Eric Vishria, General Partner at Benchmark, and Sergiy Nesterenko, Founder and CEO of Quilter, join Nathan Labenz to discuss Quilter's groundbreaking use of reinforcement learning to automate integrated circuit board designs.
They delve into the importance of thinking beyond 'co-pilots' to fully automated AI solutions, and explore the balance of research and engineering in the AI space.
TIMESTAMPS:
(00:30) Introduction
(01:07) Eric's Investment Thesis
(03:00) The "AI and Done" Approach
(06:50) Diverging from Large Language Models
(09:15) The "Idea Maze" Framework
(11:54) Disruptive Innovation
(12:47) Sponsor: Brave | Babbel
(15:09) Exploring the Data Landscape
(16:18) Data Limitations and Reinforcement Learning
(18:00) Overcoming the Sparse Reward Problem
(19:10) Human Heuristics and Intermediate Rewards
(21:06) Extrapolation vs. Interpolation
(24:37) Trusting the Reward Signal in Circuit Board Design
(25:46) Physics as an Unambiguous Oracle
(27:05) User-Defined Search Depthbility to adjust the search depth and iteration of the system, enabling flexibility based on design complexity and optimization goals.
(29:39) Compute Allocation
(31:47) Different Physics Simulation Techniques
(33:31) The Diffusion Model Approach
(36:00) Advantages of Reinforcement Learning
(38:00) Timeline and Bottlenecks
(39:37) Unsolvable Problems and Gradual Progress
(41:58) The Future of Circuit Board Design
(43:29) Benchmark Portfolio
(45:00) Challenges and Opportunities in Each Category
(51:00) The Gap Between Research and Engineering
(53:00) Call for Startups
(54:06) Wrap
What should we build next?
