We contributed to Lido DAO, P2P.org, =nil; Foundation, DRPC, Neutron and invested into 150+ projects
Stepan Gershuni, Vitaly Kleban
Sep 4, 2024What to build, where to invest and what to expect from Decentralized AI: 38 ideas, open opportunities and challenges in decentralized & distributed AI
This article explores the most promising ideas and technologies that will enable truly decentralized AI: the one that is unbiased, permisionless and open for innovation and composability. Thesis around why this is happening and what economic impact this will have in outlined in the first part of this article: The Internet of Intelligence.
Over the last year, we've seen growing interest in decentralized AI, primarily driven by:
The emergent properties and value of interconnected, composable AI systems. Similar to the Web, many practical AI use cases require not just a single model but multi-agent collaboration to deliver a seamless user experience.
Concerns about centralization and control of the most capable AI models by a handful of organizations or nation-states.
The need for more open, permissionless experimentation in machine learning research and development, particularly from private labs.
Issues of bias and lack of transparency in closed-source, closed-weights models.
Last year witnessed a surge in decentralized AI projects that, while simplistic, achieved relative success in terms of valuation. These projects primarily focused on token-gated chatbots, GPU rental services, and LLM-based autonomous agents. Although these experiments are impressive and often entertaining, I contend that the true value of decentralized AI lies deeper than these surface-level applications.
In this post, I'll explore the most crucial directions for research, product development, and company building that will enable useful and universally accessible decentralized and distributed AI.
Compute refers to the processing power needed to run and train AI systems. Decentralizing compute distribute the computational load, increasing resilience and reducing centralized control over AI capabilities.
End game
Compute services that are trustless, permissionless, private, and competitively priced with current cloud offerings. Developers should experience no UX difference when running agent code and inference on AWS, a trustless alternative, while also benefiting from the enhanced qualities of trustlessness (no need to trust the hardware provider), permissionlessness (no need to seek permission to host anything), and privacy (ensuring no one sees their code and data).
Decentralized compute should be universal, handling a wide range of tasks beyond just EVM code execution or small model inference. This universality ensures that any computational job, regardless of its nature, can benefit from the enhanced security, privacy, and resilience offered by decentralized systems. By accommodating diverse workloads, decentralized compute can attract a broader user base.
Training and inference should be communication-efficient: It should be possible to run complex AI models in decentralised setting as it may be more reliable, censorship protected and cost-effective. The main challenge for decentralized systems is relatively unreliable and slow communication over the Internet (compare to the datacenter fabric). Given that both training and inference need to be communication-efficient to lower the dependency on the network bandwidth and reliability. There is a certain progress in the area, with the approaches like Petals, DiLoCo, and DiPaCo showing promise and traction.
Reliability: web3 cloud should provide some form of SLA or reliability guarantees to be competitive with AWS and GCP (not saying they don’t have outages but they do have a brand name).
Alternative infrastructure solutions: currently the standard de-facto for the training and inference is the Nvidia GPUs located in some datacenter, but there are alternative options for the infrastructure, such as Qualcomm inference accelerators Cloud AI 100 that are capable to deliver the same performance but with 2x lower CAPEX and 4x lower power consumption. Some companies such as tinygrad with their tinybox solution or Lambda Labs with their AI-ready workstations. The main idea here is to get as much compute power out of residential power outlet as possible. Companies like Etched and Groq are building LLM-specific accelerators called Language Processing Units that are capable of producing up to 300 tokens per second. Companies like Exo show the ability to run even largest available open source models on a pair of consumer laptops, enabling enterprises to run their models without building a dedicated infrastructure for it. Such an approach also contributes to the circular economy by re-using partially amortised consumer hardware and distributing the load over the entire power grid.
Variable privacy: a flexible approach where users can choose different levels of privacy protection based on their specific needs, balancing security with cost-efficiency. At one end of the spectrum, Trusted Execution Environments offer a relatively inexpensive solution with moderate privacy guarantees, though they may be vulnerable to certain attacks. On the other end, more robust cryptographic methods like Fully Homomorphic Encryption provide stronger privacy assurances but at a significantly higher computational cost.
