governing-Open-Foundation-Models.p
df

42
12. Dragonfly - decentralized inference blog
https://medium.com/dragonfly-research/dont-trust-verify-an-overview-of-decentralized-i
nference-c471a9f7a586
13. Emad Mostaque https://www.youtube.com/watch?v=1rVihrDo3rM
14. Alex ohanian:
https://x.com/alexisohanian/status/1778448863906443752?s=46&t=vfYxiBXNgK3QZVf
nOF66Hw
15. Livepeer earning for generative compute tasks
https://x.com/petkanics/status/1778157805293248672?s=48&t=lBF42Yc1kXaAUfTLeB
kEZA
16. Why Meta’s open source is not really open -
https://www.calcalistech.com/ctechnews/article/atv6xnkya
17. Podcast: The AI Email Assistant I've Been Waiting for, with Andrew Lee of Shortwave:
https://podcasts.apple.com/gb/podcast/the-cognitive-revolution-ai-builders-researchers-an
d/id1669813431?i=1000651774407
18. ImageNet: A large-scale hierarchical image database - Jia Deng; Wei Dong; Richard
Socher; Li-Jia Li; Kai Li; Li Fei-Fei - https://ieeexplore.ieee.org/document/5206848
19. TinyStories: How Small Can Language Models Be and Still Speak Coherent English? By
Ronen Eldan, Yuanzhi Li - https://arxiv.org/abs/2305.07759
20. Will We Run Out of ML Data? Evidence From Projecting Dataset Size Trends by Pablo
Villalobos, Jaime Sevilla, Lennart Heim, Tamay Besiroglu, Marius Hobbhahn, Anson Ho
-
https://epochai.org/blog/will-we-run-out-of-ml-data-evidence-from-projecting-dataset?ut
m_source=substack&utm_medium=email
21. LLaMA Now Goes Faster on CPUs (Justine.lol, blog)
22. Announcing DBRX: A new standard for efficient open source LLMs (Databricks blog).
23. Inside the Creation of the World’s Most Powerful Open Source AI Model (Wired).
24. Microsoft and OpenAi Plot $100 Billion Stargate AI Supercomputer (The Information)
25. Anthropic AI on leaving OpenAI - quote on why OpenSource needs to be
controlledAnthropic CEO on Leaving OpenAI and Predictions for Future of AI
26. Coatue AI report 2023 -
https://www.coatue.com/blog/perspective/ai-the-coming-revolution-2023
27. ImageNet challenge 2012 -
https://en.wikipedia.org/wiki/ImageNet#History_of_the_ImageNet_challenge
28. Fei-Fei Li - How we’re teaching computers to see -
https://www.ted.com/talks/fei_fei_li_how_we_re_teaching_computers_to_understand_pi
ctures?language=en
29. AlexNet paper - ImageNet Classification with Deep Convolutional Neural Networks
https://proceedings.neurips.cc/paper_files/paper/2012/file/c399862d3b9d6b76c8436e924
a68c45b-Paper.pdf

43
30. Casey Caruso - Decentralized AI https://www.caseycaruso.com/thoughts/decentralized-ai

44