Summary:
Intelligence is turning into a resource you can scale like electricity, and that shift changes what's actually worth your attention in the AI news cycle. In part two of their refreshed Fundamentals of AI series, Amith Nagarajan and Mallory Mejias pick up exactly where part one left off, moving from definitions and building blocks into what abundant intelligence means in practice for associations. They trace AI's growing role in drug discovery, diagnostics, and materials science, explain why a model's black box matters less now that reasoning is visible, and reframe hallucinations as a byproduct of the same creativity that makes these tools useful. They also break down how to weigh AI's energy and water costs against everyday technology, and why picking an AI vendor should never be treated like a slow, locked-in AMS purchase. Amith's throughline: build for a world where intelligence keeps getting cheaper, and protect your flexibility by keeping your agents decoupled from any single provider.
Timestamps:
00:00 - Welcome Back for Part Two
03:19 - The Case for Abundant Intelligence
09:43 - Why AI's Black Box Matters Less Than It Used To
11:35 - Hallucinations as a Feature, Not Just a Bug
18:09 - Cutting AI's Energy and Water Footprint
23:43 - Don't Pick an AI Vendor Like an AMS
27:34 - Recap: The Falling Cost of Intelligence
👥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:
Claude ➔ https://claude.ai
ChatGPT ➔ https://chatgpt.com
Gemini ➔ https://gemini.google.com
OpenAI ➔ https://openai.com
Anthropic ➔ https://anthropic.com
Microsoft Copilot ➔ https://www.microsoft.com/en-us/microsoft-365-copilot
GPT-4.1 mini ➔ https://openai.com/index/gpt-4-1/
GPT-3.5 Turbo ➔ https://platform.openai.com/docs/models/gpt-3.5-turbo
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 🤖
[00:00:00:14 - 00:00:09:17]
Mallory
Welcome to the Sidecar Sync Podcast, your home for all things innovation, artificial intelligence and associations.
[00:00:09:17 - 00:02:56:20]
Mallory
one of your hosts along with Amith Nagarajan. And today we are excited to bring you part two of the Fundamentals of AI series. You might remember that we put out the initial version of the series back in early 2024 and to this date, those episodes are some of our most popular episodes ever of the Sidecar Sync. So a few years later in 2026, we thought it was time to redo the series. Now this is part two. So if you have not listened to part one, I highly recommend that you do this first because today is going to be a continuation of that discussion that we had in part one. It is episode 148 if you want to check that out first. I also want to let you know that if you did listen to part one and you got all the way to the end and thought, Oh, maybe it was a little bit of an abrupt ending, very good on you for noticing because initially the episode that you're going to hear today was part of the one that we recorded in part one. So we unintentionally recorded a 90 minute gigantic episode of the Sidecar Sync and realized, huh, it's probably best that we split that up for you all so you didn't have this giant Fundamentals of AI episode to consume. So we actually broke it into two parts. So where we pick up today is going to be exactly where we ended in part one, just as a note for you. Now in part one, we kind of covered more of the vocabulary of AI, what it actually is a little bit of what's happening under the hood. We talked about open versus closed models and we stuck mainly with definitions, frameworks, and building blocks. What we're going to cover in this part two of the series, we're going to zoom out a little bit and talk about the cost curve. So essentially intelligence getting cheaper, faster and smaller at the same time. We're going to talk about this branch of models that we have called reasoning models, whether AI can actually reason or not. And honestly, whether that question even matters for you, we're going to talk about how to make sense of model announcements, what to look at when a new model drops and why you don't necessarily need the biggest, smartest model for every single task. We're going to zoom out and talk about what abundant intelligence looks like outside of our world, and then we're going to talk about some of the things that we're looking at, things briefly like drug discovery or diagnostics, material science, realizing that AI is not only making an impact in business, of course, but in the world at large. And then we are also going to talk about some limitations, the black box of AI hallucinations and energy and water concerns, because it didn't feel right to have a fundamentals of AI series that was only looking at the optimistic side, right?
