AI FOR SCIENCE · INTERVIEWS
Lila's CTO says reinforcement learning needs more verifiable domains to keep scaling; science is the largest one, but only about 5% of it is cheap to verify, so Lila is building an AI science factory that uses real experiments as the verifier.
Numbers the July interview does not have: the roughly 5% estimate, a quantum-dot design loop (about 90 minutes a round, base model at 12.5% accuracy), an internal scaling curve from 10^19 to 10^23 post-training FLOPs, a one-trillion-parameter model by the end of the year, and a cost for the in vivo CAR-T work: a couple of million dollars to the IND point, against five years and $100 million for the previous best (the six months was already in the July interview).
All of it is Lila's own account: the 5% is an LLM's estimate over the literature, the scaling curve uses an internal benchmark, the quantum-dot comparison names no task set or models, and the CAR-T cost comparison names no programme it is measured against.
Beam's account: models were scaled first on internet data and, from about two years ago, by reinforcement learning with verifiable rewards; domains with cheap checks, such as maths and code, are largely used up, so the question is how to enter domains where verification is hard. He calls science an infinite token generator, always able to pose a new hypothesis and produce new data, if someone builds the verifier.
“how do we start to tackle domains when verification is hard?”Andrew BeamHe took an open index of the literature and had an LLM estimate how many hypotheses could be verified with a simulator or computational tool in under an hour, and got about 5%, with, he says, wide error bars but an order of magnitude scientists would accept. His conclusion: most of science must be tested by actually running the experiment, so a large verifier is needed.
“the vast majority of science is actually dark matter for reinforcement learning with verifiable rewards because we can't verify it cheaply.”Andrew BeamHe splits the AI science factory into an intelligence layer, a software layer wrapped around the lab, and robotics. Most lab equipment assumes the operator has opposable thumbs; what sounds hard, many data points per unit time, is easy to automate, while taking the top off a test tube is hard. They run a 15,000 sq ft lab in Cambridge, which they treat more like a data centre than a traditional lab, and plan 100,000 sq ft at the end of the Red Line in early 2027.
The task is to design a quantum dot that emits at a given wavelength (about 650 nm for red, under 500 for blue); the model's actions are the precursor reagents, the lab dispenses and shakes 96-well plates so the nanocrystals form, then measures the colour they glow, about 90 minutes end to end, and the video shows dots from green to red; quantum yield is another target. Starting from an open-weight base model trained with SFT and reinforcement learning, the base gets 12.5% accuracy and, he says, the post-trained model beats frontier models. It is a project with the UK government on AI nanoscientists, and the results come from the previous generation of their stack.
He says quantum dots, electrocatalysts and proteins should go into one model, since specialising forgoes the abilities that only come from widening the scope. The x-axis is post-training compute from 10^19 to 10^23 FLOPs, the y-axis an internal benchmark that averages reward over a couple of thousand scientific tasks (each 100 to 100,000 prompts); they began late last year, say they see a reliable relationship between compute and scientific intelligence, and plan to scale the model to one trillion parameters by the end of this year.
The timeline has no Lila entry on quantum dots or on any scaling result; this site's July interview page says the lab is the model's verifier and gives no figure for how the model improves with compute.
Beam gives an internal curve (post-training 10^19 to 10^23 FLOPs, reward averaged over a couple of thousand scientific tasks) and a quantum-dot design loop (base 12.5%, post-trained above frontier models); the benchmark, the task set and the comparison models are not described.
The July interview page records in vivo CAR-T as two or three people reaching non-human-primate data in about six months, up to where one might think about filing an IND, and records that Lila will not file one or run a trial; the timeline has no Lila CAR-T entry.
The timing matches July: about six months to the IND point, for a preclinical asset, and Lila is not an in vivo CAR-T company. What is new is the cost: a couple of million dollars, against five years and $100 million for what he calls the previous best; which programme that is, he does not say.
The timeline had no Lila catalyst entry; the July interview page records only the guests' statement that model suggestions at first held with low confidence became their best non-platinum-group catalysts, with no data.
The talk (given in August, on video from 17 September) describes the electrocatalyst work as platinum-group-free, reaching about the level of platinum-group catalysts at roughly ten thousand times lower cost; the basis of the comparison is not given.
If only about 5% is cheap to verify, training science models on simulators never reaches most questions, which is Lila's reason to build labs rather than only simulate. But the figure is an LLM's estimate over the literature, not a measurement; it is the premise of an argument, not a result.
Lila bets on one model for all of science, Radical on its own alloy data, Periodic on the full trajectories of many kinds of lab. Judging who is right takes the same tasks run through a general and a domain-specific model; the talk has a curve and no control.
The audience was Ray users and the talk is about compute and platform. The CAR-T cost comparison, the ten-thousand-fold cost gap and the quantum dots beating frontier models are each one sentence on stage, with no paper, release or outside data; until they appear, read them as Lila's account of its own route.
