AI FOR SCIENCE · INTERVIEWS
Lila says it trains one general science-reasoning model with real experiments as the reinforcement-learning verifier, and its CTO insists it is neither an automation company nor a biotech.
How the lab runs: instruments as nodes on a graph, plates on levitating planar tracks, people left below the API line; what the roughly ten trillion tokens are; two or three people reaching in vivo CAR-T primate data in about six months; and the virtual-startup way of charging.
Neither the CAR-T result nor 'the general model beats domain-specific ones' has an independent comparison; the model cannot yet design experiments fully open-endedly; and materials Sim2Real is, by their own account, unsolved.
Beam's argument: pretraining ran on internet text, which is spent; post-training by reinforcement learning with verifiable rewards is a way for a model to generate its own data. Maths and code have cheap checks, science needs experiments, so running the scientific method with nature as verifier is the ultimate version, and a scaled verifier is what they call an AI science factory.
“science, running the scientific method and using nature and experiments as verifier is like the ultimate version of that.”Andy BeamThe platform puts generality and flexibility ahead of raw throughput of one assay: the next datum must be informative to the model. The lab is a graph of instruments with 96-well plates levitating on planar-motor tracks, like a computer bus; steps not worth automating stay with people, below a single API the model sees.
“We're not automation maximalists. We are actually sort of like token generation maximalists”Andy BeamBeam says they are not downloaded sequence databases but model-generated reasoning traces verified by experiment, a mix of English tool calls and experimental feedback, sized against the 15 to 30 trillion of general pretraining. They start from an open-weight pretrained model (Nemotron is named, with an NVIDIA partnership). They say the general model often beats domain-specific ones sample for sample, and that about 1,000 internal scientific RL environments exist, a subset of which they will likely open.
“These are reasoning tokens, model generated, that are experimentally verified.”Andy BeamGómez-Bombarelli says chemistry learned on small-molecule drug discovery carried over to metal-organic frameworks for carbon capture. Beam tells the in vivo CAR-T case: binder design, lipid-nanoparticle formulation and mRNA design combined, two or three people reaching non-human-primate data in about six months, B-cell depletion better than Capstan's public data and untranslated regions about ten times the expression of Moderna and Pfizer references. Lila will not run a trial and used it to launch partner programs.
They say round-over-round iteration matters more than large noisy multiplexed screens. Gómez-Bombarelli's team swapped a slow gas-sorption measurement for a parallelisable proxy on 96-well plates, about 2,500 times faster. They also describe reinforcement-learning pathologies: a chain of thought collapsing into repeated answers, or skipping the experiment to give an answer; Beam calls the chain of thought an unreliable narrator, so the experiment stays the ground truth.
Materials Sim2Real: physics simulation is not predictive enough, which is why a self-driving lab is needed; instrument onboarding takes about 30 days now, with a goal of about 30 minutes; reinforcement-learning MFU is only about 5 to 6%; scale-up from small samples to manufacturing is left to customers or partners; and Ken Stanley's open-endedness team is meant to tackle models that answer questions but do not ask them.
Lila on the timeline is two capital entries from 2025: the March launch with $200 million, combining scientific models, autonomous labs and new experimental data generation, and the October $350 million Series A to scale AI science factories and external partnerships.
The CTO sets out how it runs: the lab is a token generator, the model's verifier and its moat, not automation for its own sake; people sit below the API line and onboarding an instrument takes about 30 days.
'Ten trillion experimentally validated reasoning tokens' appears in the Latent Space episode's introduction without saying what is counted.
Beam says they are model-generated reasoning traces verified by experimental feedback, not downloaded sequence databases; the counting unit, deduplication and the share tied to physical experiments are still not said.
The timeline has no Lila CAR-T entry; the public line is the episode page's one sentence about in vivo CAR-T primate data in six months.
The interview gives the method and its limits: three capabilities combined, two or three people over about six months, all the way to where one might think about filing an IND, with B-cell depletion set against Capstan's public data; Lila will not file the IND or run clinical work, and uses it to launch partner programs paid by platform fee, reagents and milestones.
Lila's bet is that what science lacks is cheap verification, so it builds a factory that produces verification at scale. Like Radical and Periodic it treats experimental data as the next training data; what differs is doing biology, chemistry and materials at once, on the bet that breadth buys depth.
They admit the model can skip experiments, loop pathologically and find holes in a reward, so the experimental result must be the final check. It draws a line under what an agent says it did: read the experiment record, not the narration.
By their own comparison a drug that passes a trial is easy to underwrite, while a material has a supply chain, devices and qualification, lab tests only partly predict lifetime, and materials firms mostly stay private. So in materials Lila leans on government and defence programs and partners; that is the interview's account, not our data.
About ten trillion experiment-verified reasoning tokens assembled, sized against the 15 to 30 trillion of general pretraining.
Company-reported
Explained as model-generated reasoning traces, not downloaded sequence databases; the counting unit and deduplication are not said and it cannot be audited outside.
