What text can’t teach
What stands out about the last few years of AI is where models got good first: code, mathematics, law and the contents of papers, all fields that already lived in text. The internet turned out to be a nearly complete record of how people write and a very poor record of what actually happens in the world.
Biology is where that gap is widest. Most of what goes on inside a living cell has never been measured by anyone, and the measurements that do exist are mostly buried. A typical experiment survives as one figure at the end of a paper, while the protocol changes, the runs that failed and the things the scientist noticed but never wrote down are lost, even though those are exactly the parts you would need in order to learn the craft. Nobody becomes a good experimentalist by reading results sections, and there is no reason to expect a model to manage it either.
Why it has to be physical
In code and mathematics a model can grade its own work by writing a program or a proof, checking it and trying again, and that free practice is a large part of why progress there has been so fast. Biology offers no such shortcut. A cell is too complicated to simulate faithfully, so the only way to find out what an experiment does is to run it in the physical world.
That makes biology both the most important field for AI and the one most starved of data. People usually treat intelligence as the hard part, but in biology we think measurement is the harder problem, and better algorithms will not close that gap without many more experiments.
The thinking moves to models
Within a few years, models will do the cognitive core of experimental biology, choosing what to test and deciding what the result means. When that happens, good ideas will be plentiful and the ability to act on them will be what limits progress. A model that thinks like a great scientist but cannot run an experiment has nowhere to go, and what it needs is the part we are building, the eyes and arms in the lab.
AI labs are becoming fabless
The chip industry went through something similar. Chip companies used to build their own factories, until the industry split into companies that design chips and foundries that manufacture them for everyone, and the foundries ended up at the centre of it. We think AI labs will go the same way in biology. They will design the experiments, someone will have to run them, and the most useful partner for that is a lab built around them from the start.
A lab built for a model
Labs today are built around human researchers and run at human speed. Many contract research organisations take about six weeks to return an experiment that takes about five days to run. A person with other work can live with that delay, but a model that could design the next experiment in seconds spends almost all of that time waiting.
A lab built for a model runs far more experiments in parallel than a normal lab would and records far more of what happens, because to a model the record of how an experiment unfolded is worth as much as its outcome. Failed runs count as data as long as the failure is visible in the record, and results stream back during the run so the model can change course before it is over, much as a good scientist does at the bench.
Whatever an experiment produces belongs to whoever ran it, and we never share it, train on it or hand it to another lab.
What changes
If this works, the pace of experimental biology will stop being set by how many skilled people a lab can hire and start being set by how quickly ideas can be tested. We think that is how biological superintelligence gets built, out of an enormous number of real experiments, and it is how models will eventually help cure disease.
The details of how Petrichor runs an experiment are on the How page.