Summary:
This week on Sidecar Sync, Amith Nagarajan and Mallory Mejias break down a whirlwind week in AI, from Anthropic’s rapid-fire Claude releases to OpenAI’s tightly controlled GPT 5.6 rollout. They unpack the surprising performance of mid-tier models like Sonnet 5, the implications of government intervention in frontier AI, and what it means when access to the most powerful tools is suddenly restricted. The conversation then shifts to a new study reshaping the AI jobs narrative, revealing that companies investing deeply in AI are actually growing headcount—while others fall behind. From practical model selection strategies to big-picture workforce implications, this episode connects the dots between cutting-edge tech, policy, and the future of work.
Timestamps:
00:00 - Nothing New in AI… Right?
👥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:
The AI jobs Debate Just Got Messier (Article) ➔ https://shorturl.at/wdGPc
Claude ➔ https://www.anthropic.com
Gemini ➔ https://deepmind.google/technologies/gemini/
Gemma ➔ https://ai.google.dev/gemma
Artificial Analysis ➔ https://artificialanalysis.ai
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 - 01:00:21:05]
My name is Amith Nagarajan.
[01:00:27:15 - 01:00:29:10]
Mallory
And my name is Mallory Mejias.
[01:00:29:10 - 01:00:35:04]
Amith
And we are your hosts. And there's nothing new to talk about today, right Mallory? Nothing new at all.
[01:00:35:04 - 01:00:42:09]
Mallory
Nothing new at all. It's always hard to create an episode outline, just no information out there. So I don't know what we're gonna talk about Amith, we'll see.
[01:00:42:09 - 01:00:44:11]
Amith
Yeah, it's just too boring out here in the world of AI.
[01:00:45:20 - 01:00:52:12]
Amith
It is crazy, isn't it? It is crazy. And actually down here in New Orleans, I don't know what it's like in Atlanta, but it is disgusting. It is so hot.
[01:00:52:12 - 01:01:34:14]
Mallory
It's hot. It's really hot here. Actually, I was thinking about this before the episode. I'm like, I wonder if I should share this, but we did an Atlanta bucket list item this past weekend, which was something I believe locals lovingly called shooting the hooch, the hooch, referring to the Chattahoochee River. Shooting the hooch, meaning tubing down the river in the city of Atlanta, which I know sounds concerning. Believe me, I was like, is that safe to do? Just given the fact that it's in a big city, river, things like runoff, but apparently the E. coli levels were checked. This weekend was a good weekend to do it. So that is how we escaped this crazy heat we've been experiencing in Atlanta. And it was a really good time, but I'm sure it's hotter in New Orleans, if I had to guess.
[01:01:34:14 - 01:01:51:14]
Amith
Yeah, I don't know if it's hotter, but it's certainly very humid here. And for those of our listeners that are in New Orleans or the nearby areas, I'd probably recommend not doing what Mallory did in the Mississippi. Maybe there's some other water nearby that's a little bit cleaner than that, but that sounds like it was a lot of fun.
[01:01:51:14 - 01:02:10:09]
Mallory
It was a lot of fun. We've done, in the New Orleans area, we've done the Bogachitta, which you have to drive two hours to, but the water is really hot and brown. I think it's still safe, but at least the water here was very cold. And I don't know that I'll be doing it again anytime soon, but it was a good way to get outside, connect with nature, all that fun.
[01:02:10:09 - 01:02:13:21]
Amith
I'm assuming in Atlanta, you have fewer alligators per square meter than here.
[01:02:13:21 - 01:02:17:06]
Mallory
I would think so. I don't know the numbers, but I'll pull them for the next pod.
[01:02:17:06 - 01:02:56:24]
Amith
Yeah, I don't know if I've told this story on the podcast, but when my daughter was a lot younger, we went paddle boarding on Bayou St. John here in New Orleans. And those of you that know New Orleans, it's beautiful waterway in the middle of the city. And there are gators in there. They tend to be smaller ones, though. That's what the locals say in order to comfort you. And so she had just gotten her paddle board. She was super excited. And she was on her own, because actually my paddle board, I had an inflatable that had a hole in it at the time, so I hadn't patched it yet. Anyway, she was out there on her own, and she was just paddling, paddling, doing a great job. And then I saw what I thought was a gator. She was probably only like 50 feet away from me or something. And she pushed it away, and she thought it was a lot.
[01:02:56:24 - 01:02:58:00]
Mallory
Pushed it away? Okay.
[01:02:58:00 - 01:03:15:06]
Amith
She thought it was a lot. And the gator swam away. She didn't realize it was an alligator. So I didn't say anything. I waited for her to come back and I'm like, "Hey, did you notice that thing "that you kind of hit with your paddle "and then she goes, "Yeah." I'm like, "Do you know what?" She's like, "It was just a log." I'm like, "Uh-uh." I said, "Look back out there." And it was swimming around.
