Lila Sciences

A laboratory that runs itself

Walk into one of Lila's labs and the strangest thing is what you don't see: nobody hunched over a bench at 2 a.m. The liquid handlers, robotic arms, imaging rigs, and chromatography columns are all still there. What's missing is the bottleneck - the graduate student, the postdoc, the single human deciding which of a thousand experiments to try next.

Lila Sciences calls the result an "AI Science Factory." Software generates a hypothesis, the robots run the experiment, the model reads the data, and then it decides what to test next - over and over, without waiting for morning. The company's phrase for the ambition is unsubtle: scientific superintelligence. The claim underneath it is simple. Discovery has been rate-limited by human attention for four hundred years. Lila wants to remove the limit.

The lab that never asks for a coffee break. Also never spills the reagents.

Science is slow on purpose, and slow by accident

Here is the uncomfortable part. Most of what makes research slow has nothing to do with how hard the questions are. It's the logistics. A hypothesis waits weeks for a free instrument. A failed run gets noticed on a Tuesday and re-designed the following Monday. Knowledge sits in a PDF nobody re-reads. The scientific method, as practiced, spends most of its time on hold.

The founders' read was that artificial intelligence had gotten very good at the thinking part - reasoning, generating candidates, summarizing literature - while the doing part stayed stubbornly manual. A model can propose ten thousand catalysts before lunch. It cannot, on its own, weigh them, mix them, and measure what happens. Brilliant ideas, in other words, still need hands.

There is a second, quieter problem layered on top of the first. Science doesn't just move slowly; it forgets. A negative result rarely gets published, so the same dead end gets re-explored by a different lab two years later. Methods live in the heads of people who graduate and move on. The institutional memory of a field is, in practice, a stack of papers nobody has time to read end to end. An autonomous system that records every run - the failures included - turns that loss into an asset.

So the bottleneck was never the imagination. It was the gap between a good idea and a measured result, and the leaky bucket of everything learned along the way. Close that gap, keep the memory, and the rate of discovery stops tracking the size of the team. That is the wager Lila is built around, and it is either obvious in hindsight or wildly premature. The next few years will tell which.

Hand the loop to the machine

Lila was incubated inside Flagship Pioneering in 2023 - the same venture-creation firm that built Moderna. The bet its founders made was that you could wire AI into both halves of the work at once: the reasoning and the wet-lab execution. Not a smarter assistant for scientists. A closed loop that is the scientist.

The people making that bet are not lightweight. Geoffrey von Maltzahn, a serial Flagship company-builder, leads as CEO and founder. Noubar Afeyan, Flagship's CEO and Moderna's co-founder, sits behind it. And the scientific anchor is George Church, the Harvard geneticist whose name appears on a sizeable fraction of modern biotechnology. When the chief scientist is a living legend, you've at least bought yourself a credible benefit of the doubt.

A founding team where "we know a guy" means the guy edited the genome.

An operating system for science

Lila's core is an Autonomous Science platform: generative AI models paired with scalable, self-running "AI science units." With human guidance, the system is built to optimize experimentation in any scientific domain - life science, chemistry, materials. Two front doors describe what partners get. Catalyst enhances and accelerates work that's already underway. Creation generates new discoveries and launches products from them.

Crucially, Lila isn't trying to sell you the molecule. It plans to operate as a platform provider - renting access to its models and robotic labs rather than auctioning off the discoveries. It's a bet on the picks-and-shovels, except the shovels are centrifuges, incubators, and imaging systems wired into a brain.

From stealth to unicorn, fast

Receipts, not just slides

Ambition is cheap; data is not. Lila says its platform has already cleared real benchmarks across several domains. Its science-focused language models post state-of-the-art reasoning on scientific problems. The platform has generated genetic-medicine constructs reported to outperform commercially available therapeutics. It has discovered and validated hundreds of novel antibodies, peptides and binders. And on the materials side, it produced non-platinum-group-metal catalysts for green hydrogen - at a fraction of the cost of current commercial options.

None of these results, on their own, settles the bigger question. A great catalyst from one autonomous loop doesn't prove the loop generalizes to the next problem. But taken together they make a useful argument: the system produces outputs you can hold, measure, and compare against the commercial state of the art - not just leaderboard scores on a benchmark someone else designed. For a field crowded with confident slideware, having something to weigh at the bench is the rarer currency.

The backers seem convinced. Beyond Flagship, the cap table runs through Braidwell, Collective Global, Nvidia's NVentures, Analog Devices, IQT, General Catalyst, the ARK Venture Fund and a subsidiary of the Abu Dhabi Investment Authority. When a chip-maker, a defense-tech fund and a sovereign wealth arm all want a piece of the same robotic lab, something is either very right or very expensive. Possibly both. What's clear is that the money arrived faster than almost anyone expected - roughly $550M and a unicorn valuation inside the company's first full year out of stealth.

Aim the machine at the hard problems

The stated mission is to build scientific superintelligence to solve humankind's greatest challenges - health, sustainability, materials. That's a big sentence, and Lila knows skeptics have heard big sentences before. The difference it points to is that its system doesn't stop at a slide deck. It runs the experiment.

What can people actually do with it? If you run R&D - in pharma, chemicals, energy, or advanced materials - the offer is to point an autonomous lab at your hardest screening problem and let it iterate around the clock. New antibodies. Better catalysts. Cheaper materials. The work that used to take a building full of people and a few years, compressed toward the runtime of a model. For a smaller team, it's leverage they could never afford to staff; for a large one, it's a way to run a hundred parallel experiments without a hundred parallel hires.

There's a cultural claim folded into the technical one. Lila keeps a human in the loop on purpose - the platform optimizes experimentation "with human guidance," and its chief scientist is a Harvard geneticist, not a server rack. The pitch isn't that scientists become obsolete. It's that they stop spending their best years on logistics and start spending them on judgment: which questions are worth asking, and what to make of the answers a tireless machine keeps handing back.

Back to the empty lab

Return to that quiet room at 2 a.m. - the one with no human in it. For most of history, the speed of science was the speed of the people doing it. Funding, attention, and the number of careful hands set the ceiling. Lila's wager is that the ceiling is about to move.

If it works, the empty lab isn't a sign that the people left. It's a sign the work didn't stop when they did.

The skeptic's questions are fair and still open. Can autonomous loops generalize past the demos? Will partners trust a black box with their pipeline? Does superintelligence survive contact with a clogged pipette? Lila has the capital, the bench, and a year of unusually loud results to find out. The rest is, fittingly, an experiment.

See it in motion

Links & sources