Spirit Airlines' Data Sale, the Model Too Powerful to Ship, & the Landlords of AI | [Sidecar Sync Episode 149]

Summary:

When Spirit Airlines went bankrupt, its planes and gates weren't the only assets on the auction block — Google paid $10 million for its data. Amith Nagarajan and Mallory Mejias use that sale to dig into what's actually valuable in AI right now: not bigger models, but the specialized data only your organization holds. They trace the Spirit Airlines deal, unpack OpenAI's Astra model solving open math problems that had stumped researchers for decades, and explain Neoclouds, the specialized companies renting out the GPU capacity behind the AI boom. They also break down Google's fast, cheap Gemini 3.7 Flash release and why it might be all the model most associations need. Amith makes the case that associations sitting on rich member and industry data have a real opportunity to build and monetize AI products, not just cut internal costs.

Timestamps:

00:00 - The Gorilla in the Room
06:17 - Google Buys Bankrupt Spirit Airlines' Data
09:09 - Why Specialized Data Is AI's Next Frontier
14:21 - Turning Association Data Into Products
23:21 - OpenAI's Astra Solves Unsolved Math Problems
33:00 - AI as a Research Partner for Scientific Associations
35:25 - Neoclouds: Renting the GPUs Behind the AI Boom
41:24 - Google Quietly Ships Gemini 3.7 Flash
44:48 - Why Gemini 3.7 Flash Is Winning on Speed and Price
49:46 - Closing Advice: Build AI Fluency, Spot the Gorilla

👥Provide comprehensive AI education for your team

https://learn.sidecar.ai/teams

📅 Register for digitalNow 2026:

https://digitalnow.sidecar.ai/digitalnow

🤖 Join the AI Mastermind:

https://sidecar.ai/association-ai-mas...

🎀 Use code AIPOD50 for $50 off your Association AI Professional (AAiP) certification

https://learn.sidecar.ai/

📕 Download ‘Ascend 3rd Edition: Unlocking the Power of AI for Associations’ for FREE

https://sidecar.ai/ai

 🛠 AI Tools and Resources Mentioned in This Episode:

Selective Attention Gorilla Test ➔ https://www.youtube.com/watch?v=vJG698U2Mvo

Sidecar Sync Ep. 110: Claiming the Right to Win ➔ https://shorturl.at/kFy1M

Claude Code ➔ https://claude.com/product/claude-code

Gemini 3.7 Flash ➔ https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash

Google Antigravity ➔ https://antigravity.google

OpenAI Astra ➔ https://openai.com/index/ten-advances-in-mathematics

CoreWeave ➔ https://www.coreweave.com

Nebius ➔ https://nebius.com

OpenEvidence ➔ https://www.openevidence.com

Skip ➔ https://askskip.ai

MemberJunction ➔ https://memberjunction.org

👍Please Like & Subscribe!

https://www.linkedin.com/company/sidecar-global

https://twitter.com/sidecarglobal

https://www.youtube.com/@SidecarSync

Follow Sidecar on LinkedIn

⚙️ Other Resources from Sidecar: 

• Sidecar Blog
• Sidecar Community
• digitalNow Conference
• Upcoming Webinars and Events
• Association AI Mastermind Group

More about Your Hosts:

Amith Nagarajan is the Chairman of Blue Cypress 🔗 https://BlueCypress.io, a family of purpose-driven companies and proud practitioners of Conscious Capitalism. The Blue Cypress companies focus on helping associations, non-profits, and other purpose-driven organizations achieve long-term success. Amith is also an active early-stage investor in B2B SaaS companies. He’s had the good fortune of nearly three decades of success as an entrepreneur and enjoys helping others in their journey.

📣 Follow Amith on LinkedIn:
https://linkedin.com/amithnagarajan

Mallory Mejias is passionate about creating opportunities for association professionals to learn, grow, and better serve their members using artificial intelligence. She enjoys blending creativity and innovation to produce fresh, meaningful content for the association space.

📣 Follow Mallory on Linkedin:
https://linkedin.com/mallorymejias

Read the Transcript

🤖 Please note this transcript was generated using (you guessed it) AI, so please excuse any errors 🤖

[00:00:00:14 - 00:00:09:17]
Amith
 Welcome to the Sidecar Sync Podcast, your home for all things innovation, artificial intelligence and associations.

[00:00:09:17 - 00:00:25:03]
Amith
 My name is Amith Nagarajan.

[00:00:25:03 - 00:00:27:09]
Mallory
 And my name is Mallory Mejias.

[00:00:27:09 - 00:00:33:22]
Amith
 And we are your hosts. And as usual, there's nothing happening in the world of AI, so I think we're done for the episode, right, Mallory?

[00:00:33:22 - 00:00:43:24]
Mallory
 Yep, 30-second episode, everybody. Thanks for tuning in. Just kidding. We've got a lot to cover today. Amith, I know you were fresh off some travel to ASAE Annual.

[00:00:44:24 - 00:02:14:13]
Amith
 I loved it. It was so great. ASAE did a wonderful job putting together-- they always do this. They're great at putting on events. But Annual Conference was a fantastic event. There were so many people there that were excited about just what's going on in general. And from an AI perspective specifically, every event that I go to, there is a progression in terms of the frequency with which I talk to association creators from large and small associations, from the most senior level positions down to folks that are earlier in their careers. And this progression I'm referring to is essentially really doing things, hands-on doing things with AI. It's more than just experiments. It's gone way past that. And I find that exciting. I mean, we've been out here beating the drum about AI for quite a long time now. It was quite lonely for a while. And now it's not. And it's pretty cool to have a raging party going where everyone's coming to get excited about AI. But the most important thing is I'm hearing people talk about how it's helping them serve their members. It's not just about internal efficiency, but people are saying, hey, I'm able to actually do A, B, and C to make my members' lives better because of AI. So technology is awesome. I'm a technologist at heart. I love the technology. It's fun. But I don't care about it at the end of the day. What matters is whether or not we as associations, like broadly, are advancing our missions or not. And that's what I'm starting to see.

