Best of September 2026
57% of the talking has a confirmed speaker.
Best of September overview 0:02
David Roberts introduces the monthly best‑of clips show and thanks listeners.
Dr. ML.
Hello, hello everyone. This is Volts for October 2nd, 2026, the best of September. I'm your host, . I am in a hotel room in . Just about to go to bed, but I wanted to get this out to you, so pardon the sound quality. Once a month, I pull one chunk out of every episode, string it together, and create a clips show to give you an overview of the month's episodes. This was a big month for Volts. A couple of technical deep dives, a couple of broad overviews. And lots more. Some of my favorite episodes of the year, I think, this month. Paid subscribers get these early. Each one opens up to everyone when the next one comes out. So the best of August is now open to everyone. If you know somebody who's been meaning to check out the show, that's the one to send to them.
All right.
7 conversations this month, one piece of each. As always, I deeply appreciate all of you who pay for subscriptions. This podcast is entirely listener supported and I couldn't do it without you. So let's get into it. First up, we begin with , CEO of , which pushes back on clean energy disinformation.
I'm curious what else you have found just in your message testing. I mean, is it just all the intuitive stuff like cost and convenience or what matter, you know, that kind of thing? Like, what have you found about messages themselves?
Yeah, definitely. I think messages, right, we have to get away from, and I think our biggest failure is sort of our own egos and wanting people to care about climate or wanting them to care about, and really essentially what it boils down to, Factwe got almost 70% of Americans that live paycheck to paycheck right now. So we're thinking, am I gonna keep my job? Everything's more expensive. What about my healthcare? What about filling my car up with gas? We do this great series, it's on our YouTube channel called What We Pay, and we're telling these stories. There's a story on there from Pennsylvania 10. This is 's district. He's a House Republican, and we just talked to this family. They have an $859 utility bill. They pull it up on their screen. We did like a mini documentary. They're like, yeah, we have other bills due tomorrow, so we're gonna have to choose which ones to pay. So essentially, yeah, economic security. Republicans and Democrats all have to fill their gas up, their car up with gas. They all have to pay an energy bill. So I would say election cycle after election cycle, having economic security and feeling like one is getting ahead, I would say really, really helps. How do we tie that to the energy transition? Oh, well, how about we stop the tyranny of far-off wars and just break our dependence on this stuff and just, you know what, charge your car in your garage. For a couple bucks. That sounds pretty great. And if you look at Australia and if you look at Europe, EV sales are through the roof right now. Yeah, I mean, the trends are everywhere globally, so I'm hoping to see that in the US. But I think the values argument around some of these technologies and how ideological they are and how polarizing they are across political lines is really, really holding us back. So I think what is working, definitely economic security. I think second, When we look at audiences in rural conservative places where most of the utility-scale renewables are being built, energy security, grid reliability, like keeping the power on around extreme weather, American competitiveness with the rest of the world and keeping us ahead and making sure that, you know, we can build these technologies here. I would say, you know, the jobs message, while I feel like we've, some practitioners have kind of turned away from that, we still see a lot of potential and influence when you show really cool people in clean energy jobs that love their jobs and they're showing beautiful vistas from the top of turbines or women in solar or people on battery energy storage sites. There's a ton of stuff on our YouTube channel about that. So I would say it depends on the audience. Uh, sometimes a climate message really works for like a young audience under 30.
This is what I wanted to ask about. One of the big fights right now is I think like everybody, almost everybody has come around basically to thinking Yes, obviously, like cheaper energy, more reliable, better for your personal immediate life is the best thing, is the top thing, is the best thing to lead with. But now there's this argument about like, some people want to say, but like, yes, bolstering that with climate will boost you a few points. And then other people will say, no, no, no. If you throw climate in, in fact, you polarize a bunch of people away. even if you've got them on your side from that, then they'll polarize them away with climate. Like, I wonder specifically about climate, what you've found.
Can I just start by saying something maybe very polarizing, but I think Opinionthis is just the most ridiculous false debate that is so dangerously distracting from the bigger issues. I'll just say that.
