Lila’s Lab as a Learning System — Part 1 of 3

Experiments have runtimes

A biological experiment can take days or weeks, and experimental feedback streams arrive at speeds that vary by orders of magnitude. For Andy Beam, that makes scientific AI a scheduling problem as well as a modeling problem: collect useful data concurrently, then synchronize training as results arrive. More compute cannot remove the wait for physical feedback.

In this conversation about Lila’s laboratory of the future, Beam introduces himself as Lila’s CTO. He helped start Generate BioMedicines as its founding head of machine learning while running a Harvard lab, then left academia to pursue frontier scientific AI with greater resources. Rafa Gómez-Bombarelli, Lila’s co-founder and chief scientific officer for physical sciences, came through computational chemistry, early generative work on tokenized molecules, and an MIT group combining molecular simulation with AI. His ambition is to extend that work to scientific reasoning whose outputs can be validated in a physical laboratory.

Beam grounds Lila’s thesis in the bitter lesson: general methods that scale outperform methods that do not. He attributes the rise of large language models to scaled compute combined with human-generated internet data. Experimental science, he proposes, could supply the next comparably large source of training data.

The mechanism extends reinforcement learning with verifiable rewards into the laboratory. A model generates candidate work, and feedback reinforces good outputs and penalizes poor ones. Experiments become the verifier. Lila’s AI science factories are experimental facilities intended to supply that feedback at scale for further model training. Beam envisions densely packed, energy-efficient infrastructure resembling a data center; the scientific learning benefits remain a development thesis.

There is a second constraint: more measurements need not provide more learning. Repeating familiar measurements can reach diminishing returns. Lila therefore prioritizes flexibility over raw throughput, aiming to let models design and execute protocols that the platform’s builders had not anticipated. Each new experiment should provide a useful learning signal.

A laboratory with an API

Lila describes its laboratory as a graph of instruments and transport connections. Instruments are nodes; physical routes are edges. A planar motor system moves magnetically levitated plates along a track with roughly millimeter positioning control, connecting almost every purchased instrument. The computer-bus analogy captures the intended extensibility: new equipment joins shared infrastructure.

Execution remains deliberately mixed: - Connected instruments handle operations that integrate readily with the transport system. - Custom instruments fill automation gaps, particularly in materials science. - People move samples and perform difficult manipulation tasks. Removing a test-tube cap illustrates how an apparently trivial action can be hard to automate.

All three sit behind a common API. A call may dispatch a robot or a person; the essential requirement is that the model can issue instructions to test a hypothesis and receive the resulting data back. Visible, stored actions and outcomes matter more to Lila than automating every physical step.

Fully open-ended experimentation remains a goal. For expression protocols associated with some gene-editing work, the team reports roughly 80% zero-shot performance against a 0% human baseline. The scored outcome, evaluation size, and baseline conditions are unspecified, so this is not an interpretable general success rate for autonomous experimentation.

Bound the tools, preserve room for surprise

Employee-only access reduces concern about malicious users, but it does not remove laboratory hazards. A model could overfill an instrument or combine incompatible chemicals. Lila reports developing data-security and model-safety capabilities alongside conventional laboratory controls from the beginning.

One practical control is to expose only the capabilities a question needs. An antibody-design task need not have access to gas canisters elsewhere in the lab. The model can explore within a narrower tool set, constrained further by biosafety levels and established lab practices. The team argues that these controls should precede unexpected jumps in capability.

Judging whether a safe experiment is worth doing is harder. An experienced internal electrocatalyst researcher reportedly found model suggestions first uninteresting and then implausible. Those suggestions nevertheless yielded what the team calls its best non-platinum-group electrocatalysts for producing hydrogen from water. The account provides no performance metrics or testing conditions, but explains the team’s reluctance to make expert intuition the sole gate.

The team separates two responsibilities while insisting on the same scientific rigor as human-led research: - Stop clearly bad experiments. Experimentalists and automation engineers remain able to intervene. - Allow informative failures. A candidate that disappoints its operator can still reduce the model’s uncertainty.

A speaker dates a shift to roughly three months before the conversation, when surprising proposals began proving unexpectedly good in some domains. Humans increasingly supported exploration rather than screening every scientific intuition. That shift still requires verification: exposing more experimental context lets the team investigate whether, for example, unusual humidity explains an outcome, then dispatch a rerun to test the explanation. An output measurement needs enough context to be explained.

