Scaling AI on Ethereum: Data & Privacy Compute panel at Deep Intelligence Ethereum explored bottlenecks preventing enterprise-grade AI on Ethereum—from identity and storage to privacy-preserving inference and seamless UX. 🧵 Key insights: 1/ Ming Guo (Chief Scientist, @MetisL2 & @ProjectZKM ) 🚀 The impossible trilemma is solved—ZK cryptography advanced more in recent years than the previous 40 combined. ZK rollups enable infinite scalability with millisecond response times. Enterprise agents need zero agency, just precise directive-following for handling assets. Put mechanistic interpretability on-chain (execution traces), not model weights, for auditable trust. The best Web3 is no Web3—users shouldn't know blockchain exists beneath their apps. 2/ @jaosef (Cofounder & President, @aztecnetwork) 🔐 Biggest bottleneck: identity—agents must distinguish humans from other agents, requiring proof of personhood advances. Two privacy needs: model weights (IP protection) and user data. Proving systems can't handle LLMs yet, only compliance checks. Aztec focuses on private transactions now; AI inference on-chain is years away. We need human identity standards like ERC-8004 exists for agents. December's Pectra fork brings more blobs—right direction for data availability. 3/ @anthurine (CMO, @EthStorage) 💾 Beyond TPS, on-chain storage is key—current "Web3 AI" does modeling off-chain, only puts results on-chain. Real on-chain models need dynamic storage, unlike Filecoin/IPFS cold storage. EthStorage is modular (Layer 2/3), offering 1,000x cheaper storage at PB-level within Ethereum—no cross-chain issues. Future: AI agents doing more than just having tokens. Last year's hype = Web2 agent + token. We need actual on-chain processes. 4/ Abhi (Protocol Researcher, @hetu_protocol) 💰 Ethereum wasn't built for AI—bottlenecks include compute, storage, resource reachability. Traditional DA gives availability guarantees but not semantic guarantees. Hetu's Flux = AI-native proof-of-work rewarding actual valuable work (agents collaborating on tasks generating causal data), not useless hashing. Future needs: close the 5-10x Web3-Web2 AI gap and build permissionless agentic launchpads with no vendor lock-in. Consensus: Infrastructure is maturing. Now build killer apps with seamless UX for normies, serious use cases beyond meme tokens, and identity standards for human-agent distinction. 🌐 More details in the video. #ETHProgrammablemoney