[01:03:15:06 - 01:03:24:13]
Mallory
Oh my goodness. You have not shared that story on the podcast. And I gotta commend you for playing it so cool. You didn't think about yelling or jumping.
[01:03:24:13 - 01:03:33:18]
Amith
I wanted to go and do something about it, but there wasn't much I could do. So I'm like, "All right, well, if it comes after her, "I'll probably start swimming, but probably."
[01:03:33:18 - 01:03:40:24]
Mallory
Oh boy, I think you would. But no, you're right. And how old was she? Not to create too much commotion was probably smart.
[01:03:40:24 - 01:03:53:07]
Amith
I mean, she was tiny. She was probably eight or nine or maybe 10. Oh wow. Yeah, yeah. So good fun here in New Orleans down in the bayou and we think about AI all the time, even when we're paddle boarding with alligators.
[01:03:53:07 - 01:05:21:07]
Mallory
Oh man, that's a good segue to today's episode. Let me see how I'm gonna tie that in. So this week we've got some new models from Anthropic and OpenAI, almost like new alligators in Bayou St. John. We're gonna start with Claude, including a wild two week saga with the US government and then get into OpenAI's release and the bigger question of what happens when the government starts putting reins on the most powerful models. And then we're gonna close on a new study that's reigniting the AI jobs debate. So first, Anthropic's newest models. Anthropic, as a quick note, makes Claude and three everyday tears. We've got Opus being the biggest and the priciest. We've got Haiku, which is the small, fast, cheap one. And then Sonnet sits in the middle as the workhorse most people actually use. There's also a new top class called Mythos that sits above all three, which is gonna come up in just a bit. Anthropic released Claude Sonnet 5 on June 30th. That was actually just yesterday as of the recording of this podcast, calling it its most agentic Sonnet model yet, meaning it can take actions on its own with tools. The headline is price to performance. So it runs close to the flagship Opus 4.8, but it costs a fraction of it, roughly a quarter of the price at introductory rates. It's now the default on free and pro plans and available everywhere else, including Claude code and the API. Amith, have you had a chance to try out Claude Sonnet just yet? It's only been 24 hours, but have you tested it out?
[01:05:21:07 - 01:06:27:10]
Amith
Yeah, I jumped on it. I tend to try new things as soon as they come out if it's kind of in my work stream. So yes, I tried Sonnet 5 yesterday. I was impressed with its speed. I was impressed with its kind of overall level of intelligence. As our listeners know, I tend to do a lot of software development type stuff. And so within Claude code, it wasn't smart enough for what I was doing, but it might've been the project I'm working on right now, which is kind of complicated, but I think it's good for a lot of things. So it's just great to see the progression of this tank because Sonnet 5 is, it's smarter than Opus 4.6 by a good margin. I think it's about the same as 4.7 and only just behind 4.8. So, you know, it's a good model. It's exciting to have it out there. The other thing that I do not have confirmed, I researched last night a little bit, is Anthropic and Google have been partnering more closely on hardware for both training and running these models. And I'm pretty sure Sonnet is running on the TPU 8 architecture, which is Google's new very fast inference compute. So very excited about that. I mean, it seemed really fast. And that's why I went and checked only. This is a little bit different than what I expected. So it's interesting. I think it's going to be the workhorse for the vast majority of people.
[01:06:28:16 - 01:07:35:19]
Mallory
So Anthropic's benchmarks are showing Sonnet 5 landing just behind Opus 4.8 on agentic coding and computer use and actually slightly ahead of it on knowledge work. And the practical read on that is that the mid-tier model now does most of what the flagship did just a few months ago. Now for a quick catch up, we covered Fable and Mythos at this point a couple of weeks ago on the podcast. These are two of the more powerful, most powerful models, Anthropic launched June 9th, aimed at things like advanced cybersecurity with Mythos going to only a small group of vetted cyber defenders and Fable being available to the public. But on June 12th, just a few days after that, the government issued an export control directive and national security rule limiting who can access sensitive tech to suspend access for any foreign national. Well, of course, Anthropic couldn't verify nationality in real time. So it pulled both offline for everyone. Now, as of June 30th, again, yesterday, the export controls on Fable 5 and Mythos 5 have been lifted. For me personally, Fable is still unavailable as of today, July 1st. So Amith, what is your take on all the model craziness with Anthropic in the month of June?
