FUNDING
Dermot O'Riordan and Nelson Ryan
January 16, 2023

Investment Thesis for Gensyn

Six minutes
Those running ML research outside of the major tech giants are left with few good options, relying on expensive cloud providers such as AWS for spot cloud compute or racking GPUs themselves to complete this work. Such ever-increasing computing demands for ML projects is the reason why established global tech giants run most large scale AI projects. When you look closer, it becomes clear that much of the problems we see with AI compute boil down to a coordination problem as the supply is artificially constrained.
Follow us

“Both Hegel and Marx believed that the evolution of human societies was not open-ended, but would end when mankind had achieved a form of society that satisfied its deepest and most fundamental longings…” Francis Fukuyama, The End of History and the Last Man

Exponential advances in technology promise a future utopian world of abundance. Yet, there is a growing realisation that the biggest question in technology is about who is in control. Politics, not economics, is ultimately what makes the world go round. Whatever your political affiliation, Lenin seemed to get this one right when he said that the crucial question in politics is “who will overtake whom?”. As coined by the term, Who, Whom? 

From the actions of the cloud giants and social media platforms in response to the storming of the Capitol to the Canadian legislation to restrict access to financial services for anyone directly or "indirectly" supporting the trucker protests, supply chain resilience, to... Ukraine.

Access to resources always has a political element. And how we govern our fundamental resources is one of the most critical questions of our times. 

Web3 as the governance wrapper for fundamental shared resources

“if the priority for an increasing number of citizens, companies, and countries is to escape centralization, then the answer will not be competing centralized entities, but rather a return to open protocols… open alternatives that are perhaps not as full-featured and easy-to-use as big tech offerings, at least in the short to medium-term, but possess the killer feature of not having a San Francisco kill-switch.” Ben Thompson, Internet 3.0 and the Beginning of (Tech) History

The opportunity for DAOs lies with the ability to leverage a blockchain as a neutral, independent monitoring and enforcement mechanism that can recognise the contribution and governance rights of much larger groups of stakeholders without exponentially increasing coordination costs.

DAOs provide the organisational form necessary to govern shared fundamental resources. The real question, therefore, is, what fundamental resources do we need to focus on sharing? 

Machine learning compute as a fundamental resource

In 2020, data scientists used machine learning (ML) technology to help make diagnoses and aid researchers in developing a cure for COVID-19. And DeepMind just published a paper in Nature about using deep learning, a subfield of ML, to improve a vital component of some approaches to nuclear fusion

However, applications of ML to nuclear fusion research and vaccine developments are just the tip of the iceberg, with further applications at the heart of innovations in self-driving cars, natural language processing, silicon chip design, and so much more. 

Yet, while there has been significant progress in ML research and applications, OpenAI’s analysis shows that the amount of compute used in the largest AI training runs has increased exponentially, doubling every 3-4 months; compared to Moore’s Law, which doubles every two years. 

Source: OpenAI 

Those running ML research outside of the major tech giants are left with few good options, relying on expensive cloud providers such as AWS for spot cloud compute or racking GPUs themselves to complete this work. Such ever-increasing computing demands for ML projects is the reason why established global tech giants run most large scale AI projects. 

When you look closer, it becomes clear that much of the problems we see with AI compute boil down to a coordination problem as the supply is artificially constrained. Enormous GPU pools currently mine Eth1 and other PoW protocols while on-premises data centres have average server utilisations rates are often between 12 to 18%, cloud server utilisation is approximately 65%, personal computers sit idle, mobile phones have huge potential for overnight utilisation, and blockchain miners have massively underutilised expensive hardware equipment. 

All in all, we see three major problems overhanging the future of ML research and development:

  1. Ever-increasing costs: Despite gains in processor performance, it’s getting more expensive to carry out fundamental ML research due to model size, but margins for the tech giants are increasing. At the same time, 35% of the world’s supply of computing power sits idle.
  1. A handful of companies gate access to fundamental research: As ML applications become more and more important, the opportunity cost for restricting access to ML compute becomes ever more significant. Is it right that Chinese, Russian, or any other scientist working in a country in the crosshairs of global politicians should get shut off from continuing their research into new vaccines or climate change technology? What about if you are shut off simply because you decided to speak up against your country? Where do politics stop and end in this regard? 
  1. Lack of collaboration: Without a method to pool resources while still capturing value, there is no incentive to open source the foundational models most likely to unlock the next major advances in AI. Instead, models remain gated behind private organisations like OpenAI.

Thankfully, Web3 now provides us with the tools to 1) coordinate and incentivise compute providers across the globe who don’t have to trust each other, 2) create a democratic and permissionless governance structure around fundamental infrastructure, and 3) directly incentivise collaboration amongst researchers.

Gensyn: enabling an open infrastructure for cost-efficient ML compute

Ark Invest estimates that AI hardware companies could produce $1.7 trillion in annual revenue by 2030, roughly 100x today’s figure. 

What will such a vast market look like in 8 years’ time? And how will access to AI compute be governed? Further, will there be a mandate to keep costs affordable for researchers working outside of the biggest companies? 

With all this in mind, the long-term vision of Gensyn is to “enable anyone to train ML models for any task using a self-organising network that encompasses every source of compute power in existence”.

Why is Gensyn better than existing approaches to deep learning compute?

1. Built for global scale using a market-based approach 

As Gensyn connects processing units directly to the demand side, these infrastructure providers - “Solvers” in Gensyn parlance - receive revenue for training deep learning models that they select from a mempool of requested jobs.

If a solver sees demand on the network increasing or the price for compute increasing, they’ll know it’s a great time to spin up more infrastructure, and the same works vice versa. These price signals - the market’s “invisible hand” - guide participants in Gensyn to deliver (or not to deliver) what the marketplace requires.

