The Lab as a Learning System — Part 2 of 3

Better early designs must survive downstream failure

A promising drug candidate can still fail on safety or in later trials. That is the opening challenge in Part 2 of the Lila Sciences conversation: if investors already manage biotech uncertainty through portfolios, how much does improving an early development step change the overall outcome?

The response centers on small improvements in success probability across a portfolio. Fewer failures can make investment more attractive and, in expectation, get medicines to patients sooner. A respondent describes a system informed by one million unique mRNA designs, optimizing properties expected to translate into therapeutic benefits. The aim is to improve the odds; a clinical-outcome improvement has not been measured here. Faster clinical development also requires cooperation with regulators, the respondent argues, citing regulatory frameworks as part of the competitive difference between US and Chinese biotech.

Another participant proposes designing against downstream failure from the outset: eventually, a model should filter drug candidates for their prospects of surviving phase III trials. The chemistry-and-materials work offers a concrete way to bring later constraints into design. Tools can call process-engineering simulators to choose pipe diameters and heat exchangers for economically viable scale-up. Physical reasoning, available data, and measurable proxies—properties that help predict later behavior—can inform those choices, although some process behavior still requires direct measurement.

The timing matters: the molecule or sequence can be fixed before later decisions about patient populations and how success will be measured. Anticipating those decisions could improve the initial design. Lila also acknowledges the risk of excessive scope. Commercial work concentrates on scientific problems the team believes it can address now; broader clinical and manufacturing contributions remain goals tied to better models and access to partner data and processes.

Breadth supports connections; open-endedness chooses questions

Lila’s argument for scientific breadth is that knowledge from one specialty can unlock reasoning in another. The discussion uses Emily Whitehead’s CAR T treatment as an example: a respondent connects the clinician’s intervention to knowledge of an antibody used in pediatric arthritis. The proposed opportunity is to make such cross-specialty connections systematic, improving development odds without depending on the right person happening to be in the room.

Open-endedness, described here as machine creativity, addresses a related challenge: exploring possible directions, judging what is interesting, and choosing worthwhile questions. Ken Stanley is described as leading Lila’s team focused on this outer loop of scientific reasoning. A respondent argues that excellent answers—even from reinforcement learning at scale—do not by themselves establish exceptional creativity. The team’s mandate is to make models ask interesting questions as well as solve difficult ones; this remains a research direction without a concrete result presented in the conversation.

Connect instruments into a programmable lab

The lab makes the computing analogy physical. A planar motor system magnetically levitates plates between instruments, while robot arms handle transfers. Carriers sometimes park until congestion clears. A speaker calls this transport layer the lab’s “PCI bus”: a shared connection between instruments that participate in a larger automated workflow.

The workflows span several kinds of materials work: - Hydrogen catalysts: an ink containing material nanoparticles goes through a spin coater to produce a scaled-up form factor; a robotic handler moves catalyst pieces for testing. - Thin films: magnetron sputtering deposits material from three or four sources across a chamber. It produces mixtures of elements using very little material, supporting rapid design–make–test cycles for catalysis, corrosion, and mechanical properties. - Quantum dots: a repurposed liquid handler supports synthesis and design. The broader argument is that a few dozen instruments can support substantial work if a model can operate them through an integrated platform.

The difficult engineering is making device automation interoperable. A commercial instrument may execute a macro entered on an attached tablet yet offer little ability to coordinate with other devices. Lila reports writing custom drivers, firmware, and coordinating software to obtain granular control. Legacy interfaces require unusual adaptations: a vision-language model controls a Windows 95 machine, and a robot has been used to press an instrument’s attached iPad interface.

The physical layout still accommodates people: instruments sit on benches where humans can service them and replenish reagents. The proposed future lab would stack equipment vertically, pack it densely, retain service access, and emphasize energy efficiency. A speaker frames the objective as experimental “tokens” per unit volume—a lab designed to feel like a data center.

Standard plate formats reduce the integration burden. Lila adopted 96-well plate form factors for materials science because they are already a basic unit of lab automation. The speakers treat 96- and 384-well formats as an 80/20 compromise, not a measured coverage guarantee. Larger deposited-material samples can retain the plate-shaped carrier while reducing the sample count; one tentative example is a four-by-three arrangement.