I mean, they love it on social media though, Leah.
They love it. They love it. And they, I've been reading all the blogs and news stories and Substacks and everything about it. I mean, how we think about it is maybe a little different, but I think we can have our cake and needed to, and to continue with the dessert analogy, right? We think about it like a pie. So Factthere's some people right now, like 8% of young people under 35 rated climate as their top voting issue in 2024, right? That's a lot of people. So we want to use a climate-first message for them. And, you know, it'd be great if people who are kind of, could be predisposed to being really driven around a climate message to get more of them to increase the size of that slice of the pie. But then there's other people who will never be driven by a climate message. I will reference my dad.
Yeah.
Who is a staunch conservative Trump supporter, but he now loves solar because his HVAC guy has a side hustle doing rooftop solar, not because his daughter works at , but because his HVAC guy convinced him, you know, he could save money. So, I mean, it really depends on the audience and this sort of crude segmentation between should we say climate, shouldn't we? When you look under the hood and really look at the nuances of how we communicate The audience segmentation, how we market, the answer is just not black and white like that. It's, you know what, use it if it's advantageous. In most cases, it's not incredibly polarizing. It's also what I said, a logistical issue. When you say climate, how do the algorithms respond? On TikTok, there's evidence that your content is suppressed. So it could be a logistical reason. It's not even a polarizing thing. I'm less worried about the polarization of saying climate. I actually don't think it's that big of a deal for most people right now. We're seeing it, we're living it. It's really a logistical thing with certain platforms deciding to suppress or serve your content to different audiences. And how do we get by the goalie?
Geothermal blind discovery 7:16
Joel Edwards explains how Zanskar’s machine‑learning models led to finding a previously unknown geothermal resource in Nevada.
Next up, and of , who use machine learning models to hunt for conventional geothermal.
And so then let's talk about this 3rd discovery. This is called Big Blind in , and this is sort of I think the news that made everybody really sit up and pay attention, because this was unlike the previous 2 we've discussed. This was genuinely new. There were no prior wells on this site to give you any data. There were no surface manifestations of activity on the surface. This was a genuine blind site. So how did you— Find it in a nutshell. How did you find it?
Yeah, this is really our bread and butter. So Joel, I'll let you kind of talk through how greenfield prospective exploration works.
Yeah, it's that concept we talked about at the beginning, David, that you build models. Everybody can build models. Zanscar builds models, but you have to pair your model predictions with tests in the field of those predictions. And so this is one of those cases where we had made a set of predictions across this part of the state, and we were out testing those predictions. And ran into this system at this location. And this is a system that's west of , , so pretty remote. And like you said, there was no prior data at the site. And if you or me or anybody were to walk across the site, there would be nothing at the surface that would clue you into that there's a massive system below. That initial discovery was made at the end of 2023, and then we went through the permitting processes and We came back with a larger rig and drilled into that resource in what year is it now? It's 2026. I think we drilled into it in 2025 and Factwe encountered commercial reservoir conditions at around 2,000 to 3,000 feet. The thing that's really exciting about this system in particular is how big it is. This system, by the surface footprint of the heat flow anomaly, is much bigger than Pumpernickel and much bigger than Lightning Dock.
Oh, interesting. I guess you've just kind of started exploiting it, right? I mean, is it commercially operational or where's the state of play there?
No, we're in the wellfield stage of the project, so we're permitting additional exploration wells right now to go back to the site. So we're still early stages and yes.
Still trying to figure out how big the site itself is, how big the resource itself is.
Exactly. We have a sense of sort of the surface area, how big it is. That's why I can make that claim relative to some of the other projects. But we do not yet know at depth how big and hot the system is. It seems quite spicy to think that this thing was sitting out there. There's lots of exploration across the West for minerals and other things. And to just think that this had been missed for so long, you know, gets us really excited about the future.
