How to Keep Novel Research Out of an AI's Training Data: Uneasy Money

Summary

The clip is a discussion about AI research privacy and model choice rather than a financial market episode. Kain Warwick, Taylor Monahan, Jon, and Alex Thorn debate using frontier versus open-weight models, the risks of training-data leakage, and why fast takeoff has not yet occurred. They also discuss private inference, custom harnesses, and human competitive incentives as the bigger risk. No concrete investable ideas are presented.

• Kain asks whether a mathematician should use a frontier model or an open-weight model for novel proofs.
• The panel doubts the reliability of 'don't train on my data' toggles and warns that novel IP can be hoovered into training data.
• Jon says open-weight models are closing the gap and private inference via Venice can protect data while delivering near-frontier capability.
• Kain argues fast takeoff has not happened because models lack long-horizon continuity and context resets.
• Taylor says humans, not robots, are the greater destructive risk, pointing to OpenAI's competitive behavior.
• The group explores custom harnesses and multi-model aggregation as a privacy-preserving way to outperform general assistants.
• No specific public securities, tickers, or trade recommendations are made in the conversation.
• A sponsor segment promotes 1inch Aqua for shared liquidity and capital efficiency.