Insomnia Labs reposted this
Privacy VS Quality In AI❓ I've spent the week trying to build out a framework for how to leverage blockchain technology to secure user's knowledge, instructions, and data when creating AI Agents. The solution I've come to is outlined below. But... the major issue I see is a trade-off between quality and privacy. In order to create an AI agent (Or GPT), a user must provide custom information to inform/instruct the output of a LLM. This could be SOPs that took years to refine, sensitive documents that include PII of clients they have worked with in the past, or stories with a highly specific tone of voice/style; information that would be objectively sensitive for the Agent creator. In order to fix this, per the diagram below, I suggest running a rule engine that would anonymize the data. A good way to imagine this is if a chef uploaded his cookbook, the rule engine would take out instructions on how to cut specific ingredients. Before Rule Engine: Mince the onion and try to maintain a cubbed shape After Rule Engine: Cut the onion into small pieces. But then, some questions are raised, "Doesn't that make the agent less effective?" or "By removing what makes the 'difference' in a dish (Taking the chef example) doesn't that inherently make the output of the agent worse?" The only solution I can think of is to provide a temperature for privacy. Just like how we can currently provide a numeric value to a custom LLM that can stop it from outputting answers when it doesn't know the true solution, can we create a temperature for privacy via the rule engine? Allowing people to decide how anonymized they want their data. I think Blockchain and AI have a role together. BUT just like most on-chain components, it comes with a trade-off. If you read this far, let me know your thoughts, curious to know 💖
Director of Partnerships & Growth @ Insomnia Labs
1moDidn’t make it to the end, but support <3