Latent Space

Lila Sciences Is Turning the Lab Into an Infinite Token Generator

Andy Beam and Rafa Gomez-Bombarelli· CTO; Co-founder and CSO, Physical Sciences at Lila Sciences
·~101 min·English·Latent Space
TrainingReasoningAI InfrastructureAI Company
TL;DR

Lila Sciences applies the bitter lesson to science: a general reasoning model proposes hypotheses, a flexible automated lab runs the experiments, and nature verifies the results, turning experiments into the next internet-scale dataset while making the model itself the product.

01The Data Wall

Science Is the Next Internet

Pre-training consumed the entire internet, so the next internet-scale dataset has to come from running experiments — making the physical lab a new scaling axis for data.

we have but one internet. It's the fossil fuel. We fracked. We got every ounce of data that we could out of the internet, but it's gone.

— Andy Beam, Latent Space
Key Insight
The framing quietly reclassifies experiments as data infrastructure rather than science. If tokens are the binding constraint, then whoever can manufacture novel, verified tokens fastest sets the pace of progress, which is a very different race than owning the single best model.
02Verifiable Rewards

Nature Is the Ultimate Verifier

RL works when a reward can separate good data from bad, and Lila makes the physical experiment itself the reward signal — turning the scientific method into a scaled verifier.

what at Lila we believe is that actually science running the scientific method and using nature and experiments as verifier is like the ultimate version of that.

— Andy Beam, Latent Space
Key Insight
Math and code RL work because answers are cheap to check. The bet here is that a lab can be that checker for chemistry and biology, but the check has a runtime measured in days, so the hard engineering problem becomes scheduling verification, not defining the reward.
03The Lab as Software

Below the API Line

Every step in the lab is an API call that resolves to a robot arm or a human arm, so Lila optimizes for flexibility rather than automating everything.

We're not automation maximalists. We are actually sort of like token generation maximalists and flexibility maximalists.

— Andy Beam, Latent Space
Key Insight
Treating the lab as an API means a person uncapping a test tube is just an unautomated function call. That keeps the data pipeline complete even where robotics fails, so a human below the API line is a design choice, not a stopgap.
04Business Model

The Model Is the Product

Lila refuses to become a drug or materials company: the model is the asset, while the lab's compounding data generation is the moat.

we're much more of like a neo lab, trying to think of a new way to push forward capabilities of a core reasoning LM-based model.

— Andy Beam, Latent Space
Key Insight
A normal biotech concentrates risk in one clinical asset and bets the company on it. Lila instead makes the model the product and uses the lab as a compounding data engine that spreads learning across programs, the same inversion that turned search and social from products into platforms.
05Cross-Domain Transfer

The General Model Beats the Domain-Specific Model

A single model trained on 10 trillion experimentally-verified reasoning tokens across biology, chemistry, and materials beats narrow domain-specific models, because knowledge spills over between fields.

if you went back 10 years and you tried to create like a coding assistant model, you would just get coding data. You wouldn't also get Shakespeare poetry, carnitas recipes.

— Andy Beam, Latent Space
Key Insight
This is the direct rebuttal to the objection that if you already have the data you do not need the model. A general model needs less data in any single domain because adjacent domains pre-pay the cost, and experimentally-verified reasoning traces round to zero on the public internet, which is exactly why they lift a weaker model above the frontier.
06Compression

The Zero-FTE Startup

Two or three scientists plus the model and lab can do five years of biotech work in six months for a tenth of the cost, pointing toward virtual startups that run on the platform instead of building labs.

a couple scientists who have domain knowledge and a combination of the model plus platform can do five years worth of biotech work over a six-month period for 10% of the total investment.

— Andy Beam, Latent Space
Key Insight
A team of two or three people reached non-human-primate data in six months, reporting results better than Capstan's after AbbVie bought Capstan for about $2.1 billion. The leverage is not just accelerating discovery; it is improving the odds across the whole downstream portfolio, one loaded die at a time.
07Physical Infrastructure

The Lab Should Feel Like a Data Center

Today's benches are built for human arms; Lila's endpoint is a lights-out, 24/7 facility measured in data per square foot.

the lab of the future should not be made for people to easily walk into it. It should feel like a data center where you go and you see the rows of server racks.

— Andy Beam, Latent Space
Key Insight
The tell is the world's largest collection of voided warranties: custom drivers and firmware, even a vision-language model clicking through a Windows 95 instrument. The hard part was never the levitating plates, it was the software wrapper that turns dozens of incompatible machines into one addressable computer.
08Open-Endedness

You Can't Just Be a Good Test-Taker

Scaled RL answers hard questions ruthlessly, but scientific superintelligence also has to ask interesting ones, which is why open-endedness is a separate outer loop.

you can't have scientific superintelligence if you're just a good test taker.

— Andy Beam, Latent Space
Key Insight
Hiring Ken Stanley to own machine creativity concedes that optimization alone plateaus into a very fast test-taker. The harder capability is taste, knowing which questions are worth asking in the first place, which is an outer loop over the reasoning rather than simply more of it.