Key Takeaways

  • Lila Sciences is building "AI Science Factories" to generate internet-scale, experimentally verified data for foundational AI models, a direct counter to slow, traditional R&D.
  • Unlike AI, scaling in materials and chemistry often reveals unworkable ideas; Lila's method integrates 'supply chain conscious' reasoning from the earliest experimental stages to avoid this trap.
  • Their core strategy prioritizes rapid, round-over-round iteration over slow, broad multiplexed experiments, as demonstrated by accelerating gas sorption measurements 1000x to a one-hour cycle.
  • The team actively works to bridge the "sim to real" gap by focusing on iterative learning, acknowledging physics-based simulations often lack real-world predictive accuracy.
  • A major bottleneck in their AI training pipelines is Mean Flop Utilization (MFU), currently operating at a dismal 5-6% of theoretical GPU power, meaning 95% of paid computing power is wasted.

The Method: Building AI Science Factories with Iteration-First Design

Rafa Gomez Bambarelli, co-founder of Lila Sciences, points out a bitter truth for founders in deep tech: “In chemistry and materials scaling is a spooky thing because it turns out only the things that that you can scale matter.” This stands in stark contrast to AI, where scaling often provides a clear roadmap. Lila Sciences isn't just aware of this; they've engineered their entire operation, the "AI Science Factory," to circumvent it.

Their method starts with "supply chain conscious" reasoning. From the first experiment, Lila integrates techno-economic analysis, asking: can this design actually scale? This front-loaded filter prevents costly dead ends. Instead of developing assets then trying to scale, their business model centers on the AI model itself, trained on the internet-scale, experimentally verified data their factories produce.

Andy Beam, also from Lila, emphasizes their strategy for navigating complex R&D: “round-over-round iteration is more important than like a broad, hugely multiplexed, uh highly like noisy kind of thing.” This means favoring fast, tight feedback loops over slow, sprawling experiments. For example, they've dramatically accelerated gas sorption measurements, a critical process in materials R&D, making it “a thousand times faster than the way we were doing it,” now completing a cycle in about an hour. This directly tackles the "sim to real" gap, where physics-based computational models often fail to predict real-world experimental outcomes. By iterating rapidly, they generate real-world data faster, closing the simulation-to-reality feedback loop.

This iteration-first mindset extends to their AI infrastructure. Beam highlights a critical, often ignored, bottleneck in their reinforcement learning training: Mean Flop Utilization (MFU). “MFU is always a very small fraction of peak theoretical flops,” he notes, revealing that their current RL pipelines hover “around 5 to like 6%.” That means they're getting only a tiny fraction of the GPU power they pay for. Addressing this low MFU isn't just about efficiency; it's about accelerating the core AI learning cycles, making their entire factory more productive.

Where This Breaks Down

Lila Sciences' method shines when you can actually accelerate the experimental cycle. If your fundamental R&D process inherently requires weeks or months for a single meaningful data point due to biological growth times, extreme environmental testing, or complex multi-step syntheses, then “round-over-round iteration” at high frequency isn't possible. This approach assumes you can design experiments that yield meaningful, non-noisy data quickly. If initial data generation is noisy or sparse, rapid iteration might just cycle through bad data without true learning. It also demands significant upfront engineering investment to automate labs and integrate robust techno-economic analysis from day one, a luxury many early-stage startups might lack. Fixing low MFU requires deep system-level optimization expertise, which isn't always readily available.

What to Do With This

This week, analyze one core experimental or data generation process in your own startup. Ask: can you accelerate this cycle by 10x, even 100x? Don't focus on broadening the experiment; instead, narrow it to the fastest, most informative single iteration possible. Then, pull up your cloud compute bill. If you're running any ML training, investigate your Mean Flop Utilization; if it's below 50%, you're likely wasting capital and slowing your iteration loops. Find an engineer whose sole job for the next month is to identify and fix these infrastructure bottlenecks.