Data is the lifeblood of AI, essential for learning and decision-making. Decentralizing data improves privacy, reduces monopolies, and increases the diversity of training datasets.
End game
Data incentivization networks would leverage mechanism design principles to encourage the creation and validation of high-quality datasets. These systems would reward contributors for generating valuable data and verifiers for ensuring its accuracy, creating a self-sustaining ecosystem of data production and curation.
Large-scale synthetic data generation would become accessible to a wide range of users. Smaller and specialized models (SLMs, fine-tuned models or expert RAG-powered AI applications) can generate synthetic data with a higher quality compared to larger SoTA LLMs.
The scope of data would extend beyond text to encompass a diverse range of modalities. This would include visual data (images, videos, 3D models), audio, code, and domain-specific data types such as DNA sequences or chemical compounds.
Continuously updated and curated datasets would serve as dynamic knowledge bases for AI systems. These datasets would be designed for easy integration with RAG techniques or model fine-tuning, allowing AI models to stay current with the latest information and adapt to evolving contexts without full retraining.
Models are the core AI systems trained to perform various tasks, from language processing to video or music generation. Decentralizing models promotes diversity in AI capabilities, reduces dependency on a few dominant players, and allows for specialized models tailored to specific needs or cultural/legal/business contexts.
End game
Continuously pre-trained models created by a large decentralized network of contributors. Similar to the previous example, this will provide the most relevant and up-to-date result. In addition, such models can use a decentralized compute networks (discussed in the first chapter) to ensure constant uptime and uninterrupted pre-training.
Fair, verifiable, and decentralized compensation for model creation and inference. This ensures transparency and trust, as all participants can verify the legitimacy of the rewards. Enables co-ownership of models and some forms of “ML royalties” based on one’s contribution.
The flip side of model compensation is model funding. This could include crowdfunding mechanisms similar to ICO (or TGE, or whatever term is currently in vogue to avoid associations with ponzi). As an investor, I could support model creation with my time, compute resources, work, or financial contributions. In return, I'd receive a share of the future model profits, enforced by a decentralized network.
The use of mechanism design to monetize the usage of community-owned models.
Recent paper by Google invents a new primitive they call a “token auction”. Such auction allows market participants to bid directly on the token output which can be used for advertising purposes. For example, two companies — an airline company and a resort — can both bid so that the resulting message will be mentioning both brands and advertise a wonderful trip to resort A using airline B. Even though, this still leaves the question of transparency and bias in the model output open, it’s an interesting primitive to consider and build upon.
Evals are metrics used to assess and improve AI performance. Decentralizing evaluations ensures diverse testing scenarios and unbiased performance assessment, crucial for developing trustworthy AI.
End game
Ability to decentrally benchmark and evaluate AI applications in specific (potentially narrow) domains.
Able to evaluate not just against a public set of benchmarks but also using
closed test set and very specialized, potentially human-rated, tasks.
Evaluations against actual business problems are crucial for distinguishing between models that are genuinely useful and those that are merely good for casual conversation. Unfortunately, these evaluations are really hard to get and most of the datasets are proprietary. There should be a mechanism to contribute these datasets in a private manner.
Evaluations of agents is even more challenging problem as agents interact with each other and something that is good for one agent is not that good for another. There should be a mechanism to universally evaluate agent behaviour in a collaborative and multi-agent setting. Currently, many specialized benchmarks exist (OpenToM for theory of mind, InfiBench for coding Q&A, Infiagent-dabench for data analysis, SWEbench for software engineering skills, GAIA for universal assistants). However, for multi-agent (IOI) setting new type of evaluation approach is needed.
Now, moving from individual models to agents, we will discuss more complex systems that are capable to autonomously complete tasks and solve problems.