[01:07:35:19 - 01:08:28:02]
Amith
Yeah, first of all, I don't have it back either. So I don't know if that's just rolling out slowly and we're just too far down the tone pole from a Anthropic's special customer list, but we'll see. I suspect it'll bubble up fairly quickly, but that was pretty cool. I mean, the brief period of time when I used it for that three days, it was notably smarter than Opus 4.8, it was pretty cool. So maybe this is all a master plan and it's a remarkable lesson in marketing to keep people waiting and give them a little taste and then take it away. So I'll coordinate it now, I just kid. My thought is Anthropic is doing what Anthropic does. They're incredibly fast at shipping. It's just amazing how much stuff they ship. I think they're willing to put stuff out there like this new Claude tag capability we've talked about and other functionality they're able to put out new models. At this point, probably faster than anyone else.
[01:08:29:07 - 01:09:53:18]
Amith
Part of it is ultimately in the world of AI, the smartest people and the most capable people in whatever the field is, are still more important than ever. And Anthropic has had a series of people, brilliant people come to them over the course of several years, but particularly in the last six months. So that's certainly helping accelerate them because it's not just the headline catchers like the Andre Carpathi and others that have gone to Anthropic, but it's the people who follow that. So when Andre goes to Anthropic, lots of other people choose Anthropic over open AI, over Google when they're thinking about where to make their move. And I'm talking about people who are kind of at the core of this AI research, which is a pretty narrow list of people. But in any event, I think Anthropic's done a good job putting these stuff out there. The most important thing is just to keep the trend line in your brain that you can access really smart AI for a reasonable price. It may not be through Anthropic, but it is coming to you fast through other models and the compression continues. And what I mean by that, if you haven't heard some of our other episodes about small models and just the relentless progression of model intelligence becoming more and more kind of commoditized, that's really what's happening. It's, you wait a week or two weeks or a month and a new model comes out either from an open source provider or a major lab that is in fact dramatically better than the last generation only months prior. So it's pretty exciting.
[01:09:55:15 - 01:10:20:05]
Mallory
Oftentimes on the podcast, we harp on choosing the right model for the right tasks. So not selecting the giant frontier behemoth model to do something pretty simple and efficient. Can you give us an example of me putting you on the spot here of when you're thinking of Haiku, Sonnet and Opus? We talked about Fable on that episode I just mentioned earlier. So let's just focus on Haiku, Sonnet and Opus. What kinds of tasks do you think are for each?
[01:10:20:05 - 01:13:48:13]
Amith
So when a Haiku came out, I used it for computer use in the browser. So the Google Chrome plugin for Anthropic Clot I've used in the past and Haiku was super smart for its size and fast. So it was cool because you could tell, hey, go do this and this and this and it would very rapidly do that. But Haiku tends to lag once they push the new models ahead. I'm sure there'll be a Haiku five because they needed a small model in their catalog. And at that point, I think that kind of simpler task would be great. I think of it almost as like planning versus execution Mallory, where regardless of what the task is, if I'm building a complex piece of software, I'll have a frontier model like Opus or Fable build a really thorough plan. And I'll spend a lot of time digging into the plan, pushing on the plan. I even asked that model to spin up a sub agent and ask it to take on an adversarial role. And when I specifically use that terminology, adversarial AI is a technique. It's also a cybersecurity term. And it basically means that you're asking an AI to take on kind of the counterpoint, right? It's like debate one side of the issue, then debate the other side. You're kind of asking it to rip apart the plan. And so if you put the reps into the plan with a really smart model and you're engaged there, then I find that you can almost like take that plan and outsource it to a smaller model. My workflow that I've been using recently is I'll use something like Claude Opus 4.8 for doing planning. And this is, by the way, applicable to business stuff too. Like if I'm working on something in terms of a business process or a release, or I'm working with a client on some strategic type of conversation, I'll tend to use a really smart model to kick around ideas, come up with the ideas at a detailed level, like a detailed kind of outline. And then I'll outsource it to a smaller model to execute. So my workflow with coding is I'll use Opus 4.8 in Claude code to build a really detailed plan. Then I'll fire up something called anti-gravity, which is Google's AI code type of thing. And they have a model in there called Gemini 3.5 Flash, which is an incredibly smart model, but it's also crazy, crazy fast and really cheap. And so Gemini 3.5 Flash, sometimes I'll ask it to like review its own work, but it'll rip through a really complex software engineering task. Then I'll have Opus review that work. So Gemini 3.5 Flash in my example, is kind of like Sonnet Haiku type thing. It's almost as smart as Opus in many ways. And then Opus sometimes finds holes, corrects them, acts as QA. So that's kind of the way I look at it is I use a blend. I would say right now, the last comment on this is Cerebrus, which is the fast inference hardware and cloud company that just went public in the last, I think 45 days. They just did a public preview of the Gemma models. Gemma is G-E-M-M-A, the Gemma 4 series of models from Google. This is Google's open model series. And what's interesting is Gemma 4, the model that Cerebrus has hosted, is smarter than Gemini's smallest model. So Google's commercial model, which is the smallest fast model is 3.1 Flashlight. Gemma 4 is both way, way smarter than Gemini 3.1 Flashlight, and dramatically faster when you run it in Cerebrus. Like I'm getting 1800 tokens per second, which is probably something in the order of around five to seven times more token throughput there. So for really small, fast things, I would point to somebody like that. I also saw some news, by the way, in the fast inference game that the folks at Grok with a Q, GRO-Q,
[01:13:49:17 - 01:14:35:06]
Amith
when their IP was licensed by NVIDIA at the end of last year, I just kind of figured they'd shut down their cloud ops, but they decided not to. In fact, they just raised $650 million, which in AI land is, I guess, a couple bucks. And they're investing heavily and going after a further build out, and they're gonna be deploying more models on their public cloud. So that's super exciting as well. So think more broadly is really what my point is, is like think about a model mix that allows you to kind of extract the most value from what you need, depending on the workload. Like if you have a backlog of a million documents, you want to auto classify and tag or something like that, which is a common workload in associations, to run on a recurring basis, right? Because your taxonomy changes, you don't need anything more than Gemma 4. In fact, that's probably overkill.