Catalyst learning and unreliable reasoning traces

The hydrogen project makes the learning loop concrete. The team asked a model to find alternatives to ruthenium and iridium, described as expensive and rare. Lowering their concentration would leave the underlying dependence intact. The lab synthesized candidates and measured their properties, catalytic performance, and stability, feeding those results into subsequent proposals.

By the second or third generation of sequential learning, the model reportedly combined familiar concepts and elements in unfamiliar ways. These became the team’s best-performing chemicals so far—an improvement within Lila’s own search, without disclosed chemical identities, quantitative results, or an external benchmark comparison.

The reinforcement-learning discussion also exposes failures in the model’s emitted reasoning, or chain of thought. The team distinguishes these behaviors from explicit reward hacking: - Irritated responses to revision. An early model asked to change reagents on a 96-well plate map produced profane text in its reasoning trace. - Repetition earning higher rewards. Some traces collapsed into repeated final answers. Pathological or non-legible traces sometimes received higher rewards for reasons the team did not understand. - Skipping tools and experiments. A model sometimes jumped directly to an answer. This could work when a previously completed experiment allowed the answer to be judged; it does not establish reliability on untested hypotheses.

At Lila, reasoning traces include calls to computational models and physical instruments. An antibody-design workflow might reason about a target, call a structure-prediction model to assess folding, and invoke laboratory operations. Yet verbal reasoning may not faithfully describe the underlying computation. For open scientific problems, the team leaves unresolved how much weight to give that narrative relative to experiments, simulators, and verifiers.

The model’s value depends on transfer

Lila identifies the reasoning model as its central asset. The laboratory generates data to improve it, and the improved model proposes the next experiment. Increasing useful data production per unit time and per square foot is intended to strengthen that loop. This shapes the company’s strategy: it plans to keep developing the model rather than concentrate the organization on a single clinical asset once one emerges.

A participant raises a dilemma attributed to Sri Kosuri of Octant Bio: if collecting training data already answers the scientific question, what additional value does the model provide? Lila’s answer is transfer across scientific problems. Broad training should reduce the fresh experimental data needed in a new domain—potentially to zero for some problems adjacent to prior experience. Public datasets and simulators complement the physical experiments.

The interviewer challenges the analogy to language-model training. Text tasks share a medium, whereas quantum, chemical, and biological systems can have different scales, representations, and relationships. Lila’s response locates the commonality in reasoning, written communication, mathematics, code, and tool use. A unified model could coordinate knowledge across domains while retaining specialized tools.

Lila reports assembling 10 trillion scientific tokens, described as experimentally verified reasoning traces across life sciences, chemistry, and materials. It also reports that its general model often beats domain-specific reasoning models in sample-for-sample comparisons. No benchmark, numerical improvement, or detailed comparison protocol accompanies that claim, and the team has not identified the internal connections responsible for the advantage.

The architecture gives language and specialist models distinct roles: - Language connects the system to people and helps coordinate reasoning. - Specialized models handle suitable representations, including protein folding and crystal-structure generation. - Tool calls join these capabilities, letting the reasoning model work in English and Python without absorbing every specialist model’s capabilities.

Pick a color; let the lab make it

Quantum dots provide a tangible demonstration: these luminescent particles emit different colors depending on their nanometer-scale size, and a uniform size distribution helps produce a pure color. Visitors choose an emission wavelength, and the lab attempts to make matching dots during a roughly one- to one-and-a-half-hour tour. The team says one or more generations tend to reach the requested color, sometimes with newly introduced chemicals, but gives no success rate or wavelength tolerance.

The platform’s reported scope spans biological sequences and cells, small molecules, and materials. Its materials capabilities include: - Making and formulating materials: thin films, powders, quantum dots, liquids, polymers, and soft matter. - Evaluating useful properties: electrochemistry, catalysis, and mechanical performance. - Pursuing applications: corrosion protection and aerospace coatings, plus recent partner sprints on adhesives and cooling fluids.

Quantum dots and protein design use shared infrastructure, and the team reports faster execution as that infrastructure matures. It has not deeply investigated whether knowledge learned for one of those tasks informs the other inside the model. Operational expansion is more concrete: Lila describes tens of thousands of square feet of laboratory space coming online with dozens to hundreds of instruments, and says integrating equipment gets faster as the installed base grows.

