AI Starts Running Experiments by Hand: Scientists Spend 80% of Their Time Building the Rig
Scientists spend 80% of their time building rigs, tuning hardware and debugging instead of doing science. Model Hardware Standard lets Claude drive microscopes and robotic arms directly, compressing two years of setup into two months.
The argument · tap a timestamp to hear it
Scientists spend eighty percent of their time not doing science
The judgment offered at the top of the episode: proposing a theory isn't enough, you have to build an experimental apparatus to verify it, and building that apparatus eats roughly 80% of a scientist's time — setting up equipment, debugging hardware, debugging software. None of this has anything to do with doing science itself, but it is exactly what makes science actually run. This is the foundation for every argument that follows in the episode: if AI can take over that 80%, what accelerates isn't one experiment but the throughput of the entire scientific process.
Every instrument speaks its own language
Neuroscientist Arco Bast is building a custom microscope in the lab, scanning with a laser beam to form images, requiring every component to be precisely aligned. The real difficulty isn't the optics, it's that every instrument speaks a different language, and getting them to talk to each other is extremely hard. Arco found a way of connecting any two instruments generically. This detail explains why MHS is a protocol layer rather than a custom script for one lab — what it solves is device heterogeneity, not one specific experiment.
It only clicked standing in that room
Standing in Arco's lab watching him run experiments, the speaker had an epiphany: what Arco built doesn't only apply to this one lab, and the idea could be used to let AI run any scientific experiment in the world. So the two started collaborating on a general way for AI to interact with instruments, named Model Hardware Standard. This is the episode's turning point: from a concrete problem in one lab to a general protocol.
The safety boundary is refused at the protocol layer
Once the prototype worked, they tested it with a robotic arm. The first thing was to define the safe range — how far the arm is allowed to reach beyond the table. When they asked Claude to deliberately move outside the safe range, the action was refused by MHS. This is the key design of MHS as a protocol layer: safety constraints aren't written in the prompt, they're intercepted by the hardware standard layer. There's a finer observation later: when asked to switch to a higher magnification, Claude clearly realized high magnification could crash into the sample, showing it has an understanding of microscope operation.
Building it from scratch took only minutes
Live, they had Claude grab something from scratch, something they'd never tried before, and nobody knew what it would come up with. It did it, and the reaction in the room was ‘that was sick’. The speaker's judgment: being able to build this from scratch on the day, and hit the goal within minutes, is itself crazy. At the same time they're working with instrument makers, for example with Danaher to get Claude connected to their Leica microscopes. The attitude given here: Claude has never seen this application scenario, it will definitely make mistakes, so treat it as an iterative process.
Crash the sample and the experiment is gone
There's a live reminder about cost: scientists spend weeks keeping a sample alive to this observation state, spend thousands of dollars on reagents, and if the sample is crashed, the experiment is gone. This segment makes the risk of ‘AI operating physical equipment’ concrete — not abstract safety, but irreversible sample loss. At the same time you can see on screen Claude trying to change microscope settings to get an image, and the speaker thinks it needs a hand.
On day two, treat it like a colleague
Day one's conclusion: Claude walks in, you tell it nothing, it figures things out on its own. Day two's approach changed — treat it like a colleague. They then had Claude write a program to track swimming algae, because manual tracking means sitting and staring for hours. Claude finished the initial script, but the speaker pointed out that only running a background script is unacceptable, and it needed to build a UI itself so people can see what it's doing. This is the shift from ‘AI executes commands’ to ‘AI delivers a usable tool’.
Two years of setup compressed into two months
Once the tracking program was running, they watched it chase algae for a few minutes. The speaker's judgment: this is a prototype system, or rather a very dynamic system, that can build things fast, and may already be good enough for some scientific applications. The most concrete payoff — a PhD student doesn't have to spend two years getting the system running, only two months, and the rest of the time can go to biology questions. That's the goal of the whole thing: give engineering time back to scientific questions.
A bubble decides whether the pipetted volume is right
The Genentech scenario: making drugs requires testing thousands, hundreds of thousands, even millions of molecules to find the right one. They want Claude to run a series of operations and interpret the data. The specific mechanism: if you aspirate from a well with a bubble, the transferred volume is wrong — wells without bubbles have the correct volume, wells with bubbles don't. With Claude, you can check during the production run whether bubbles appear, and hopefully it has the context and knowledge to adjust. Claude executes, reads, fine-tunes parameters, forming a closed loop.
The faster the loop, the more hits
The core logic of the Genentech segment: speeding up this loop means more attempts at the target and answers faster. The live verification result was fewer bubbles — two wells had bubbles, the rest didn't, better than before. The speaker's characterization: this is the first time in history AI interacts with the physical world in drug discovery, and says explicitly ‘definitely historic’. This is where the episode lifts MHS from a lab tool to industrial significance.
Nobody can predict what thirty years will bring
The ending returns to motivation: it's very hard to predict how AI and Model Hardware Standard will affect the world in 30, 50 years. The speaker's driving force is those best moments — things that couldn't be done before can now be done, things that couldn't be seen before can now be seen. He wants to understand what he doesn't yet understand. Finally he widens the scope: drug development, quantum computing, nuclear fusion — these big technologies that can change the world, and what scientists will see once they have this technology in hand.
In their own words · checked verbatim
That process of building the experiment takes maybe 80% of a scientist's time.
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.
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.
The mere fact that I was able to build this from scratch today and it achieved it in a matter of minutes.
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.
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.
This is really the first time in history where we are enabling AI to interact with the physical world in drug discovery.
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.
Figures
| Share of scientists' time spent building experiments | about 80% | 0:02 |
| Time for a PhD student to get an experimental system running (with MHS) | from two years down to two months | 7:18 |
| Number of molecules drug R&D needs to test | thousands to hundreds of thousands, even millions | 8:22 |
Glossary
- Model Hardware Standard (MHS)
- A general protocol letting AI interact with any experimental instrument, proposed by Anthropic in collaboration with neuroscientist Arco Bast.
- aspirate
- A lab operation that draws liquid from a well plate; a bubble causes the transferred volume to be inaccurate.
- closed loop
- A cycle in which AI performs an operation, reads the result, fine-tunes parameters and executes again, used to automatically optimize experiments.
How to listen
Engineers and founders working on lab automation, scientific instrument interfaces and AI for Science; plus investors who want to know how AI gets into the physical experiment loop.
The Genentech bubble pipetting details from 8:22 to 9:22; skip ahead if you don't have a lab background.