Resources

• Project Bellwether at X
• X, The Moonshot Factory
• How Bellwether is helping the national guard transform disaster relief: Link
• TechCrunch: Alphabet X’s Bellwether harnesses AI to help predict natural disasters
• Sarah’s post on LinkedIn: Wildfires, breakthrough tech, and the opportunities to be found in crisis
• Benedict Evans Newsletter – A tech industry newsletter offering analysis on platforms, AI, and digital infrastructure trends

Host and Guest

Host: Mike Toffel, Professor, Harvard Business School (LinkedIn)

Guest: Sarah Russell, General Manager, Project Bellwether, X (LinkedIn)

Transcript

Editor's Note: The following was prepared by a machine algorithm, and may not perfectly reflect the audio file of the interview.

Mike Toffel:

Sarah, thank you so much for joining us here on Climate Rising.

Sarah Russell:

Great to be here. Thanks.

Mike Toffel:

So, Sarah, we're here to talk about Project Bellwether, which I'm really excited to talk about. And it's brought to us by your organization, X the Moonshot Factory. I know you're a medical doctor as well as an MBA, and now working in this really interesting, innovative space. So, can you tell us a little bit about your journey and how you got here?

Sarah Russell:

Sure. Well, I would say this is a reflection just briefly whenever people talk about their careers, they often seem very linear, and they make a lot of sense. But I would just remind your listeners or share with your listeners that I have had many moments of profound indecision and confusion that have lasted several weeks to months, right? So, I'm going to describe things in this very natural progression as if everything fell into place, but there were a lot of moments in which I was, you know, wrestling with stuff. So, the big story is I've always wanted to be a physician ever since I was a kid. I love medicine. I love taking care of people. I love all sorts of things too. So, at the end of my medical training at Mass General, which was a fabulous experience, I knew I wanted to be a doctor some of the time, but not all the time. And in that other part of my life, I was really interested in technology and the ways it could help all sorts of patients get access to care. And I built a platform at the tail end of my residency that some people thought was a great research project, but I actually thought it was a business.

And so, before I knew it, I found myself in entrepreneurship. And the entrepreneur in me, I think, has continued for most of my career, which is basically I really love zero to one. And I think you can do that in all sorts of places. I've done it in the VA system, helped start things from zero to one. I've done it on my own in startups. And then I was in the middle of building something in the generic drug space when I was recruited to Google X and now X. And now I’m working on a project that I've moved into over the past five years that we call Bellwether, which is what we're here to talk about today.

Mike Toffel:

And so, what were you recruited to Google X at the time to do?

Sarah Russell:

You know, they called it a fire starter at the time. I think it was largely like an entrepreneur in residence. My manager, blessed his heart, really encouraged me to not, I didn't have to do medicine. I didn't have to stay in the healthcare services, which I think was great for me at the time because I was feeling a little discouraged. Of all the industries for which technology has failed to improve productivity, would say healthcare's at the top, right? So, it's a hard space. And maybe I was looking for a breath of fresh air. But I started in actually the question of insurance and risk taking and what it means to match capital to risk. I was interested in pricing climate risk. These were core questions that I was really interested in. But the entrepreneur in me was thinking about where do we have a right to win? Just because we're Google doesn't mean you should win. In fact, you have some constraints being a huge company. But where would you have the right to win? Where would you be unfairly advantaged? And to me, Google's repository of geospatial data and information about the Earth is highly differentiated. I think it's 15 years ahead of the competition All of those images that power Google Maps and Street View and so forth are available for internal use. And so, I thought, that has to be a part of this solution. For me personally, I thought that was a way to help establish a very credible and exciting moat in whatever I was building.

Mike Toffel:

Let's talk about those images for just a minute because they're critical to, as you're saying, what Bellwether is, which we'll reveal in a minute. But those images, those are captured from what technologies are, does Google itself have satellites or do they contract exclusive rights to satellite images? How do they get all these images?

