AI models can now help run physical science experiments

Transcript

[00:02] Scientists come up with theories
about how the world works.
[00:09] But just coming up with a
theory isn’t good enough.
[00:14] And so we have to build experiments,
[00:16] physical measurement devices,
to test our theories.
[00:21] That process of building the experiment
takes maybe 80% of a scientist’s time.
[00:28] They’re building devices,
they’re setting things up,
[00:32] they’re debugging hardware,
debugging software.
[00:37] These are things that aren’t
really related to doing
[00:40] science, but it’s what
makes science actually work.
[00:46] When I joined Anthropic, I had a
vision for using AI to accelerate
[00:52] running scientific experiments,
But I thought it was a pie in the
[00:57] sky, crazy idea, until I saw the
work of neuroscientist Arco Bast,
[01:03] who studies how memories are
formed in the brain in real time.
[01:09] I’m in the lab for a year now, and I’m
setting up a very difficult experiment.
[01:18] So the stuff is over here.
[01:23] That’s a custom -built microscope.
[01:25] You see the laser beam in here.
[01:29] That’s actually what’s happening
when you’re imaging in the
[01:31] brain, that you have this laser
beam that scans and moves.
[01:35] It’s really, really important that
everything is precisely aligned.
[01:40] There are so many components,
and I just want to
[01:42] have them talk to each
other in a seamless way.
[01:45] The problem is that each device has
a different language that it speaks,
[01:50] and getting the devices to talk to each
other in their languages is very difficult.
[01:56] But Arco figured out a way to do this
that could work between any two devices.
[02:02] Set beam one to 50% power.
[02:05] Yeah, we got a beam.
And the beam is there.
[02:08] When I was standing in that
room watching him run his
[02:11] experiment, I had kind of
an epiphany in that moment.
[02:15] what he had built wasn’t
just applicable to this lab.
[02:22] This idea could be used to have AI run
any science experiment in the world.
[02:29] I was basically speechless.
[02:33] Is this ours? I believe
everything on the table is ours.
[02:37] We’ve got to start putting this together.
[02:38] So Arco and I started working together
to create a general way for AI
[02:45] to interact with devices, which we’re
calling Model Hardware Standard.
[02:50] Look, with a very sophisticated
microscope, and this microscope has
[02:54] so many degrees of freedom, you
should see it moving now, right?
[02:58] Yay, moved. It moved.
[03:01] It did. Yes, I think
everything looks good.
[03:04] Okay. Okay.
[03:07] Once we had a working prototype,
we had to test this on other
[03:11] devices to see how it worked
beyond just neuroscience.
[03:24] Go to the left boundary first.
[03:27] One of the first things that I did was
to define the safe range of this arm.
[03:31] How far is it allowed
to go out of the table?
[03:33] And so if we ask Claude to maybe
intentionally move out of the safety range,
[03:39] you can kind of see that this motion,
this movement was refused by MHS.
[03:45] Oh, that’s incredible. Can
it grab anything right now?
[03:48] Oh, great question.
[03:51] We’ve never asked Claude
to do this from scratch,
[03:53] and so I have no idea what
it’ll cook up for us.
[03:56] There are turns that I need to… Oh.
[04:02] Wow. Oh.
[04:04] Wait, wait, wait, what?
[04:08] What?
[04:11] Oh my God.
[04:13] Okay, that was sick.
[04:15] That was really cool.
[04:17] The mere fact that I was able
to build this from scratch
[04:21] today and it achieved it
in a matter of minutes.
[04:26] That’s insane.
[04:29] And of course, we also have to
work with the manufacturers
[04:33] and the vendors who build the devices.
[04:35] So we started working with
Danaher to get Claude
[04:38] to connect with their Leica microscope.
[04:44] We have to remind ourselves
that Claude has never
[04:47] seen anything related to
this application before.
[04:50] It will make mistakes. So we can think
about this as an iterative process.
[04:55] You have to imagine that I as a scientist
spent weeks making this sample alive
[05:00] to this point of view and I spent
thousands of dollars in ingredients.
[05:05] I see. Okay. If you break
them, your experiment is gone.
[05:15] You can even see how it’s thinking.
[05:18] It’s now brightly trying to change settings
of the microscope to get an image.
[05:25] and I think we need to help it.
[05:28] What’s interesting here is
you can see that when I asked
[05:31] it to switch to a higher
magnification, it is very aware
[05:35] of going to a higher mag can
crash it into the sample.
