AI + a16z

Joel de la Garza engages with Dylan Airy from Trunkle Security and Firas Abukudijay from Socket to unpack alarming trends in AI-driven cybersecurity threats. They reveal how advanced AI models are not just identifying vulnerabilities but actively exploiting them, with software supply chains becoming increasingly susceptible to attacks. The conversation also touches on the moral responsibilities of AI labs in mitigating the risks they create and the urgent need for organizations to adapt their security strategies in an AI-first world. For consultants: Understanding these evolving threats is vital for advising organizations on how to implement robust self-management and distributed authority in their cybersecurity practices.

Simon Mo, co-founder of InferRack, emphasizes the pivotal role of VLOM as an open-source inference engine that powers a growing number of AI applications. He explains how VLOM addresses the unique challenges of serving large language models, distinguishing itself from traditional machine learning workloads. The conversation also touches on the transition from open-source curiosity to critical infrastructure, highlighting how companies are increasingly relying on open-weight models to innovate beyond closed-source solutions. For consultants: Understanding the evolution and infrastructure of open-source AI models like VLOM is vital for guiding organizations in adopting self-management and distributed authority principles in their tech strategies.

This episode explores the concept of self-accelerating AI and its potential to revolutionize scientific research by enabling AI systems to improve their own capabilities. The co-founders of Mirndl discuss how their technology can streamline the process of scientific discovery, reducing the need for large teams and resources by allowing AI to conduct research and engineering autonomously. For consultants: Understanding the implications of self-accelerating AI is crucial for guiding organizations in adapting their structures and practices to leverage these disruptive technologies effectively.

Mohammad Norouzi, Ideogram's CEO, explains why his company released an open-weight image model and why controllability has become central to generative design, moving beyond generic prompts toward systems that let creators specify exactly what they want. The conversation covers the technical challenges of generating readable text and layouts within images, the tradeoffs between universal models and tools optimized for specific creative tasks, and how agentic workflows might reshape design collaboration. Norouzi argues that creators and enterprises will increasingly demand agency over their outputs rather than accepting what a general-purpose model decides. For consultants: Norouzi's insistence on user control and transparency reflects the Semco principle that decision-making power belongs closest to those doing the work, here, designers wielding real authority over their tools rather than serving systems built to constrain their choices.

Will Bryk built Exa on a single insight: traditional search engines were designed for humans reading results, not machines making decisions. When the user becomes an autonomous agent, retrieval shifts from information delivery to infrastructure. The conversation maps this change across coding agents, data access patterns, and why search may become foundational for the agent economy. For consultants: Building organizations around distributed authority means designing systems where autonomous agents operate with minimal human oversight. How those agents access and act on information is no longer a technology detail. It's a foundational design problem for your organizational model.