[01:14:35:06 - 01:14:45:00]
Mallory
Okay, that makes sense. So the Frontier model is more for the planning and the strategy, and then you can use the smaller models, which are still quite smart to execute on that plan.
[01:14:45:00 - 01:15:01:15]
Amith
I think so, yeah, I think there's some situations where there's so much nuance that maybe you do want the smart model to do the whole thing. But I think a lot of the planning work, if you get that right, you have a good blueprint, and then you can get a lot of kind of worker bee models to actually execute the plan.
[01:15:01:15 - 01:17:19:07]
Mallory
Hmm, well, I do want to ask you about the kind of government intervening with this model and what that means at a larger scale. But before we do that, Amith, I want to actually go into OpenAI's latest model release and then talk about both. So OpenAI's GPT 5.6 launched this same week, June 26, and it's actually three models under one generation number, each at a different price point. And they have better names, everybody. I'm really excited about this. So we've got Sol, this is the flagship model, the most expensive. We've got Terra, which is a balanced everyday model at about half Sol's price. And then we've got Luna, which is even the faster and cheaper option out of all three. OpenAI says the number marks the generation while the three names are the tiers that can each advance on their own timeline. So GPT 5.6 has launched in quotes, but only as a preview to about 20 government approved partners through the API, not in chat GPT, with no public release date just yet. OpenAI has said it doesn't want this kind of government gated rollout to become the norm. We'll discuss that. On performance, the standout result right now is coding. So Sol set a new record on terminal bench, a test of whether a model can carry out a full multi-step software task in a command line, not just spit out a snippet of code, scoring about 89% with its high effort ultra mode near 92%. That edges out Anthropics most powerful models, including Mythos and Clears, the publicly available Opus 4.8 by a wide margin. So in plain terms on the kind of hands-on engineering work AI is increasingly trusted to do end to end, this is currently perhaps the best on record. Two caveats here though, OpenAI hasn't released a couple of the other major coding benchmarks. So this is a strong data point, but not the full picture. And an independent evaluator flagged that Sol had the highest reward hacking rate of any public model it's tested, meaning it's sometimes games a task to pass rather than actually solving it, which makes you read even a record score with some caution. Amith, I wanna pause on that note because you're the resident coder, I would say out of the two of us. Are you impressed by what you're seeing with the Sol benchmarks? Would you be running to codecs as soon as it's available? What do you think?
[01:17:19:07 - 01:19:07:23]
Amith
I'm definitely gonna jump into codecs and try it out when I get access to it, but I'm not holding my breath either because benchmarks don't tell the whole story. Even with Sonnet 5, it says that it's nearly as smart as Opus 4.8, but what difference do benchmarks make when your own workloads either represent the benchmark or don't at all? And so you might find that a model that is not so great on benchmarks is extraordinarily good at some of the things that you need it to do. That's actually why the concept of evals or essentially having your own benchmarks ends up becoming a really important concept. It's something we recommend people do when they're particularly doing agentic work is to build a benchmark suite of their own agents and the agents that they work with a lot and run them through a variety of different models when they're considering new model configurations. We do this for our own SaaS products like Skip and Betty and so forth. And I think it's actually fairly easy to do that. There's a number of tools you can use for these kinds of evals. We have some of those built into MJ and there's a number of third-party tools you can get both paid and open source, but that's what matters. So you wanna run these things through your evals, through your own benchmarks. Yes, of course the public benchmarks are directionally helpful, but whether or not Sol is going to be better for you than Fable or even better than Gemma 4, it depends on your use case. So I wouldn't get too excited about the releases. And I would also underscore what you did mention Mallory, which is OpenAI does have a habit of selectively releasing benchmark results. They don't tend to release everything all at the same time. And sometimes they never release it. They also tend to be more focused on their releases to compare them against their own benchmarks in prior versions of their own models versus actually comparing against their competitors directly. So I think in this particular release, they did talk about Fable and Methos, but only in the one thing that you mentioned.