So you had your models and they gave you a sort of probabilistic sense, and then you go out and do some exploration in those probabilistic areas and you just got lucky on this one. Or I guess not lucky, the model, the probabilities paid off in this spot.
Yeah, I know. Maybe the models got lucky. Maybe we got lucky. There's luck in all this. And there's also bad luck when we miss. And to be clear, we've missed way more than we've hit. Like to make this sound like this thing has been working better than the legacy explorers. In the early days, Factour models were frankly much worse than expert explorers. And we do benchmark against expert predictions in our modeling space. And it's really only, our models really only in the last year or two have started to outperform. And every model that you build, sometimes your models get worse. So it's not like it's up and to the right. It's like there's a wobble. It's sort of like the climate, right? Like some years it's wobbling hotter and cooler, but you're generally up and to the right. And that's very true for our modeling space.
So you think from now on over time, the more exploration you do, the more data you feed into the models, the better the models get over time. You think you're going to get better and better at finding these things?
And that's actually quite clear already from the data we've collected. Like Joel said, they started out much worse than humans. They caught up and they're performing better in many locations and areas. And we've seen it improve in 2 ways. As you put more data into the models, the same architecture ends up yielding higher precision, you know, better performance. But in other cases, actual changes to the architecture, the way we've designed the model itself with the same data yields an improvement in prediction. So the teams are constantly driving improvements on both of those fronts.
Interesting. So maybe not up and to the right for every increment, but over time, In the fullness of time, you're going up into the right with your models.
Model uncertainty step back 12:11
The speakers explain that taking a step back improves the model by providing a more accurate estimation of uncertainty.
Yeah. FactAnd really those sort of sidesteps that Joel's talking about, or what might feel like even a step back, is often in a certain type of metric, like precision or accuracy, but it usually reflects that we weren't realizing that there was something less generalizable about the prior model. And so this step back is getting you to a more accurate representation of the uncertainty. OpinionSo it's still a better model. You're just being more correct in your estimation of uncertainty for it.
Right. Interesting. Interesting.
You know, this manifests, it's funny, Factyou would expect our sort of hit rate, like when we find a system to be like consistent and linear. But what I've found is that our hit rate is like, you know, like in sports you get into like a streaky zone, like home run after home run or whatever, you're hitting the threes like back to back to back. That's the same thing in our models. Like we go through periods where bang, bang, bang, like survey after survey after survey, we're hitting, hitting, hitting. And then we go through like a cold slump for a period of time and then we push out a new one. So it's been a very like nonlinear sort of spikiness to the discovery rate.
And that's the end of the free preview. If you'd like to hear more, become a paid subscriber at . Thank you.
“we got almost 70% of Americans that live paycheck to paycheck right now.”
A large majority of Americans are financially strained, shaping energy‑messaging strategies.
“this is just the most ridiculous false debate that is so dangerously distracting from the bigger issues.”
The climate‑messaging debate is seen as a distracting false argument.
“there's some people right now, like 8% of young people under 35 rated climate as their top voting issue in 2024, right?”
Only a small fraction of young voters prioritize climate as their top issue.
“we encountered commercial reservoir conditions at around 2,000 to 3,000 feet.”
The geothermal well hit commercial conditions at depths of 2,000‑3,000 ft.
“our models were frankly much worse than expert explorers.”
Early ML models underperformed compared to human experts.
“And really those sort of sidesteps that Joel's talking about, or what might feel like even a step back, is often in a certain type of metric, like precision or accuracy, but it usually reflects that we weren't realizing that there was something less generalizable about the prior model.”
Metric‑driven step backs reveal lack of generalizability in the prior model.
“So it's still a better model. You're just being more correct in your estimation of uncertainty for it.”
Stepping back yields a better, more accurate model.
“you would expect our sort of hit rate, like when we find a system to be like consistent and linear. But what I've found is that our hit rate is like, you know, like in sports you get into like a streaky zone, like home run after home run or whatever, you're hitting the threes like back to back to back.”
Hit rates are streaky rather than linear, resembling sports streaks.