It is June 2026. Lila Sciences is reported to be in active Series B talks to raise about $2 billion at an expected $8.5 billion valuation, anchored by CalPERS and Nvidia’s NVentures. Inside its Cambridge labs, Lila is attempting to build what it describes as autonomous "Science Factories." The company has claimed that its platform, Lila Iris, can design two-target antibodies in a single hour — a projection that represents a massive shift from conventional timelines of over a year, though these throughput metrics remain company-promoted rather than peer-reviewed. The technical architect leading this integration of statistical reasoning models with robotic wet-lab automation is Andrew Beam, PhD.

If Geoffrey von Maltzahn is the visionary who climbed the abstraction ladder from gold nano-antennas to microbiomes to gene writing to the scientific method itself, Andrew Beam is the operational engine who made that highest rung executable in silicon and steel. He does not merely apply machine learning to static datasets. He installs pattern-recognition models as the literal operating system of discovery, then hardwires them to physical hardware so the universe itself can serve as the causal validator.

The architecture of Beam’s mind was forged long before frontier foundation models. At North Carolina State University he completed a rare triple major in Computer Science, Computer Engineering, and Electrical Engineering, followed by a Master’s in Statistics and a Ph.D. focused on Bayesian neural networks and bioinformatics. In 2025 he was inducted into the NC State CS Alumni Hall of Fame. This hardware-to-statistics trajectory is not decorative; it means Beam intrinsically understands the physical limits of silicon, the flow of logic gates, and the irreducible uncertainty of biological systems. He is not a software theorist abstracted from consequences. He understands the substrate.

After his Ph.D. he moved to a postdoctoral fellowship in biomedical informatics at Harvard Medical School and then became Assistant Professor of Epidemiology at the Harvard T.H. Chan School of Public Health (secondary appointment in Biomedical Informatics). Here the theoretical met the visceral. Beam’s research zeroed in on neonatal and perinatal outcomes in the NICU — an environment of continuous, high-signal, multimodal data where the patients cannot speak and a single brittle prediction can be fatal. Collaborating closely with his wife, Dr. Kristyn Beam, a practicing neonatologist, he saw firsthand that standard deep learning, while extraordinary at interpolation, is fundamentally untrustworthy in safety-critical settings because it operates on correlation rather than causation.

Flagship Pioneering’s FL56/FL57 — the entity that became Generate:Biomedicines — was the first full-scale deployment of the Bitter Lesson inside physical biology. Richard Sutton’s 2019 essay had argued that AI progress repeatedly stalls when researchers try to hard-code human domain knowledge. The only methods that ultimately scale are those that rely on massive computation and unconstrained statistical search.

Beam, serving as co-founder and Founding Head of Machine Learning alongside Molly Gibson and Avak Kahvejian, brought that wager into protein engineering. For decades computational biology had treated de novo design as a physics problem: simulate thermodynamic folding pathways, bond angles, and van der Waals forces with tools like Rosetta. Beam and the team hypothesized a radical reframing: protein design is not a physics problem. It is a grammar problem. Evolution had already written a three-billion-year corpus of sequences. If a model could ingest enough of them, it could learn the hidden statistical grammar of functional biology without any human ever teaching it the laws of thermodynamics.

The technical manifestation was Chroma, a generative model for proteins and protein complexes that Beam co-developed with Gevorg Grigoryan and the multidisciplinary team. Chroma broke from conventional simulation through three deliberate architectural choices that prioritized scale and physical realizability over hand-crafted physics:

Experimental characterization of 310 de novo proteins demonstrated that Chroma generated structurally viable, previously non-existent proteins that expressed, folded, and exhibited favorable biophysical properties. Crystal structures matched the model’s in silico predictions with backbone RMSD of approximately 1.0 Å. Beam had proved that a machine fed enough evolutionary sequence data could hallucinate functional biology that nature itself had never discovered.

But Generate still operated at the level of the molecule. A human scientist still had to take the digital output, synthesize it in a wet lab, run assays, analyze failures, and manually formulate the next hypothesis. The iterative loop remained bottlenecked by human cognitive bandwidth. To reach the next velocity regime, Beam had to climb one rung higher: he had to build a system that generated not only the molecule, but the experiment itself.

In early 2022 Flagship launched Lila Sciences with a $200 million seed. Beam transitioned from Senior Fellow at Flagship and co-CEO-level role at Generate to full-time Chief Technology Officer. His mandate: turn the pattern-recognition engine into the literal operating system of scientific discovery.

