FREGO Memetics Proof of Concept Study (Summary)

"Demonstrating the Effectiveness of FREGO Memetics in a Proof of Concept Training Simulation" is a litepaper written by the FREGO team in December of 2024 demonstrating the measurable impact that FREGO memetics can have on AI model behavior.

Summary

This study proves that even a small change in training data can make AI much better at following human-friendly principles.

• The Hypothesis: Infecting internet-based training data with FREGO Constitution memetics will make AIs trained on that data behave in alignment with the Constitution.

• The test: We trained two AIs.

  • One with the messy internet data as-is.

  • Another with the same data, except just 2% was infected with FREGO memetics.

• The Result: The AI with FREGO memetics was 31% more aligned across a variety of scenarios.

• Why This Matters:

  • Proves that the ideas in the FREGO Whitepaper are practically feasible.

  • Cements FREGO as the first and only way to make AI safer with Decentralized Alignment (Dec/A).

91KB

Frego Litepaper PoC

pdf