Help is on the way
Transforming biology with smarter robots
I wrote this post after a conversation with my friend Stephen, who asked whether a camera recording everything a biologist does could revolutionize biological automation. My short answer is yes. My longer answer follows below.
Biologists strive to automate tasks and have achieved some success. Most high-throughput platforms and clinical labs worth their salt employ some form of pipetting robot that replaces one or a few monotonous critical tasks. These tasks would otherwise fall to a bench scientist who, over time, might make a mistake. I enjoy lab work, but given the choice, I would never set up a PCR, grow colonies, or reorganize my primers. I suspect most biologists aim for an ideal where they can spend most of their time thinking of cool experiments, leading me to my first point:
Current lab automation has (mostly) been a failure.
What I mean by this is that most robotics focus very heavily on accomplishing some single thing, like PCR, but force the user to fit their experiment to match the parameters of the machine. Unlike a human brain, which can take goals – “I want whole genome sequencing results from all patient samples and 1 control included” – and execute many different aspects of a workflow, lab equipment has been built in a kind of way that limits what people envision is possible. Put differently, every piece of "automated" equipment I have ever used in a lab rested on an understanding between me and the original designer that I would supply the reagents, ideas, experimental set up and interpretation of the result, and they would supply the mechanical hands to do it. Taking this a step further, the pipetting robots of today are essentially built to fit established protocols of the field, and are by definition not capable of anticipating the needs of the future of scientific development – they cannot really speed up discovery that much. This is not ideal, and is not something scientists should collectively be content with, which brings me to my next point, which I make rather begrudgingly:
Tesla actually might have something to teach us biologists.
Historically, Tesla never really impressed me, mostly because I don’t care much for nor understand cars (other than a profound appreciation for first-gen Miata featuring highly fun pop-up headlights), but a recent update on their Optimus bipedal robot caught my eye. Using vision and joint encoders, their bipedal robot is now able to sort objects independently. Though it’s true that executing biology at a high level involves much much more than sorting blue and green legos, Tesla successfully demonstrated a suitable approach that can unlock genuinely creative science. In terms of fitting software with hardware, there are new players entering the field, most notably DeepMind, who announced partnerships with a bunch of academic labs covering a variety of robot types to enable general-purpose robotics learning. If we take this concept, and we overlay it with what seems to be working for building extremely safe and capable self-driving cars, we arrive at what I think is actually a reasonable course of action to generating the right kinds of data for training smart lab-ready robots:
Let’s start (video) hoarding
Imagine a scenario where a bright scientist picks out her favorite Ray-Ban Meta Smart Glasses, and records everything she does from the moment she enters the building until she leaves. She can very quickly amass a horde of data covering an enormous number of operations spanning cell culture, cloning, freezer organization and even proper note taking. The specific advancement that enables real progress towards smart robots today that did not exist 10 years ago is not just the ability to give modern machines the ability to learn spatial reasoning, but also to apply the power of LLM’s or small industry-specific models (I’m not picky, I’d take either or both). What makes the all-encompassing video data so important towards achieving this goal is that it breaks big tasks into small digestible ones. Cell culture or PCR are not just one unified thing – they are many steps that each contribute step-wise towards a finished product. While current pipetting robots are very adept at accomplishing a small number of steps in a larger workflow, they force the user to fit the machine’s limitations, leading down a never-ending path of single-task robots. Giving a machine the data necessary to learn that doing cell culture involves much more than following a protocol in a paper will get us closer to fully automated labs.
Smart robots can speed up the engine of biology
There are additional advantages to this approach that contribute beyond the concept of free-thinking robots. For starters, the act of using video to record lab protocols will capture all of the little micro-steps that typically don’t make the final cut of a protocol. Anyone who has done lab work knows that what you read is only some small part of what it actually takes to bring an outside protocol in-house (think: what does “thaw on ice” mean – does it mean in ice, on top of ice?). Multimodal models seem suitable to translate this raw video data into richer natural language protocols that can serve as actual templates of action for others. Progress towards better automation would create good text search of videos, enabled by people narrating what they’re doing on a small amount of video that gives a joint encoding, thereby leading to the world’s best searchable database of lab protocols and content.
Like language models, I anticipate these types of machines will not be infallible. They will hallucinate, they will do weird things, and they will likely break stuff. But the benchmark they are working against – humans in the lab – are not perfect, particularly newly trained ones who are doing the bulk of discovery work. When I think about the 10X biologists, I think of a scientist who does a lot of high quality experiments that are born of good ideas and rarely die due to bad execution, which is really what this type of automation promises. Ultimately a fully autonomous lab robot delivers one thing: a faster path from concept to clinic, an objective that is deeply important to me. I get pretty excited about the idea of a fully autonomous and capable machine that can run broader experiments on a continuous basis. Where we are today is robots that execute the will of scientists as a very limited extension of their hands, but where we can be tomorrow is physically versatile robots that turn the average scientist into a well staffed PI.
A note on safety
I understand technology like this might enable a heightened concern for biorisk. I don’t have much to comment on, other than to say there are some really smart people who are thinking about what AI has or hasn’t enabled, and you should read and listen to their thoughts on it.
Here’s a list to consider:
Kevin Esvelt on mitigating catastrophic biorisk; more on bioweapons
Eric Schmidt, to the National Security Commission on Artificial Intelligence
On the other hand, to the degree that this kind of lab setup lends itself to centralization by a few providers, it may end up concentrating lab expertise in fewer hands which might be a net positive. Unclear!