Scientists come up with theories about how the world works. But just coming up with a theory isn't good enough. And so we have to build experiments, physical measurement devices, to test our theories. That process of building the experiment takes maybe 80% of a scientist's time. They're building devices, they're setting things up, they're debugging hardware, debugging software. These are things that aren't really related to doing science, but it's what makes science actually work.

When I joined Anthropic, I had a vision for using AI to accelerate running scientific experiments, But I thought it was a pie in the sky, crazy idea, until I saw the work of neuroscientist Arco Bast, who studies how memories are formed in the brain in real time.

I'm in the lab for a year now, and I'm setting up a very difficult experiment. So the stuff is over here. That's a custom -built microscope. You see the laser beam in here. That's actually what's happening when you're imaging in the brain, that you have this laser beam that scans and moves. It's really, really important that everything is precisely aligned. There are so many components, and I just want to have them talk to each other in a seamless way. The problem is that each device has a different language that it speaks, and getting the devices to talk to each other in their languages is very difficult. But Arco figured out a way to do this that could work between any two devices.

Set beam one to 50% power. Yeah, we got a beam. And the beam is there.

When I was standing in that room watching him run his experiment, I had kind of an epiphany in that moment. what he had built wasn't just applicable to this lab. This idea could be used to have AI run any science experiment in the world. I was basically speechless.

Is this ours? I believe everything on the table is ours. We've got to start putting this together. So Arco and I started working together to create a general way for AI to interact with devices, which we're calling Model Hardware Standard. Look, with a very sophisticated microscope, and this microscope has so many degrees of freedom, you should see it moving now, right? Yay, moved. It moved. It did. Yes, I think everything looks good. Okay. Okay.

Once we had a working prototype, we had to test this on other devices to see how it worked beyond just neuroscience.

Go to the left boundary first. One of the first things that I did was to define the safe range of this arm. How far is it allowed to go out of the table? And so if we ask Claude to maybe intentionally move out of the safety range, you can kind of see that this motion, this movement was refused by MHS. Oh, that's incredible. Can it grab anything right now? Oh, great question. We've never asked Claude to do this from scratch, and so I have no idea what it'll cook up for us. There are turns that I need to... Oh. Wow. Oh. Wait, wait, wait, what? What? Oh my God. Okay, that was sick. That was really cool. The mere fact that I was able to build this from scratch today and it achieved it in a matter of minutes. That's insane.

And of course, we also have to work with the manufacturers and the vendors who build the devices. So we started working with Danaher to get Claude to connect with their Leica microscope. We have to remind ourselves that Claude has never seen anything related to this application before. It will make mistakes. So we can think about this as an iterative process. You have to imagine that I as a scientist spent weeks making this sample alive to this point of view and I spent thousands of dollars in ingredients. I see. Okay. If you break them, your experiment is gone.

You can even see how it's thinking. It's now brightly trying to change settings of the microscope to get an image. and I think we need to help it. What's interesting here is you can see that when I asked it to switch to a higher magnification, it is very aware of going to a higher mag can crash it into the sample. It's amazing. It knows how to operate a microscope. Do you think we're at a point now, I could enter a bunch of commands and get a focused image and query what is the image and add false color and all that type of stuff? We should try it. That's incredible. That's awesome. What is the different colors right now? Could we ask it? Magenta red slash pink. Lignified cell walls. These things are the cell walls. That's great.

What we accomplished in a day is pretty transformative. Claude walked in. We told him nothing. He's just trying to figure it out. Tomorrow I think we should try to give it, you know, treat it as a colleague.

Oh, there's a lot going on in there. Let's say a scientist wanted to look at this, they would have to sit and wait and keep tracking it. Correct. For like hours. Correct. Yeah, I had a nice one, but it swam away. I'm wondering if Claude can write a program that can track that. Claude is almost done with the initial script to do the tracking. That's off. Oh, shoot. Yeah. Yeah, damn. I actually think it needs to build a UI so we can see what it is doing because just running a script in the background is not acceptable. This looks good. Yo, that's awesome. It's doing what it should. Yeah, it is. It's tracking. It's tracking. Three minutes, four minutes. It's been tracking for a few minutes. No way. We've just been watching it. Chasing algae. Amazing that we managed that. It managed it.

This is a prototyping system somehow, or a very dynamic system where you can create something very quickly. Maybe it's even good enough for some science applications. The PhD can do his PhD work faster because he doesn't need to spend two years to get it running. He only needs two months. And then he can focus on biological questions, which is what our goal is. I still can't believe it. We're very impressed.

I think Model Hardware Standard opens the door for a lot of new types of science. The obvious example is in pharmaceuticals. Here at Genentech, we make medicines for patients with serious and even life-threatening diseases. It takes many, many iterations to make a drug. We will test thousands or even hundreds of thousands or even millions of molecules to find the right molecule that will really help patients. So we are going to start an experiment where Claude will run a series of operations and then interpret the data. If you aspirate out of a well that has bubbles, you're not getting the correct transfer amount. If we were trying to aspirate out of this, we're going to get the proper amounts in the wells with no bubbles and then improper amounts in the ones with the bubbles.

With Claude, we could potentially check during the production runs for these bubbles to see if they are happening, and hopefully it has the context and knowledge to make adjustments. Claude will do some execution, take the reading and then change the parameters of the execution slightly to see if it can improve the experiment overall in a closed loop. Speeding up this loop means we are just able to make more shots on the goal and can get to the answers faster.

This is fewer bubbles. It's got bubbles in two, but not the rest. So this is better. This is really the first time in history where we are enabling AI to interact with the physical world in drug discovery. It's definitely historic, yeah.

I think it's very difficult for us to predict how AI and model hardware standard will affect the world 30, 50 years in the future. These are the best moments when there's something I couldn't do before and now I can do it. Something I couldn't see before and now I can see it. That's the drive. I want to understand things we don't understand right now. Imagine what we'll see in drug development, in quantum computing, in nuclear fusion, in big technologies that could change the world when scientists have access to this technology.