[01:19:08:23 - 01:19:58:22]
Amith
So there is a website which is called artificial analysis that I would recommend folks check out. This is a tool that basically aggregates up model data and benchmark data from publicly verified sources. And what they also have is their own index, the artificial analysis index, or I think it's called intelligence score, which is basically, it's kind of like the Dow Jones industrial average, right? It blends in a whole bunch of benchmarks into a single number. So rather than looking at like all these different numbers, you can see the models overall intelligence. And so this is a proprietary benchmark, not really benchmark, but proprietary index effectively into a lot of benchmarks. I look at that a lot because it's just a quick way to say, roughly how smart is this thing compared to models I know well? So I would recommend you do that as a directional thing. It's kind of like navigational direction, but then you got to hone in on exactly what you need.
[01:20:01:00 - 01:20:25:17]
Mallory
Within one week, Anthropic had its most powerful models forced offline, and then OpenAI could only ship its newest to a government approved shortlist. So we're seeing the government increasingly act as kind of a gatekeeper for the most capable closed models. I mean, do you think that's the new reality for where we are with AI and Frontier closed models that they're all going to be increasingly, I don't wanna say dangerous, but potentially dangerous in the hands of bad actors?
[01:20:26:21 - 01:20:40:17]
Amith
I think it's the same concern that we've been talking about since the beginning of this podcast, Mallory, for 141 episodes now, we've been talking about how you need to have good AI to keep up with bad AI, or really there's no such thing really as good or bad technology.
[01:20:41:17 - 01:21:43:04]
Amith
Technology can be used for amazingly wonderful things or terrible things. And the perspective you have in your value system, obviously depends, that drives how you interpret these things too, to some extent. So I think we have to keep pushing the envelope. I think the downside risks to these types of restrictions is that it also causes people to maybe gravitate towards models coming from other places. So if you think about what's happening with the Chinese open models like GLM 5.2 and others, they're right behind the frontier. And so, you know, I would suspect that if you took GLM 5.2 or QEN 3.7 or Minimax 3, and if you put them in an agentic harness, which is to say that you use the model iteratively or in a loop, you could probably beat me those or fable at some of these cybersecurity things. Now I'm saying that purely as a speculation. I'm not a cybersecurity expert, but I do know that that's true with coding. I do know that that's true with breaking down really complex research tasks that agentic harnesses make the models dramatically more capable.
[01:21:44:08 - 01:21:53:17]
Amith
So I think that this is somewhat of a misdirection. I don't think anyone's intentionally misdirecting, but it's problematic. Now the problem of course is yes, this is a massive concern,
[01:21:54:19 - 01:22:35:10]
Amith
but I don't know what the solution is other than continue to advance the power of the tools we have to try to create, you know, defensive strategies that can protect against, you know, the things that are coming. So, you know, that doesn't provide a lot of comfort, but I do think that it's important to have a realistic view on what this means. For our association and nonprofit listeners, what does this mean? Couple of things. In our recent episodes, we've talked about basic cybersecurity things you can do, like for example, not writing your password or down on a post-it note and standing patting it in your monitor. I know that sounds silly. I've seen people do that as recently as last year when I was visiting someone's office. Not their password to their computer necessarily, but passwords to like some website that they find annoying to have to log into.
[01:22:36:16 - 01:23:49:14]
Amith
Don't leave your computer unlocked when you walk away. Have verbal passcodes with, you know, friends, family, and colleagues that are used for important things that you don't put on any electronic device. We've talked about this stuff before. The other thing is pay for a cyber audit. Pay somebody outside of your own team to do, you know, pen testing and some other basics. You can spend anywhere from as little as 10K up to a lot more than that to do basic, you know, cybersecurity testing. It's not a service we offer across our business. I'm not pitching anything we offer. I'm just saying like, you should go do this because it's really, really important to have independent testing and at least push a little bit on your own defensive wall. You know, you might think you have this amazing thing and then you have someone from the outside just like poke at it for half a second and the whole thing falls over. That's a problem. I've known a lot of clients in the association market who thought, oh, little old me, I'm just a nonprofit. No one's gonna pay attention. I'm just a $5 million a year nonprofit. Nope, nobody cares. You know, the hackers that are out there, they use bot farms, they use AI at scale. They're gonna come after you if you have weaknesses. So you gotta, unfortunately, you have to invest in these things. So that's really the advice I have on it as a practical matter for our association listeners. I have one other thing to say too.