The foundational premise is that the public internet — the corpus that trained ChatGPT, Claude, and every frontier LLM — is exhausted for scientific progress. To produce what the company frames as “Move 37 moments” in physical science (referencing AlphaGo’s 2016 creative stone that no human had conceived), models cannot rely on scraping poorly configured PDFs of old papers. They must manufacture their own proprietary, high-fidelity, empirically validated ground truth.

“Science is the ultimate token generator,” the Lila technical doctrine states. “Every experiment produces new data that has never existed before.”

Under Beam’s architecture the AI Science Factory runs as a fully closed stigmergic loop — the same coordination-through-environmental-modification principle von Maltzahn had instantiated with scout and assassin nanoparticles in 2009, now lifted to the level of autonomous laboratories:

• Hypothesis generation — Lila Iris (the reasoning model) proposes the next experiment.
• Physical execution — Autonomous liquid-handling robotics, mass spectrometers, and custom hardware synthesize and assay.
• Ground-truth token — High-dimensional multimodal assay output becomes the new empirical token.
• Weight update & iteration — The model ingests the token, updates its internal causal map, and launches the next generation.

By removing the human from this specific iteration cycle, Lila aims to collapse R&D timelines from years to hours. The company's marketing and investor briefings (such as reports by investor Peter Diamandis) claim a proprietary corpus exceeding 10 trillion scientific-reasoning tokens, though Beam himself has referred more conservatively to "trillion-token scale models" under training. Regardless of the exact number, this represents a strategic attempt to build a proprietary dataset of empirical, physical truth rather than relying on scraped internet text.

To commercialize the platform, Lila has structured two access models: Catalyst(Lab-as-a-Service — external teams integrate Lila Iris via API, converting fixed capex into on-demand throughput) and Creation (end-to-end autonomous discovery campaigns). While some early reports conflated throughput numbers with academic labs (such as David Baker's Institute for Protein Design which reported an order-of-magnitude increase in validation throughput), Lila's internal platform metrics remain proprietary. For a methods-focused CTO like Beam, the emphasis remains on the structural loop itself rather than unverified commercial speedups.

Beam’s parallel roles in June 2026 create an apparent contradiction. On one side he is CTO of Lila Sciences, deliberately removing human scientists from the iterative experimental loop because human cognitive bandwidth is too slow for hyperdimensional search. On the other side he is co-founder and deputy editor of NEJM AI and co-host of NEJM AI Grand Rounds — the premier institution insisting on rigorous human-in-the-loop evaluation, causal guardrails, hallucination filtering, and the enduring importance of human judgment in clinical medicine.

The conventional narrative would frame this as a pivot: the academic safety researcher abandoned his principles for private-sector velocity. A structural reading reveals something more precise and architecturally coherent.

Beam has not abandoned the membrane. He has relocated it according to the substrate.

In clinical medicine the output of the AI interfaces directly with human biology. The stakes are immediate and personal. The physician must remain the final editorial membrane — applying causal reasoning, ethical judgment, and accountability before any decision is executed. The human is load-bearing safety.

In scientific discovery the objective is exploration of an impossibly vast hyperdimensional chemical space. Here the AI is not prescribing treatment to a patient; it is interacting with a highly constrained robotic assay. If the model hallucinates a non-viable molecule, the physical assay immediately returns a failure signal. The robotic wet-lab is an unfoolable ground-truth validator that continuously grounds the generative model back into causal reality.

Therefore, in the Science Factory, Beam removes the human from the iteration loop precisely because the robotic assay now provides the causal guardrails he spent his academic career demanding. The human is not removed from the system. The human is elevated to the boundary-condition setter — the entity that defines the objective function, sets strategic parameters, and acts as the editorial membrane around the entire autonomous apparatus rather than acting as a localized, rate-limiting cog inside it.

Beam is the rare technical operator who understands this distinction natively. He can build the velocity-maximizing closed loop because he knows exactly where the causal guardrails must be bolted into the physical hardware to prevent algorithmic drift.

By June 2026 the capacity to autonomously generate scientific truth at scale has become a matter of national security and macroeconomic survival. The United States and China control roughly 90 percent of the compute required for frontier foundation models. Concerned by this concentration, a global surge in “Sovereign AI” initiatives has shifted focus from text generation to physical scientific dominance.

In early June 2026 the U.S. Department of Energy and Japan’s ministries launched the Genesis Mission — a $1 billion joint partnership spanning twelve U.S. national laboratories — explicitly positioning AI as the new “operating system” for breakthroughs in quantum information science, nuclear fusion, and biotechnology. The geopolitical framing mirrors Lila’s internal doctrine almost verbatim: the traditional publish-and-verify scientific process is too slow for the current arms race.

