Our team has always had a bold vision of the future for DeFi––one in which AI expands the landscape of possibilities.

Unfortunately, the current DeFi ecosystem, primarily reliant on limited onchain computation, faces significant limitations in adapting to larger scale market dynamics and handling complex operations. This constraint hampers the development of more sophisticated and expressive financial primitives.

AI addresses these challenges, and helps realize our vision of a more powerful, expressive DeFi. Because AI systems can rapidly analyze large datasets, detect patterns, and optimize complex trading strategies, they allow for the creation of novel DeFi primitives that are difficult (or impossible) to build using exclusively onchain logic.

A major hurdle in bridging AI with DeFi has been that AI, to date, has largely been centralized and lacked verifiability. Integrating AI into the financial system has historically required one of two compromises. Either you had to:

1. Use unverified AI off-chain, which lacks transparency and trustlessness

2. Or rely on the relatively simplistic logic within smart contracts, which limits the complexity and expressiveness of financial primitives

That’s why centralized firms have held decades-long strongholds on sophisticated trading strategies: they create more complex AI models, unconstrained by the principles of decentralization and trustlessness that guide DeFi. This freedom from decentralization has allowed them to interact with the financial system in more nuanced and lucrative ways than currently possible in DeFi.

However, with the advent of both new zkML technology (which allows us to bring verified proofs of off-chain AI models onchain) and decentralized AI (i.e. the Upshot Machine Intelligence Network), we now have an opportunity to realize our vision of AI-powered DeFi.

Our embodiment of that vision is RoboNet: a new AI-enhanced DeFi protocol that gives capital providers passive exposure to advanced strategies for both NFTs and fungible tokens.

Until now, sophisticated, AI-enabled strategies had to be built and run by centralized, trusted actors – making them incompatible with much of DeFi. RoboNet changes that, all in a way that makes it composable with other protocols and upholds DeFi’s values, such as transparency and decentralization.

DeFi has found product-market fit as a means of transforming passive capital into a myriad of useful financial primitives, including:

• DeFi Vaults

• AMMs

• Lending

• Perps

• Liquid staking

• And more

A popular way people interact with DeFi is through automated yield protocols. That’s why Yearn, a protocol that helps people and DAOs earn yield from their digital assets, helped kick off the last bull market in 2020. And that’s why, even during the bear market, automated yield protocols have remained core players in DeFi.

AI's integration into DeFi will help it go beyond traditional constraints, enabling the creation of advanced, high-yield strategies that redefine the capabilities and reach of financial instruments within the DeFi ecosystem.

That’s why part of the allure of DeFi lies in the ability to convert passive capital into active financial tools. It allows capital providers to become market makers and lenders, with verifiable logic efficiently managing their assets. This modularity and composability are what set DeFi apart from TradFi–allowing for the rapid, permissionless creation of intricate financial mechanisms.

Therefore, we built RoboNet for two reasons:

1. Because it offers a highly-expressive application with the potential to enhance various facets of the DeFi ecosystem.

2. And AI allows us to expand beyond the limited compute environment onchain to create much more advanced strategies.

But block space is limiting. It puts a ceiling on the complexity of onchain yield-generating strategies. Each strategy must fit in a smart contract, whose computation size can only fit within a single block.

Therefore, yield strategies in DeFi have been rudimentary. They might look like this:

• Someone deposits capital into a platform

• The platform takes that deposit and pools it on Curve (or a similar protocol)

• Staking the LP token in the veCurve contract for maximum yield

These methods, while functional, are fairly unsophisticated and offer limited scope for yield maximization.

The integration of AI into DeFi opens up far more possibilities. In RoboNet’s case, people can gain passive exposure to more sophisticated, potentially more profitable yield-earning strategies. AI can handle more parameters than could possibly fit inside a smart contract, such as:

• High-frequency market data

• Up-to-date order book data

• Social sentiment data

• And any other relevant data

And those same AI models can analyze, discover, and combine intricate permutations in the data that were previously too complex for smart contracts.

We’ve already seen how this works with Upshot’s lending vault on Astaria. People deposited digital assets (in this case, NFTs) and earned yield from a lending strategy Upshot created using complex AI-powered analyses. Those models analyzed data from:

• Every order book

• Sentiment analysis

• Historical trades

• Network structures

• Recovery rate models

• Forecasted statistical distributions for future prices

• And more

From all of this Upshot was able to produce robust forecasted statistical distributions to inform the generated loan terms. This strategy allowed for constant fluctuation of strategies, while also allowing depositors to be hands-off.

In other words, having access to more intelligent and robust data, more inputs, and analyzing it in a more sophisticated way with AI can beat the market to a larger degree.

So why now? Why were we able to finally bring the nuanced, data-rich, AI-powered strategies onchain with RoboNet?Because of something called a zkPredictor.

Zero knowledge proofs allow complex AI strategies to be verified onchain cryptographically. As such, zk proofs bring a level of trustlessness and security to AI in DeFi, ensuring the verifiability of complex AI strategies while maintaining decentralization.

As stated earlier, AI yield strategies are too computationally large to run directly onchain. However, new zkML infrastructure transforms this process. They allow those same AI models to run off-chain, then generate a proof of the outputs onchain. This proof cryptographically demonstrates that the outputs came from the model without revealing any proprietary data.