[00:02:15:18 - 00:02:31:00]
Mallory
 Did you hear about any use cases from association leaders that surprised you? Because I feel like as much as we talk about and you think about this, it would be pretty difficult for you to hear about a use case that you hadn't already thought about. But I'm curious. Have you heard anything recently where you said, oh, that's a good idea?

[00:02:31:00 - 00:02:35:12]
Amith
 Have you ever seen the video where there's

[00:02:37:04 - 00:05:22:20]
Amith
 basically a bunch of distractions going on in this video? I'm trying to remember exactly what it is. But essentially, you're watching a landscape, and you're watching all sorts of people moving around. And they're saying, oh, count the number of people that enter and leave the movie's frame or the video's frame. And people are busy counting the number of people coming in, coming out. And there's like, oh, was it 17 people? Was it 19 people? And see, you play this in a room full of people, and the video goes by. And people are arguing about 17 versus 19 or 21. And then someone asks after about a minute or two of discussion, hey, did anybody see the eight-foot-tall gorilla walk through the room and wave at everybody? And the answer is like, what? We didn't see that. It's in the video. And then they play the video again. Instead of focusing on counting all the different individuals coming and going from the video frame, there's this giant gorilla in the background, about halfway through, that's like waving its arms like this and being silly. Presumably, it was a person in a gorilla suit. I never actually checked that out. But it's a gorilla-esque creature. But people didn't notice it is the point. So I guess the reason I bring that up, aside from the fact that it's just kind of fun, is when you are distracted day to day with a lot of tasks that are keeping you underwater, you miss big things. And so what is happening a lot of times in associations is that you talk to people, let's say, who are in member services roles. And they are inundated with one email after the next email, after the next email. And these are smart people who care about your mission, who know your members. They know a lot about the domain. And all they do all day long is count people coming in out of that video frame. And they don't have time to think at all. And so they miss the gorilla. They miss the big opportunity. And what's exciting is that if you can save some time, increase efficiency, and then give your people a little bit more breathing room to take a breath and actually look around and say, hey, what's going on with our members? What's happening? They actually realize there's opportunities to serve your members better, right? There's a lot of bottom up organic idea generation because you no longer starved of oxygen in that creative process. That is part of what's happening that I find exciting is people saying, you know what? This thing, this board deck that used to take me 40 hours to prepare, once a quarter, it took up a whole week. Now it takes me 40 minutes to prepare it. And it's better quality. And I spend that 40 minutes vetting the content. Previously it was kind of this horrible like process. And now it's so automated that I can actually think about it. And then I'm able to reclaim that week, do more actual thinking about what I should be talking to the board about, and then maybe even have some time left over to do other things, right? So that to me is the opportunity is you make some savings happen and then you invest that in improving the experience.

[00:05:22:20 - 00:05:37:07]
Mallory
 Wow, I love a good analogy, Ami. I have not seen the video you're talking about. I'll have to go find it after we record this, but I'm gonna try to carry this through line throughout the episode of finding the gorilla and making sure you're open enough to see it when it walks in the room.

[00:05:37:07 - 00:06:00:06]
Amith
 I just remembered the more specifics if you wanna search it on YouTube. We'll find the video and put it on the show notes as a link for our listeners and viewers on YouTube. But it's specifically a number of, it's a basketball court and there's a number of players of each uniform. So you're counting like red versus blue jerseys or something and then there's somebody in a gorilla suit walks by in the middle of the basketball court and

[00:06:01:15 - 00:06:15:19]
Amith
 it's good fun. But it goes to show people the point that multitasking is not a thing. Our brains only can do one thing at a time. And that's exactly what's happening to us in our roles and trying to serve our members. So we miss out on the bigger signals.

[00:06:17:00 - 00:07:47:01]
Mallory
 Well, speaking of big signals, I think our first topic is gonna be a good relational point to the gorilla. But first, let me talk about what we're covering today. One, a bankrupt airline selling its internal data to Google to help train its AI. We've got OpenAI, the company behind ChatGPT revealing a model so capable it solved open math problems. That takes us to companies actually renting out the computing power behind all of it, a category of company called Neo Clouds. And we close with Google shipping a fast, cheap, new model, 3.7 flash while all of that is playing out. So first and foremost, when a company goes bankrupt, its assets get auctioned off to pay creditors, usually planes in the case of an airline, gates, real estate. What's interesting here is that data was one of the assets on the block. So Spirit Airlines, the budget carrier shut down operations back in May and filed for bankruptcy with billions in debt. As part of winding down, its assets went up for auction. And according to an August 14th court filing, Google won the bid for a chunk of Spirit's internal enterprise data and software code for $10 million. Google says it'll use the data to help improve its products and AI models. Amith, Google already has YouTube and this huge monopoly on search. Arguably, I feel like they have some of the biggest troves of data we have out there on the planet. Why then help us understand why Google would pay $10 million to get Spirit's data, which surely is a lot, but compared to what they have, why is that?

[00:07:47:01 - 00:08:55:12]
Amith
 I don't know what they have. I mean, if you assume it has to do with airline operational data about plane operations, like maintenance schedules and flights and flight scheduling, and then on the customer side ticketing information and CRM type data, that could potentially be interesting in a variety of domains, like logistical domains or CRM domains. Now, they did say that they were going to scrape this data through an intermediary so that Google would not get personally identifiable information, because that would potentially pose all sorts of ethics issues. So that is a factor. But ultimately, even absent the ability to link it to people, if you have operational data from years of flying planes around, I think there might be something in there that potentially Google thinks it could help train better models in some of these domains. But I don't know. I'm speculating. I think the trend line, though, is data is valuable. And I think that even a dead airline's data is valuable to a company like Google. And $10 million, of course, is kind of like pocket change to Google. It's probably smaller than pocket change. It's pocket lint to Google. So

[00:08:56:14 - 00:09:08:17]
Amith
 it's not a lot. But-- and maybe they're speculating. Maybe they're saying, let's just take a look at it and see what's in there. Maybe there's something useful. Or they might have a very, very strong hypothesis that there is something in there operationally that would be of benefit to them.