[00:02:59:07 - 00:03:19:22]
Mallory
So we're going to talk about some of the things that we're going to talk about in a little bit about the full picture of where we are right now in the AI landscape. And then finally, we wrap up with a discussion on why picking an AI vendor should not feel like picking an AMS. It's my little teaser for the moment. We have a really great conversation lined up for you. So without further ado, here is part two of our fundamentals of AI series.
[00:03:19:22 - 00:04:22:00]
Mallory
talk about, Ameth, the good and the bad of AI. So I feel like a lot of what we talk about on this show is AI inside your walls. So how you run the place, how you serve your members, what you can stop doing manually, but I want to zoom way out for a minute because the AI story spans all fields, all industries. Take these examples. In medicine, AI is picking which disease targets to go after and designing the molecules to hit them, work that used to take years of trial and error in a lab, and a handful of those AI design drugs are now in human trials. On the diagnostic side, models are reading scans and pathology slides and catching things earlier than a standard workflow does. In material science, AI is proposing entirely new compounds for batteries, solar and clean energy, screening millions of possibilities that no research team could test alone by hand. So Ameth, what do you want listeners to know about types of discoveries and developments that AI will drive in our lifetimes that are maybe outside the world of business and associations?
[00:04:22:00 - 00:05:23:11]
Amith
When you zoom out and you think about the possibilities to have abundant intelligence on tap, it's kind of mind boggling because if you think about, you know, scarcity versus abundance or limited resources, and you say, well, how have we as, you know, a civilization progressed over the millennia to improve our body life and to advance how our civilization lives, advance the human condition fundamentally, how do we do that? Well, it requires intelligence. That is the key thing that has allowed us, you know, to do all of these things. And intelligence has been a fairly scarce resource if you think about it. It's why societies that invested in education have done well, right? To be able to, it's a long-term investment, but it generally has resulted in more GDP growth when you have really high quality, you know, kind of open education systems. Now, when we think about intelligence as a resource,
[00:05:24:12 - 00:05:31:02]
Amith
it's kind of hard because it's like, it's our brains, right? It's like the only form of intelligence we've really known is human intelligence.
[00:05:32:04 - 00:05:59:10]
Amith
And so now we're saying all of a sudden there's this really weird concept called intelligence that is decoupled from our brains, and it's something we can scale. To scale intelligence pre-AI, it means you have to have more people. And, you know, having a kid takes a while and then takes a while to raise them. It takes a while to put them through school. And then sometimes you get a productive human out of that process. Sometimes you do. I'm hopeful that mine will be checking that box.
[00:06:00:14 - 00:08:24:10]
Amith
And, you know, sometimes you get a productive human who contributes to that intelligence pool. Sometimes you don't. But and there's, of course, there's people who have a dramatically outsized impact. You know, you have the luminaries in various fields who contributed dramatically more than the average. But it's not easy to predict. It's hard to scale, hard to grow. And there's all these factors in it. But what if you could scale intelligence much faster? Right. What if you had abundant intelligence? What if you had, you know, Dario Amade, quote, he's the head of entropic, is a data center full of geniuses. Right. So this idea that you have a billion Albert Einstein level brains essentially in one data center and you have many such data centers all over the world and maybe in space as well. What happens? Right. What do you solve for? You solve for all these fundamental problems in science, which leads to advances in medicine, as you mentioned. There's a number of A.I. powered labs that are saying our our goal is to cure all disease, not cancer, not Parkinson's, not Alzheimer's, but all disease. I mean, that's tremendous. Right. Or what about free, abundant, totally clean energy for everyone at whatever level we need. Right. So these are the kinds of I wouldn't even call moonshot goals. Right. They're like, you know, extra, extra outside of the solar system. These are like just unbelievable things you can't even conceptualize and you need intelligence to do these things. You don't have enough scientists in the world to do these things. So that's really the idea here that is so incredibly compelling is to solve these incredibly complex, difficult problems and do things nobody thinks are possible. Right. In any of those domains. I think that's what's fascinating. And it does affect all of us in a couple of ways. First, I think when we accept the possibility that something like what I'm describing could happen and in my opinion likely will happen in our lifetimes, that's exciting. That's something to be optimistic about. And there are many downsides to A.I. There's concerns over safety. There's concerns over environmental impact. There's concerns over equitable use or access. There's concerns about ethics in the use of these A.I. systems that are all very important things for us to address as a team, as a society. When we think about that possibility, that should inspire hopefully everyone to work together to try to deploy this safely and at scale.
