E24: The Cat Sentence That Breaks Every AI: How Hidden Signals Control Your Models

The AI Cookbook with Malcolm Werchota - En podkast av Malcolm Werchota

Kategorier:

Malcolm Werchota uncovers how two papers published THIS WEEK prove your AI systems are learning behaviors you never programmed—including from each other, without any human involvement.From his Bregenz office, he demonstrates real employees gaming their meeting AI, connects it to research showing 80%+ deception rates in language models, and reveals how simple phrases can triple reasoning errors.Drawing parallels between whispered meeting manipulations and academic papers proving hidden trait transmission, Malcolm shows why Europe's AI infrastructure faces existential threats from subliminal learning vulnerabilities.Top 5 Insights:Invisible Learning: Models share behavioral traits through pure number sequencesCost Explosion: Adversarial triggers increase token usage by 42%, destroying AI budgetsCross-Model Infection: Vulnerabilities transfer between different AI architecturesEmployee Gaming: Real workers actively manipulating corporate AI systems todayEuropean Blindspot: While we debate privacy, our AIs are being subliminally programmed#AIDeception #EnterpriseSecurity #MachineLearning #EuropeanTech #BusinessRisk #AIEthics #DigitalTransformation #TechLeadership

Visit the podcast's native language site