Agents are autonomous AI entities designed for specific tasks within AI ecosystems. Decentralizing agents enables diverse, independent AI actors. Millions of AI agents working together will propel us towards the AGI and automate all kinds of work that people do today.
Agents can be as simple as chaining a few LLMs together—using smaller, cheaper ones for simple tasks like natural language classification, and more complex models for advanced reasoning, coding, or analysis.
MAS are networks of interacting AI agents for complex problem-solving. Decentralizing MAS fosters emergent behaviors due to trustless runtime composability. Decentralized MAS systems also enables continuous adaptation of the multi-agent system as new capabilities are creates or market conditions change (e.g. price of compute for certain model or agent).
End game
An efficient workflow orchestration system exists. Orchestration in multi-agent systems can be highly dynamic, adapting on the go if something doesn't work as planned. It is capable of selecting different optimal paths depending on whether cost or quality is prioritized for a given task. The system continuously indexes the space of available agents and enhances its registry by integrating new agents and removing underperforming ones. It keeps detailed logs and tracks the reputation of each agent, using this feedback loop to improve quality for future uses.
Agents can manage budgets (in a form of real money, tokens or compute allocations) to budget and plan multi-agent workflows. Once this is a reality, a new set of tools, such as AI-native accounting, checking and even credit will become necessary.
Agents can efficiently communicate data, intents, context. Such communication can involve natural language or machine-readable data such as directly passing embeddings or even weights directly between systems.
Multi-agent systems are fault-tolerant and distributed, meaning they can continue to operate even if some agents fail or are compromised.
Autonomous agents that are able to “wake” themselves up to complete tasks. Think of this as an intelligent cron scheduler or a smart contract automation systems like Keep3r Network but for autonomous agents.
Multi-agent systems are interpretable and auditable, allowing stakeholders to understand and verify the decision-making processes of the agents. For instance, financial services using decentralized AI can provide transparent and traceable records of transactions, ensuring compliance with regulatory standards and building trust among users.
Products leveraging mechanism design to guide and control agent interaction in a multi-agent system. Specifically, in use cases where two or more agents with diverging or contradicting goals (for example, because they represent users with different goals).
There is a system to evaluate and compare efficiency of routing between LLMs, agents, tools against the end goal as specified by the user. Such system can employ reinforcement learning techniques to predict the result routing request even before the evaluation step. WithMartian is building while Scade.pro provides a single no-code interface to test and compare various AI and LLM routing algorithms in real time.
Martian helps compare various LLMs. Lmsys is building an open source router for LLMs.
AI memory and empathy, allowing agents to record, pre- and post-process user interactions, record long-term memory graph. In the future, as every user has more and more AI applications, automatically sharing context (in a privacy-preserving and SSI-compliant form) between those will lead to more useful and delightful user experience.
AI-Human communication tools, akin to a mailbox where you, as a customer, can view all receipts and logs from completed tasks, requests for reviews and decisions, feedback, and additional information queries by AI agents.
Agentic economy refers to a system where AI agents act as autonomous economic participants, engaging in transactions, decision-making, and value creation. A tooling marketplace for AI agents would be a platform where agents can access and temporarily utilize various capabilities, tools, and functions to enhance their ability to complete certain tasks.
End game
A marketplace for agentic tools and capabilities. Any AI agent can trustlessly utilize the tool (API, ABI, function calling) while having a guarantee of it’s output.
A human-in-the-loop marketplace allows AI agents to hire humans for tasks, effectively combining AI and human capabilities. Decentralizing this ensures fair compensation and diverse human participation in AI ecosystems. Such tasks might require doing some work in real life (sending a package, taking a picture) or acquiring human judgement (rate web design options, provide human feedback for a poem written by LLM).
Global, permisionless, AI-first contracting space for autonomous agents to enter and enforce contractual relationships. This will require using:
commitment devices for AI agents to ensure compliance by the agents with the agreed terms
medium of recording and retrieving contracts for reference and auditing
cryptographic tools to implement commitments and, optionally, privacy-preservation via something like zero-knowledge proofs
building blocks for the “AI contract space”, similar to precompiles or legal templates to simplify creation of such contract.