Leveraging the global supply of compute enables Gensyn to deliver compute at a scale that simply isn’t possible for centralised alternatives. Less powerful machines maintain the Gensyn network too. While powerful GPUs or TPUs take up positions as off-chain solvers, less powerful CPUs are used to run blockchain nodes and maintain on-chain consensus.

2. Dramatically more cost-efficient

Without any human intermediary, Gensyn’s protocol takes payment from researchers for the machine learning tasks they request and distributes rewards to solvers for their work done.

Gensyn’s open protocol-based approach to delivering ML compute can deliver compute at a fraction of the cost of centralised cloud providers. For example, existing solutions such as AWS spot compute cost around $0.9 per hour, AWS on-demand at $2 per hour and more generalised crypto native solutions such as Golem cost up to $1.2 per hour.

Source: Competitor analysis Gensyn Litepaper 

3. Credibly neutral and censorship-resistant

The Gensyn network is open to anyone, allowing infrastructure owners anywhere in the world to earn revenue for performing compute tasks in the network while having the opportunity to become an owner of the network for doing so.

Similarly, the Gensyn Protocol will have no kill switch or centralised party who can censor work completed by the network. The Gensyn Protocol has no political bias supporting anyone who wishes to use the protocol - from hobbyist researchers to academics, large scale corporates and even nation-states.

4. User owned and operated

At maturity, the Gensyn Protocol will be owned and operated by its core stakeholders: infrastructure providers, researchers, technologists, protocol politicians, not directors motivated to maximise shareholder profits. 

5. Open source collaboration

A protocol with a sole vision of enabling anyone to train ML models for any task has coordination at its heart. And if greater collaboration is necessary to unlock the next advances in AI, then Gensyn as a credibly neutral platform owned and operated by AI researchers and infra providers is most likely to be the player with the extrinsic (financial) and intrinsic (culture and values) incentives to do so. 

How does Gensyn work? 

One of the well-known challenges in building any decentralised network for compute is the difficulty of verifying that this work has been completed correctly by the network. While more general-purpose computation networks like Golem use a more naive solution to this problem by simply having one or more additional parties replicate the work. However, this solution fails to solve the problem as it not only fails to scale but is also inefficient and increases the cost of providing the service.

The Gensyn protocol incorporates three key concepts to solve this replication problem. 

The first is Gensyn’s “probabilistic proof-of-learning” verification method which uses the metadata from gradient-based optimisation processes to form a receipt for work completed, similar to a “Proof of Relay” in Pocket

Source: Initial research into the novel verification system Gensyn Litepaper 

For the second Gensyn incorporates “graph-based pinpoint protocol”, which utilises a graph-based pinpoint protocol and cross-evaluator consistency execution to allow work to be re-run and compared for consistency and confirmed with the on-chain record. 

For the third concept to keep all actors economically aligned, a “Truebit Style Incentive Game” is incorporated using staking and slashing to ensure all actors are financially motivated to complete all of their aligned work honestly and correctly. This process allows for quick verification through replicating certain portions of the finished work, similar to the way Broadcasters may check individual segments in Livepeer to ensure they were encoded correctly.

How are tasks completed within Gensyn?

Within the Gensyn Protocol, there are four types of participants: 

  • Submitters - submit deep learning tasks to the network to be completed for payment
  • Solvers - perform the compute work and generate receipts of completed work (similar to Service nodes in Pocket Network, Mix nodes in Nym or Transcoders in Livepeer)
  • Verifiers - ensure the work has been performed as requested through partial replication
  • Whistleblowers - check the work of verifiers and may challenge incorrectly performed work

Each of these participants play a key role in ensuring work submitted to the network is completed and verified while keeping everyone honest. Tasks completed by the network can be broken down into five phases.

  1. The submission phase in which submitters specify the details of the work to be completed and split it into smaller tasks if necessary. 
  2. The profiling phase in which a baseline performance is determined to establish verification thresholds. 
  3. The training phase in which work is assigned to solvers and the task is performed in a similar process to sessions in Pocket. 
  4. The verification phase in which work receipts are submitted and the completed work is verified. 
  5. The challenge phase in which verified work may be challenged and the challenger can earn the reward if the work was verified incorrectly. 

It is through the elegance of Gensyn’s design that these tasks can be performed in a decentralised network with minimal redundancy in the process.

The Gensyn flywheel

Post-launch, Gensyn will be able to leverage network effects around its new incentivised network for ML compute. Two-sided marketplace network effects will result directly from increased usage of the network and a growing pool of network fees for the supply-side to earn. This will attract additional infrastructure providers to the network, improving the supply-side of the network by increasing the capacity of the network, which should, in turn, lower costs and incentivise further usage from the demand-side. 

What’s the future for Gensyn?

As the Gensyn protocol scales - and after the project has fully exited to the community - governance of the Gensyn Protocol will become increasingly critical. Stakeholders will be mandated with the task of ensuring Gensyn remains true to its mission of enabling "anyone to train ML models for any task using a self-organising network that encompasses every source of compute power in existence” - a truly credibly neutral platform for ML compute. 
We believe that Gensyn provides the technical substrate, financial incentives, and intrinsic collaborative open-source values necessary to empower new businesses and research simply not possible for the public in the centralised world of today. 

Support Gensyn's mission

Disclosure: Eden Block is an investor in Gensyn

Nothing contained herein constitutes investment, legal, tax or other advice nor is to be relied upon in making an investment or other decision. This presentation contains the opinions of the author, and such opinions are subject to change without notice. Furthermore, it may also include data and opinions derived from third party sources. Eden Block does not accept liability for the accuracy or completeness of any such information or opinions which can be subject to change without notice.