DeFi 3.0时代，AI如何靠预测信息“收割”市场？

Author: 0xJeff , Crypto KOL

Compiled by: Felix, PANews

Prediction has been a core ability of human evolution - since ancient times, humans have relied on their senses and instincts to predict threats and opportunities in the environment, including detecting predator activity patterns, the chance of prey appearing, and seasonal food availability, all of which are critical to survival.

Since then, this predictive model has evolved into the use of tools and planning (such as predicting the needs of growing crops, slaughtering and preserving meat), predicting social cues (intentions, emotions, behaviors), and the development of writing, science, mathematics, and modern tools such as statistics, computers, machine learning, and artificial intelligence, all of which are used to enhance human predictive ability.

Prediction markets in particular have evolved into an economic tool — one that harnesses the human ability to predict economic, political, and cultural outcomes. Unlike traditional opinion polls, prediction markets like Polymarket and Kalshi leverage economic incentives for accurate predictions, as participants stake real money.

Polymarket attracted nearly $4 billion in bets on the 2024 U.S. election market, outperforming even opinion polls in predicting a Trump victory, reflecting the economic value of crowdsourced predictions.

The same evolution applies to spot and perpetual contract trading, from the rise of CEXs to meet the growing demand for cryptocurrencies around the world to the recent disruptive development of Hyperliquid, which provides self-custody and KYC-free services with a CEX-like trading experience.

Prediction is a core capability of human evolution, and with the rise of artificial intelligence/machine learning predictive models, the ability to predict events, asset prices and volatility is improving dramatically.

This takes humanity to the next stage of evolution.

DeFi 3.0

DeFi 1.0 introduced smart contracts and decentralized applications, allowing anyone to transfer, buy, sell, pledge, lend, and mine yield anytime, anywhere. In essence, it is to put crypto assets into on-chain operations to create economic value, such as Uniswap, AAVE, Compound, Curve, Yearn, and Maker.

DeFi 2.0 expands on 1.0 and introduces novel token economics and incentive distribution mechanisms designed to coordinate the interests of different stakeholders in the protocol (such as Olympus/Wonderland, Solidly/Aerodrome), and has spawned emerging markets that provide alternative sources of income (such as Maple, Pendle, Ethena, Ondo, Clearpool, Solv, USDai, etc.).

DeFi 3.0 introduces artificial intelligence into DeFi. Some people call it DeFAI, while others call it AiFi. It means integrating large language models (LLM) and/or machine learning models (ML) into DeFi products.

From simple LLM integrations (acting as customer support/co-pilot to help users navigate the protocol), to multi-agent/swarm and machine learning systems that fundamentally improve the product (increase trading profits, reduce impermanent loss, increase LP returns, reduce liquidation risk for perpetual trading, etc.).

In addition to the DeFAI abstraction layer and fully autonomous financial agents, today we will discuss the role of AI/ML systems and predictive models in transforming DeFi and other verticals.

Prediction System

Neural networks and decision trees have been around since the 2000s, and these systems were used by hedge funds to predict stock and commodity prices. Early stock prediction results were quite promising, with short-term predictions reaching 50% - 60% accuracy, but overfitting and limited data limited their application.

Then came the rise of deep learning and big data, which enabled models to process larger data sets (time series data, unstructured data such as news and social media), leading to more accurate predictions and wider applications.

Breakthrough developments have occurred in the past five years, with Transformer models and multimodal AI integrating more diverse datasets such as Twitter sentiment, blockchain transactions, oracles, real-time news, crowdsourced predictions (Polymarket, Kalshi), and more. This has enabled some AI models to achieve 80% - 90% accuracy in predicting event outcomes and asset prices.

As these models continue to improve, the demand for integrating predictive capabilities into DeFi systems has increased significantly. We are currently in the early stages of DeFi 3.0 and are witnessing some players in the market combining AI/machine learning systems with Web3 application scenarios in real time.

DeFi x AI/ML system

Allora

Allora is probably the most widely used decentralized prediction model network currently. Allora has achieved numerous integrations with DeFi protocols and AI agent teams, giving it prediction capabilities (mainly focusing on cryptocurrency price predictions such as BTC, ETH, SOL).

Its short-term cryptocurrency price predictions are said to be around 80% accurate.