[00:09:08:17 - 00:09:52:24]
Mallory
 Yep. I did, too. And I should have listed this. I have the list of what it includes. It's finance and accounting records, their operations data, pricing and revenue management models, their internal emails and Teams chats. Interesting. HR and payroll records going back decades. But what's not included per the reporting is customer and loyalty data. The passenger profiles and the frequent flyer records are being kept out of the deal. And then, as you mentioned, a third party will scrub it of any personally identifiable information. So I guess the reason why I thought this might be interesting for our podcast is because we know associations sit on tons of valuable data. And I'm just wondering what you think the bigger signal is there for associations that have this incredible value locked in.

[00:09:52:24 - 00:12:53:02]
Amith
 Yeah, I was really glad you included this, Mallory, because it was a very small piece of news that was flying by alongside many other bigger things. But this does highlight the fact that specialized data is very valuable. All of the models we have from all of the major players-- include the ones that you know from Google and from Anthropic and from OpenAI and from other companies, as well as dozens and dozens of fantastic near frontier enterprise models or capable models from Chinese Labs and others-- they're all very smart. And they've all been trained on basically everything that's publicly available. So the question is, how do you make these models smarter? One potential path is improving the algorithms, making the actual AI models do things they didn't do before. A good example of a step change in that just from a couple of years ago was when the codename Strawberry became 01 from OpenAI. That was the first reasoning model. And a reasoning model simply is a model that has the ability to edit its thoughts and to keep thinking about things for a period of time until it gets to a better answer, somewhat similar to if you were to just have to submit the very first thing that came to mind in answering a complex question, you wouldn't have reasoned through it quite as well. And models used to do that before reasoning models. Models would literally give you their very first spot. But with reasoning models, OpenAI first, and then shortly after basically everyone else, had an algorithmic improvement to allow models to take different budget sizes, like low, medium, and high essentially, and reason for longer and longer if needed. And the more time you give an already intelligent model to think about something, the better it can refine its answer. So that was an algorithmic improvement that dramatically the power of our AI systems. So other algorithmic improvements certainly are on the table. There's a lot of things being investigated by the research community. But at the same time, we know that data is a big, big, big part of why these models are as smart as they are today. And so specialty data is kind of the next frontier to take these models and, on the one hand, the frontier companies like Google want to make their models better at everything. On the other hand, there are people in very narrow domains that are saying, I'm going to make models that are taking a frontier lab's best model or near best model and tuning it to work really, really well in a particular domain like industrial engineering or nuclear science or some branch of medicine or law. So this is going back to your point, a really interesting thing for associations to think deeply about, which is your data is valuable. You're probably sitting here going, duh, I know that. I've been saying that data is one of my most valuable assets for eons. And you've never been wrong about that. But as a practical matter, you've never had a way to unlock the value of that data. And now you potentially do. So not by the way, by going under and then selling your data to Google, I wouldn't recommend that as a business strategy. It is a strategy. It just may not be the best one.

[00:12:54:08 - 00:14:19:24]
Amith
 But there are other ways to think about how you might monetize that data, right? So you can put a layer of AI on top of your data that only you have access to and then make that a paid service to the community, right? Something that nobody expects you to have the world's greatest intelligence in your particular narrow domain, or an AI that's able to research against structured data that you have in something like a medical clinical registry, or to be able to do research using the synthesis of both structured and unstructured data. These are tools or services you could build or you could partner with someone to build that your data is kind of like the special fuel for. And the key to that though, in my mind, is generally to make sure that you keep that data under lock and key, which is a really important concept. That's not the only way to monetize it. You could, for example, medical societies left and right are partnering with Open Evidence, which is a platform for the medical community. And they're getting big checks a lot of times to license their content. Sometimes it's not even their choice because they already have licensed their content to a publisher and the publisher makes a deal with Open Evidence. Then the aggregator, in this case, Open Evidence, basically sucks in that data into their environment. And even in those situations though, there's aspects of your data that you probably didn't license through that agreement. And that's one example. But in those situations, you do get money, but you lose control. So the question is, what makes more sense for you? And there's no one size fits all answer to that.

[00:14:21:05 - 00:14:54:24]
Mallory
 I feel like the treasure trove of your association's data is kind of like the gorilla going on the basketball court and waving to you. But maybe you as a leader are already at the point where you know, say, "Mallory, I see the gorilla. I know the gorilla's there. It's more so what do I do with the gorilla?" So Amis, I'm curious in your conversation with leaders, do you feel like most association leaders are in the camp of regarding our data with our life? We don't know what we're gonna do with it yet. Are they considering AI products that they could build using their data? Are they considering the licensing route of the conversations you have? Where are people at with this?

[00:14:54:24 - 00:15:04:20]
Amith
 It's all of the above and it depends on the individual. And it partly is based on their fluency level with AI, how comfortable they are with ways of safeguarding the asset while leveraging it.

