Shaping Science with Nick Dirks
Oct 21, 2026 · 47 min · 12 segments
Alek Kemeny, a member of Anthropic's Beneficial Deployments team, was trained as a quantum physicist and has spent the past year focused on a novel question: What can AI models safely do in the…
Nick DirksHostAnd so, you know, one thing I'll say is I think products like Cloud Science are very important, right? So essentially, Cloud Science is like Cloud Code, where we developed what we like to call an agent harness, or maybe some folks are familiar with agent scaffolding.
You know, it's all the things that you put around the model to get it to be able to do tasks in, you know, kind of a like productive context.
And so, you know, we kind of took the same principles as we took when we built Cloud Code and applied them to the scientific context and optimizing this agent harness to do science.
And, you know, so traditionally what this looks like is, you know, you are focused on Being able to help someone do literature review much faster.
Really dig in and get what you need out of analyzing thousands of papers at once.
Similarly, when you have data that you've collected after an experiment, what math, what statistics can you do to really juice as much out of that data as possible? These are core pieces of the scientific method that, you know, AI historically, you know, it has the promise of being able to help with.
And what we've seen is my background as a grad student, I had way more ideas that I wanted to test than I was able to actually run on the experiment.
Now that problem is becoming even bigger because AI is helping people design more experiments that are even more likely to succeed.
And so, you know, cloud science Helps with all these parts of the scientific method, but it is missing this one key piece of running the experiment.
And so now, you know, you see scientists have all these ideas and they're really good ideas because, you know, they're working with the models to help develop them.
But it's like, well, now I've got, I had 10 experiments I wanted to run before, but I could only run one in a year.
And so, you know, my view is that what MHS does, it comes in, it really lowers the barrier to entry to what you can run on your same experimental platform.
And this kind of gets into that first part of your question around robotics, right? What we're trying to do with MHS is bring what has traditionally been limited to a lab automation setting that you would see in a pharmaceutical company, you know, like the work that we did with Genentech, and bring that into, you know, kind of a flexible, you know, grad student run lab, right? And so in a pharmaceutical setting, you have robotics that are optimized for chemistry experiments.
But, you know, when you are thinking about these academic experiments with hundreds of different components and some of them are handmade and all the wiring is bespoke, it's like, you know, you're kind of the person who's building the thing from the ground up.
And so, you know, one thing I'll say is I think products like Cloud Science are very important, right? So essentially, Cloud Science is like Cloud Code, where we developed what we like to call an agent harness, or maybe some folks are familiar with agent scaffolding.
You know, it's all the things that you put around the model to get it to be able to do tasks in, you know, kind of a like productive context.
And so, you know, we kind of took the same principles as we took when we built Cloud Code and applied them to the scientific context and optimizing this agent harness to do science.
And, you know, so traditionally what this looks like is, you know, you are focused on Being able to help someone do literature review much faster.
Really dig in and get what you need out of analyzing thousands of papers at once.
Similarly, when you have data that you've collected after an experiment, what math, what statistics can you do to really juice as much out of that data as possible? These are core pieces of the scientific method that, you know, AI historically, you know, it has the promise of being able to help with.
And what we've seen is my background as a grad student, I had way more ideas that I wanted to test than I was able to actually run on the experiment.
Now that problem is becoming even bigger because AI is helping people design more experiments that are even more likely to succeed.
And so, you know, cloud science Helps with all these parts of the scientific method, but it is missing this one key piece of running the experiment.
And so now, you know, you see scientists have all these ideas and they're really good ideas because, you know, they're working with the models to help develop them.
But it's like, well, now I've got, I had 10 experiments I wanted to run before, but I could only run one in a year.
And so, you know, my view is that what MHS does, it comes in, it really lowers the barrier to entry to what you can run on your same experimental platform.
And this kind of gets into that first part of your question around robotics, right? What we're trying to do with MHS is bring what has traditionally been limited to a lab automation setting that you would see in a pharmaceutical company, you know, like the work that we did with Genentech, and bring that into, you know, kind of a flexible, you know, grad student run lab, right? And so in a pharmaceutical setting, you have robotics that are optimized for chemistry experiments.
But, you know, when you are thinking about these academic experiments with hundreds of different components and some of them are handmade and all the wiring is bespoke, it's like, you know, you're kind of the person who's building the thing from the ground up.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.