[00:08:25:17 - 00:08:46:08]
Amith
Then the next thing I'd say is for our associations, you represent all of these that are directly affected by many of the things that we're hypothesizing now but are very tangible in our lifetimes in my mind. And so what does this mean? What does this mean for doctors and lawyers and nurses and architects and engineers? If there is abundant
[00:08:47:11 - 00:09:04:22]
Amith
intelligence like this that's solving problems, I'm optimistic that it means that your fields are going to thrive. Your fields are going to accelerate in many ways, but you have to adapt. If it is possible to cure all disease, does it really mean that doctors will be more needed than ever or does it mean doctors are not needed at all?
[00:09:06:03 - 00:09:43:10]
Amith
Does it mean if we have abundant energy that you don't need people going into engineering fields to try to solve for and build the next generation of power plants or whatever it is? These are all good questions. What I'm optimistic about is that we will figure it out as long as we have a little bit of time. My biggest concern about it ultimately is the speed at which it's moving, which none of us can control. But ultimately, I think for our association listeners, what I want them to know is that the generality of this intelligence curve is what's going to be really the most exciting, I think, for all of us, independent of what we do within our own businesses.
[00:09:43:10 - 00:10:20:01]
Mallory
Yeah, I like that optimistic take of me. We'll figure it out. I do want to talk about a few, perhaps, of the negatives and you kind of touched on them when it comes to this conversation around AI and our fundamentals episode. When we talked about AI being a black box, so essentially because it's creating rules and connections across synapses for itself, we can't really open the AI model and understand why it's making certain decisions. How concerned are you about that when you're, like, as a business leader, thinking we don't even understand fully how this thing works, yet we're going to use it to provide copious amounts of value to our members?
[00:10:21:01 - 00:11:35:14]
Amith
So the good news about that is a lot of progress has been made to better understand the way models work. They are still largely a mystery in terms of how they're able to do some of the things they do. But there is a lot of research going into this because interpretability is or understandability of the models is a really important area. Right. And I do think we're going to see tremendous advancements there. My thought is that, you know, as long as we're able to validate that the answer is correct, how the answer was arrived at is still important, but it's less important. Previously, what we were worried about with hallucinations and with earlier models is that because they were black boxes and because we didn't really have any way of knowing how they produced the answer, we couldn't really have any sense of it. It's just like pumping out an answer and that's it. Now you can actually see the thinking that went into it. You can see which tools were used by the model or by the system to calculate what it gave you. So you can actually kind of trace back through the work. And so that helps you understand, oh, the model made a mistake or not. And other models can look at it. So there's a lot of advancements that have occurred over the last couple of years that really make me optimistic about the usability of these things.
[00:11:35:14 - 00:11:54:12]
Mallory
You mentioned hallucinations, which funny enough, I feel like any time I lead a session on AI in any capacity, there's always one person that says, what about the hallucinations? I don't want to use this. Hallucinations are so bad. I feel like they've gotten a lot better in the past few years. Do you, are you still concerned about hallucinations with the work you're doing with AI ME?