[00:15:05:21 - 00:19:08:14]
Amith
 You know, Disney, the very first time that Disney took a film from their vault and made it a home movie release, it was a big debate because classically what Disney did is every seven years or so, they would re-release a film from the vault in theaters. And that film would generate a tremendous amount of very high margin revenue for them. But they decided right around the time that VHS became the obvious standard, that they would take a film that had been in the vault for a long time and release it on home video. And they tried it first with an old film Pinocchio, which at that point in time had been, you know, re-released I think four or five times already and they had kind of milked it. And they said, "Let's try it here, let's make it and see what happens." And in fact, it didn't take away at all from the IP because the VHS tapes simply built more durable demand for people wanting everything else that Disney offered. And then of course they ended up doing that with lots of other things. And I guess the point I would make is, is that there are ways of leveraging IP or slicing IP that aren't necessarily obvious that may seem as though you're packaging it up and giving it away, but are not. And there are ways to give away your IP without realizing it. So this is something that requires really careful consideration. This is a strategy conversation. It's got nothing to do with technology, other than the fact that the technology enables you to create new ways of leveraging that asset. So this is where I'd say, what I'd start with first and foremost is to actually set aside the concern about the association losing control. That is important, but I temporarily set that aside. And what I would go do is go talk to members and go talk to people who you want to be your members. And I'd find out where they have pain in their lives. I'd find out where they need help. And not about what they need help with you about, cause if you ask them that, they'll probably say, make your website not suck as much. Make your website possible for me to register for a webinar with fewer than 18 clicks. And that is valuable input to hear, but that is tactical. That's more about reducing friction in your current offerings. What I'm talking about is asking them what their lives look like. How do they do their business? How do they practice their profession? Where is their pain? And where would they like to see improvements? Not from you necessarily, just in general. And they might not know about the improvement side. They might only know about the pain cause that's what they can relate to. And where you find pain and where you find pain that's a pattern of pain, you will find opportunity. And then you might think, hmm, interesting. Our members are having a hard time solving cases that are kind of on the edge. The things that are in the 80 percentile that they see every day in whatever field of law engineering or medicine, easy. It's like basically almost instant. But the things that are outside of that, they spend a massively disproportionate amount of time doing research, trying to figure out the right solution, vetting it, working with colleagues, which by the way, this is a pattern I've seen in lots of professions. How do you help that? Well, and does the association have an angle on how they can help with that? What it turns out that associations have knowledge that represents kind of the collection of everything, the sector knows, what if you could activate that to help people in their day to day lives, not just be an episodic partner to them once a year or a couple times a year when they attend a webinar or an event, but give them a tool they can use every single day, day in and day out, that's indispensable for them to do their job dramatically better that basically crushes that pain and opens up new doors. That's the kind of thing you can start to think about doing with AI, right? But you have to think then, okay, but how do I protect the asset? But the point is, what's the value creation that you're doing in your market? And then you have to think about an equally important concept, not just value creation, but value capture. So when we talk about creating value in a market, it's like saying, hey, I've got this ability to create this unbelievable tool that's gonna dramatically reduce the amount of time it takes for people to solve problems A, B and C, cool.

[00:19:09:21 - 00:20:02:19]
Amith
 Do they expect that from your association as part of the package of services and goods they get from membership? Maybe some people do, but largely you would probably find that they didn't even think the association would be part of that discussion. Because they think of you as the brand they trust, but the brand they interact with once a year, somewhat painfully, to renew their credential or their membership. But if all of a sudden you came out and said, hey, I've got this tool that you can use every day, it's in your pocket or on your computer, and it can create this value, people are willing to pay for that. So yeah, maybe you have a tier of that type of service included in membership, or maybe not, but you monetize it like crazy because you're creating an immense amount of value. The value capture that I'm describing is a big part of what people forget to do. Both in associations, by the way, and businesses. There's lots of people out there who've created enormous value in the world who fail to capture much at all of it. So I would really focus on those three things.

[00:20:02:19 - 00:20:18:16]
Mallory
 So would you say, and we're speaking very generally here, that value creation and the sense of maybe creating some sort of AI product that's leveraging your unique data set to help your members in their everyday lives should not be just added on to basic membership?

[00:20:18:16 - 00:20:44:08]
Amith
 Yeah, I mean, how do you value that? Because if I am, let's say I'm serving CPAs, and CPAs are spending 75% of their time solving the 10 percentile cases, right? Most of the work just flows right through, and I can find a way to save them 20% of that time. I can directly calculate the ROI of that. And if I'm capturing zero of that, I know that that's the wrong business model.

[00:20:45:10 - 00:21:21:11]
Amith
 Associations are not expected to provide this kind of service, and to the extent that they are, you can reframe the conversation even if that's the mindset of your industry, because the value you're creating is so immense. This is much more of an entrepreneurial way of thinking about it, it's much more of a quote unquote business minded way of thinking about it. My argument is that's exactly what associations are. They need to be entrepreneurial businesses to solve problems in their market and monetize them. And the monetization in the case of associations simply feeds that funnel right back into the top, which is to create more value for their community. It's not a profit motive ultimately, but you have to be profitable in order to be sustainable.

[00:21:21:11 - 00:21:30:16]
Mallory
 Hmm, so it sounds like there's a lot of value in your data. Got to figure out a way to leverage it for value creation and then ultimately capture it as an entrepreneur would.

[00:21:30:16 - 00:21:31:21]
Amith
 Absolutely.

[00:21:32:22 - 00:22:29:21]
Mallory
 Moving to our next topic today, on August 1st, OpenAI revealed its next major model, which it's calling Astra, I approve of the name, but it didn't reveal it the way we're used to. So we haven't quite seen a launch or pricing or the way to use it. Instead, OpenAI published 10 solutions to longstanding open problems in mathematics and theoretical computer science that an internal version of Astra had worked out. And it shipped each one with a machine checkable proof so the math can be independently verified. The reported compute costs for all 10 problems was only around $2,000. And just as a note, an open math problem is one no one has solved. So not a homework question with a known answer, but a genuine unknown at the frontier of the field. That's the part that got many researchers' attention. I mean, sometimes we talk about AI solving math problems on the podcast, it's come up probably at least a handful of times. Can you help our audience understand why is that such a big deal?

[00:22:29:21 - 00:22:45:13]
Amith
 Well, there's two reasons. One is that there's still a perception in our market. And I think more broadly that these AI models are so-called stochastic parrots, which means they're probabilistic machines that can only repeat what they've heard before, right? They parrot back what they've been taught.