[05:39] It’s amazing.
[05:42] It knows how to operate a microscope.
[05:45] Do you think we’re at a point
now, I could enter a bunch
[05:47] of commands and get a focused
image and query what is the
[05:51] image and add false color
and all that type of stuff?
[05:54] We should try it.
[06:00] That’s incredible.
[06:01] That’s awesome. What is the
different colors right now?
[06:05] Could we ask it? Magenta red slash pink.
[06:09] Lignified cell walls.
[06:11] These things are the cell
walls. That’s great.
[06:13] What we accomplished in a
day is pretty transformative.
[06:18] Claude walked in. We told him nothing.
He’s just trying to figure it out.
[06:21] Tomorrow I think we should try to give
it, you know, treat it as a colleague.
[06:38] Oh, there’s a lot going on in there.
[06:42] Let’s say a scientist
wanted to look at this,
[06:45] they would have to sit and wait
and keep tracking it. Correct.
[06:48] For like hours. Correct.
[06:49] Yeah, I had a nice one, but it swam away.
[06:53] I’m wondering if Claude can write
a program that can track that.
[07:08] Claude is almost done with the
initial script to do the tracking.
[07:13] That’s off. Oh, shoot.
[07:15] Yeah.
[07:17] Yeah, damn.
[07:18] I actually think it needs to build
a UI so we can see what it is doing
[07:22] because just running a script in
the background is not acceptable.
[07:28] This looks good. Yo, that’s awesome.
[07:32] It’s doing what it should. Yeah, it is.
[07:35] It’s tracking. It’s tracking.
Three minutes, four minutes.
[07:37] It’s been tracking for a few minutes.
No way. We’ve just been watching it.
[07:41] Chasing algae.
[07:42] Amazing that we managed that.
[07:46] It managed it.
[07:48] This is a prototyping system
somehow, or a very dynamic
[07:51] system where you can create
something very quickly.
[07:54] Maybe it’s even good enough
for some science applications.
[07:57] The PhD can do his PhD work
faster because he doesn’t
[08:01] need to spend two years to get it
running. He only needs two months.
[08:04] And then he can focus on biological
questions, which is what our goal is.
[08:11] I still can’t believe it.
[08:13] We’re very impressed.
[08:15] I think Model Hardware Standard opens the
door for a lot of new types of science.
[08:22] The obvious example is in pharmaceuticals.
[08:27] Here at Genentech, we make
medicines for patients
[08:31] with serious and even life
-threatening diseases.
[08:36] It takes many, many
iterations to make a drug.
[08:41] We will test thousands or even hundreds
of thousands or even millions of
[08:47] molecules to find the right molecule
that will really help patients.
[08:55] So we are going to start an
experiment where Claude will
[09:00] run a series of operations
and then interpret the data.
[09:04] If you aspirate out of
a well that has bubbles,
[09:07] you’re not getting the
correct transfer amount.
[09:09] If we were trying to aspirate
out of this, we’re going
[09:12] to get the proper amounts in
the wells with no bubbles
[09:16] and then improper amounts in
the ones with the bubbles.
[09:18] With Claude, we could potentially
check during the production
[09:22] runs for these bubbles to
see if they are happening,
[09:25] and hopefully it has the context
and knowledge to make adjustments.
[09:31] Claude will do some execution,
take the reading and then
[09:36] change the parameters of the
execution slightly to see if
[09:40] it can improve the experiment
overall in a closed loop.
[09:44] Speeding up this loop means
we are just able to make more
[09:52] shots on the goal and can
get to the answers faster.
[09:57] This is fewer bubbles.
[10:01] It’s got bubbles in two, but not the rest.
[10:04] So this is better.
[10:08] This is really the first time in
history where we are enabling
[10:13] AI to interact with the physical
world in drug discovery.
[10:21] It’s definitely historic, yeah.
[10:26] I think it’s very difficult for us
to predict how AI and model hardware
[10:32] standard will affect the world
30, 50 years in the future.
[10:38] These are the best moments
when there’s something
[10:41] I couldn’t do before and now I can do it.
[10:43] Something I couldn’t see before and
now I can see it. That’s the drive.
[10:47] I want to understand things we
don’t understand right now.
[10:51] Imagine what we’ll see in drug
development, in quantum computing,
[10:57] in nuclear fusion, in big
technologies that could change the
[11:01] world when scientists have
access to this technology.