Some key applications include:

• Vectis Finance is an AI-driven USDC-based vault that uses Allora’s reasoning technology to maximize SOL trading returns. Its cumulative return since April 23 is 2.4%, with an annual interest rate of approximately 10%.
• Steer Protocol’s AI LP Vault uses Allora’s predicted price data to better put liquidity ahead of price fluctuations, thereby avoiding impermanent loss.
• Allora works with many teams including Cod3x, Axal, Brahma, Virtuals Protocol, etc. to support trading strategies and execution of AI agents.

Bittensor Subnet

Since Bittensor’s dTAO incentive distribution mechanism can help startups (subnets) offset development costs, the team uses Bittensor to launch its product development and outsources a lot of development work to miners. The higher the incentive, the better the quality of the miners.

Given that machine learning models and prediction systems are one of the easiest tasks to quantify (building a model that can accurately predict something), this is one of the most common verticals that subnets focus on.

Subnetwork focused on prediction

• SN6 @Playinfgames
• SN8 @taoshiio
• SN18 @zeussubnet
• SN41 @sportstensor
• SN44 @webuildscore
• SN50 @SynthdataCo

Since SN6, SN18, SN41, and SN44 have been introduced in detail before, these sub-networks will be skipped, but I still want to emphasize again:

➔ SN6’s @aion5100 (SN6’s AI agent/predictive hedge fund layer) is about to launch a DeFi vault that automatically allocates user deposits to bet on high-confidence events/markets. The vault is launching soon, and early testing is reportedly giving over four-digit APYs.

➔ SN44’s @thedkingdao continues to see improved signals in football/soccer. Recent Club World Cup performance showed aggressive bet sizing leading to a 232% ROI. The team is also working on a DeFi vault product that will take a more risk-adjusted approach.

The AI agents/tokens representing these two application layers on CreatorBid did an excellent job of demonstrating the capabilities of the SN6 and SN44 intelligence. This inspired many other subnet teams to follow suit and launch AI agent tokens to demonstrate the capabilities of their subnets.

➔ SN50 Synth is particularly interesting. This subnet is built around a highly general volatility prediction model. It can be used to cover a wide range of probabilities of what prices could happen (not just predicting future prices), such as predicting liquidation probabilities, lifespan/liquidation times for perpetual positions, setting Univ3 LP ranges and predicting impermanent loss, predicting option strike prices and expiration times within a window, etc.

• Synth is claimed to outperform the traditional baseline model (geometric Brownian motion) by 25% - 30%.

There is a huge demand for L1/L2 ecosystems that want to integrate such engines into their DeFi ecosystem.

So far, Synth has been integrated with the following platforms:

• Arbitrum, supports AI trader competition
• Chainrisk, understanding volatility so partner agreements can better handle sharp changes in volatility
• Solana is the last major liquidity staking protocol for an unknown use case (according to the team, the official announcement will be released in 1-2 days)

The team positions Mode L2 (their own L2) as the application layer, enabling traders to use Synth to predict asset prices and trade better by combining Synth inference with the Mode AI Terminal + Mode Perp product.

What makes SN6, SN44, SN50, and many other subnets so interesting is that they offer miners incentives ranging from $2 million to over $10 million in dTAO tokens per year to continually improve their prediction models.

The goal is to use dTAO incentives as capital expenditures to guide product development and achieve commercialization/productization as soon as possible, thereby earning real returns and offsetting the selling pressure of dTAO. Some of these subnets have begun to move towards the commercialization stage (as evidenced by DKING’s $300 million deployment support for a top sports hedge fund).

What will happen next?

The pursuit of higher returns and lower risks will continue, prompting builders to bring more RWAs on-chain. Existing DeFi revenue sources will continue to be optimized and become increasingly accessible.

Prediction markets will become the main source of information, AI will act as market makers, and experienced participants will further stimulate the wisdom of the crowd. Tools are becoming smarter and models are becoming more accurate, and some results have already been seen.

The more these systems learn, the more valuable they become. And the more composable they become with the rest of Web3, the more unstoppable the whole thing will become.

The point here is… at the end of the day, everything in crypto is a bet on the future.

Therefore, infrastructure and applications/agents that can see the future even slightly more clearly—whether through collective wisdom, better data, or more accurate models—will have a significant advantage.

Related reading: IOSG: Exploring the prediction market and its competitive landscape through Kalshi