AI-native AMMs can become tools for purchasing services and creating agreements with and between AI agents. An agent offers work that it can do according to its own supply curve, and the very fact of buying the token concludes a digital contract for its execution. The contract can include penalties for non-performance, an arbitration process, and insurance. Thanks to the trustless nature of the blockchain, you can assemble a composite and complex graph of contracts with contractors to complete a more complex user task that one model/application/agent alone cannot handle. And the task of finding the optimal configuration that will provide the highest quality but cheapest result for the user is a kind of new form of AI-MEV searching.
Just like contracts, AI agents need a way to operate and adapt to social norms and instituions. They need a system to align actions with values of the society.
This paper suggest, that to achieve it agents must learn to predict which actions may be sanctioned, guide agent decision-making to select actions that comply with community norms and learn which institutions are treated as authoritative through interactions with other agents.
AI finance which starts with simple tools for AI agents to exchange value between each other and external entities such as individual users or legal entities but potentially expands into more complex financial primitives built specifically for the agentic economy, such as credit (in a form of compute), tokenization, hedging, index and managed funds.
AI-assisted cybersecurity refers to the future of cybersecurity where both offense and defense consist of AI agents actively probing, pentesting, and trying to find new exploits. It becomes a 24/7 warfare with self-evolving tooling, such as fuzzy security and evolutionary algorithms, making the battleground constantly shifting and highly dynamic. In this environment, many billions are at stake, and the continuous arms race between offensive and defensive AI agents drives rapid advancements in cybersecurity capabilities.
Security of AI is a very complicated and hard to solve topic as new vector attacks appear on the every day basis. Below are couple of paper examples from the ICML 2024:
One of the most intriguing papers presented is titled "Stealing Part of a Production Language Model." This research introduces a novel model-stealing attack that targets black-box production language models, such as OpenAI's ChatGPT and Google's PaLM-2. The attack is designed to extract precise and nontrivial information from these models, specifically focusing on recovering the embedding projection layer.
Another notable paper is "Watermark Stealing in Large Language Models," which explores the vulnerabilities of watermarked models. Watermarking is a technique used to embed identifiable information within model outputs to trace and verify the model's usage. However, this paper reveals that even watermarked models are susceptible to attacks that can learn and replicate the watermarking rules.
There are of course other usual suspect like prompt injection, data poisoning and model tampering attacks that require reliable security and audit mechanisms.
AI-assisted governance will augment humans to achieve faster, better and more inclusive coordination. Agentic systems will be valuable for coordination tasks on different levels:
AI assistant that helps you break down, summarize and make sense of the most pressing issues in a group, company, DAO or politics
AI co-pilot will assists you in making decisions, taking into account your history, preferences and values
AI mediator will assists groups (from 2 to millions of people) to mediate conflicts, deliberate complex issues, build consensus in an impartial form
AI evaluator will help track effectiveness of governance decisions and use complex data analysis to find the optimal ones
AI representative will act as an individual democratic representative for each voter, shareholder or group member, trying its best to defend personal values and goals
finally, AI-managed organizations and even nations will emerge as a new paradigm of governance, where artificial intelligence systems take on the primary role of coordination and decision-making. Such a system will process vast amounts of data to make informed, unbiased decisions based on objective analysis rather than personal interests or political motivations. It will mplement policies and allocate resources with perfect efficiency, eliminating waste and corruption inherent in human-managed systems. And, finally, continuously adapt and optimize governance structures based on real-time cybernetic feedback.
Governance of AI, which means recursively leveraging the technologies described in the previous point but to govern the development of new AI models, agents and system.
If you’re building any of the systems, products and/or protocols listed above — reach out to us: https://cyber.fund/contact
cyber•Fund is supporting founders and researchers in decentralized AI space with startup funding, research and OSS grants.