Investing in AI for Hard Tech, with Eric Vishria of Benchmark and Sergiy Nesterenko of Quilter

Episode Summary

Executive Summary: The episode examines how Benchmark evaluates AI startups and why Quilter’s AI-driven circuit board layout stands out as a durable, “AI-and-done” problem rather than a co-pilot use case. Eric Vishria argues that the best AI investments ride structural shifts and solve enduring problems, while Sergei Nesterenko explains Quilter’s reinforcement-learning approach, physics-based validation, and phased roadmap toward superhuman PCB design.

Main Topics: Why Benchmark backed Quilter (Priority: 5/5): Eric Vishria explains that Quilter fits a major world shift: electronics are spreading into everything, while board layout remains slow and manual. He was especially drawn to the idea that this is not a co-pilot problem but a full automation problem. Circuit board design as an AI-and-done problem (Priority: 5/5): The discussion frames PCB layout as work that should be fully automated by AI rather than merely assisted by humans, because the final 10% of dense layouts is especially hard for people to finish without breaking constraints. Reinforcement learning over LLM-style supervised learning (Priority: 5/5): Sergei argues Quilter does not rely on large open datasets or language-model patterns. Instead it uses reinforcement learning, physics simulators, heuristics, and self-play-like iteration to improve beyond human performance. Sparse rewards, physics, and evaluation (Priority: 4/5): The team discusses how Quilter handles sparse rewards by decomposing board design into intermediate steps and using fast, conservative physics-based proxies, with full numerical simulation reserved for validation. AI investing strategy and moats (Priority: 5/5): Vishria divides AI startups into foundation models, infrastructure, and vertical applications, arguing that foundation models depreciate rapidly, infrastructure may be temporary, and vertical apps need strong durability to avoid easy copying. Benchmark’s idea-maze and founder depth (Priority: 4/5): Vishria emphasizes that great founders have lived with a problem long enough to have tested dead ends and refined their thinking, which he sees in Sergei’s SpaceX background and repeated exploration of solution paths. Future of electronics tooling and abstraction (Priority: 4/5): The conversation looks ahead to a world where layout compilers enable higher-level workflows for electronics, analogous to how compilers and programming abstractions evolved in software.

Key Arguments: Electronics are increasingly embedded in everyday products, expanding the market for circuit-board automation. PCB layout is so manual and constraint-heavy that a co-pilot is insufficient; the right solution is full AI automation. Quilter cannot depend on abundant open-source board data because much of the best data is proprietary and, more importantly, human-designed boards are often electrically flawed. Reinforcement learning is better suited than supervised learning because it can optimize toward superhuman outcomes rather than reproducing average human performance. The reward signal can be trusted in Quilter’s domain because physics is deterministic and computable, unlike subjective human feedback. Sparse reward is mitigated by breaking board design into stages and using heuristics and approximations as intermediate signals. The compute burden is shifted toward evaluation and validation, especially physics simulation, while model sizes remain much smaller than LLMs. Benchmark sees AI opportunities in three buckets: foundation models, infrastructure, and vertical applications. Foundation models are described as extremely capital-intensive and rapidly depreciating assets, making them a risky venture category. Vertical AI can scale quickly but may be vulnerable to commoditization unless it has a durable moat. Great AI founders tend to have deep “idea maze” experience, repeatedly stress-testing assumptions before building. The long-term vision is a compiler-like stack for electronics, starting with layout and eventually moving upward toward easier schematic-level design.

Data Points: Benchmark partner meetings: 4–5 AI startups per week - Vishria says Benchmark partners are constantly meeting new AI startups. Initial board complexity target: Simpler boards first; not a super-complicated iPhone A14-class board at launch - Quilter starts with easier designs and works upward. Typical board design time: 2–4 months for a complicated board - Used to illustrate why human designers accumulate margin-on-margin errors and why faster iteration matters. Board failure rate: 80% faulty in some way - Sergei says humans are not good at judging board correctness, citing high fault rates. Production evaluation latency: Within a few minutes to an hour for some designs - Ideal fast-return mode when users want a quick faithful layout. Overnight turnaround: Within 12 hours - Current working constraint for Quilter: upload by end of day, review result next morning. High-value search horizon: A month - For very high-volume products like an iPhone motherboard, the system might search much longer to save per-unit costs. Board current limit: Up to 4 amps - Example of an initial physics/board class Quilter is targeting before expanding into harder domains. Simulation cost for advanced boards: A few GPU days per board - Estimated cost for validation runs on more complex high-speed digital boards. Foundation model cost example: $150 million - Vishria’s example of the cost to build a model that can depreciate quickly. Foundation model follow-on cost example: $5 million - Vishria says a comparable model could potentially be built later for far less, illustrating rapid depreciation. Time to copy a vertical app: 3–6 months - Used to explain why vertical AI can be quickly copied and may lack durable moats.

Pivotal Quotes: "this isn't a co-pilot problem, it's just not for co-pilots" — Eric Vishria: He explains why PCB layout should be fully automated rather than human-assisted. "AI should just do it, do it entirely" — Eric Vishria: Vishria describes Quilter’s full-automation thesis for circuit board layout. "we don't have to negotiate with humans" — Sergei Nesterenko: He contrasts trustworthy physics-based validation with subjective human feedback signals.

Implications: The episode suggests that the biggest AI winners may be full-automation systems in hard, deterministic domains where physics can verify outcomes. For founders and investors, durable advantage will come from deep problem understanding, not just shipping fast wrappers.