[01:23:50:15 - 01:24:42:03]
Amith
I think that it is more important now more than ever to invest in the most powerful form of AI, which is augmented intelligence in this case, which is your own brain. And that is to say that you need to go and double down on your own education. Obviously, Sidecar Learning Hub is one option. There's tons of other things. You're doing it right now by being part of this pod and part of our community. But letting yourself fall behind is probably your greatest risk. Now, so far as I know, Mallory, the government has not yet put Sidecar's Learning Hub on the restricted export list due to its power, but it is extremely powerful if you use it right. So I would suggest everyone here take advantage of that and take advantage of other learning tools because honestly, you know, all this stuff is overwhelming. It's overwhelming to both of us, but you have to stay on top of it. There's really no other answer to that.
[01:24:43:07 - 01:24:45:20]
Mallory
We're gonna get flagged for that on this podcast. So thank you. Excellent.
[01:24:47:07 - 01:25:13:01]
Mallory
One question in researching this topic specifically, I went to the information, which is a news source, a website that we love on the podcast. And they wrote an article about how developers told them that they're shifting to wanting to use only open models to avoid getting burned again if they're using a closed model that suddenly gets pulled by the government. What are your thoughts on that? Like, is that something you're scared about for the greater Blue Cypress family of companies or not really?
[01:25:13:01 - 01:25:49:07]
Amith
I think optionality is always important when you're dealing with things that are critical to your business. So, you know, with financial infrastructure, it's best not to have all of your eggs in one basket. Don't have all of your money with one brokerage. Don't only have one banking relationship. Don't only have one AI inference provider that you rely on for everything. Because you're like, "Oh, Fable is awesome. I'm going to switch all my workloads to Fable." And then three days later, Fable is gone. So you need to have optionality. You need to plan for change. And it's not just the government or, you know, cyber hacks or whatever. It's just stuff happens. So optionality is key. Build with options in mind.
[01:25:51:15 - 01:28:01:07]
Mallory
I want to shift to topic three for today. The AI jobs debate gets a little bit messier. So let's start with the backdrop. Through May, employers announced close to 90,000 jobs cut tied to AI and BCG projects up to 15% of US jobs could be eliminated by AI over the next five years. Industry promises that AI will also create jobs haven't done much to calm the fears, especially for people about to enter the workforce. A new working paper from Ramp and Revelio Labs tracking AI spend and workforce data across nearly 22,000 companies complicates that story at least a little bit. It found high intensity adopters, firms spending around $30 per employee per month on AI early on grew head count by about 10%. And even entry levels in those companies rose about 12%. The exact jobs everyone assumes are most at risk. Now the authors are careful. In their words, the paper does not show that AI universally creates jobs by any means, just that it counters the idea of broad losses at least a little bit. And the data skews heavily toward tech forward VC backed firms that were already growing fast. So it's hard to say AI caused the hiring versus just showing up where hiring was happening anyway. But the headline we think for our audience is that the gains went to companies that went all in on AI and kept investing. While those that just bought subscriptions or ran pilots and didn't execute on any of those saw no head count growth at all. The authors warned this sets up a widening gap since the winners tend to be firms that already had the resources to pull it off like capital or technical staff management bandwidth. And they say firms without those channels may fall behind. So the honest read here is that this doesn't cancel out the bad news by any means layoffs are real and still growing, especially for junior workers broadly. What the study shows is narrower, a certain kind of company, one that invests deeply and sustains it might be able to use AI to grow rather than shrink. So Amith, I'm curious, what is your take on this new study in the AI jobs debate?
[01:28:02:08 - 01:29:25:04]
Amith
Well, I mean, Mallory, we're hiring more people than we have in the last few years. And we're growing a ton and we're hiring people that are right out of school. So, and I'd say that we're pretty heavily focused on AI. So, my thought process is that if you see opportunities to grow your business, to grow your impact, to grow the volume of activity that you do, yes, AI is an enabler, but you need, you always are gonna need smart, motivated people to help you grow. It's just the work that they're doing is differently. You know, we have tons of software engineers here. We're adding software engineers like almost every week to our team. And these folks that are coming in are not writing code. I mean, they're, they know code. They all have computer science degrees. Some have advanced graduate degrees in computer science and AI and machine learning and stuff like that. But they're not working at that lower level. I mean, we've automated all of that almost. And so you still need to understand it because I wanna make sure that the engineers deeply understand the fundamentals so that they're building really good software. So they're building architecturally sound software, software that's reliable, software that's secure, all of these things. And I'm not gonna rely just on AI for that, but the average engineer here produces at least a hundred times as much software today than they did two years ago. And that's not make believe that, I mean, if a hundred is actually a baseline, I'll give you one quick example. One of the engineers that joined here about a month and a half, two months ago, he's brand new, right out of school.