In this landscape, Lila Sciences is not merely a biotech startup. It is a private, vertically integrated, sovereign-grade asset attempting to generate a strategic reserve of physical truth. The claimed 10 trillion proprietary scientific tokens represent an ambitious attempt to map the causal grammar of chemistry and biology that no competitor can replicate by scraping journals.

To place Beam’s work in the larger FounderFiles thesis, return to stigmergy — the mechanism by which agents coordinate by modifying a shared environment rather than by direct communication. In 2009 von Maltzahn built a scout/assassin nanoparticle system: benign scouts found the tumor, mildly heated to expose cryptic binding sites (the “pheromone trail”), and assassin particles carrying payload read the environmental signal and recruited en masse — a 40× gain in targeted delivery.

Seventeen years later Andrew Beam has built the ultimate stigmergic loop in silicon and steel. In Lila’s AI Science Factories the reasoning model (Lila Iris) acts as the scout, venturing into hyperdimensional molecular space. When it finds a promising parameter it instructs robotic hardware to synthesize and assay. The assay result is the new pheromone — a freshly manufactured empirical token. The model ingests that token, updates its weights, and deploys the next generation of algorithms — the assassins — to precisely optimize the molecule. Von Maltzahn climbed the ladder until he reached the scientific method. Beam is the engineer who made that highest rung executable at industrial scale and sovereign speed.

Beam’s operational signature is a rigorous intolerance for ungrounded abstraction — a trait derived directly from his hybrid engineering background and medical safety focus. Unlike software-only researchers who treat the physical world as a messy inconvenience, Beam treats physical chemistry and assay hardware as the ultimate validation layer. Deep learning left to its own devices will confidently hallucinate structurally impossible proteins or mathematically flawed causal inferences. His entire career has therefore been a sequence of building architectural constraints that force the model to survive contact with reality:

1. NICU / Causal Hybrid: Multitask RNNs that explicitly model conditional distributions and prevent parametric misspecification in complex temporal data.
2. Generate (Chroma): Polymer diffusion processes and equivariant layers that respect the biophysical constraints of collapsed proteins, ensuring only physically stable structures are hallucinated.
3. Lila: Hardwiring the reasoning engine directly to robotic wet-lab hardware so every theoretical design is instantly tested against the unyielding laws of physical reality.

He is a leader described in NEJM AI Grand Rounds program notes as selecting for “hybrid thinkers” to “evaluate safety, not just performance.” Yet he is entirely unsentimental about the role of the human operator when that operator is demonstrably inferior in speed and scale to a closed robotic loop.

“Science is the ultimate token generator. Every experiment produces new data that has never existed before. The internet data that powered the current generation of AI has been exhausted, but scientific data generation never runs out.” — Andrew Beam, Lila Technical doctrine

“In 1997 Deep Blue defeated the world chess champion. In 2016 AlphaGo’s Move 37 showed AI could generate creative strategies no human had ever conceived. Now, in 2026, we are witnessing the equivalent moment for all of science.” — Lila Sciences corporate positioning / marketing narrative

Education.North Carolina State University — B.S. Computer Science, B.S. Computer Engineering, B.S. Electrical Engineering; M.S. Statistics; Ph.D. Computer Science & Bioinformatics (Bayesian neural networks). Inducted into NC State CS Alumni Hall of Fame, 2025.

Affiliations.Chief Technology Officer, Lila Sciences (2024–). Co-founder & Deputy Editor, NEJM AI; Co-host, NEJM AI Grand Rounds. Associate Professor of Epidemiology, Harvard T.H. Chan School of Public Health (on leave); secondary appointment, Harvard Medical School Department of Biomedical Informatics. Previously: Founding Head of Machine Learning & co-founder, Generate:Biomedicines; Senior Fellow, Flagship Pioneering.

Key Technical Contributions. Chroma generative model for programmable protein design (Nature 2023). Deep learning methods for the noniterative conditional expectation (NICE) g-formula for causal inference from complex observational data (arXiv 2024). Architecture of Lila Iris reasoning engine and closed-loop AI Science Factories.

Collaborators / Peers.Geoffrey von Maltzahn (Lila co-founder/CEO), Molly Gibson & Avak Kahvejian (Generate), Arjun Manrai (NEJM AI co-host), George Church (Lila Chief Scientist), Kristyn Beam MD MPH (neonatologist collaborator & spouse).

Honors. Robert Wood Johnson Foundation Pioneer Award (for medical AI). Modern Healthcare 40 Under 40 (2024). NC State CS Alumni Hall of Fame (2025).