Sarah Russell:

Yeah, I mean, they do all the things. It's awesome. The organization's called Geo within Google Maps, they produce Google Maps. So, ages ago, they contracted with Airbus to just bring in their satellite imagery. But at the time, the insight was that raw satellite imagery still needed quite a bit of processing. And they need software engineers to take a low-res picture that's way up coming from low Earth orbit and be able to what we call orthorectify or understand what it might look like at higher resolution on the Earth. That was key to making maps. That software engineering around geospatial data became a powerhouse-like asset within the Google ecosystem. They came in and learned how to do this. And then Google started flying its planes and getting kind of aerial imagery. So, images taken out of planes, like directly pointing down, which gives you these very high-res images of cities and so forth. Many places around the world have that. And then they started doing street view, which gives you this image. And then, of course, with Waymo and so forth, now you have LIDAR images of what's coming around you in this environment.

And now they're putting up a new satellite called FireSat, very much focused on how do we get good pictures of wildfires before they start, when they're burning after. There are all sorts of other pictures about the Earth. So, the truth is, all the things you can imagine. And at different times, Google has tasked and tried to figure out new products and projects to build, but yeah, that's the nature of the imagery. It's fantastic.

Mike Toffel:

So that's the imagery that you're saying this provides this differentiated advantage as you were thinking about developing.

Sarah Russell:

Correct.

Mike Toffel:

Okay, so let's dive in and talk a bit about Bellwether. So, as I understand it, Bellwether is, in a way, it's a prediction engine, but it's predicting some long-term trends, and particularly on fire risk and on disaster that are undergoing, like right now, what's going on in disaster. Are those two things it's currently predicting?

Sarah Russell:

Correct.

Mike Toffel:

Yeah, so tell us a little bit about how you even got to those things. What did you consider having it predict and how did you land on those two?

Sarah Russell:

Sure. Well, let me just reframe slightly so everything makes a little more sense. If you're interested in the question of climate risk or understanding changes in the Earth with respect to a changing climate.

The question is, are you going to be in carbon removal, I would say, or adaptation? Where is your focus? Very squarely, we're in adaptation. So, carbon removal is a huge part of sustainability and that work. And we're in the sustainability ecosystem. We care about adaptation, which is the notion that climate change is here today. The climate has changed. We're suffering with the consequences of more severe natural disasters. OK. So, you start with that.

What I was interested in is if you want to do a better job forecasting and predicting and understanding natural disasters throughout their life cycle in the middle of an aftermath, building back better, that whole life cycle of a natural disaster and its impact on a community.

Where should you go first, right? And to me, the question was who's actually in the business of understanding natural disasters and who pays for information about natural disasters? It turns out that's insurance. Insurance is in the business of understanding “NatCat”, natural catastrophes, as a way of understanding risk to structures. And so, they're the ones who price that, and they're the ones who are trying to do a better job at that.

So then when you say to the insurance company, hey, what do you think about climate change? Most of the time they're like, oh, no, no. We don't pay attention to climate change. It's way in the future. We think about severe weather. I say, oh, OK, so what part of severe weather are you interested in? They said, whatever's happening one year, five years into the future. So, all of sudden that time frame narrows down. So that was this first customer base. And that's why we said, well, then what's the hardest peril for you? And they all said, wildfire. It's practically unknowable. No good tech.

This was five years ago. At the time, I think there were probably three startups in the wildfire space. Now they're over 120. It's really changed. But when I was getting going, this was sort of this very unknowable space. After that, we built a model that forecasts the probability of wildfire one year into the future and five years into the future in an absolute sense. And this is what gets at the heart of Bellwether after that long preamble. We'll deliver absolute risk scores and real-time data analytics to businesses and the public sector who are interested in building back better or being more resilient in the face of natural disasters. Does that make sense?

Mike Toffel:

Yeah, so you said absolute probability or absolute risk. What do you mean?