[00:22:46:14 - 00:22:57:17]
Amith
 And that is clearly false because if they're creating novel solutions to problems that have never been solved by us, there is nothing that they're repeating. So that's worth noting is that AI is creating

[00:22:59:05 - 00:24:36:02]
Amith
 significant new things, not just silliness, which is plenty fun, but not really of any enduring value, but rather here, we're creating something that is net new to the world. And if we can do it in math, then the idea is, is that that is kind of the fundamental language of everything else above it, whether it's physics or biology or something else. And so if we can solve open problems in math, then that can ripple across to other things as well. But the bigger point is that a lot of people have this frankly, ill-conceived perception that probabilistic machines cannot create new things. And that may have been true many generations ago, particularly before models had reasoning abilities, but we're long past that actually in a lot of fields. And this is actually a very clean way of proving it because if you were to say, well, no AMS has ever done X, well, okay, that's kind of silly first of all, but also it's one of these things where can you really prove that or not? Whereas math is something that is an open field. It is something where there's known problems that have not been solved, that have been posited as potentially solvable or potentially unsolvable for in some cases decades. And now you have a machine solving them. So it's kind of indisputable in a sense that it's creating something new for the world to benefit from. Whereas there's a lot of other fields where you might say, well, it created a novel. Is this novel really new work or is it not? With the math side of it, there was no solution for this problem, now there is. So it's a pretty clear indicator that the AI is indeed creating something that's net new.

[00:24:36:02 - 00:24:58:14]
Mallory
 Wow, it's almost unreal to think about because I understand your point of AI, not creative or generative AI, not creating anything new because it's been trained on data that we already had. But the fact that that is able to provide, give it the ability to create novel solutions right now is almost unbelievable but very exciting for I think the future of science, medicine, all those things.

[00:24:58:14 - 00:29:36:17]
Amith
 One of the things that I think I talked about this on the pod before, but I'll mention it briefly and I won't go too far down this rabbit hole, but it might be interesting to quite a number of our listeners is that AI models have kind of a way of adjusting how free form they can flow or how creative they can be. And it's something called temperature. And so the temperature basically reflects how wide of a distribution probabilistically the model is emitting results from. So I'll give you an example. If you were to say the temperature is zero, meaning it's as cold as possible, the model will always give you the highest possible probability answer to your question. So when you think about like in the training data, if there was, you have like a set of words and what's the next word, right? And the classical way we've talked about this and this is true for images and audio and so forth as well, there's a distribution that comes back. So you run these 10 tokens through the machine and it says, hey, for the 11th token, it is one of these 20 possibilities and they're ranked in order by probability. This one's 0.95 likely, this one's 0.75, this one's 0.35, et cetera. So you have what's called a probability distribution of the next token or the next tokens. So why does this matter? If you have a temperature of zero, you will always get the very top choice, always the highest probability, which means it's effectively deterministic relative to that model. It'll always spit out the same answer for the same input. Same inputs, always get the same outputs. Whereas if a temperature is much higher, you're giving the model more dynamic range essentially to be able to say, hey, sometimes pick the top one, but sometimes pick something a little bit lower probability, mix it up a little bit, make it fun. So that's kind of what you're telling the model. And you can experiment with this yourself, not in the consumer grade tools like Clod or OpenAI's chat GPT, but if you go to a tool called Playground, which all the major labs offer this, you can just Google search OpenAI Playground or Anthropic Playground, you log in with the same account, but here you go to a little bit, it's a little bit overwhelming looking, but there's a Playground, I think Mallory, you might've played with the Playground before, right? I have played. It takes a minute to adjust to, and I'll let you talk about your experience with that in a second, but you do have more settings there. So you do have like a little control panel where you can say, now I'm kind of like peering into the cockpit of a jet aircraft, and there's all these little dials and switches. First thing is, is you're not flying a plane, so don't be afraid to like, you know, mess around with them, you're not gonna crash anything, just test out and see what happens. But the one I'd point you to is called Temperature, play around with it, and try the same question over and over with different temperatures, and you'll see different results. Whereas if you have a very low temperature, zero, you'll get the exact same results every time. Now, why I point that out, even since much earlier generation models, we've had the ability to use AI creatively. Some called it hallucinations, which is actually what it is if you think of it as emitting truth or fact. But if you are actually looking for a creative partner to come up with a range of possibilities, hypotheses, for example, to a scientific dilemma, hypotheses to a open math problem, it's actually quite helpful to have a wide range of temperatures so that you get more creative hypothesis generation, which then can result in other downstream processes testing out all these creative hypotheses that our brains probably would not ever come up with. And that is some of how these systems are working is that they have kind of this creative process where they come up with ideas and they have so much power and so much capability, they didn't test all these ideas in parallel. So they kind of brute force a lot of these creative solutions. And that's a big part of what we're doing here with AI systems is we're using them in ways that seem totally counterintuitive. We wanna solve math problems which definitively can either be proven to be correct or incorrect. So we start off with this really wide creative space, right? Well, why haven't these problems been solved? Well, part of it is because we train ourselves to think like all the mathematicians of years past, that's how we learn math, that's how we learn physics, that's how we learn how to write, that's our training set. And we have limited bandwidth for this. So we necessarily have a low temperature to use the language of AI. So that's a long way of saying you can apply this novel creative process to your own work. Just remember if you increase the temperature and this is true even if you don't, don't treat AI answers as absolute. You have to fact check them even today, but you can use these tools in different ways to get very different outputs and that can allow you to look for novel solutions.

[00:29:37:18 - 00:29:55:24]
Mallory
 So it sounds like if you were doing a creative brainstorming session, you would crank the temperature up and get some good results. However, if an association had a knowledge assistant that was trained on its scientific journal publishings, you would probably want your temperature at what? At zero or all the way down? If members were asking questions of the content?