At larger scale, scheduling must coordinate experiments, transport, and instrument access to generate both customer value and information for the model. The speakers compare this to a Slurm scheduling queue that maximizes global throughput under constraints. The current lab is essentially one fully connected graph, a topology they say cannot scale indefinitely; operations producing hazardous fumes already require isolation. A multilevel, millions-of-square-feet facility running continuously with data-center-like uptime is the long-term target. Its layout remains unresolved, and the team describes that reliability goal as exceptionally difficult.

Shorten the learning cycle by redesigning measurements

When asked which kind of experimental scaling matters most, a respondent prioritizes successive rounds of experimentation over one broad, highly noisy experiment. Speed must be judged against the field’s baseline: reducing a month-long cycle to four days could substantially accelerate progress.

Breadth and speed can coexist in pooled assays: many candidates share an experiment, and the readout identifies successful candidates without requiring a separate measurement for each one. DNA-encoded libraries are the example offered. The structure of the assay makes broad screening compatible with a fast result.

Another route is to change what gets measured. The team reports building instruments for coolants and catalysis that measure alternative properties with responses roughly a thousand times faster. Gas sorption—the uptake of gas by a material—provides a more detailed example. The conventional workflow described adds a known amount of gas, waits for uptake, infers uptake from the pressure change, and repeats at higher pressure. The speaker estimates about a day per sample. Parallelizing it requires another gas line and canister for each additional setup.

The team’s parallelizable proxy assay processes 96 materials in about an hour. It tests metal-organic frameworks in a 96-well plate by measuring another property as a readout for what pressure would reveal. Against the stated day-per-sample baseline, the speaker estimates roughly 2,500-fold faster throughput. That comparison combines batch parallelism with a changed measurement; the alternative property and its validation are unspecified, so equivalent accuracy is not established.

The experimental strategy also depends on the model’s starting knowledge: - Little relevant knowledge: a large, slower data collection can supply a foundation the model lacks. - Existing competence: rapid serial iteration lets each round inform the next. The team intends to use both. Its bet is that better models become more sample-efficient—learning more from each experiment—so gains from repeated rounds can compound enough to outweigh a large, noisy, broad dataset.

Let the platform outgrow its instruments

Fast learning raises another concern: what if the model exhausts the problems suited to expensive equipment? A respondent welcomes that possibility, hoping some binding measurements could become unnecessary within two years because the model had mastered the relevant behavior. That is an aspiration. The platform should accommodate obsolete instruments and changing measurement methods; the team explicitly hopes it will not still be doing 2026 science in 2036.

The goal is plug-in instrument onboarding: a model reads manufacturer specifications and abstracts the appropriate software interfaces, aided by work with vendors. A speaker hopes to reduce onboarding from thirty days to thirty minutes within two years, while acknowledging that the team may not achieve it. Swapping an existing instrument—including replacing a Hamilton liquid handler—is reportedly quick already. Integrating an entirely new instrument remains a distinct capability.

Instrument vendors are also expanding the available science. A speaker describes measurements that a decade earlier required scheduled access to major neutron or X-ray facilities becoming available beside experimental workflows. Miniaturization, better resolution, and brighter sources support automated, high-throughput work, while existing instruments are expected to keep supporting new questions. The nearer-term facility plan is a 100,000-square-foot space in Massachusetts, with autonomous mobile robots handling some transport.

Ten trillion tokens of reasoning, tools, and feedback

The reported ten trillion scientific tokens come from reinforcement learning (RL) across scientific environments. A respondent frames RL as a data-generation mechanism that steers models toward more valuable tokens. The resulting material mixes English, tool calls, and experimental feedback.

The key unit is a model-generated reasoning trace with experimental verification. The count does not describe downloaded scientific databases tokenized at the sequence level, although DNA tokens can appear when a model addresses a DNA question. A speaker compares the scale with pretraining corpora they put at 15–30 trillion tokens, motivating an expectation of mastery and emergent capabilities. That is a scaling rationale, rather than a demonstration of those outcomes.

Lila builds on pretrained open-weight models, assuming their exposure to the internet and scientific literature provides useful starting knowledge. A speaker reports frequent use of Nemotron as a base model through a partnership with Nvidia, without identifying a specific version.