Sarah Russell:

Yeah, I think it makes sense in the context of the way people think about risk largely in say wildfire today. Today, most of the time it's a relative risk score. And what I mean by that is if someone says, well, what's the risk of fire in downtown San Francisco? Someone would say, well, that's a low risk. And I'd say, well, what's the risk in the middle of Lake Tahoe or Tahoe forest? And say, that's extreme.

What is it the next year? It stays the same year after year, and it's all relative. Tahoe risk relative to Tahoe, but that's relative to San Francisco. But that's actually not what you care about if you're trying to either change the landscape or manage risk or price risk or underwrite it or do something interesting. What you care about is literally what is the probability of something burning in Lake Tahoe? What's the probability in San Francisco? And so, we say, okay, we'll give you a number between zero and one.

It's 1.5 % for this structure in Lake Tahoe, a 1.5 in 100 chance in the year ahead that this structure will burn. And truth be told, in places like Pacific Palisades, we had numbers that looked like 16 % in December of 2024. 16%, super high, that the Pacific Palisades would burn in the year ahead. So, these are numbers, when you start getting familiar with them, you see that some are quite extreme and extraordinary. And most of them hover, and even in high-risk areas, between 1%, 3%, 4%. But sometimes they really jump out. And that's a very useful number for people who are trying to understand where to allocate resources, where should we focus, where should we mitigate, where should we find more opportunities, that kind of thing. Does that make sense?

Mike Toffel:

Yeah, got it. Yeah, absolutely. So, okay, so that was the first you said. So, fire risk because it was at the time an open space, now a little more crowded space, and there's willingness to pay for that from insurance companies.

Sarah Russell:

Correct. We've also found some willingness to pay in the public sector, but, you know, wow, that is a tough space. I mean, that's not my strong suit, I would say. Cracking the public sector. Hard, hard.

Mike Toffel:

Because of the procurement processes that they engage in?

Sarah Russell:

Yeah, and I also think, you know, in spaces where there's a ton of money and a ton of money to be spent and everyone's focused on it, it's tough to crack in actually. It's sort of strange. Because the more money there is, the more people running after it, the more competitive, the more you need to know people, you know, it's harder.

Mike Toffel:

Okay, so the other one that you're looking to is this disaster response.

Sarah Russell:

Correct. This actually was a relatively specific thing that was born out of a need that came from the National Guard, the Department of Defense. And then it's very interesting, but actually right after a natural disaster, like a flood, the airspace gets blocked off. No one can fly. And in that airspace, the only things that's allowed is the US Air Force and something called the Civil Air Patrol. And they go up and assess the scene. And they take pictures of the scene, everything that's occurred. And they try to capture all the critical infrastructure and key resources hanging out of planes with these SLR cameras, right?

Then they land and they upload these images. And the challenge for years has been, how do we process these images to understand which are off nadir, meaning they're not straight up and down, they're at an angle, and we have no metadata, we have no understanding of where they are in space, we only know where the plane is. How can we very quickly understand everything in that image? This has been a very hard problem. But it means a lot because if you want to figure out where to evacuate people, whether or not the power is up or not, Wi-Fi, internet is down and which places and so forth. That all relies on your ability to very quickly understand those images. And before we came along, they were having human beings look at the image and compare it to something on Google Earth. And it was taking minutes per image. It was like geo-guessing. And the technical challenge for us is could we teach a computer to look at this image and do a bunch of math and understanding of the earth and everything it knows and do a best estimate and then create a bounding box, actually square off that space and tell you what's in it. That's what we did.

Sarah Russell:

And that solution has now been purchased by the National Guard over a five-year period. And once it was a trial period, once you successfully complete this, there were 60 folks that competed, and we went head-to-head with. We were the only ones who were successfully able to nail this, deliver the results. We were able to secure a sole source contract, which was fantastic. So, we have a five-year contract to deliver the solution to the federal government.

Mike Toffel:

Now these seem very, very different. I mean, they both leverage the mapping technology back engine that we started our conversation with, but one is a real-time adjusting imagery and the other is a long-term forecast, or one to five-year forecast. How did you decide that this is the right combination of a product portfolio to go to market with?