[00:29:55:24 - 00:31:13:24]
Amith
 It depends. If you're doing something that's a very simple knowledge retrieval, absolutely. You wanna ground it in fact, you wanna make sure it's 100% vetted and you actually go back and essentially fact check the work. And that's actually how knowledge assistants like Betty do their work is they're doing multiple passes at it to guarantee that the answers are grounded in truth. But if you ask that same knowledge assistant to help you solve a problem and that problem isn't clear in terms of like if you say, hey, what's the answer to this problem? And it's like a known way of doing something. Of course, there's probably lots of content that says this is how you solve it and you can get the correct answer. But that's obviously something that is much more of like a Google search. It's like, it's just a better retrieval. But if you want to have the knowledge assistant actually work with you the way I'm describing, you could have knowledge assistants that have more creative modes allowed. You just have to make sure that you ultimately come back to vetting and validating that the process that was used to create the answer is actually verifiable. Some domains are more naturally suited for this. Math and physics are ones that are more obvious because there's ways of actually proving these things out specifically with math. You can have proofs attached to these things as open AI did. In other domains, it's less obvious. You have to be thoughtful and careful about it. But I wouldn't be so strict as to say a knowledge assistant should never have degrees of creative freedom.

[00:31:15:13 - 00:31:40:17]
Mallory
 So it seems obvious that at this moment in time, AI is a really great assistant in the research process. But given the advancements we're seeing with Astra, it sounds like AI is and will be producing original research now and very soon. So for associations that are scientific medical engineering, how do you recommend that they think about the idea of AI producing original research?

[00:31:40:17 - 00:32:55:07]
Amith
 I mean, it piggybacks Mallory and our last topic in that you have a data set. In this case, it's the unstructured data as well as perhaps some of the structured data that your association holds. Could it be a research partner for people in your community to come up with novel solutions? And what's the financial model around that? Is there a licensing role? Is it like the way universities deal with commercialization where if you have essentially this workbench that people are using to create novel compounds for drug research, for example, based on years of clinical data, your association and only your association holds, do you have some upside in that? Perhaps the way university might have a small piece of a biotech venture that gets spun out and commercialized from its research. So there are thoughts like that that come to mind, but ultimately it comes back to the same principles that we've talked about a number of times. One is control and safety around your data. Another one is really understanding what you're doing. So if you're providing a tool that's intended to be solely used for correct information retrieval, it's not gonna do some of this stuff. But if you want to have a research partner, maybe it's a different agent that has a different level of creative freedom that has the ability to hypothesize and test ideas, that could be perhaps a different service you offer.

[00:32:56:17 - 00:33:13:11]
Mallory
 And if this topic is interesting to you all, we did an episode with IFT, Institute of Food Technologists, a while back, I'll add the link to the show notes where we talked about them doing exactly what we're saying, which is creating something called co-developer, which is an AI research assistant within the field of food science.

[00:33:13:11 - 00:33:17:18]
Amith
 So we should check back in with IFT and see how they're doing with that. That was a little while ago now.

[00:33:17:18 - 00:33:26:14]
Mallory
 I was gonna say, it's probably a bit outdated at this point, but yes, we should. I would be very, I just looked it up to make sure it's still around, it's still kicking, so we should bring them back on.

[00:33:26:14 - 00:33:35:10]
Amith
 Well, knowing the folks at IFT, I'm sure it's evolved a ton since then. They're a very dynamic lot and doing creative, fun stuff, so it'd be awesome to have them give us an update on what's going on.

[00:33:36:16 - 00:34:57:01]
Mallory
 Moving to our next topic, Neoclouds. Astra's a good ramp up to this one because that $2,000 figure that I mentioned is the cost of running an already trained model. Training a frontier model like Astra takes enormous fleets of specialized chips, and increasingly companies don't own those fleets. They rent them, and the companies they rent from are what people are starting to call Neoclouds. You've heard of the big cloud providers like Amazon's AWS, Microsoft Azure, Google Cloud. Those are hyperscalers, and they offer thousands of different services. A Neocloud is a newer, more specialized company that basically does one thing, rents out high-end AI chips, mostly in videos, as a service. Why did this category even need to exist when we already had giant cloud companies? The short version is demand for AI computing has outstripped what even the hyperscalers can supply. So a whole tier of specialists sprang up to fill the gap, and interestingly, the big hyperscalers are even renting from the Neoclouds to expand their own capacity. So, Amit, I came across the term Neoclouds probably a few weeks ago and realized, what is this? And I had no idea, so I thought it would be interesting to include on the pod, what do you think the existence or creation of Neoclouds as a whole tells us about where AI is right now?

[00:34:57:01 - 00:35:39:05]
Amith
 Well, one is just demand. I mean, when you have this much demand for something as powerful as emerging AI capabilities, how do you serve that demand? So every time there's been a boom cycle in any industry, there's all sorts of new players that are in the market. This is just the term of art people who have attached to this idea. And here, what we're dealing with this focus, what we're dealing with is, yes, AWS, GCP, and Azure have enormous infrastructure, but they're also massive organizations. They can't move nearly as fast as these organizations that are starting largely from scratch in the last several years. So that's a lot of what it is. Hyper-specialization, hyper-focused, a lot of capital, moving quickly, setting up big data centers.

[00:35:40:10 - 00:35:45:09]
Amith
 And then, yeah, like you said, renting that capacity out, that's the business model these guys have.

[00:35:45:09 - 00:36:31:23]
Mallory
 Yep. So the scale of this is staggering. I was impressed. CoreWeave, one of these companies, reported a backlog of committed future business worth of $100 billion, and Meta has committed something like $48 billion combined across CoreWeave and Nebbius. And despite the explosive revenue though, the major Neoclouds are all losing money under standard accounting, largely because the chips are wildly expensive and they lose value quite quickly. They've taken on heavy debt, often borrowed against the chips themselves, and then they have big repayments coming due over the next couple of years. So, Amit, you kind of talked about the demand for compute. Associations obviously don't buy from Neoclouds directly, but do you think there are any lessons to be learned or takeaways from the idea of GPUs as a service?

[00:36:31:23 - 00:39:30:22]
Amith
 Well, you know, this is a lot like when people were putting down a ton of fiber optic cable at the beginning of the dot com boom in order to anticipate all the things that would come. And the people who made those investments largely weren't the ones who realized the benefit from them. And so, you know, you had a lot of people go belly up because the capital required to do that kind of work was enormous. But actually putting all that fiber down really did help the world because it's given us the bandwidth to do all the crazy things we do on the internet now. And of course there's been continual investment in growing that. Here what we have is a similar concept, but fiber is a lot more like real estate. It has some enduring value. You know, new lines and new types of fiber need to be put down to have bigger and bigger pipes, but the old pipes still work. And that's somewhat true for GPUs that, you know, like A100 chips from NVIDIA, which are a couple generations old now, they didn't stop working generally. They still run, but they're a tiny fraction of the power of the newer chips. So there really is a practical matter not gonna be used for very long. So these chips might have maybe a five year lifespan. Typically I think they're being depreciated on a two or three year cycle. So in contrast, what we're dealing with here is the real estate investment, to use that analogy, actually is the physical real estate of the data centers, which is also a depreciating asset. You know, the land is potentially an appreciating asset, but the improvements on top of it are always depreciating, but the chips are depreciating at a very rapid pace. So the point would be the financial engineering going on here between, you know, commitments coming from different companies that then, you know, buy like NVIDIA sells the chips and then rents them back and rents them to someone else. There's a lot of things going on right now that are very fast moving. They seem a little bit loose from a regulatory perspective. They seem, I don't know, just, there are things that remind me of eras in the past when things have moved really fast in order to fulfill demand or fulfill investor demand in some cases. And in this case, the revenue is essentially capacity constraints still. So at some point does that level off and then there's a big crash. I think that is possible, but my suspicion is actually it's more that not that revenue levels off for these firms in collectively, but the scale players, whoever gets to some economy of scale sooner that actually can reach operational profitability survives. And then everyone else goes away. And then the infrastructure is still physically there. So it gets sold off in bankruptcy or acquisitions occur, pennies on the dollar. Ultimately, it's good news for the consumer side of the market. It's, I don't know how you manage the risks and probabilities of these various investments, but I think it's generally a good thing that this investment is happening because it makes it possible to run the inference workloads we have right now. But I would just be very thoughtful about how you think about one Neo cloud versus the other versus the other. It's like, those are things that I'll leave to others to speculate on, but the capacity is needed is the short version of the story. And these guys have sprung up to fill the need.

[00:39:31:24 - 00:39:34:17]
Mallory
 All right. And now I know what that term means. So thank you for that.

[00:39:35:21 - 00:40:11:11]
Mallory
 I wanna move to our last beat for today. We're seeing obviously the whole industry pouring capital into ever bigger training runs, but maybe we're seeing Google go in the other direction, at least partially. They quietly released Gemini 3.7 Flash on August 13th, only a few weeks after its last workhorse update. Flash as a reminder is Google's line of smaller, faster, cheaper models, not the biggest brain in the lineup, the one built to be quick and inexpensive for everyday high volume tasks. And as a note, I would say, Amith, you are a big fan of the Flash models as well. So have you tried out 3.7 and what do you think?

[00:40:12:14 - 00:42:58:24]
Amith
 Yeah, I tried it out the day it came out and I've been using it as a daily driver for software development pretty much ever since. So it's been a whole week. Actually just posted some thoughts. Yeah, it's been a long time in AI land. I posted a comment on LinkedIn this morning about my use of it. And what I would say is it's incredibly smart. It's probably like Opus 4.8, Opus 4.7, 4.8 kind of caliber intelligence. But as a practical matter, I've said this a number of times, even those models were plenty smart to do the vast majority of things. So as frontier models keep getting smarter and smarter and smarter, the question is, okay, but it doesn't matter. Do you really need that to draft that email? Do you really need that to write that code? Do you really need it in order to make that next board deck? Or as a practical matter, most of the tokens being consumed now are actually agents making decisions or agents doing iterative workloads. And so for most things, you're somewhere kind of in the middle in terms of the intelligence relative to cost and power and all that. So Flash is a great model because what they're essentially saying is, listen, we're gonna come after the fat middle of the market where you need plenty of intelligence, but you want something trustworthy from a vendor that you know and trust, presumably, and at scale, because Google is Google and they have compute capabilities that are really hard to match, both because they're Google, but also because they have the TPU architecture, which we have spoken about briefly on the pod in the past. But TPU is their tensor processing units. They're on the eighth generation of the TPU, and they have broken it out into two subcategories, TPUI and TPU for training. TPUI is their inference chip. The newest two Gemini 3.6 and 3.7 Flash models are on the new TPUI8 chip, which is their best, fastest chip. What matters to you as an association leader is any of that nerdiness, but rather the fact that Gemini 3.7 Flash is incredibly fast. It is so fast, 300 and something tokens per second, which is on the order of somewhere in the range of five to seven times faster than what you're gonna get from Clod or OpenAI. It's noticeably faster. They've upgraded a number of their consumer facing products to use this model for coders, their anti-gravity agent, which is their Clod code type thing, is incredibly good and really, really fast. The other side of it is they cut their price in half. They said it's through the end of the year, but basically it means it's forever because by then there'll be like three more models and 3.7 Flash will be old news. So I think it's a really important thing to pay attention to because for most of your work, Gemini 3.7 Flash will be far enough along in terms of its intelligence that it'll be more than enough actually for what you do. So I'd recommend people check it out.

[00:42:58:24 - 00:43:06:06]
Mallory
 And where are you using 3.7 Flash? Is that the Gemini, the regular user interface or are you using it through the API? How do you use it?

[00:43:06:06 - 00:43:40:07]
Amith
 I've used it through the regular user interface, just the Gemini website, which is great. It's lightning fast there, but I've primarily used it through anti-gravity, which is Google's cool name for their agent harness, which is basically like their Clod code equivalent. And in that environment, I did post this morning on LinkedIn that it's a little bit buggy. It's not as well rounded, I'd say, or as polished as Clod code in certain ways. I mean, it's a newer product, I kind of expect that. I think very quickly it'll be very good and at parity with Clod code. But other than that,

[00:43:41:07 - 00:44:03:11]
Amith
 Gemini 3.7 Flash has been truly extraordinary this last week. I've gotten way more done with it than I would have ever gotten done with Clod because it's just so instant and I can run multiple things in parallel. I'd also say this about Google is their plans are very generous in terms of the token limits they give you. I haven't even paid attention to that because I've never been even told I was coming close to my limits. And I've hit this thing pretty hard.

[00:44:03:11 - 00:44:16:04]
Mallory
 So you talked to a lot of association leaders about the projects that they're working on, the experiments they're running. Would you say most of what you hear about from association leaders right now could be done with 3.7 Flash?

[00:44:16:04 - 00:47:56:05]
Amith
 100%. I think there's very few workloads that associations do that even need 3.7 Flash's intelligence. It's an incredibly smart model. It's really fast. It has, I think, a million token context window. It's fully multimodal natively. And something I love about the Google models is Gemini has been multimodal since the beginning. You mentioned at the beginning of the episode that Google owns YouTube and they have a lot of other data. And so Google from the beginning with their Gemini models, or maybe it was from version two, but early on, they've been natively multimodal. So that means that these models reason across image, text, audio, video, all at the same time. It's really kind of amazing how good they are. Like you give it a screenshot of something and it immediately understands it at a native level. The other models, by the way, from Claude and OpenAI do that as well these days, but they have a history of doing this really, really well. I guess the way I tie it together is this. So yes, for your, what I call consumer use cases, as an individual association leader, solving one problem on your desktop, Gemini is definitely worth checking out because of its speed and its intelligence. Definitely check it out. But if you plug it into an agentic system where agents are running thousands of prompts per minute or per hour, this is where it really starts to pay off. I'll give you one example, Mallory. One of the agent products that we have in the association market is called skip. Skip is an AI analyst. So what you do is you have a chat with skip and skip is connected securely to your database or databases, any amount of data you wanna throw at it. And skip is able to create these beautiful dashboards and interactive analytics essentially. It's really quite amazing. Well, we've had skip for a while and we've worked through all sorts of different models with skip. Skip does work with open AI and entropic as well, but with Google, we've been mainly focused with Google in 2026. We plugged in 3.7 flash shortly after it came out and a typical request with skip when it's producing a complex report might take anywhere from in the past five to 15 minutes. 3.7 flash is twice as fast as its predecessor and considerably better. It produces way better output with the exact same harness that we have set up. So it's a notable improvement because that means it's gonna get used more. It's also half the price, so that's a nice benefit. And so aside from your internal workflow, if you were to take a tool like skip and say, you know what, part of the value prop we have going back to the earlier discussion of your data is maybe having something like skip or skip itself available to your customers, to your members, to be able to do research, to be able to do certain things with some of the databases you have. Well, that is really cheap to run, really smart, really fast. That makes it more attractive to scale up in a big way. And that's what we're seeing here in the trend line. I will say one thing that's negative about Google, it does suck that they have not released a pro edition of their model in some time. I'm hopeful that that will get resolved soon and they'll wow us all in terms of whatever the next Gemini pro model will be. It'll probably be 4.0, but I'm hopeful. I'm actually optimistic because I do think they still have an incredible advantage in a number of areas, data being one. Not because of Spirit Airlines, by the way, just in general. So I'm hopeful that they'll remedy that. But independent of that, Google is going to have a fantastic business model around Flash because it is a far stronger offering in terms of cost and performance than the loads here, like the Luna offering or the Haiku offering. So the only thing that's actually even somewhat close on other providers, but even that, in terms of speed and cost, I think Flash would be worth considering.

[00:47:57:21 - 00:48:33:02]
Mallory
 I feel like we've talked about on the podcast sometimes, Google is late to the party or it seems that way within the crazy world of AI. But when they show up, they show up ready to go. So maybe they're making us wait for good reason. But Amith, I was going to ask you, the takeaway from this episode like I normally do, but I think maybe a better question is going back to the gorilla at the beginning. For the person listening to this podcast, that feels like they are way too focused on the counting, people coming on and off the court, and they're really struggling to free up time to notice the gorilla, the opportunity to serve their members and their association. What is your advice for that person?

[00:48:33:02 - 00:49:06:19]
Amith
 Well, the first thing is continuously work on your AI fluency. If you're here with us on the sidecar sync, first of all, we appreciate you. It takes time to listen to a podcast, and you have a lot of choices out there. We appreciate you for being part of our community. But being part of this community means you're continually working on your fluency with AI. And the more fluent you are in any language, be it AI or something else, the more likely you are to use that skill, and the more likely you are to recognize when new things come out that are really helpful. And so you're likely to not

[00:49:08:04 - 00:49:43:09]
Amith
 be unaware of these important developments. The other thing is that if you improve your fluency, you're more likely to knock out some of the busy work. All of us, and I mean all of us, it doesn't matter if you're the president of a country, or if you're an entry-level employee who just got your first association job, you have busy work. I do as well. How can we squish that down and eliminate more and more and more of it while still thinking about it, but not doing the actual task? That will help you get the oxygen back in your tank so that you have energy, you can be creative, and you'll spot the gorilla and take advantage of those gorilla moments.

[00:49:44:16 - 00:50:01:10]
Mallory
 Gorilla moments. Everybody, to recap today, a bankrupt Spirit Airlines selling its internal data to Google to help train AI. We've got OpenAI revealing Astra, a model that solved open math problems. Then the NeoCloud specialized companies renting out the computing power behind all of it,

[00:50:01:10 - 00:50:07:08]
 (Music Playing)

[00:50:18:01 - 00:50:35:00]
Mallory
 Thanks for tuning into the Sidecar Sync podcast. If you want to dive deeper into anything mentioned in this episode, please check out the links in our show notes. And if you're looking for more in-depth AI education for you, your entire team, or your members, head to sidecar.ai.

[00:50:35:00 - 00:50:38:06]
 (Music Playing)