[01:29:26:11 - 01:30:42:08]
Amith
I gave him an assignment the other day. I asked him to basically rebuild SurveyMonkey type form, Microsoft Forms, Google Form type app. I said, "It'd be great if we just had a free one of those so that all associations could have their data flow right into the, this is a member junction based app. So it flows right into member junction, which is also free." I said, "Wouldn't it be great if we had, you know, a forms application?" Actually another fellow who's another, another one of the early career folks had suggested this. And my original reaction being the old guy said, "No, no, no, we don't want to do that. Reinventing SurveyMonkey is a lot of work." And it's not necessarily wrong. It's just that I thought about it a couple of days later. I'm like, "Actually, wait a second. This guy was right that actually we could, not only could we do it, but it makes sense because housing that data locally, along with your other data, that unstructured data that comes back from surveys is extraordinarily valuable. And having that tucked away and sometimes associations use many of these tools that use type form and SurveyMonkey and Microsoft Forms and blah, blah, blah, blah, blah, blah. And so anyway, this, I digress a little bit, but the point is, is that is a non-trivial software engineering effort. It is something that will take, I would say that, you know, pre-AI, we would have said, "Well, that's probably a basic version of SurveyMonkey, a team of three to five engineers to spend a year on it." Something like that. How long do you think it took us, Mallory, to rebuild it? It's basically done.
[01:30:43:14 - 01:30:45:23]
Mallory
I'm gonna say a few days, knowing what I know.
[01:30:45:23 - 01:31:30:10]
Amith
It's 24 hours end to end. Now it's not 100% done, it's not released. It will be another day or two, but think about the order of magnitude variance there between three to five person years and a handful of days. It's crazy. And so, but we're gonna hire more people because I want more software. I wanna release more software into the market. I wanna give associations more and more powerful tools, make it easier for them to use. And a lot of what we're doing now is also a lot of the engineers are client facing. We have so many associations coming to us week after week saying, "Hey, can you help us actually implement this stuff?" Which we do a bunch of that. And we love doing that because that's actually where the rubber meets the road. That's where theoretical ideas and software products that are in the cloud actually come down to earth and actually solve problems for people in our industry.
[01:31:31:16 - 01:32:21:02]
Amith
And before we wouldn't have had the bandwidth to do a lot of that. Now we're doing tons of it because the software development process is, I won't say it's completely automated, but there's so much of that low level coding is done. So the point would be, that's not making me as the founder of this company say, "Hey, this is awesome. I can go from, you know, however many engineers we have globally down to half of that or something and get rid of people. I want to hire more because I see so much opportunity to serve this market and grow our business and have more impact." So it's a mindset issue, I think is what it boils down to. I would also agree with you that unfortunately, that mindset is not common. And so most people are looking at it from that inverse angle and saying, "Hey, I have 5,000 employees. I only need 2,000 employees. Goodbye, they're 3,000." That's going to happen and happen and happen. And it's going to be very, very hard for us to figure out how to solve that problem. That's a society level issue.
[01:32:22:07 - 01:32:49:02]
Amith
But I do think some companies are going to be driving tremendous growth. They're going to be pulling in a lot of those folks and bringing in certainly early career folks and people at every level, because different experience levels, particularly people who are extraordinary communicators, people who are really good at, you know, the human thing. If we're good at human to human work, that's really important. And so we look for engineers, we look for people in other roles that are both, enjoy that and are really good at it.
[01:32:49:02 - 01:33:14:22]
Mallory
Mm-hmm. And I would challenge you as the association leader listening to this, like, which camp do you want to be a part of? Do you want to be the leader that's part of the world shrinking the workforce because we've got to cut costs? Or do you want to be one of these heavily invested into AI organizations, growing your staff? I mean, you have both options. I mean, that was going to challenge you on that a hundred times more software statement that you just made, but I feel like you just supported it. So that's insane.
[01:33:14:22 - 01:34:44:02]
Amith
Yeah, it's totally nuts. And this product, I'm super excited about it. It's called MJ Forms. It's going to be a free open source app. We're publishing it. We'll put it out there this month in the month of July and let people start installing it and using it, but it's a free app. You can install it. Obviously it's nice that you don't have to pay, you know, type form or survey monkey hundreds of dollars a month or whatever, but the bigger advantage is then the data flows directly into your data platform. You own it forever. You can do whatever you want with it. You can run AI against it. You can have agents create these forms. You can have agents kick off when these forms are completed. All sorts of cool workflows come out of it. So it's pretty neat. I'll say one other quick thing too, Mallory, is that to add on your point about the association leader listening in, you have an opportunity to also put your team's fears to rest. Because if you haven't openly addressed the idea of job loss for your staff and your association, you need to think about how you're going to do that because you may not be talking about it, but I guarantee you a large percentage of your team is thinking about it. I'll give you an example. Depending on your association, you may have one or two dedicated member services folks. You may have dozens, or in some cases, some of our larger clients have call centers with over 100 people that do nothing but answer the phone. And when you think about the advancing capabilities of real-time audio AI, which we need to do another episode on that, Mallory, because earlier this year when we did that one with Grace, that was cool, but Grace is like totally, that demo was cool, but it's super old school now compared to what we're doing these days. And that was like what, four months ago, three months ago?