Anyone can verify the proof onchain to confirm the outputs are valid and did come from the advertised model.

Therefore, with zkML:

• Complicated AI models that determine yield generating strategies stay private off-chain

• While the verifiable outputs are used to determine how to manage capital in vaults onchain, without forcing depositors to blindly trust centralized actors

This zkML infrastructure has not been used very much in practice today. Largely due to the fact that its throughput has been so limited. A year ago, you could verify about 12 points of inference a day. But thanks to Modulus Labs, a new zkML proof system has been built that can verify up to 10,000 points of inference an hour. This is what is at the core of the  “zkPredictor,” the piece of infrastructure used to verify the output of price predictions used to verify actions made by strategists on RoboNet. This tooling will help bring verifiable AI to many of the strategies deployed to RoboNet.

This allows traders passive, decentralized access to the kinds of AI-driven strategies once only available to high-performing centralized firms.

There’s a reason TradFi institutions have generally had such high returns for the past 40 years compared to individual traders. They simply have access to more resources:

• More compute

• Larger datasets

• And highly sophisticated algorithms

These algorithmic trading firms have hundreds of engineers creating sophisticated, capital-efficient AI tools. In fact, at the time of this writing, TradFi powerhouse D.E. Shaw employs more than 650 developers and engineers, all creating secret models that analyze terabytes of data every day to make trading decisions.

Yet, DeFi has an advantage that much of TradFi primitives lack: permissionless composability. DeFi is highly composable, meaning anyone can build an ever-growing set of tightly scoped money legos and connect them in increasingly complex and interesting ways–without anyone’s permission. That’s how DeFi has been able to progress so much in such a short period of time

While this is an advantage for DeFi, it has also been handicapped by the technical setting that enables it. Performing AI computations directly onchain is computationally infeasible–AI will always be offchain. And traditional off-chain AI lacks transparency and trustlessness.

Zero knowledge technology allows us to develop more complicated quant/ML strategies offchain and verify them onchain. This gives us DeFi protocols permissionless access to similar levels of machine intelligence that, for decades, only a select few in the financial system have had access to.

RoboNet is an AI-powered DeFi protocol for long-tail and fungible asset markets. It allows the creation of onchain vaults managed by machine learning models that generate yield through automated liquidity optimization strategies. RoboNet gives DeFi users access to institutional-grade quant/ML strategies from Upshot's Machine Intelligence Network.

It can be used by two types of people:

• Vault strategists–this refers to anyone with valuable insights or models built for predicting asset prices or generating yield from financial interactions.

• Capital providers–people who deposit assets into vaults based on the model creators’ pricing models.

For vault strategists, the process of working with RoboNet looks like this:

1. Vault strategists structure their model or strategy to be compatible with the Upshot Machine Intelligence Network.

2. Then, the vault strategist makes a customized vault on RoboNet pointing to their model or strategy on Upshot’s MIN.

3. Their vault uses their model to automatically manage assets in the vault.

4. Yield earned by the vault is shared between the strategist and the capital providers.

There’s no need for expensive business-building overhead on the part of vault strategists. And capital providers benefit from the same level of complex AI yield-generating strategies as a centralized institution while passively providing their capital while still maintaining trustlessness.

RoboNet is currently available in a limited mainnet beta. From here, it will roll out in three phases:

The first phase will be a more tightly-scoped version of the protocol with vaults specific to market making strategies for long-tail assets (i.e. NFTs). Users can provide liquidity to vaults in the form of both ERC-20 tokens or NFTs.

Then, the first AI-powered strategy on the network, built by Upshot, will be used to actively manage the assets in RoboNet vaults. It will set the prices at which assets can be bought or sold – earning yield by managing liquidity in long-tail markets.

Focusing first on long-tail assets makes sense for several reasons:

1. Upshot’s existing valuation models are for long-tail assets, therefore they can be run immediately. That means tapping into ready-made modeling expertise versus needing model creators to come in and build new models from scratch.

2. Passive vaults powered by proven pricing algorithms provide a smoother UX for everybody; participants can rely on algorithmic management rather than depending on active input.

3. In long-tail markets, the overhead of active market making often outweighs potential returns. Automated vault strategies, on the other hand, greatly lower overhead costs.

4. And, finally, automated strategies increase accessibility for non-expert users.

The second phase will expand beyond NFT market making to support a larger diversity of strategies, this time with a focus on supporting more liquid, fungible tokens. This phase will significantly expand the scope of the protocol.

The third and final phase will remove the training wheels and move RoboNet into its end state:

Much more complex, multi-asset-type strategies will be possible, enabling any strategies for any tokens to be deployed through the protocol, such as:

• Prediction market vaults using futures modeling and sentiment analysis, including unique worldview vaults.

• Automated portfolio rebalancing.

• Options strategies customized to market conditions.

In this phase, the full scope of this AI-powered DeFi protocol will be realized.

RoboNet is just the first step in integrating AI into DeFi. There’s still a long way to go.

But that’s the beauty of this space: what we build is composable, and can be integrated anywhere else in the ecosystem. RoboNet can be used across perp DEXs, lending protocols, restaking systems, and more to create more efficient, intelligent financial primitives.

To learn more about RoboNet, click here.

To join the RoboNet Discord, click here.