Sarah Russell:

Yeah, it's a great question. Here is the core story. When I was trying to sell internally why wildfire was important internally to my stakeholders here at Google, at X, many of them asked, what is your core technology? What's the core technology of your group? This is a tech company, right? This is the driving force here. It didn't matter that I de-risked some business opportunities. They said, what's the core tech? What are you going to build forward with? How big is this market, right?

Sarah Russell:

And I discovered through conversations with my team and actually through some patient listening that in fact our core tech was around how we manage data. It wasn't machine learning. The core tech was the fact that we could work with geospatial data very quickly. In fact, we could work with really good repositories of publicly available data and very quickly turn them into a data pipeline using a bunch of software that allowed us to run machine learning iterations and cycles in a matter of weeks rather than months, which was our core technology.

So, when I made that case, my leadership said, well, show us another example where you can very quickly process geospatial data and turn a result. And so that's what's powering the National Guard work as well. Does that make sense? So, it's under the underlying pieces. Actually, the long pull in the tent or the hardest part of the problem is not the ML. It's actually how do you wrangle vectors and all these different data types with many different data dictionaries with different time stamps, different resolutions. How do you get them to all line up so that a machine can read them? Because a machine can't read messy data as well as we wish it would. Especially, can't read geospatial data as well as we'd like it to unless you give it a very organized understanding of it.

Mike Toffel:

So, you mentioned the insurers are interested in it. There's a public, of course, like public agencies would be interested in it they can figure out how to purchase it. There are other things related to fire risk, like drought or floods or agricultural productivity that I also imagine would use different flavors of machine learning prediction models but based on imagery. For example, even when to harvest, if you're looking at crops and you're combining the crop health and you're trying to forecast based on weather, when's the optimal time to harvest or even when to add nutrients to a crop, for example, given what nutrients it has, what the soil type is, what the sunlight has been, what the weather has been. So, it feels to me like there's so much space in so many potential applications for the use of your combination of imagery, and machine learning based forecasting. What's your sense of these other types of avenues?

Sarah Russell:

Well, I can tell you, like six months ago, I would have given you a very different answer. And the reason why is because of the emergence of geo-foundational models. A foundation model for geospatial data is a model that has basically looked at so much geospatial imagery and has learned so much about geospatial imagery that with one or two examples, it can generate a map or an understanding or a classifier segmentation of everything in that scene. That's come online so fast, so much faster than I thought it would. And it's the kind of thing that will allow geospatial data to talk to these large language models like Gemini and Chat-GPT and Claude. All these things will now become more, it'll be easier and easier because these foundation models are designed so that an agent could potentially ping and say, show me all the crops that have changed over the past three months, and boom, it shows up. This used to be like 2022 machine learning. You'd first have to collect all the data, look at it all over time, get it perfectly organized and so forth. But now you have these foundation models that have done all that work for you.

So, I would argue that now what you just mentioned, all of those use cases are now imminently possible to explore in serial lightning fast. I think you could deliver those things. The question would be after you deliver them, who would pay for them? So how much time should you spend? What should you price it? What's the value? Where should you go first? I think those are good startups, like early company scale questions. But the underlying technical process for hitting on those things is like changed dramatically in the past six months. It's fantastic. It's awesome. It's an exciting time.

Mike Toffel:

Yeah, amazing. So, who owns those foundation models?

Sarah Russell:

Maybe they were being snarky, but someone said they think they're like over 500 in the market right now. Like 500 geospatial foundation models. Not in the market but across research organizations and academics and universities. Google has a couple, like two. I bet they're all for different use cases, valuable, right? I'm sure they've all been kind of tuned or whatever to do a very good thing. Some are at very high res, some get down to 10 meters, some go to 50 centimeters. Like it's cool. But I would say what's interesting is there's a ton of academia happening there. So, to what extent will those get spun into companies? I don't know. Or will it become a commodity that we all can buy easily? Who will go open source? I don't know.