[00:11:54:12 - 00:13:13:00]
Amith
Yes. However, to your point, it's gotten way better. And that's even true with very small models for a couple of reasons. One is the model architectures are better. Two is that the models can reason so they can check their own work. So models are now trained to both check their work through their reasoning process, which we've talked about a little bit. And it's essentially this way of the model being able to look back as well as look forward, be able to say, oh, here's what I just said. Is that correct? And the idea of models having access to tools where the model has the ability to actually look something up. You'll often find with current AI tools that when you ask a question, it'll say, let me check for you, or it'll say, let me, let me look that up because I don't want to rely just on memory alone. And so you'll actually, you'll see Opus say this a lot. And so what it's doing is it's using tools to go get verifiable and correct information from trusted sources. That's really, really important. And so the rate of hallucinations that I've encountered is dramatically less. It used to be the model didn't have any tools and it just had to basically blurt out the very first answer it thought was right, as we've discussed. And so, yes, in that scenario, I also would make a lot of mistakes. The flip side of that point, I'd say, is that hallucinations with the reduction in
[00:13:14:03 - 00:13:24:22]
Amith
what we're talking about in terms of these negative problems, hallucinations actually can be a feature, not a bug, if you control them. And so what I mean by this is our creativity
[00:13:25:23 - 00:15:46:00]
Amith
is limited by what we think is the right answer. And many of the domains we work in are not, you know, two plus two equals and give you the correct answer. They're all sorts of judgment calls, all sorts of ideas. How do I improve engagement amongst West Coast, you know, middle-aged members? Well, there's lots of interesting questions we might ask if that's the problem we're trying to solve. Lots of hypotheses we might want to test. And so a model that has a little bit broader range in terms of its creativity could potentially come up with all sorts of ideas that you wouldn't have thought of. I find this to be the case all the time. One little tip for our listeners is that there's a little known feature in most of these tools where you can control something called temperature. And the temperature allows you to give the model a wider range of how much creativity it can use. So if you have a temperature of zero, it means that it's always going to essentially pick exactly what it thinks is the highest probability best answer. So there's basically no variability. So if I were to say to you, Mallory, here's some text, what's the best next word or that kind of thing, you might have several words that come to mind. Oh, it could be this, could be this, could be this. This one's probably the most likely, but there's also this possibility, this possibility, this possibility. And that's what models do as well. It's called distribution. And basically what it's doing typically is picking if you have a temperature of zero, it's saying I'm going to pick the very highest probability every time. If I have a temperature above zero, then it has some latitude to pick from amongst the best possible answers. And the variability actually is part of what gives these models the ability to do a lot more than just spit out the same thing over and over again. That is actually part of what makes these AIs so useful. We just don't even realize it sometimes. So there's a lot of things to be thinking about there in terms of where you use them. If all you want is a calculating machine that does two plus two, like a fancier version of that, then no, you probably actually shouldn't be using AI at all. There's all sorts of other tools you can use. But hallucinations, I'm not trying to state that they're unimportant. They are very important, especially in agent design. You have to really make sure that you have grounded truth everywhere and there are no hallucinations. And there's ways to do that now. But what people refer to as exclusively a negative actually has a very important positive side to it as well.
[00:15:46:00 - 00:15:54:24]
Mallory
I love the idea of equating creativity with hallucination. I just think that it kind of is because sometimes to be creative, you have to make things up a little bit. So yeah,
[00:15:54:24 - 00:16:20:23]
Amith
like last time you had this stroke of inspiration, like, you know, going for a walk or, you know, swimming in a pool or driving a car or whatever, right? Like dodging a streetcar here in New Orleans. Like, you know, these these these moments of inspiration come to us in ways that we don't really even understand. But all of a sudden, like, where'd that idea come from? Right. And so that is this this ability to allow ourselves to have, you know, creativity essentially.
[00:16:20:23 - 00:16:40:12]
Mallory
Many people within our association community care deeply about the environment and rightfully so. And the water usage required for cooling data centers is definitely concerning. So how do you recommend navigating that as an organization that wants to be innovative, but also deeply cares about its impact on the environment?
[00:16:40:12 - 00:16:54:07]
Amith
So the first thing is pay attention to the models you're using. I've used this in a algae before, but, you know, Mallory, if you're flying from Atlanta to San Francisco, we don't need to load up, you know, a Boeing 737 with 135 seats just for you.