About 1,000 internal scientific RL environments, a subset of which may be open-sourced.
Company-reported
A plan; not reproducible outside today.
The general science model often beats domain-specific models sample for sample.
Company-reported
No tasks, numbers or comparison are given.
mRNA untranslated regions give about ten times the expression of Moderna and Pfizer references; two or three people reached non-human-primate data in about six months, up to where one might think about filing an IND, with B-cell depletion and durability better than Capstan's public data.
Company-reported
No independent head-to-head; the characteristics looked at are not specified; Lila will not file an IND or run a trial.
The lab is a graph of instruments linked by levitating planar tracks; instruments are mostly off the shelf with custom drivers and firmware, and some still run Windows 95 under a vision-language model.
Current system
The guests' description, with a lab video shown.
People stay below the API line for steps not worth automating; the model gets about 80% of one gene-editing expression protocol zero-shot but cannot yet design experiments fully open-endedly.
Current automation
Beam's description, with this one example.
Replacing a gas-sorption measurement with a parallelisable proxy, about an hour for 96 MOFs, roughly 2,500 times faster.
Company-reported
Gómez-Bombarelli's statement; how well the proxy agrees with the original measurement is not given.
Platinum-group-free electrocatalysts: model suggestions that were at first low-confidence became the best non-platinum-group catalysts they have made.
Company-reported
No data given; a later catalyst result is under After the interview, and its best material is mostly palladium.
Instrument onboarding takes about 30 days now, with a goal of about 30 minutes; a new 100,000 sq ft facility is planned in Cambridge.
Target
A target, not a demonstrated plug-and-play standard.
AbbVie bought Capstan for about $2.1 billion, cited to show what in vivo CAR-T preclinical data is worth.
Industry context
Industry context from the interview, which the episode page repeats.
Materials Sim2Real is unsolved; reinforcement-learning MFU is only about 5 to 6%; scale-up and clinical work are left to partners; reward hacking and loops remain.
Not yet closed
Said plainly by the guests themselves.
Whether Lila opens a subset of its RL environments and evaluation protocol.
Would show up as Public environments and an evaluation that others can run their own models in.Still open Lila's news index (19 posts, read 2026-10-02) lists no post releasing RL environments or an evaluation protocol.Whether the CAR-T primate result is confirmed by a partner or paper, with the measures and the comparison conditions.
Would show up as Side-by-side data on the same species, dose and measures, ideally from an independent party.First evidence Lila's post of 2026-08-28 restates the result and links ASGCT 2026 abstract 3251 (12 authors, all at Lila). The abstract gives doses: one 0.5 mg/kg dose in monkeys gave near-complete B-cell depletion through day 7 and about 50% of baseline at day 14, when the benchmark had returned to baseline; in mice 0.25 mg/kg beat the benchmark at 1.0 mg/kg. The abstract says three months, the post under six; neither names the benchmark or gives animal numbers or statistics, and no outside party confirms it.A controlled comparison of the general model against domain-specific ones.
Would show up as Results on the same tasks with the same amount of data.Still open Lila's news index lists no post with such a comparison.How far instrument onboarding time falls from about 30 days, and the unattended runtime and cost per useful token.
Would show up as Figures Lila publishes on the same basis.Still open Lila's news index (20 posts, read 2026-10-07) and Beam's Ray Summit keynote (on video 2026-09-17) give no onboarding time, unattended runtime or cost per token.The first named virtual-startup partner, with milestones.
Would show up as A release by the partner or Lila naming the program and milestones.Still open Lila's news index lists no post naming a virtual-startup partner.Later evidence that tests or limits what was said, and targets whose date has passed.
The in vivo CAR-T the guests said two or three people took to primate data in about six months now has a Lila post and an ASGCT 2026 abstract, all authors at Lila: one 0.5 mg/kg dose in monkeys gave near-complete B-cell depletion through day 7 and about 50% of baseline at day 14. That adds the doses and measures the interview lacked, but the two sources disagree on the time (three months or under six) and on the benchmark, neither names that benchmark, and no outside party confirms it.Timeline entry →
Lila reported an iridium-free oxygen-evolution catalyst: from 2,942 oxides over 53 systems, InMnPdOx, 99.9% palladium by metal, stayed below 0.5V overpotential for 1,000+ hours in acid while a ruthenium oxide reference failed at 615. It is a published electrocatalysis result from Lila's platform, and its best material is mostly palladium, itself a platinum-group metal. The paper's strongest platinum-group-free comparator, CoFeZrOx, started about 90 mV worse, and that cobalt-zirconium family lasted under 4 hours in acid, the closest the paper comes to the guests' best platinum-group-free catalysts.Timeline entry →
Read 2026.10.01 The full machine transcript from a third party (PodScripts), checked against the Latent Space episode page; the transcript does not label speakers, so who said what is judged from content. The first version, of 2026-09-10, was checked against the YouTube transcript. Open the interview ↗Latent Space episode page ↗