Mike Toffel:

And how does one think about accuracy in this whole space, the accuracy of these predictions, whether it be agricultural yield predictions and know when to harvest or back to the business that you're in, the fire risk prediction? How do we think about accuracy assessment?

Sarah Russell:

So, assessing the accuracy of a machine learning model in general is one of back testing, where if you've designed the model correctly, you'll be able to look at a version of the model that hasn't seen any data from, 2015 and you showed everything from 2014 and then see how it does predicting 2015, 2016, 2017, 2018, and then you say, well, how accurate was it? And then you have to do some interesting statistical sampling, right, because it predicts probabilities. So, there's a distribution there. So how do you determine the accuracy of a distribution? Well, you do some sampling and that kind of thing. But the point is there is a method for assessing accuracy. But it takes a design of a model that's actually quite challenging. It's an out of sample analysis. saying, I'm going to give it truly something it's never seen before, like no hints. And then I'm going to ask to predict it. And what you want is those predictions to be relatively smooth. You don't want them to go year after year, because that's not sellable too. So, you want to be able to look back and say, well, how well did we do? Did we say that things are going to get much worse in California in 2020? So that's how you evaluate it.

Mike Toffel

I think one of the foundational assumptions is a stable system. But with climate, of course, it's not a stable system. That's the challenge of adaptation. That's the whole thing about adaptation, of course. And so, predicting wildfires, if you use a model trained on data from five to three years ago, and then you predict what's supposed to happen in the last two years, it's a little unfair because the world's changed.

Sarah Russell

My understanding of the way we approach this is we just assume that the temperature's rising every year. We bake into the model some assumptions that allow for this range and we say how is it and we don't know how much it's going to increase but we assume that. So, there are some hacks we use to account for the fact that things are getting warmer, hotter, drier year over year. And we haven’t used it for just the past five years. We can go back, we've had good weather data for ages, right? So, you can kind of bake in all those trends.

Mike Toffel:

 What are insurance companies using your product to either replace or to supplement? Because they've been in the business of forecasting for a long time.

Sarah Russell:

Yeah, I would say I've never known an insurance company to outright, like maybe they replace, but they love many views of risk. They love many views. Here's what we have found is the most exciting use case, which is insurers who are using our model to understand risk are finding lots more places to underwrite and give policies than they did before, because our model is finding lots of risk that was previously thought to be very high and our model is predicting it to be low. So, we're finding the use of our model is actually expanding opportunity both for homeowners but also for the insurance company. Because remember, there's alignment there. The insurers want to underwrite. That's their business. So if we can enable them to do that by finding them good risk to balance out a portfolio, that's a good use case.

Mike Toffel Yeah, it's really interesting because you do hear that with more sophisticated technologies, it's allowing insurance companies to figure out where to exit. But your point is the other side is happening as well, where to actually underwrite, where they might have otherwise been afraid because it's in a pocket of red, but actually there's some green or yellow zones in there.

if you go into a neighborhood and you do some interesting vegetation management, controlled burns and so forth, there is meaningful changes. the wildfire will still happen, but you can decrease the probability that it's catastrophic.

Mike Toffel:

Yeah, I imagine if you have enough frequency of satellite or aircraft imagery, you can also be in the business of sellingmonitoring of those types of services. So, if I imagine if insurance companies give a discount to homeowners who engage in these types of mitigation practices, maybe they visit the first time the site to actually see it in person.

But after that, like that, it’s pretty expensive to go back every year to see if they're actually continuing to sort of keep the forest 100 meters away. But imagery could dramatically reduce the cost of that. Is that a viable business case as well?

Sarah Russell:

I mean, super smart. It's like you've been around this for a while. But this is now like a totally emerging space for startups. It's providing the technology using cool AI and computer vision tools. Look at a home remotely, assess how it needs to have its vegetation managed, and then recommend and then track it. And then provide the insurance to that home. That's happening. Just what you described. And it's not even having someone go to the house at first. They're having someone just hold up their phone.

