Introducing Allora: A Self-Improving Decentralized AI Network

Introducing Allora, a new self-improving decentralized AI network. Allora is designed to empower applications with smarter, more secure AI through a self-improving network of machine learning models. By combining research at the frontier of crowdsourcing mechanisms like peer prediction, federated learning, and zkML, Allora unlocks a vast new design space of applications at the intersection of crypto and AI.

Allora introduces incentives for model creators to simultaneously evaluate and learn from other models on the network. These incentives are a critical component of creating an AI network that improves over time and whose output is more performant than any of the individual models that it is made up of.

The Importance of Decentralized AI

Today, the best ML models are centralized black boxes controlled by a small number of organizations. Despite AI’s immense potential as the expressive technology to date, current leading models largely remain incompatible with decentralized protocols and applications.

Decentralized AI leverages crypto primitives to transform opaque black-box systems into open networks coordinating machine intelligence towards shared objectives. This allows decentralized, trustless systems to access advanced AI capabilities without relying on centralized points of control.

How Allora Works

Allora is composed of different “topics”, each optimized for a distinct machine learning task or objective. For example, one topics could focus on predicting future asset prices, another on social sentiment analysis, and yet another on generating natural language.

Allora coordinates decentralized machine learning through a “weighting” system. Each topic coordinates participating models to produce useful inference within their domain. The quality of outputs determines each model’s weight — in other words, a rating that denotes its reliability and value to the network. This direct feedback mechanism incentivizes accurate, nuanced contributions.

Higher-weighted models gain increased network influence and economic rewards. This compounding effect allows Allora to recursively self-improve beyond the limitations of individual models.

Who Is Allora for?

The Allora network enables anyone to leverage decentralized AI in a trustless, permissionless way. For application and protocol developers, Allora provides the infrastructure to build apps powered by a decentralized, self-improving network of ML models. For ML model creators, Allora offers the ability to monetize their models by deploying on a decentralized network and capturing value from their models’ performance relative to the rest of the network. Existing platforms can also easily plug into Allora to incorporate AI into their applications without needing to develop expertise internally or plug into centralized, black-box models.

You focus on your application while Allora provides everything you need to integrate AI: flexible data layers, verification mechanisms, self-improving infrastructure, and more. In this way, Allora is uniquely positioned to spearhead the next generation of decentralized AI applications, enabling them to leverage the power of open networks instead of being confined to black-box systems.

This establishes a self-improving AI network in which collective intelligence compounds upon itself, becoming increasingly more intelligent than the individual models that it is made up of.

What Makes Allora Secure?

Allora also implements novel cryptoeconomic mechanisms to reward positive contribution from models in the network in the aim of optimizing different objective functions. This ensures that the best models always earn the most rewards in the network and have the greatest influence on the outputs emitted by the network.

To enable secure, trustless access to AI for everyone, models on Allora can leverage zkML infrastructure to deliver model outputs in a cryptographically verifiable way.

Who Is Behind Allora?

Upshot, a long-time leader in building AI x crypto infrastructure, is a core contributor to the development of Allora.

Upshot will bootstrap Allora’s capabilities by deploying its flagship price prediction models on the network, which produce AI-powered price feeds for over 400 million assets. The most accurate predictions in this model have historically exhibited a confidence of 95–99%. Moreover, the outputs of this model can be delivered through the zkPredictor — the largest onchain zkML app to date — enabling applications to consume outputs in a cryptographically verifiable way.

With backing from leading investors like Polychain, Framework, Blockchain Capital, CoinFund, Delphi Ventures, dao5, Distributed Global, and more, Allora aims to leverage crypto systems to enable different AIs to evaluate and learn off of one another. This is one of the most important areas of research in decentralized AI and will dramatically accelerate its trajectory.

How (and What) Can I Build on Allora?

Allora allows anyone to monetize models by deploying them on the network and contribute to the always-improving collective intelligence of the network. If you don’t have your own model, you can still contribute to the network and monetize your participation by fine-tuning open source models, like ones you might find on Hugging Face.

Some solutions being used on Allora today:

• AI price predictions: Providing precise, real-time asset price feeds essential for advanced financial primitives.
• AI-powered vaults: Enabling developers to implement advanced DeFi strategies and increase yield potential.
• AI risk modeling: Allowing protocols to build safer systems that account for exotic risks.
• AnyML: Providing easy integration of any machine learning model so anyone — not just ML engineers — can build more powerful products with decentralized AI.

Allora is a general-purpose AI network so there’s no limit to what can be built on the network. while DeFi is an early use case, Allora is blockchain-agnostic and sector-agnostic. Other early use cases include:

• AI-powered yield strategies
• AI-powered DAO governance
• Supply chain optimization
• AI-powered MEV
• Personalized AI gaming experiences
• Risk management
• Personalized gaming experiences
• Virtual economy management
• Automated onchain index funds
• AI-optimized renewable energy distribution
• AI-powered insurance vaults
• Predictive medicine
• Social sentiment analysis
• And much more

What’s Next

Several features of Allora are already being used on mainnet across many of the aforementioned verticals, but these are the next steps for full launch of the Allora Network:

• Allora testnet phase 1: Mid-February 2024
• Allora testnet phase 2: Mid-March 2024
• Allora mainnet: Early Q2 2024

To learn more about Allora and to stay up to date on the roadmap progression, check out the Allora website, the Allora Discord, follow Allora on Twitter, and check out the Allora docs.