Evaluation operates at two scopes: - A broad environment suite: roughly 1,000 scientific RL environments accept frontier or internal models. The team intends to release a subset using generated and curated data. - Product-specific benchmarks: multidisciplinary teams compare models trained from scratch, off-the-shelf frontier models, and internal models on narrowly defined problems tied to intended products.

Lila reports that its scientifically trained model with tool access typically substantially outperforms the alternatives on those internal tasks. It also reports immediate gains from experimentally verified traces despite having fewer parameters than frontier models. The claimed advantage is domain-specific and unquantified here: scores, comparator identities, and evaluation protocols are not given.

A shared scientific platform creates several commercial paths

Generate’s protein-engineering origins introduce the commercial discussion. It began with an effort to apply machine learning to protein engineering, supported by Flagship seed funding. The typical Flagship path is described as sole investment at Series A, followed by outside capital at Series B, often producing businesses organized around therapeutic assets. At the time of the conversation, Generate reportedly had a phase III monoclonal-antibody trial for asthma and a phase I trial for COPD.

One rationale for Lila was avoiding repeated duplication of machine-learning and platform teams across companies. A shared platform could support several scientific applications. A speaker recalls Lila’s first year coinciding with the arrival of o1, which helped clarify the opportunity for a new kind of scientific model. The early conviction was that science could become an effectively unlimited source of training tokens; the commercial strategy took longer to clarify.

Lila also took a different financing path: outside investment arrived before Series A, and Flagship did not lead that round. It retained access to Flagship’s network and company-building experience.

Its CAR-T work illustrates the commercial choice. Lila reports using one therapy to launch several partnerships for further development, including changes to therapeutic properties and new indications. When the interviewer suggests that the original therapy was merely a proof of principle, the response clarifies that licensing or partnering that specific therapy was an option. Several development partnerships were judged the better route. In one arrangement, a partner creates a virtual startup around an indication and pays Lila for further development with associated milestones.

The mission extended beyond biotech from day one, requiring different hiring networks and instrument vendors. Lila reports a substantial software engineering presence, competition for candidates considering frontier AI labs, and GPU spending atypical of biotech. This breadth is part of the scientific thesis: the team expects capabilities to emerge from work across fields.

Biology and materials face different routes to value

Asked which field is harder, the participants offer competing answers: - Small-molecule discovery combines chemistry and biology: it requires reasoning about both synthesis and adverse biological effects. - Materials lack biology’s mature automation: another speaker emphasizes less developed high-throughput infrastructure and a more heterogeneous collection of problems, despite strong simulators. - Both require downstream validation: supply chains matter in both, and materials qualification—validating and verifying a product—plays a role analogous to clinical trials.

The economics differ even where the development stages have parallels: - Materials must work over a service life: lab tests only partially predict performance inside devices and supply chains. - Biological assets are described as easier to finance: the participants argue that successful clinical trials establish substantial value, with manufacturing cost less often the blocking factor for those therapies. - Scientific importance need not produce a recognizable supplier: superconductors illustrate how a material can support a successful business with good cash flow while the company remains little known.

Finding the valuable materials problem can itself be commercial work. Materials companies are described as having a less open innovation ecosystem, with engagements often beginning by persuading a company to disclose what matters to it. Government funding and national-security demand provide another route. Lila reports British government partnerships, work and awards involving the US government, and existing national-laboratory relationships. Asked about Mission Genesis, a speaker describes Lila as a named partner and reports submitting 25 Genesis Lighthouse proposals in the preceding week.

The prediction gap that keeps the lab necessary

The closing question asks which bottleneck a guest would remove by fiat. Rafa chooses the simulation-to-reality gap: the mismatch between simulated behavior and what happens in the physical world. He explains it through robotics, where training in virtual 3D environments does not transfer perfectly to movement in physical spaces, then applies the same idea to materials.

Molecular and electronic-structure simulations, Rafa argues, are useful but insufficiently predictive for materials discovery. If they were predictive enough, the team might not have needed to build a self-driving materials lab: prediction could have replaced that experimental effort. Part 2 ends during this explanation, with measurement filling the gap left by simulation.