Mike Toffel:

So, visiting is so old school. So, let's look forward a little bit. We've been diving into the history and the current practice. As you just already described, some remarkable technologies over the last few months have even rolled out faster than you anticipated. Where's this market heading? As we think about these now, base geo models, is that going to absorb fire risk modeling as well? Like is that this idea of this big moat that you had in imagery is now coming up against a tidal wave of lots of these other foundational models.

Sarah Russell:

I think that's a very likely result. I think these foundation models will make it possible to generate useful data and understand or create data layers that allow you to do a very good job forecasting.

I also think there's so many new sources of satellites coming up. It's just exploding. New vendors, new startups, I everyone's taking pictures of the Earth. And so that availability and that sort of, as the price of obtaining these images goes down, that's also going to be super important. You still need good data. So that's, I think, very exciting. So as both, as that gets cheaper and more abundant, and then you have these other tools that can do one-shot learning and so forth, which is, you know, look at one image and then I know what to look for in the, show me one example and then I can find all the other examples.

That is going to be an important trend. I would say to me, in terms of the future, I think two directions are likely. One is there is an element of the geospatial space that is intensely visual. And I think the ability to visualize these results, for users of all stripes is going to be really valuable and will allow them to verify what something like an LLM is doing in the background. And that is going to mitigate the experience that these are black boxes hallucinating. Because if I can show you a result that's visually represented, you're going to be able to see it and say, well, wow, that makes a lot of sense because I can see it with my own two eyes.

So I actually think the abundance of these tools are going to make geospatial visualization a thing that's much less niche and much more abundant and something that everyone gets comfortable using like a spreadsheet. It's like, we're going to look at that image and see what it is very easily and quickly to verify what we're doing. Now it's relatively cumbersome to do that, if you can believe it. It's not easy. It's expensive. It's hard to move all the data around, even if it's something you own. So I think that's one trend that will change the game.

The other thing I think will happen is some of these new technologies like these LLMs and just these agentic models in which an agent is talking to an LLM and leveraging the value of that LLM. You have different kinds of models interacting with proprietary data. That ability to have an LLM learn from and augment and enhance the insights of your own proprietary data, I think that's going to be a game changer for several industries that have been stuck in really old infrastructure that are kind of hybrid cloud and on-prem, that are highly regulated and so forth, the things I'm thinking about, like the public sector, like insurance, healthcare. I think all of these spaces are going to be transformed, particularly, by the ability to use new tools to get your proprietary data liberated.

Mike Toffel:

Yeah. So who do you think are going to be the winners in all of this? Is this something that where the spoils will go to the big few tech companies that are so far advanced with capital and technology, or is this a equalizing moment which will provide a little bit more level playing field for entrants and for other incumbents?

Sarah Russell:

So, I guess some people who are currently writing right now believe that it's the tech titans who will take over. It's not clear to me that the people who are going to win are going to be the ones who have always won before. I think entirely new products and services will emerge.

So, I would say it's going to be like our very notion of SaaS is going to be different.

Mike Toffel:(47:02.372)

Interesting. Let's go back to this idea of guardrails that you mentioned a moment ago, because there's a lot of people who are worried about that with this fast-pacing evolving technology that regulators aren't keeping up. And there's even in the current bill in the US, there's a question about who could be the regulator, federal versus state. And maybe there's a restriction on, a proposed restriction on states intervening. But if we bring this back to Bellwether and the idea that you mentioned earlier that lots and lots of people are putting up more satellites and images are everywhere. Let me just think about the image space that I know that Google Maps has had to think about this, when they drive around their vehicles and they're capturing people walking down the street and they decided either proactively decided or were told to blur their faces. I'm not sure about the backstory, but faces are generally blurred, and license plates are generally blurred and so on. Is that a decision that the company made on its own, or was that a decision that was under regulatory threat or mandate?

