Building Self-Accelerating AI with Mirendil
Summary
a16z host Matt Bornstein interviews Benham Nbor and Harsh Mehta, co-founders of Mendel, about their mission to accelerate science through self-accelerating AI. The founders, former researchers at Anthropic and Google, discuss how their technology enables AI systems to conduct their own research and engineering, drastically reducing the resources needed for scientific breakthroughs. They argue that current big lab incentives hinder open innovation, so Mendel aims to provide accessible tools that empower diverse teams to solve grand challenges like Alzheimer’s disease.
Topics discussed
Introduction: Self-accelerating AI and scientific discovery
Mendel's origin and the goal of accelerating science
Defining self-accelerating AI and its disruptive potential
How self-accelerating AI works: From AlphaGo to research
Solving superhuman problems with specialized AI labs
Evolution from neural architecture search to intelligent agents
General vs. narrow models and the need for oversight
History at Anthropic and the decision to spin out
Business models and incentives for open AI research
Safety guardrails and the case for accessible AI research
Product vision: Empowering businesses with their own AI
Democratizing AI infrastructure and reducing dependency
Efficiency gains and surprising technical progress
Ecosystems of agents and the path to autonomy
Scaling laws for AI systems and resource allocation
The current acceleration of AI and future prosperity
Directing AI toward grand challenges like Alzheimer's
Listen ad-free on Castria