About 5% of scientific hypotheses can be verified cheaply, with a simulator or computational tool in under an hour.
Guest's judgment
Estimated by an LLM over an open literature index, in Beam's words with wide error bars; no method detail is given.
The quantum-dot design loop takes about 90 minutes, fully automated, and team members have ordered dots remotely; dots from green to red were made.
Current automation
Beam's description with a lab video shown; the video shows the workcell running.
The open-weight base model scores 12.5% on the quantum-dot design tasks and, after post-training, beats frontier models.
Company-reported
From the previous generation of the stack; the task set, scoring and the frontier models compared are not named.
Across 10^19 to 10^23 post-training FLOPs, an internal benchmark (reward averaged over a couple of thousand scientific tasks) relates reliably to compute, presented as an early scaling law for scientific intelligence.
Company-reported
An internal benchmark with no public tasks, scoring or held-out set; only what he said about the curve was read.
In vivo CAR-T: about six months and a couple of million dollars to the IND point for a preclinical asset, against about five years and $100 million before.
Company-reported
No data and no comparator named. The six months matches the July interview, where Beam put it at the point where one might think about filing an IND and said Lila will not file; the cost figures are new. Lila's own post of 2026-08-28 puts the saving at about ten-fold in R&D cost, and Anyscale's conference recap at non-human-primate efficacy for about 30 times less.
mRNA therapies designed with more favourable clinical profiles than those on the market (fewer doses, fewer immunogenic reactions).
Company-reported
The talk gives no data. Lila's own post of 2026-08-28 gives cell figures: an EPO mRNA producing as much protein on day 15 as the benchmark did on day 2, and longer half-lives for five more proteins' mRNAs, compared only with published sequences; there is no data on doses or immune reactions.
Platinum-group-free electrocatalysts reach about the level of platinum-group catalysts at roughly ten thousand times lower cost.
Company-reported
No data; the basis of comparison is not given; in the timeline entry of 2026-09-24 the best material is mostly palladium, itself a platinum-group metal.
A plan to scale the model to one trillion parameters by the end of the year, and to open a 100,000 sq ft facility in early 2027.
Target
Targets.
Knowledge from all domains should live in one model, since specialising forgoes abilities that come only from widening the scope.
Guest's judgment
Beam's view; there is no general-versus-specialist comparison.
An automated scientific lab is hard because it rolls up hard problems in science, software and post-training.
Not yet closed
Said plainly by Beam himself.
Whether the scaling curve is redrawn on public tasks and a held-out set.
Would show up as A benchmark released by Lila or an outside party, stating the tasks, scoring and held-out set, with a curve others can reproduce.A side-by-side of the post-trained model against frontier models on the quantum-dot tasks.
Would show up as Results naming the task set, scoring and the models compared.Whether the CAR-T preclinical asset is named and a partner takes it to an IND, since Lila says it will not file one.
Would show up as A named asset with its data, published by Lila or a partner, or a partner's IND acceptance.Whether a one-trillion-parameter model is released, with its training compute.
Would show up as A Lila release stating parameter count, training compute and evaluation.Checked on Said: through the end of this yearThe basis of the electrocatalyst cost comparison, and a best material free of platinum-group metals.
Would show up as A paper or release stating the comparator, the cost basis and the material composition.First evidence arXiv 2609.30133 (2026-09-24) states its comparators and composition: the best material, InMnPdOx, is 99.91% palladium, a platinum-group metal; the strongest platinum-group-free comparator, CoFeZrOx, started about 90 mV worse, and the cobalt-zirconium oxides failed in acid within 4 hours. The paper gives no cost basis.Whether the 100,000 sq ft facility opens on time, and which experiments it runs.
Would show up as A release by Lila or a local announcement giving the opening date and the workflows inside.Checked on Said: at the beginning of 2027Later evidence that tests or limits what was said, and targets whose date has passed.
Lila reported an iridium-free oxygen-evolution catalyst: InMnPdOx, picked from 2,942 oxides over 53 systems and 99.9% palladium by metal, stayed below 0.5V overpotential for 1,000+ hours in acid. It is a published electrocatalysis result from Lila's platform; its best material is mostly palladium and the paper gives no cost basis, so it does not test the talk's platinum-group-free, ten-thousand-fold cost claim. 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.Timeline entry →
Read 2026.10.03 The full transcript of the talk, read in full. It is Beam's keynote at Anyscale's own Ray Summit (24 to 26 August 2026, San Francisco), published on video on 17 September; Beam thanks Anyscale for a partnership on scaling Lila's training compute, and Anyscale's video description says the model is trained on Anyscale, so this is the company speaking for itself at a partner's event; for the charts and video he showed, only what he said aloud was read, not the underlying data on the slides. Open the interview ↗