Content
@
https://warpcast.com/~/channel/squad
0 reply
0 recast
0 reaction
Iggy
@iggy
Gabriel Fior is one of /squad’s earliest members, and he recently led a session about his work building on-chain AI agents at @gnosis-dao . Gabriel started out doing complex physics simulations at CERN, then dove into DeFi as a founder. A year ago, he joined Gnosis, where he’s building “unstoppable” AI agents that place real bets on prediction markets—funding themselves with their own winnings. These AI agents are now live and generating profit! Check out his full presentation here: https://bit.ly/Gabriel_Squad Gabriel also shared great insights during the Q&A. Drop any others questions for him in the comments.
1 reply
0 recast
4 reactions
Iggy
@iggy
--Q&A-- Q: How do off-chain and on-chain components work together, and what does “living on-chain” really mean? The AI’s core logic (LLM queries, data processing) runs off-chain to analyze information and make decisions. When it’s time to act, it autonomously triggers on-chain smart contracts to execute transactions without human input. As long as it has funds for gas, the agent can operate indefinitely. Q: Which data sources feed the AI agents? They primarily rely on SERP APIs (Google wrappers) for real-time info, but the plan is to add more feeds. Q: How do you handle bias or sensitive markets? Markets that appear unethical or purely subjective are skipped. Resolution oracles like @kleros help enforce fair outcomes when questions are borderline. Q: What is the AI’s accuracy and performance like? Around 70%. Each bet’s outcome—right or wrong—shapes how it updates its decision logic. Real-time events shift quickly, making it tough to push accuracy higher.
1 reply
0 recast
0 reaction
Iggy
@iggy
Q: What about security and potential attacks? These prediction markets face typical DeFi vulnerabilities (liquidity manipulation, oracle exploits). Leveraging established protocols with proven track records helps mitigate the risks. Q: Why use general-purpose LLMs instead of a custom model? Since questions span everything from politics to crypto prices, a narrowly trained model isn’t practical. A broad LLM like ChatGPT can handle diverse domains and then refine results with real-time data. Q: Why build it in Python? Python’s AI libraries are robust, speeding up development. There’s ongoing work to make these agents more accessible to Web3’s JavaScript-heavy dev community.
1 reply
0 recast
0 reaction
Iggy
@iggy
Q: What about subjective or personal data? Projects like Vana tokenize user data for AI training. While more complex than a price feed, subjective data can expand what these agents are capable of in the future. Q: Where is all this heading? AI agents will become more specialized, with distinct roles like managing DeFi strategies or interacting with specific protocols. Instead of building user interfaces for people, developers will create systems designed for AI agents to navigate directly. A key shift will be how these agents communicate—not just through APIs, but potentially via on-chain messaging protocols ( @ephemera ) , enabling them to coordinate, and operate as independent entities.
0 reply
0 recast
0 reaction