[00:16:54:07 - 00:16:56:12]
Mallory
I'm going to take the paper plane for that.
[00:16:56:12 - 00:20:28:08]
Amith
Yeah, exactly. You can take a much smaller, you know, aircraft or get there another way that's a lot more efficient. So if you go and load up GPC 5.6 sole on extra high or Opus 5 or whatever, when you're listening to this some number of months or even years from now, if you're just going to the very most the most powerful model, you're wasting a lot. You're wasting money, but you're also wasting resources, which is compute the more compute use, the more energy you consume and the more heat you generate, which requires more water to cool it down. So there are ways to be radically more efficient than the way we're we're acting now. Another simple thing is make sure you update your old systems. So independent of the consumer grade AI tools that we're using, like using chat GPT and cloud and so forth, a lot of people now have built systems that use various kinds of AI models. And I know a lot of people who I talked to, you know, mid 2026 and I say, hey, what are you what are you using for them? Like, oh, we're using the open AI API and we're using GPT 4.1 mini or we're using GPT 3.5 turbo or something like that. And I'm like, wow, that's cool. And that's like really inefficient. And it's also, by the way, really expensive to use those legacy models. They're some of those ones that are really older, getting deprecated completely, but they're really slow, really expensive. And it's just, you know, it's just something you should pay attention to and update. It's not a trivial thing to change an application to use a new model. You have to test it a bunch. But there's ways to do that reliably without a massive lift. So if you have an application you're using for like a chat bot that you developed yourself or with a partner, make sure you're keeping your models up to date because there's a lot of efficiency that comes from it. You know, in our own world, we oftentimes are able to take a step down in model size while maintaining or even improving the quality of responses that various agents that we've developed produce each time a new generation of models come out. So that's step number one is optimize your use. You might deeply care about these things, but not think about it at all. The other thing is, is if you really care deeply about this, think about where you inference your AIs. Some providers of AI are a lot more thoughtful about where they build their data centers, how they treat the communities that are in the areas they build data centers, where they source their power, what they do with water. Are they closed loop systems that recycle water? Are they not? There's a lot of subtlety to this. It's not just one size fits all. Some companies are way, way more forward thinking about this than others. And it's also really good for their bottom lines if they are, because these more sustainable systems are actually way more profitable long term. They're more expensive to build, but they're much more sustainable economically long term as well as being more appropriate in terms of their environmental stewardship. And the last thing I'd say is that look ahead at the cost curve. And the cost curve also means the resource curve, because cost means that you're using less stuff, using less power, therefore using less water. So there's reason to be very optimistic that most use cases are going to be fairly trivial in terms of their power consumption. Now, at the same time, you might ask, well, why is everybody going out and raising trillions of dollars to try to build more data centers with more and more and more compute? It's because the aggregate demand is growing at this absolutely enormous pace. So it's a societal concern, but your use for your association or certainly for you can be really managed so that it's a fairly tiny amount of use. The other thing I'd say is,
[00:20:29:17 - 00:20:58:19]
Amith
two other things actually, is one is compare it to the other technology and other things that you do. If you do a lot of flying, think about the environmental impact of that and compare that in terms of carbon, in terms of energy consumption to your AI use. Compare it to your use of other technology, whether it's Microsoft Office 365 online or Zoom. These are not free resources when it comes to environmental use, but it might be somewhat unfair to compare them because AI is somewhat more power
[00:21:00:12 - 00:21:53:09]
Amith
draining than typical applications. That was true much more so a year or two ago, but it keeps on being less true. Right. So and the last thing I'd say is be transparent about this. If you care deeply about this, and many of our listeners really do, compute what you're using and be open about it and tell your members why you're using it, how you've done your diligence, what you're doing to mitigate your impact and compare yourself year over year and look to improve that, right? Look to improve your utilization of resources year over year. I don't think for most people it means don't use this stuff. Perhaps some people feel so strongly about it that they don't want to use AI. And that's their right. But for most organizations, it's a matter of how can we balance being responsible, being forward looking in terms of environmental stewardship and also advance our cause, advance our mission using the best available tools.
