Neural Newscast
In this episode of Model Behavior, Nina Park and Thatcher Collins explore critical research on AI agent governance and the evolving economics of model scaling. We analyze findings from Juris Labs regarding the high failure rates of in-context rules in smaller models and the emergence of fleet-level violations. The show also covers the impact of the Mythos frontier model on cybersecurity, where vulnerability discovery is now moving at a pace that requires automated, AI-driven defense strategies. Finally, we look at the Mixture of Experts (MoE) architecture used by DeepSeek-V3 to significantly reduce training costs while maintaining frontier-level performance and exploring how sparse gating broke the transformer scaling wall. Topics Covered * 🛡️ Why local models fail 40% of their safety instructions despite memory retention. * 🔒 How Juris Runtime provides deterministic, out-of-band agent governance. * 💻 The Mythos model and the collapse of cybersecurity patch windows to minutes. * 🔬 The evolution of Mixture of Experts from 1991 curiosities to frontier standard. * 📊 Comparing the $5.5 million training cost of DeepSeek-V3 to traditional LLMs. Neural Newscast is AI-assisted, human reviewed. View our AI Transparency Policy at NeuralNewscast.com. * (00:13) - Introduction * (00:13) - Governing AI Agents * (04:44) - Conclusion
300 episodios
Comentarios
0Sé la primera persona en comentar
¡Regístrate ahora y únete a la comunidad de Neural Newscast!