The real AI native step isn't just calling centralized APIs, it's in training and running models under blockchain rules so that no single part will say that they own the brain.
AI and machine learning can help ensure that it's easier to detect on-chain DeFi attacks. A 2025 AFT paper proposes a fully on-chain learning framework. This means that there would be heavy AI training running on L2, and updates pushed back to Ethereum L1. In addition smart contracts will do gas bounded inference before approving a. DeFi transaction. AI can also introduce a Proof of Improvement mechanism in which a model update is only accepted on-chain if it provably improves accuracy on a public test set. This means that bad updates will be slashed financially. Models are also optimized to fit inside Ethereum’s block gas limit and yet match off-chain accuracy. Ethereum has a decentralized security operations centre where the model itself is governed by consensus.
Bittensor on the other hand runs a decentralized market place where ML models compete to answer queries and get paid in TAO based on usefulness. This is a kind of proof of intelligence mechanism. In 2025, DeFAI write ups highlight Bittensor as the backbone of AI driven DeFi. Models specialise in trading, risk or analysis. Then their signals feed into DeFi strategies via EVM compatibility and dTAO delegation.
Instead of one company owning the trading algorithm, the network owns and trains it. And Ethereum becomes both the ledger and the AI coordination layer.
