Lila Sciences is building what it calls "AI Science Factories" - automated labs designed to generate experimental data at scale, treating the scientific method itself as the next major data source for training AI models. CTO Andrew Beam and co-founder Rafa Gómez-Bombarelli laid out the plan on the Latent Space Science podcast, arguing that the exhaustion of internet-scale text data forces a shift toward using real-world experiments as a near-infinite token stream.
The bitter lesson and the search for a new dataset
Beam pointed to the "bitter lesson" in AI - the pattern that general, scalable methods beat specialized ones over time. That principle drove large language models to consume the entire internet. Now that much of that data is used up, the field needs a new source of scale. "We think that actually science, running the scientific method and using nature and experiments as verifiers is like the ultimate version of that," Beam said. Lila's factories are designed as scaled verifiers, enabling the kind of post-training and reasoning that pushes models beyond static datasets.
Gómez-Bombarelli, whose background spans computational chemistry and early generative AI for materials, said the shift was inevitable. "We have seen the bitter lesson come to computationally generated data," he said, "and that's the reason why Meta and DeepMind and Microsoft, they have teams doing AI for computational materials science. But it was clear that we needed to reach out and get this thing all the way out and make AI for actual materials science and not just the computational version." The company's platform aims to let models design new experimental protocols, run them, and learn from outcomes - generating incremental data that was never part of a pre-existing corpus.
Labs that look like data centers
The physical design mirrors the ambition. Beam described the lab of the future as "rows of server racks, as densely packed as possible, and also as energy efficient as possible." Instruments are connected via a physical transport layer - what the team calls a "PCI bus for the lab" - so that AI can orchestrate workflows across nodes. This setup treats hardware as a shared resource, much like compute clusters, and is meant to produce validated data across modalities at a rate no human-led team could match.
The approach reflects a broader view within AI for Science & Research circles: that data, not compute, is now the bottleneck. By building a system that generates its own high-quality experimental results, Lila hopes to add a new scaling axis that sidesteps the limits of static databases.
Human-AI collaboration and safety constraints
Automation isn't absolute. Humans still handle tasks where dexterity trumps efficiency - removing a cap from a test tube, for example. This pragmatic division of labor keeps the system flexible without chasing full robotic autonomy for every step. The AI acts as an orchestrator, while people fill in the gaps.
Safety measures are layered. Beam and Gómez-Bombarelli said they constrain the problem space by exposing only the experimental capabilities relevant to a given scientific question. That, combined with standard lab safety practices, aims to prevent real-world harm. The team also insisted on maintaining scientific standards. "We cannot relax our standards of scientific rigor because it's AI," Beam said. Every measurement must meet the same bar as human-led research, with transparent and reliable outcomes.
Why this matters for science and research professionals
If the model works, it changes where research bottlenecks appear. Instead of waiting for published datasets or running experiments manually, scientists could query a lab that designs, executes, and validates experiments on demand. The immediate takeaway for researchers is that the next generation of AI tools won't just mine existing papers - they'll generate new data, and the quality of that data will depend on the same rigor expected in any wet lab. Keeping an eye on systems that combine automation with active learning will be essential for labs that want to stay competitive in hypothesis generation and testing cycles.
Your membership also unlocks:
