Summary:
In this episode of the Sidecar Sync, Amith Nagarajan and Mallory Mejias dig into the hidden world of “dark databases”—messy, undocumented systems quietly powering many associations—and explore DBAutoDoc, a new open-source tool from Amith and Thomas Altman that uses AI to map database structure, infer relationships, and generate documentation at scale. Amith & Mallory also unpack the state of AI video after OpenAI’s Sora shutdown, explaining why the category is not cooling off, which models are leading the race, and where associations might find real value in video generation, personalized learning, website assistants, and member communications. Along the way, they touch on data readiness, AI audio versus video, governance, privacy, and why focus may be the most important leadership skill in the AI era.
Timestamps:
00:00 - Welcome to Sidecar Sync
👥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
📕 Download ‘Ascend 3rd Edition: Unlocking the Power of AI for Associations’ for FREE
🛠 AI Tools and Resources Mentioned in This Episode:
Amith & Thomas's Paper ➔ https://arxiv.org/abs/2603.23050
DBAutoDoc ➔ https://github.com/MemberJunction/MJ/tree/next/packages/DBAutoDoc
MemberJunction ➔ https://memberjunction.org/
Google Veo ➔ https://deepmind.google/models/veo/
Kling AI ➔ https://kling.ai/
Seedance 2.0 ➔ https://seed.bytedance.com/en/seedance2_0
Runway Gen-4.5 ➔ https://runwayml.com/
https://www.linkedin.com/company/sidecar-global
https://twitter.com/sidecarglobal
https://www.youtube.com/@SidecarSync
⚙️ Other Resources from Sidecar:
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
🤖 Please note this transcript was generated using (you guessed it) AI, so please excuse any errors 🤖
[01:00:00:14 - 01:00:09:17]
Amith
Welcome to the Sidecar Sync Podcast, your home for all things innovation, artificial intelligence and associations.
[01:00:09:17 - 01:00:27:11]
Amith
My name is Amith Nagarajan.
[01:00:27:11 - 01:00:29:15]
Mallory
And my name is Mallory Mejias.
[01:00:29:15 - 01:00:48:19]
Amith
And we are your hosts. And we have very interesting topics to cover with you today all about the latest in database technology and AI video and probably some other fun things that we'll throw in there just for the heck of it and see what's going on in the world of AI. Before we do that, though, how you doing Mallory?
[01:00:48:19 - 01:01:14:02]
Mallory
I'm doing pretty well, Amith, a little bit under the weather this week. But actually today I woke up feeling good, feeling much better. So I'm happy about that. As always, I use my handy medical assistant, Claude, which maybe I shouldn't do to say, "Here are my symptoms. What do you think?" But Claude always says, "You know, I'm not a doctor, just so you know. Always consult a medical professional." But it's always helpful to have someone talk me off the ledge when I'm having symptoms. How are you?
[01:01:14:02 - 01:01:56:11]
Amith
I'm doing pretty good. Most of the time if you have a cough, it's not tuberculosis, but it's good to check. Yeah, it could be. It could be a lot of things. I've been using Claude and actually Google's built-in AI search mode just in the browser a lot related to puppy health issues. I've been feeling pretty well, but we have a now almost 13-week old golden retriever. He's a handful. He's awesome. He's sleeping through the night, which is great. But he's had a couple of different incidents recently, as little puppies do. And so I've been all over AI trying to figure out if the dog's okay. So at this point, I think Claude has a pretty good idea that I've got some pretty bad OCD around my dog.
[01:01:56:11 - 01:02:12:08]
Mallory
Yeah. I think one of our most common, probably AI questions in my household or Google questions is can dogs eat bananas? Can dogs eat apples? Can dogs? Because they can eat most things, but then there's those random things like grapes that you're not supposed to feed to a dog. So I totally understand that.
[01:02:12:08 - 01:02:56:22]
Amith
Yeah. And it's an interesting topic because the knowledge base behind those answers that we're seeking is interesting because we're using broad-based consumer AI tools. And this is, of course, one of the plays that associations have is that we're speaking about consumer-type questions, but as professionals in their domains have deeper questions on topics that you as an association of expertise in, that's an untapped opportunity for most of our listeners. There's lots of ways to use AI to activate the latent asset that is your historical content. So just something to think about because these questions are voluminous out there and most of them go unanswered or they're answered in lots of different ways. But you want to be the place where the questions get answered for your profession.
[01:02:58:12 - 01:03:32:07]
Mallory
A lot of times too, what I'll do with Google searches, say, because I've told you, Ametha, and I've said on the podcast before, I enjoy using Reddit when I'm trying to crowdsource information about brands and how people really feel about consumer products. So I'll often search something, "Best paint color for this Reddit," and then Google will give you the AI overview of all the Reddit threads. So thinking about that, if it were your association, what is the best procedure for X and it gives you the AI overview from the association of these medical procedures? I don't know. It could be useful.
[01:03:33:20 - 01:03:34:00]
Amith
Definitely.
[01:03:35:18 - 01:04:21:07]
Mallory
Well, Ameth, we have two interesting topics for today. First we're digging into a brand new research paper that you just published along with Thomas Altman on what you call dark databases, the messy undocumented systems that almost every organization is quietly running on, including a lot of association tech. And then I wanted to talk about AI video today. I feel like it's been a minute since we've done that. Open AI shut down Sora last week. And on the surface, it might look like that field is cooling off, but maybe that's not the case. So we'll walk through what's actually leading right now and what that means for you as associations. So first off, dark databases. Ameth, I'm going to give my overview, but you're the one who published this paper, so please stop me if you have anything to add in.
[01:04:22:12 - 01:05:12:11]
Mallory
Most of the databases running businesses today are a mess. Column names are shorthand that made sense to a developer who left years ago. The connections between tables have been stripped out for performance and documentation, if it ever existed, is out of date. Researchers call these dark databases, and they are the default state of most enterprise systems, including a lot of the AMS, CRM, and LMS platforms our listeners are using today. So Ameth, you and Thomas Altman, one of our colleagues from Tassio Labs, another Blue Cypress company, just published a paper introducing DB AutoDoc, a system that automatically figures out what is in one of these databases and writes complete documentation for it. The paper went up back in March. We'll include that in the show notes. And the system is open source as part of the Member Junction AI data platform.
[01:05:13:11 - 01:06:21:06]
Mallory
The core idea here is that understanding a messy database is not a one-shot task. It works the way a human expert would. Take a first pass, form some early guesses, then go back through and refine them. DB AutoDoc combines statistical analysis of the data with an AI model reasoning about what it means, and the two correct each other across passes. The numbers are striking. On a standard benchmark database, the system identified 95% of the table relationships correctly and wrote accurate descriptions for 99% of the columns. It did this in about two passes for under a dollar per hundred tables. Manual documentation by a human expert might cost between $12,000 to $48,000 for a database of a similar size. The most useful part of the paper for our audience is probably the two real-world case studies. One was a professional membership association on a Salesforce-based AMS with 36 tables, 1,800-plus columns, and zero documentation. The system mapped it almost completely. The second was a trade association on a custom system with 125 tables. Same result.
[01:06:22:10 - 01:06:27:24]
Mallory
So, Amit, how did I do? You are the publisher, the researcher, so let me know. Did I summarize this well?
[01:06:29:04 - 01:06:41:10]
Amith
I think you did a great job, Mallory, and this is a big problem. Databases exist in our world in lots of flavors. A lot of them are poorly documented or just completely without documentation entirely.
[01:06:42:12 - 01:07:02:12]
Amith
The reason for this is databases sometimes initially start out in life with some level of documentation, but oftentimes it's not something that business users ask for or consider historically. It's been kind of the back end of the system. It's something that developers might have done a little bit of work on, but most of the time that's not the case.
[01:07:04:17 - 01:07:42:05]
Amith
The issue is this problem compounds over time. If you had a system that was built five years ago, maybe some of that knowledge is relatively fresh, but if you built it 25 years ago, a lot of times the people who built the system are long gone and the people who are maintaining it for you probably have a very limited understanding of the data itself. It becomes a big problem in a number of levels. First of all, just maintaining those systems is difficult, and so documentation helps improve the probability of being able to successfully maintain them. The other problem that is, I think, the biggest issue for association people from the business side is being able to understand their data, use their data at scale.
[01:07:43:07 - 01:10:10:11]
Amith
Thomas Altman, one of the founders of Tassio, as you mentioned, one of the companies in the Blue Cypress family, he and I have been working together for many, many years, even well before Tassio Labs and Blue Cypress, talking about data, enterprise data, this issue. In the last roughly four years, we've been working actively on this other project called Skip. For some of our listeners, you're familiar with Skip. Skip is an agentic AI, very advanced AI, that is capable of just interacting with you through a normal chat and producing stunning dashboards and charts and graphs and all sorts of things in order to really give you incredibly deep insight into your data, not just presenting pretty pictures, but also be able to explain them to you, give you ideas for research, and then actually research those ideas for you. To solve that problem, we had to go back and say, "Well, look, the AI is not going to understand databases that have no documentation." Just like a human, if an AI is presented with this really big, complicated database and you say, "I want to know how many members I have. I want to know how many members I have renewing year over year. I'd like to know if I have a problem with retention. Is there a geographic trend line in there somewhere?" All these questions that business people tend to have, you can't answer those if you don't understand the way the database is organized, which is somewhat of an obvious statement, I guess, but people haven't really thought about this too deeply because they've always been constrained by a different choke point. The choke point people have been constrained by is just the amount of human labor available that's expert enough to write reports in Tableau or Power BI or SSRS or whatever the tool is. Using a report writing tool, even the fairly easy to use ones like Domo or Power BI, these are tools that still require a lot of knowledge of the database. What we're trying to solve is to fully automate the whole concept and to do that you have to have good documentation of the database. That's really where the problem started. This applies, what we're talking about in this particular paper is about structured relational databases primarily, but the concept also can apply to unstructured data sources. That's just not the particular focus of this paper. Our hypothesis was could you use an AI to fully document a completely undocumented database and a database that was extremely poorly conceived of, meaning a database that wasn't built with all the features and thought processes that a well-designed database would have, but rather real world databases.
[01:10:11:12 - 01:10:52:22]
Amith
We even have a test database we call lousy database or lousy DB. It's designed to intentionally be amongst the most horrendous databases you've ever put your eyes on, the table names, the column names. They're all really cryptic and nothing is documented and there are no keys, which are the links between tables. This is very, very hard to figure out and we've run that through AutoDoc as well and been able to get pretty stunning results. Very excited about this because this free tool enables any organization to now document all of their data and to continually keep it up to date, which in turn of course allows AI systems as well as humans to intelligently interact with that data at scale. Really pumped about this. It's a big unlock.
[01:10:54:01 - 01:11:00:21]
Mallory
Ameth, with the real world case studies, how were you able to verify that the AI was correct? Was there a human verification piece?
[01:11:00:21 - 01:11:21:10]
Amith
Yeah, that's a great question. Think of it almost like go back in time to when you were in school and you were given a worksheet or a workbook and the teacher had a very similar workbook, but they had a special version of that book that had an answer key in the back, which of course, as students, we would have loved to have had the answer key version.
[01:11:22:10 - 01:12:11:10]
Amith
But the AI is the student essentially. So the AI doesn't have the answer sheet. So what we're able to do is we do take some systems we have absolutely no documentation on and we review them manually, but we also have some systems that we do have extensive correct documentation on. So what we're able to do is run the AI systems on these databases that do have documentation. But what we do is we feed the AI a stripped down version of the database that has no documentation, has no linkages or keys. And by doing that, we're giving the database really in its raw form to the AI and we're able to validate did it get everything right, you know, field by field, table by table, key by key, and then the macro view of how the whole thing fits together. Because databases are kind of like giant jigsaw puzzles. You have to really see the whole picture in order to get the most value from it.
[01:12:11:10 - 01:12:21:19]
Mallory
Mm hmm. For an association leader who hasn't thought about their AMS this way, for example, how would they know if they have this problem or is it safe to say everybody's got this problem?
[01:12:21:19 - 01:13:01:14]
Amith
I was going to say, I think probably everyone has this problem and feels the pain regularly. Part of what happens is you hire someone to build, you know, the classical way you do is you hire a consultant or you have a staff member building a report, building a dashboard, or just extending your system. The absence of good documentation causes all sorts of problems. First of all, it slows everything down. So instead of being able to do something fairly quickly, if they knew where everything was, it takes the report writer a long, long time to know where everything is and to get the data organized. And then you're more likely to have all sorts of incorrect information and all sorts of defects in the database and in the reports that you write.
[01:13:01:14 - 01:13:25:17]
Mallory
Mm hmm. I know when we've talked about the Member Junction open source free AI data platform for associations in the past, something that we've heard often is, well, my data is not ready, my data is too messy. And we've always said AI can help with that. AI can identify duplicates in your data. Is this just another kind of version of how AI can assist you with taking your messy, undocumented data and getting it to a good place for an AI data platform?
[01:13:25:17 - 01:13:38:21]
Amith
Yeah, so it's part of the story. So if you understand the database structure, then you can start to understand if you do have these other problems like duplicate data and duplicate data you tend to know when you see it.
[01:13:39:22 - 01:14:25:23]
Amith
And it's one of these things that you know, your member data has duplicates, you know, your product data or your meeting data may have duplicates. So they're kind of parallel problems, but the lack of data based documentation is a big, big problem. Now, if you want to have any shot at using AI to do analytics or to write code for you to build applications, you need to have your data house in order, as we like to say, that includes both having at least somewhat cleaner data, but to have cleaner data, you need to understand the database structure. And the absence of understanding that structure means you can't really clean it up that well. You know, if you take the analogy of a house, you know, the dirty data that exists all over the place is kind of like having, you know, teenagers live in your house and there's just garbage everywhere.
[01:14:27:04 - 01:14:56:19]
Amith
But the absence of database documentation is akin to not knowing where certain rooms are, right? You don't know that, oh, OK, my son lives in this room or there is a bedroom over there. It's like completely, you know, dark and dusty and I don't know what's behind that door. And sometimes that's actually what his room is like. However, the essence of the problem is knowing the floor plan of the house is kind of like the documentation. It's your roadmap. It's your blueprint. And then, you know, having junk in the rooms is still a problem. That's like what's in the database. If that makes sense.
[01:14:56:19 - 01:15:04:19]
Mallory
It does. So this is like the essential first step. If you want to get any value out of your data with AI, you've got to make sure you have documentation in place.
[01:15:04:19 - 01:15:27:09]
Amith
Totally. And, you know, you mentioned in the overview that, you know, the equivalent economic cost, you know, was basically almost free to run Dbautodoc, which is true because you can run it with you don't even need Frontier for this. You don't need Opus 4.7 and Gemini 3.1 Pro. We ran this with Gemini Flash, which is a good model, but it's quick, as the name would imply. And it's also fairly inexpensive.
[01:15:28:16 - 01:20:48:12]
Amith
And it was it was extraordinary in its results. In fact, we did run it with some Frontier models, Opus 4.6 and 4.5 at the time we were developing this and earlier versions of GPT like 5.3 and 5.4. And they did marginally better, but not meaningfully better. So I guess the problem is such that the absence of certain information that would take you from the 95 percentile, which is roughly where Dbautodoc takes you to 100, still seems to be a bit beyond the domain of what current AI models can do. So there's not much more that you get out of throwing more dollars at it right now. But even humans who produce database documentation tend not to be above the 95 percent accuracy, particularly if they're starting with a database that has zero documentation in it. So the the key point here is both what I just said, but also that the absence of documentation is generally a problem people don't try to solve. They know it's a problem, but they also look at it and say, well, can we get by without it? The cost is so high to try to document their database. And it seems like a bit of an esoteric task. It's not like, hey, I'm going to build a new website for membership, use processing or something that's going to make life better for my staff or my members. It seems like a really kind of boring task, frankly. And it's probably not the most exciting task to actually execute or perform. But people don't do it because the economic cost is quite high and it's very, very difficult to do. I actually think the estimates we threw out there were very, very low. I think that on average, you probably spend a much, much higher amount on documenting your database. And this is not a one time task. Again, it's a reminder that your data is constantly changing. So it's not just understanding that your AMS has this certain structure, but it's also understanding how your AMS relates to your LMS, how it relates to your CMS and your custom systems and on and on and on. So it's a quite it's quite a complex problem. We wanted to give the community something that was open source and free and that would be able to just solve the database documentation problem. So this is this is just version one of DB AutoDoc. And our intention long term is to continue to improve it every time new classes of models come out. For example, something like a Claude Methos when that caliber of model becomes generally available, we're going to obviously run DB AutoDoc again with both current and more complex tests. Then we're going to look to the architecture of the AutoDoc tool itself and see if there's things we can learn to improve the tool and the way it works inside DB AutoDoc. Ultimately, we'd love to get to 100 percent, but the utility is already enormous. To give you another example, we've taken some of these databases that were described and we have our skip AI data analysts I've talked about and we ran skip without AutoDoc and we asked a whole bunch of different questions of skip and said skip here's 25 questions. Go answer all these. It's like a test suite, if you will. And in this test suite, skip did reasonably well, just kind of making smart guesses. It's kind of like you hire a really smart employee, don't give them any documentation and say, hey, Mallory, you're a world class database administrator and coder. Mallory, can you please write these 25 reports for me? And you're a smart person, you have great background, but you know nothing about my database. You'll get some of them right and you'll probably figure out some of it as you go just by trial and error, like literally bumping into the walls that you're discovering. However, you would what you would think that if I gave you great documentation, you'd perform better. And in fact, that's of course exactly what skip did. On average, skip was getting somewhere in the order of 12 to 15 out of 25 reports correct on the first pass with zero documentation. But with good documentation, skip was in like the 23, 24 out of 25 because that documentation allowed that AI to be much, much smarter. Plus it was also faster at producing those reports because the documentation, of course, was like that blueprint. So skip knew exactly where in the database to pull data, how to query it, how to aggregate it, and of course how to present it because the documentation was really robust. So there's immediate utility from this. Other examples that I would throw out there for associations is people tend to be really, really afraid of migrating from custom systems. If we had a live audience, I would say, hey, show me your hands if you have any type of custom software in your organization that is plaguing you today. It might be your AMS, but most likely it's probably for the majority, it's not the AMS, but it's some system they built at some point. It might be the website, right? There might be some custom app on the website and the data is like stored in some weird CMS repository rather than even a good database. And they just have no idea what's in this database and the system is kind of fragile and they're frankly afraid of it. And the first step in not being afraid of it and being able to move past these things is to understand them better. And this is a big unlock for that. So we're really pumped about this. It's a topic that I would say at surface level if I went and talked to the average association CEO and described, they would really be bored. They would ask me why I'm even bothering them with this. I haven't tried that yet, but I might. It's be kind of fun. But I think if you understand the value prop and how it can unlock the business outcomes I'm talking about, that gets really, really exciting really quickly.
[01:20:49:14 - 01:21:10:16]
Mallory
I disagree with you. I mean, I think if you led a session on this, people would be quite interested. Or maybe if it was self-selecting, maybe a breakout session or something. But if we do have an executive director listening to the pod right now, maybe on their morning walk or on the Stairmaster, nodding along to what we're saying, what would be the first step for them? So it sounds like this is out there in the world. It's open source. What can they do?
[01:21:10:16 - 01:21:40:17]
Amith
It is. It's a free tool. It doesn't require Member Junction. It's part of kind of the broader Member Junction project in that, you know, Member Junction is this entirely open source AI data platform. It's MIT licensed, which means it's free to use for any purpose, commercial, non-commercial, internal, whatever purposes you have. It's our gift to the world, so to speak. We're trying to contribute something valuable to our community. And this is a tool that's part of that project, but it's a standalone tool. So to use DB AutoDoc,
[01:21:41:20 - 01:21:52:10]
Amith
you basically go to the GitHub repository, you grab the tool, you install it and you run it. And you run it against your database. It works against Microsoft SQL Server, against Postgres, against MySQL.
[01:21:53:11 - 01:22:27:19]
Amith
Several other databases, I think, are supported as well. And so it's fairly straightforward. I wouldn't say it's something that a person who has a challenge using desktop tools in their computer would probably find particularly intuitive. It's still a bit of a power tool, but it is something that the average association IT person could definitely run. It doesn't require any programming skills at all. It's just a little bit of setup and off you go. And there's really good documentation on how to use it. Of course, we have teams of people that do stuff like this and are always happy to help out. But the reason it's free and open source is we wanted everyone to benefit from this thing.
[01:22:29:01 - 01:22:33:05]
Mallory
Well, listeners, if you test this out, please let us know. We'd love to talk about it on the pot.
[01:22:34:10 - 01:25:53:22]
Mallory
Moving to our next topic for today, we want to talk about AI video right now. So last week on April 26th, OpenAI shut down Sora. The web app and the mobile app for the video model went dark and the API will follow in September. The one billion dollar deal with Disney that was signed only four months ago to license characters into Sora collapsed with it. On the surface, it looks like a story maybe about AI video cooling off, but that's not necessarily the case. It is a story about how fast the field is moving fast enough that one of the companies that define the category got passed and discontinued in 18 months. The reasons reported in the Wall Street Journal and elsewhere converge on three things. So one, Sora was reportedly burning around 15 million dollars a day in computing costs. Downloads had fallen about 66 percent from the November peak and competitors had matched or exceeded Sora's quality on independent benchmarks. With OpenAI heading toward an IPO, the math just wasn't working. So originally I wanted to introduce this topic on the pod because I started to think, "Is AI video cooling off? What is the deal with AI video?" But then I did a perplexity kind of research report on it and I realized it's not cooling off at all. It's just the case with OpenAI for now. So the rest of the field is actually in its most active period it has ever had. And four models are kind of running the race right now. One is Google VO 3.1. It's the consensus all around leader. You get true 4K video and it generates the audio, so dialog sound effects and ambient sound at the same time as the video in one step. We've also got Cling 3.0. This is from QuiShow, launched in February and was the first to generate true 4K without upscaling. It can also produce a series of connected shots in a single prompt while keeping the same characters and setting consistent across cuts. We've got Seed Dance 2.0. That's from ByteDance, also released in February, and it can take up to 12 different reference inputs in a single prompt. So that's images, video and audio, and that ultimately guides your output. And then we've also got Runway Gen 4.5. That remains a favorite for professional filmmakers because it gives precise control over camera moves and motion. Worth noting, two of those four, Cling and Seed Dance, are Chinese AI models. And they're currently competitive with or beating the US frontier on independent benchmarks. Open source is also a serious option, too. Alibaba's WAN and Tencent's Hunyuan lead the open source rankings with Lightrix's LTX2, the strongest non-Chinese option open source. Three things changed in early 2026 with AI video. First, which I've already mentioned, true 4K became standard. Second, audio is now generated with video instead of added afterward, which means a clip comes out finished. And then third, multi-shot videos work now. So the same character walking through different camera angles and a single generation. And none of that existed at production quality just six months ago. So, Ami, Sora's shutdown is definitely the headline. When you read about that, did you see it more as an AI video story or an open AI story?
[01:25:53:22 - 01:27:17:07]
Amith
You know, I think it's more of the latter, honestly. Open AI had their hands in every single thing, every category of AI and even in hardware and robotics or things. Open AI has active projects in. So that leads to a lack of focus. And even an organization with considerable resources and talent and dollars and all that has to find their way through a focused lens. And so the market's maturing slowly, but it's starting to mature. And you see some real category leadership emerging. So, Anthropic with Claude has clearly, clearly focused on B2B. They have consumer as a segment, of course, but they're much more focused on business, both with Claude code, of course, but also co-work and now Claude design. They're heavily focused on the enterprise. And frankly, they're eating open AI's lunch in those categories because they're hyper focused. Anthropic, as a comparison, has no image generation model. They have no video generation model. They have many. They don't have any many of the things that open AI has. But what they have is really, really good stuff in the categories they're in. And so open AI did essentially say, hey, we're going to narrow the lens. We're going to focus. And Sora being shut down is a very natural thing to do, both because of the cost. But if you look at the potential for revenue there when the competition is so strong and they also have a structural disadvantage, I'll explain in a second, compared to Google in particular, as well as the Chinese companies you mentioned.
[01:27:18:22 - 01:27:24:10]
Amith
But open AI made a smart decision, in my opinion. It's good to cut your losses in a category that you're not going to win.
[01:27:25:14 - 01:27:43:04]
Amith
So I'm actually really happy that they did that. I think it's good for open AI, which we want open AI as a community. We all want open AI to not go away. We'd like for them to be successful and be a going concern. And the absence of focus would likely lead them to bankruptcy eventually. So I'm happy that they decided to do this.
[01:27:43:04 - 01:27:47:10]
Mallory
And you mentioned the structural disadvantage with open AI. What did you mean by that?
[01:27:47:10 - 01:32:05:07]
Amith
Well, if you think about the way AI models work, generally speaking, the organization that has the best training data is most likely to be successful. You have to have a lot of compute, a lot of money, really smart, talented researchers to build the best models. But if you don't have great training data, you're at a pretty big disadvantage. So then the question you should be asking yourself is, in each modality, who has the best training data? Well, so for video, our friends at Google have something called YouTube. And YouTube has a bit of video. And so being able to train Vayo 3.1 and future models and nano banana on images out of the videos, it's a pretty tremendous advantage. Google also has Google Photos. So that's an interesting thing to remember, is that Google really has an amazing structural advantage because YouTube is also the gift that keeps on giving. People don't realize it's the number two search engine in the world. Right. So they have this sense of not only what the world's videos look like, but they also have some very interesting data in terms of like think of it as like telemetry and what people are interested in with the continual amount of search between the number two and number one and number two search engines in the world. When you think about people who produce social media apps, there's a couple of companies in the world of China that have incredible apps for social. There's a lot of video there. The other thing about China is China has a different perspective on things like copyright, but also specifically privacy and video imagery and surveillance and all this are far more apparent and visible and video is commonly shared. There's a lot of state support in China for startups. And so being able to access video at scale is likely to be a structural advantage for the Chinese companies that you mentioned as well for different reasons. So in comparison, if you think about open AI as powerful and as advanced they are, they don't have that training data so they can have even if they had better researchers, which I don't think they do. I think these other companies have tremendous research talent and a lot of dollars, too. But even if they had an advantage in that area, they don't have the structural advantage of Google and they're continuously accelerating flywheel of YouTube. They don't have a social media platform that keeps feeding live images to them. So image data and video data is really valuable. I think that another player that may come into the fold in this area long term would be Tesla. Tesla or maybe it would be apparent through Grok of XAI, which is now part of SpaceX, a constantly emerging malleable world of Elon Musk's brain. But basically what Tesla has is also incredible real world 3D video data. Their video data is a little bit different. So a lot of people don't realize this, but many, many years ago, going back to I think the first Model S, which was the first sedan that Tesla sold, but maybe not the very first one, but early on in the Model S's development and every car sold since, they've had video. And the video has been and the video is not internal to the car, by the way, it's the video. I think they do have some cameras like that now to make sure the driver is actually paying attention when they're in full self driving. But the video I'm talking about isn't of the cabin, it's of the surroundings of the car. And they've known for a long time that their play on full self driving was going to be purely based on video processing. They didn't use radar or light or anything like that. And so they put cameras in every car and they capture that video by driving, buying a Tesla and driving it. You are consenting to all of that real time footage from your vehicle driving through the world as something that Tesla owns. So Tesla has an enormous proprietary database of video coupled with the telemetry from the car, you know, the acceleration, the speed, the braking, cornering, all this stuff, which is, of course, what allows them to train models that can do full self driving and future iterations of that. But also it's super interesting for world models, which we're not really talking about here right now, but it ties into video because world models help us understand how to generate video as well. So there's some interesting stuff there. And the last thing I'll quickly say is guess what? Google also has a play there because through Alphabet, the parent company of Google, they own Waymo and Waymo is starting to scale tremendously. So the same type of video coming into Tesla's universe is also coming right back to our good friends back at Google.
[01:32:06:23 - 01:32:30:16]
Mallory
So maybe OpenAI read that landscape and said, you know, maybe we should focus a little more. I mean, I feel like we talk about audio, AI audio so much on the podcast and why there's immediate value there for associations who want to go that route. Are you still bullish on the impact that video could have for associations in general? Yes. Very impressive. Lots of meaningful use cases in the world. What do you see for video and associations?
[01:32:31:19 - 01:35:28:18]
Amith
I think it's a medium for communication that is tremendous. I think if you can communicate at scale through video that's deeply personalized, that's tailored for each individual member, that's tailored not just for their preferences, but for their language and their dialect and their preference for brevity versus depth and all of these things, you can really deliver a different experience than without videos. You have to kind of imagine how you'd use it. So if all you're doing is saying, well, we produce 50 videos a year and they cost us ten thousand dollars each. So instead of doing that, we're going to spend way less money. OK, fine. That's great. And that's it's not uninteresting. It's it's great to find efficiencies. What's more interesting is if we said, OK, instead of 50 videos a year, what if we produce 50 videos an hour or 50 videos a minute? Because we have abundance, we have the ability to produce videos at scale. And you might ask, well, why would we produce that many videos? What if we wanted to do a video newsletter that was literally a video that explained to each reader and now viewer what they needed to know about the association every morning? Right. Audio format of that. It's kind of like what you have in notebook LM, where you can generate a podcast from any topic you're interested in. You can personalize it. What if you did that at scale of video? What about a live real time video interaction and assistant on your website or through apps that could guide people through anything that they wanted to get done in your association? Truly an amazing video driven virtual concierge that knew your association better than anyone. That would be another application of a real time version of this. So you have to get creative to really see the full potential. And I certainly don't claim to know all of the possible applications. I'm like everyone else in this area. I'm just starting to explore this. But I am very excited by it because we communicate much more deeply through the modality of video than we can through audio or text. So I think it's I think there's tons of upside. It's kind of like this, though. It's like you find a platform and you're looking for the killer app. You know, and Sora was definitely not the killer app. Sora was just like, oh, this is a cool novelty. There was no business value in Sora. So the question is, where does the business value come from? You know, at Sidecar, we've experimented a ton with a generated video through our learning. And for those of you that have been through the Sidecar learning hub, you know that we have an avatar. That's your instructor in each of the courses. That's all generated through platforms like this. And we're constantly looking for ways to raise that bar. I do think that's an area that associations could immediately play with, which is, you know, thinking about their course development. How do we build way better courses, be way more responsive to the market to build new courses? We went from, I think, 12 courses that we had in the Sidecar learning hub at the beginning of the year to over 70 courses in the Sidecar learning hub now. And, you know, this is a big part of that secret sauce. How do we do that? We didn't grow this. We didn't scale the team by six acts. We grew the team a little bit, but we had a lot of video tools.
[01:35:28:18 - 01:35:59:06]
Mallory
This makes me think of the phrase just because you can doesn't mean you should, Amith, and not saying that it will be an either or with audio and video, but with the example that you gave. Having an assistant on an association website that doesn't necessarily need to be an A.I. video. That could be A.I. audio. I know for us with Grace, our A.I. audio assistant on the Sidecar website right now, she's audio only. But I'm wondering, did you consider adding in video and do you see it as an either or or is video not there yet?
[01:36:00:13 - 01:36:42:10]
Amith
So we did consider, but only briefly for the for the immediate launch of Grace, which is back in Q1 of 2026. We thought the audio was ready for prime time, but not quite on the video front, both because of compute cost and latency. Really mainly because of quality. We wanted to put something out there that delivered value to the users and would work consistently in video. Real time video interaction is not quite there at this point in time. I think it will be there probably within 12 months, but I would focus on audio only for real time applications. So for real time, live, synchronous interaction, audio is really, really good already. And there's lots of ways to do it economically and at scale. Video is very expensive still for real time and not quite there quality wise.
[01:36:43:16 - 01:37:27:13]
Amith
That will change quickly like everything else in the world of A.I. I think for offline video generation, more of like, you know, you create a video ahead of time and then send it to somebody and you're less worried about latency. The quality is extraordinary. You can do things with the tools you mentioned earlier, Mallory, that are amazing. So imagine a scenario where you automatically generate an amazing 4K video for every blog post that you publish. I could see that as being a really complimentary way of publishing a lot of the content, a lot of the messaging. Or imagine, you know, messaging coming from leaders in your organization that you could put through video like that, as well as obviously learning content. There's a lot of opportunities around that. So but I think for real time, I would stay away from real time video for the moment.
[01:37:28:14 - 01:38:54:19]
Amith
Just because of the quality and speed thing that I mentioned. The last thing I'll say is I think when we come to video, which we need to start thinking about now, when it comes to real time two way video, there's some interesting ethics questions to ask ourselves first. Should we disclose that the A.I. you're speaking to as an A.I. We believe here at Sidecar that you absolutely should disclose that. I know that there will be many people who don't do that and try to pass off an A.I. avatar as a real person. And it will probably work very, very soon, if not already. But we think that that's it's really critical to disclose that to your communities. I think most associations would be nodding their heads to that one. But then the flip side is, do you require or do you make it optional for the other party that a human, presumably a human, to share or not share their video? So great example is earlier this morning, I was interviewing a candidate for Blue Cypress for our labs organization. We hire people. We hire a lot of different people for Blue Cypress Labs. That's our research arm. And in our labs organization, we're hiring a whole bunch of people for May, June grads called Tech Fellows, which is our program for hiring and onboarding people right out of school. And so I was interviewing a candidate this morning and we were talking about the exactly this topic. So we use A.I. much like Grace, actually the same basic tech stack we use for Grace. We use that for interviewing the first round of interviews with technical screens for these tech fellows. And it's extraordinary. It works extremely well.
[01:38:55:19 - 01:39:32:05]
Amith
We have an A.I. audio agent that is designed to work like a kind and encouraging computer science professor, speaks to the candidates, isn't trying to stump them, is just trying to ask them questions. And when they get stuck to help them along, nudge them along in the right way, it's it's there's a lot of nuance to that. There's a lot of knowledge required, a lot of nuance. And the candidate asked me, well, would you consider doing a video version of that? And it was a great question because the question was related to me showing part of our system saying, hey, this is what we do with this audio. We show that's all the candidates when we actually do a live interview with them. And I was asked, well, if you had video, you'd have even more information coming in of the candidate.
[01:39:33:10 - 01:39:58:23]
Amith
And I said, yeah, that'd be great. We'd have to be very thoughtful about that. We want to make sure that that was an art. In my mind, my initial reaction is make that optional because some people aren't comfortable with that. And obviously, you'd have to be very thoughtful about how you safeguard the data. We already think that way about just audio, but with video, it's even even more important. So there's some questions like that around governance and safety, privacy and obviously, AI ethics that are really important to answer when you get into video.
[01:39:58:23 - 01:40:13:19]
Mallory
For sure. And that has me thinking, too, about the Eric O'Neill episode when he talked about companies never having video calls with their employees and ending up hiring basically bad actors from other countries or AIs altogether, which is kind of crazy.
[01:40:15:00 - 01:40:29:17]
Mallory
Amitha, I want to go back to where we started this discussion around focus, because I feel like this is a good kind of example episode. We talked about audio, we talked about video, we talked about data cleanup, internal applications, productivity and also member facing applications.
[01:40:30:19 - 01:40:41:00]
Mallory
How do you recommend the executive director listening to this right now who feels like, oh, it's hard to focus when there's so much opportunity out there? What do you say to that person?
[01:40:41:00 - 01:41:00:13]
Amith
I would say that that is the key ingredient that's going to separate the organizations that succeed from those that will not succeed. So my way of looking at that is there's a lack of focus partly because of what you said, which is kind of the positive view on it, which is we have so many opportunities, we want to chase them all.
[01:41:01:17 - 01:42:49:11]
Amith
Versus the other problem, which is a lack of focus due to a lot of debt. And I don't mean financial debt necessarily, usually it's more of organization and cultural debt where you can't kill off any programs or any projects because there's political ramifications. There's a board member who cherry picked that project and forced it through, or there's a consensus driven culture where you have to get everyone to agree that projects go away. And so you have associations that, you know, they basically end up being these really weird organisms that have appendages representing this historical collection of past projects that they've never been able to kill off. Sounds really weird, but that's actually how a lot of these organizations operate. So having some clarity of mind around the fact that you need to part ways with a lot of the portfolio of projects you have is really key. That was true pre-AI. AI just is a magnifying glass. All the weaknesses you have in your organization are going to get amplified, and this is a common weakness. So nobody's perfect at this, right? Like I'm certainly not. Our organization has similar problems. But when it comes to looking at the portfolio, picking a small number of critical priorities, demanding that the organization perform and align the team members with that set of limited activities and measuring those results and then communicating those results. These are some basic practices. We're big believers in using a system that largely comes out of Andy Grove's teaching that were documented by John Doar in the Measure What Matters book. This is the OKRs book, as a lot of people refer to it. And it basically preaches this stuff. And there's a lot of other systems like that, like EOS and scaling up. But ultimately, it's about picking a system that works for you, but really narrowing the lengths. The fewer priorities that you can actually act on, the more likely you are to get them done. It's really that simple.
[01:42:49:11 - 01:42:50:08]
Mallory
Mm.
[01:42:51:18 - 01:43:01:16]
Mallory
Everybody, as a review on today's episode, we know AI is doing the iterative, structured work that used to require human experts. DB Autodoc reasons across a database the way a
[01:43:01:16 - 01:43:07:02]
(Music Playing)
[01:43:17:20 - 01:43:34:19]
Amith
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.
[01:43:34:19 - 01:43:38:00]
(Music Playing)