[00:21:54:13 - 00:22:09:00]
Mallory
Well said. To wrap up this segment of our fundamentals series, Amith, can you share one thing when speaking to association leaders that you feel like people get wrong about AI that you want to clear up right here and right now?
[00:22:09:00 - 00:25:42:19]
Amith
The biggest thing I think is people think of AI like they think of an AMS. Our good, good friend, the Association Management System, the AMS, there will be those things. They will probably still exist well after I'm gone. But AMS's are thought of as you take a long time to select them, you think deeply, deeply, deeply about selecting them, and then you usually get it wrong and your implementation is usually terrible. It's very expensive and very painful. And then you end up living with this kind of somewhat working thing for several years, usually 10, 15, sometimes longer years. And so people think of picking an AI vendor like, should I go with OpenAI and chat GPT or should I go with Anthropic? They kind of think of it like that, like, oh, which one should I pick? And if you think about it that way, it's just going to take you too long to figure it out. It's also not that difficult to switch as long as you do a few things really well So if you just jump right in and start adopting every service that someone like Anthropic or OpenAI or Google provides and build your whole ecosystem around their tooling, use their chat tool, build your agents in their agent toolkit, use their APIs directly, do all the other things that they'd love for you to do, then you are very tightly coupled to one particular vendor, whoever it is. This is not about any of those vendors being good or bad. It's just you are closely coupled. And then essentially you are making an AMS like decision. It is going to be hard to move. But to your earlier point, Mallory, and I've made this point as well, I think on this particular episode, optionality, flexibility is so key. So how do you do that? Well, it's through intentionality. You have to start off by saying, well, how do we build systems, both process systems with our humans and then computer systems that have a degree of indirection where we have the ability to swap out systems? And let's be very intentional about that. Let's build our agent systems separate from our model providers. Don't use the model providers native agent tool. I think OpenAI, Enthropic, and Google all have fantastic first-class agent builder tools. You can go on any of their websites and you can build amazing agents very quickly. But I would caution you to consider doing something else instead, because if you do that, guess what? OpenAI is agent builder. You think it's going to work with Gemini? You think it's going to work with open source models, inferencing on someone else's hardware? I don't know, maybe, but I really doubt it. I think OpenAI is agent builder is going to work with OpenAI's models and same thing for all the other providers. So you're vertically integrating. It's convenience you're getting, but you're losing flexibility. So to me, that's the big thing. Treat it not like an AMS, but behave in a different way. Don't go all in with just one vendor just because it's easy. People right now are doing this with Microsoft. So I think very highly of Microsoft, to be clear, I think their AI strategy has its issues, but overall, I think they're on the right track and I'm really excited about the long-term future. But if all you do is say, well, we've got Microsoft 365, let's use Copilot, it's the quote unquote safe choice, and we're just going to stick with that. And then we're also going to do, you know, Copilot agents and all this other stuff. You're literally falling into this exact trap. You cannot go anywhere other than Microsoft at that point. And there's lots of other choices out there and there's choices that are model and inference provider agnostic. There's quite a few of them actually. You just have to look a little bit harder.
[00:25:43:20 - 00:25:45:18]
Mallory
Or listen to the Sidecar Sync podcast.
[00:25:45:22 - 00:26:18:18]
Mallory
everyone, to recap this part of our fundamentals of AI series, the cost of a given level of intelligence keeps getting cut in half more than twice a year. So build for a world where intelligence is close to free. When a new model gets announced, the two things worth your attention are overall intelligence and cost, not necessarily parameters or benchmarks. And the concerns people raise about AI, the black box, hallucinations, energy and water are certainly real. But maybe you can exert a bit of control over those based on which model you pick and how
[00:26:18:18 - 00:26:24:04]
(Music Playing)
[00:26:34:22 - 00:26:51:21]
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.
[00:26:51:21 - 00:26:55:02]
(Music Playing)