AI Science Requires Physical Labs To Scale | Nenad Tomašev
MTS
October 20, 2026
Google DeepMind Senior Staff Research Scientist Nenad Tomašev discusses agentic AI in the public interest, the need for physical laboratory infrastructure to validate autonomous scientific discoveries, and addressing value lock-in and social norms in multi-agent systems.
Nenad Tomašev
The core premise of that paper is that intelligence, while it is an enabler for advancing research, opening up all sorts of discoveries, whether it's in areas like biomedicine or material science or other domains of science, that it needs to be followed by similar investments in the infrastructure to deliver on the promise. But then for many sciences that are not strictly computational, you need to have physical laboratories to validate those findings, or disprove them. So if we put all of the investments on the AI side of the equation, we're going to have a world where we have many more hypotheses and conjectures that we can actually process downstream. So the broader scientific community needs also to see investment, more people, more experts, more laboratories, just a broader ecosystem.
SPEAKER_2
All right. Welcome back. I'm live with Nenad Tomašev, who is a Senior Staff Research Scientist at Google DeepMind, working on agentic AI, multi-agent systems, AI for scientific discovery, etc.
So we'll be talking about the virtual agent economies and autonomous science, things like this.
Nenad, we were just talking about how you've been at DeepMind since 2015 So I guess my first question is, how has the work that you're doing evolved and how has DeepMind's work at large evolved and the attitudes evolved?
Nenad Tomašev
I would say in some ways it's changed dramatically. In other ways, it's the same as it ever was. One of the reasons why I joined in the first place, one of the reasons why I've stayed here for so long is the mission that we have, which is, as we say nowadays, to develop a responsibility to benefit humanity. But in any case, seeing intelligence as this driver of progress, enabling people to, as you say, advance science, advance the economy, and generally help people achieve more. So this has been the case even 10 years ago.
It's just that the technology we're building then is not exactly the same as the technology we're building now.
SPEAKER_2
Yeah, totally. So you guys wrote a paper called Agentic AI in the Public Interest, which is a really huge topic right now. We had a lot of philanthropic news come in today.
A lot of this public good, public benefit thing is very on the top of our minds. So you guys argued here that there's a lack of societal and institutional AI readiness. And I'm curious what you guys discovered through writing this paper about some of the things that could affect the future and this.
Nenad Tomašev
Yes, so that paper is a viewpoint. It's a personal viewpoint, but I sort of believe that things are changing very rapidly. We were in the midst of it, or something struggling to keep up with all of the updates. So it should come as no surprise that the broader society also cannot follow every single development to the same extent. So there is some sort of a lag maybe behind the frontier. It's very normal. It's been the case for every single technological revolution. But the core premise of that paper is that intelligence, while it is an enabler, and what we argue for in the paper is in particular, that it's an enabler for advancing research, opening up all sorts of discoveries, whether it's in areas like biomedicine or material science or other domains of science, that it needs to be followed by similar investments in the infrastructure to deliver on the promise.
In particular, we've done lots of work here, and lots of work has been done elsewhere in the community on agentic AI to autonomously derive some new hypotheses and insights, many of which can be used to advance discovery and practice. But then for many sciences that are not strictly computational, and the strictly computational science or things like mathematics or computational physics, you need to have physical laboratories to validate those findings or disprove them, because at the end of the day, some of them will also be false. So if we put all of the investments on the AI side of the equation, we're gonna have a world where we have many more hypotheses and conjectures that we can actually process downstream from that. So we're trying to make an argument that one needs to have both. So the broader scientific community needs also to see investment. More people, more experts, more laboratories, just a broader ecosystem. And I think that's one illustration of how actually the entire society can benefit from the progress as it's happening because the growth should be happening in all sorts of places.
SPEAKER_2
Yeah, I totally agree with this. So I think one thing that's very on our minds right now, as we see AI able to make its own mathematical discoveries, for example, is how are we going to ensure that there's a good body of human experts that are able to take, you know, like a whole community of mathematicians or scientists that are able to take the learnings and not only like understand them and make them easier to understand for the public, but also apply them. So I still feel like there's, you know, regardless of how many discoveries AI is making, there is this whole problem, like one or two layers down of how does this actually get translated into physical discoveries and, you know, economic growth after that. So what are your thoughts around that?
Read the other 17 minutes of this episode here.
$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.
Not what you expected? Full refund within 14 days.
Thousands of transcripts fetched by people building searchable podcast archives
Fetch transcripts with code
The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.
From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.
Using your own key:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/1000797188667