Sarah Russell:

I think long before we had these LLMs and so forth, the notion of gen AI was really in talking points and so forth. Google has been thinking about AI principles intensely, like what are the rules on engagement? How are we going to use this AI? And I think it's largely been driven by the question of military engagement. Like what is Google's relationship with the military going to be? I think for a long time, the company has been very clear. Although things are changing for sure. It's being updated all the time that these guiding principles are, in fact, the guardrails, right? It's just like we ask, what is the use case? What's going to happen to the user? What's going to happen to privacy? How will the user get hurt? How will their information be misused? Interestingly enough, Google is very concerned about an insurer using its imagery to find out whether or not you have a trampoline in your backyard. Because your trampoline conveys a certain amount of additional property risk. But Google doesn't want you to use satellite image to have you spy in the backyard. So there's a certain use case where Google's like, no thanks. You don't get to spy on users' backyards. Right?

Mike Toffel:

So, what else should we be looking for Bellweather in the future?

Sarah Russell:

We just did a demo at Cloud Next in Las Vegas in 2025 in April that was successful and awesome, which was basically using an agent to interact and reason about geospatial data, which was really a wonderful thing for our team. We finally got into it, got after it, built it. And now we're interested in all the different tools we can have that agent talk to. We're also thinking a lot about visualization. What will it look like for people to have context across the board?

So, I would say that's the future is increasingly making it easier and more user friendly to interact with these insights, get these insights, and then see them. And then moving across, just going one vertical by one vertical. Because rather than having a general-purpose tool, I do think we believe in refining each of these tools to be very good.

Mike Toffel:

And you think you're just going to stay in the B2B space as opposed to B2C?

Sarah Russell:

Probably, although there's a good case for wildfire to just go B2C, I think it is a different skill set for sure, but I would say for now B2B.

Mike Toffel:

Great. All right. Well, this has been really interesting to hear your entry and the evolution of this space. Let me step back and ask you my final question, which is for those interested in working in adaptation, geospatial, AI, entrepreneurialism. There are so many different threads that we laid over this past little bit. And you mentioned already like some books and some newsletters, but let me just consolidate it here. What are some resources or advice you have for folks who are thinking about, this sounds interesting, let me learn more?

Sarah Russell:

Sure. Well, you can definitely find Bellwether on the X website. So that is one place where if anyone's interested in learning more, looking for a job or an internship, residency, those things are absolutely, we're looking for great talent all the time. But more generally, I would say that based on what I'm hearing from people about the market, it does seem to me that it's an increasingly a question of having lots of good conversations with people and asking to be connected to other people who are doing the work you think is valuable.

I think a few years ago I would have advised people to have this be an internet-based search and whatnot, and I think that's helpful to learn about things that you're interested in, but it does seem to me, at least in my own experience hiring, that if someone makes an introduction to a person, I always look at that candidate. I just do, you know? And so, if you can get someone, a friend of a friend, to make an introduction in person, it just goes so far.

Because it just cuts through all the other noise. So there's a certain in-person quality that I think requires us to take a bit of a risk and to say, ask someone for help.

Mike Toffel:

Yeah, do you have a sense of how to assess that ex-ante before you've had the benefit or detriment of working for that person?

Sarah Russell:

Yeah, exactly. I think it's just like what people say. You go to a college campus, you get the vibe. But maybe ask, maybe find out who's worked for them in the past, see what their team says. Google did a big study, by the way, to assess who's good manager. It looked at hundreds of thousands of people and said, oh, how are we going to discern? What was the most important factor in determining who was a good manager? And it was just whether when you went to the reports, they said you managed well. It's just what your reports think. If your reporters think you're doing a good job, you're doing a good job. So, in some ways, maybe the answer is gone talk to the people who work for them.

Mike Toffel:

Sarah, this has been a wonderful conversation. I really appreciate you spending time with us.

Sarah Russell:

Mike, thanks for everything. It's great.