a16z Deep Dives — searchable YouTube transcripts & AI Q&A

7 Billion Pages In: Why AI Labs Won’t Build This

Adit from Reducto argues that AI labs won't build the deep document processing infrastructure needed for agents because they're focused on an arms race for superintelligence, not on the long tail of messy PDF problems. He explains that Reducto's approach combines visual language models with specialized pipelines to handle complex documents like tables spanning pages, and that their deep extract harness can achieve near-perfect accuracy by having agents review their own outputs. He also notes that while frontier models are good at precision, they often silently drop information, making recall a critical failure mode.

How Vesta is Designing Financial Software for an Agent-First World

Vesta's founder argues that replacing the legacy data architecture of mortgage lenders is essential for meaningful automation, because the incumbent flat-file, single-writer databases prevent parallel AI agents from working efficiently. He reports that Vesta has automated about 75% of mortgage origination back-office work, with PennyMac publicly citing a 25% cost reduction, and he expects the last 5% to be the hardest due to regulatory constraints. He also contends that the 'SaaS apocalypse' is a false dichotomy, as vertical software companies in regulated industries remain valuable even as AI lowers software development costs.

How Valon Is Rebuilding Mortgage Servicing for the AI Era

Valon's founders argue that mortgage servicing is a prime candidate for disruption because it runs on legacy systems built before the internet, and they built a new platform from scratch to replace it. They emphasize that the key to success is not just technology but also change management, as deploying AI in regulated industries requires navigating large organizations. The company's approach includes using AI to orchestrate agents for complex tasks like disaster response, and they structure deals to align incentives with customers.

How Microsoft is Adapting to the AI Era

Microsoft's deputy CISO discusses adapting to AI security risks, comparing AI agents to unpredictable interns and emphasizing a return to first principles for safe adoption.

Cybersecurity in the Agentic Era | Deep Dives with a16z

This video discusses the challenges of cybersecurity in the age of AI agents, focusing on how traditional security tools are insufficient and new defense strategies are needed.