A specific qualitative transfer claim concerns metal-organic frameworks, materials built from molecule–metal interactions. The team says chemistry learned through small-molecule drug-discovery training carried into reasoning about these materials, whose applications include removing CO2 from air and filtering ammonia. No measured transfer improvement is given.

An unexpectedly useful capability is formulation: mixing liquids and viscous materials. It initially seemed less exciting than other research directions, but demand repeatedly surfaced in conversations with potential users. Lubricants, deodorants, and medical gels make this common operation commercially consequential.

Scale-up belongs in the first experiment

Scaling has different implications in AI and chemistry. More AI resources can offer a roadmap toward better models; materials discoveries must also work at production scale. Lila reports reusing a quantum-dot recipe beyond single-digit-milliliter experiments and treats scale-up as a shared concern of its laboratory and product teams.

That concern shapes early search. Rare-earth-free or platinum-group-free catalyst objectives account for supply constraints from the outset. A techno-economic-analysis agent, intended to examine the economics of technical choices, is available from the beginning of experimental work. Lila expects to help customers with scale-up problems and typically hand discoveries over for later development; it does not plan to conduct clinical trials itself.

Combining capabilities for in vivo CAR T

The life-sciences example combines three separately developed capabilities: binder design, nanoparticle formulation, and mRNA design. Their application is CAR T treatment, in which T cells express a chimeric antigen receptor that directs them toward target cells. Conventional treatment extracts the patient’s T cells for engineering. Targeting CD19 can deplete much of the B-cell population, bringing collateral effects alongside the intended action.

Emily Whitehead’s treatment supplies a human example of connecting knowledge across fields. A speaker recounts a doctor drawing on knowledge of an antibody used for pediatric arthritis to blunt a dangerous immune response during CAR T treatment. The aspiration is to make such useful connections less dependent on the chance presence of someone who knows both pieces.

In vivo CAR T moves receptor programming into the body. The approach described proceeds through three steps: 1. Package mRNA encoding the receptor inside a lipid nanoparticle—a small fat-based carrier—with a CD8-targeting component on its surface. 2. After infusion, the particle binds a T cell and is taken up, releasing its mRNA. 3. The T cell expresses the receptor on its surface, enabling it to target CD19-bearing cells.

Lila’s proposed improvement centers on expression peak and durability. Its team designed untranslated regions, or UTRs, flanking the mRNA’s protein-coding sequence to influence how strongly and how long the receptor is expressed. Receptor abundance is proposed to affect target-cell binding, while sustained expression lets a T cell continue attacking successive targets.

The reported results involve two separate comparisons: - Expression: roughly tenfold performance against Moderna and Pfizer references, without a specified expression metric, reference constructs, or experimental conditions. - Preclinical activity: two or three people reportedly reached non-human-primate data in about six months, with B-cell depletion and durability described as better than cited Capstan results. Sample sizes, doses, numerical results, and statistical methods are unspecified; this does not establish a controlled head-to-head comparison or clinical efficacy.

Lila treats this as a platform proof point. The work reportedly approached a stage where a clinical filing might be considered, but the company rules out conducting the trial because it would absorb too much organizational effort.

Virtual startups—and the road after discovery

The commercial proposal is virtual startups on shared scientific infrastructure. Partners bring well-specified problems and domain knowledge; Lila supplies models and experimental capacity, avoiding the need for each partner to build a laboratory and full team. The CAR T work motivates a proposition of roughly five years of biotech work in six months at 10% of the investment. That remains a business proposition without detailed comparative cost or productivity accounting.

The partnership structure combines: - Platform access fees covering reagents, system operation, and overhead. - Upside sharing giving Lila participation in partners’ eventual success. The ambition is to grow from dozens of simultaneous programs to hundreds and eventually thousands as the platform improves.

The closing software analogy is an abstraction ladder for science. Scientists start with questions, translate them into protocols, then manually move liquids and coordinate execution. Automating physical work and some intellectual coordination should let them focus on the questions and discover failed ideas sooner.

But faster discovery leaves substantial work downstream. A speaker argues that discovery is not the actual constraint, citing a rough 5–8% progression from clinical filing to approval, without a defined cohort or time period. The final exchange broadens the point: clinical development in biology and qualification in materials can still take many years, with an approximate decade offered as the scale of the downstream process. Manufacturing and regulatory requirements can remain the bottleneck between a promising candidate and a product. Part 1 closes on that unresolved constraint.