Geoffrey von Maltzahn, Lila's co-founder and CEO, says post-training on under 2% of the data from Lila's own autonomous labs lets its models pass household-name models and PhD scientists working open book; the first scaled AI science factory covers about five football fields, and the quarter of capacity opened to partners has sold out.
Post-trained on Lila's own lab data, he says, its models pass household-name models and PhD scientists with under 2% of it
Others are stuffing the internet into bigger models, von Maltzahn says; Lila post-trains models that have read the literature on data from its own autonomous labs, and sees a new scaling law for scientific intelligence. On the hardest benchmarks Lila has built for materials, chemistry and life science, under 2% of that data takes its models past the household-name models and past PhD scientists in each domain working open book. He calls this conclusive evidence that scaling the labs and the compute will predictably yield higher scientific intelligence.
The road to scientific superintelligence runs like games, he argues, and the largest verifier wins
After ChatGPT, von Maltzahn says, some assumed that with unlimited compute and parameters, unlimited intelligence would flow from the internet; his view is that the internet yields something like human intelligence, and models trained on cave paintings would not have got far in science. Superhuman chess and Go did not come from training on the best human games but from letting the model play itself. Maths and coding advance the same way, because results can be verified on a computer over and over. Most scientific hypotheses cannot be, so whoever has the largest verifier will have the highest scientific intelligence, he says, and building one means automating the scientific method.
It can be done now because a general scientific model and dexterous robots have arrived together
Asked why now, he says automating the scientific method needs two things: a broad, general scientific mind, which is now feasible, and a way to do what his hands did in grad school, the dexterous steps done in every lab in the world, which only robotics at this moment allows. Lila connects frontier and applied AI to a full range of frontier and conventional robots, aiming to turn any scientific endeavour into, in effect, a list of instruments: if the instruments are on the platform, the science can be set in motion.
The lab you could talk to three years ago now runs its own experiments and updates the model
The episode was recorded at Lila's temporary site in Cambridge, which he calls the Kitty Hawk of autonomous science: three years ago the team built its first lab you could talk to there, a kind of autocomplete on the scientific method that suggested the next step and analysed data on its own. Since then they have hardened it into what they call AI science factories: when past data sets cannot answer a question, the model calls up an autonomous lab through software, runs the experiment, analyses the data and updates the model the hypothesis came from. Handling liquids is at the centre, shared across life science, chemistry and materials, and the setup now synthesises quantum dots, nanomaterials whose discovery won a Nobel Prize a few years ago, over and over.
The first scaled factory covers about five football fields, one platform for life science, chemistry and materials
The first scaled AI science factory has just entered its first stage; he puts it at about five football fields, and the host says it opens formally in January. Historically, he says, tackling any one of life science, chemistry or materials took a specialised approach; Lila has built one general stack of models, software and robots, so that under one roof it can work with customers on cancer medicines and on new superconductors.
Running around the clock gives over 8,000 hours a year, four times a nine-to-five: his compound interest machine for science
He takes two lines from Charlie Munger: understanding the power of compound interest and the difficulty of achieving it is the heart of understanding a lot of things, and the first rule of compound interest is never to interrupt it unnecessarily. The scientific method is a form of compound interest, he says, each turn adding to our mastery of the universe, and he interrupted his own all the time in grad school, because people sleep and take weekends. An institution's science runs something like nine to five, about 2,000 hours a year; around the clock it is more than 8,000, and decisions move from meetings onto GPUs, a million times faster. That is the compound interest machine for science he is building.
The quarter of capacity opened to partners sold out fast, and he says the platform needs hardening for it
Lila opened its commercial doors only a few months ago and has said little about it outside, he says: about six months ago its president of corporate development, who came from General Catalyst, took about 25% of lab capacity for commercial work and swiftly sold it out, and a meaningful share of capacity is close to booked for the foreseeable future. Partners are looking for superconductors, magnets and quantum materials, and for medicines that could reach clinical trials within a couple of years. That is the first business, access to the full platform for R&D; it is strengthening the platform, he says, and showing where it needs hardening, which a two-day leadership offsite had just been about.
In two years, he predicts, a scientist will spend tens of thousands a year on compute, paid to the strongest scientific models
The medium-term business, and the most important, is selling the scientific intelligence itself, he says. Two years ago the market for agentic coding was zero, and now there are hundreds of billions of dollars of demand; the market for agentic scientists is zero today. In about two years, he predicts, you will not get to be a scientist without spending tens of thousands of dollars a year on compute, which across ten million scientists is hundreds of billions flowing to whoever has the strongest scientific models. Human science is a $10 trillion market in aggregate, he adds, so intelligence above ours should be worth far more, a market he half-jokingly calls semi-infinite.
The scaling law and the sold-out capacity are his own account, given to one of Lila's investors
No number backs the scaling law: the benchmarks are Lila's own, with no word on which models were compared, on how many questions or by how much, and no paper or outside evaluation; Andrew Beam's Ray Summit curve on the record is an internal figure too. The sold capacity comes with no customer, value or throughput. The host, Jamie Montgomery, co-founded March Capital, which took part in Lila's seed round and Series A, and he says he is one of Lila's investors.