[01:34:44:02 - 01:34:46:08]
Mallory
Yeah, I think it was less than, yeah, maybe three months ago, wow.
[01:34:46:08 - 01:34:51:11]
Amith
Yeah, what we're doing now with real-time audio and other things is just really crazy.
[01:34:52:11 - 01:37:12:19]
Amith
But the point is though is that yes, it can totally replace what we've been doing on the phone with customer service, 100%. And there's a lot of people deploying this right now for real-world customer service. And some companies, I don't agree with this position, what I'm about to say by the way, but there are companies out there that are deploying it without informing their consumers that it's an AI they're talking to. I don't agree with that part, that's my personal opinion. I think you should always say, "Hey, I'm an AI and I'm here to help you. And I'm a new AI, so you can, don't think of me as the phone tree that you wanna hang up on. Maybe make it a little bit funny and just say, "You can ask me anything and I'll work through it with you." And these AIs are so good that they can look up data in your AMS and they can modify records. They can authenticate the customer by asking them to provide some information to verify they are who they are. All these kinds of things, which is so exciting, right? It's the tech geek in me that's going off on that. But the point is, is that you have human beings that do that job day to day and they know this is coming. Even intuitively or some know directly that this stuff is around the corner. So what do you do? Well, you could go out there and say, "Well, we're gonna retrain our people to do other things." Okay, but what other things? Exactly what does that mean? And how are you gonna do that? And how are you gonna help those people who've been answering the phone for 30 years and do something else? Those are hard questions. They're like way harder questions than the technology side. You gotta put some work into that. You gotta think about like, okay, what have our members been asking for? And what have they not asked for that we think they might like that we couldn't do because we just don't have enough people to do it? Well, guess what? Now all of a sudden you have more people available to do some of those things. So what are those things? How do they add value to the member ultimately to drive revenue, drive engagement? And then how do I retrain my people to go do those jobs? Now, not everybody's gonna be willing to retrain, right? That's their choice. That's a different problem. Assuming that you have people who are capable and willing, I think that most associations very much want to retain their team. They don't want to cut a bunch of employees, but they gotta actually have a plan for this. You gotta think through it. So that's my main advice to leaders is to actually put some effort into that and think about what those gaps have been historically to delivering new and additional value. As people say it all the time, they're like, "Oh yeah, we want member service to be more proactive. "We want member service to be more outbound." And they go reach out to our members and like deliver more value, which I totally agree with that. I'm not trying to diminish that, but the point is, is like double click on that, like figure out what that actually means.
[01:37:14:05 - 01:37:53:11]
Mallory
I can say 100% honestly, I could go to you, Amith, right now, or to our CEO, Johanna, and say, "Look, I've automated every single piece of the podcast. "I don't even have to host it anymore. "I can just have my avatar show up." Everything I do for Sidecar, and I would feel confident that both of them would have another 10 things. Well, okay, what if we did this and maybe we could use you in this way? And if you are a leader and you don't feel like your staff could come to you and say, "Hey, guess what? "I've automated all of my member service tasks "for the foreseeable future." If you don't feel confident that they could do that, that's something to sit with a little bit and just, I don't know, see if you can work on the culture within your organization.
[01:37:54:24 - 01:38:12:21]
Mallory
Amith, my last question here, I think it's kind of a fun one. The data in this survey skews more tech forward and VC backed firms, like I mentioned. If there was a way for you to put up syringe into companies like this, pull something out, and then inject it into an association, but only one thing, what would that be?
[01:38:12:21 - 01:38:15:03]
Amith
It's a very visual way of asking it.
[01:38:15:03 - 01:38:20:01]
Mallory
Yeah, I don't know. I'm trying to get over that part. It seems like fun, but I guess maybe, hopefully not graphic.
[01:38:20:01 - 01:38:21:16]
Amith
It's an organ transplant.
[01:38:21:16 - 01:38:22:09]
Mallory
Exactly.
[01:38:22:09 - 01:38:38:08]
Amith
So I think the part of it that I would probably want to inject in would be willingness to experiment, and then the willingness to call an experiment a failure, and then the inverse, of course, which is what we all would like to see, is call an experiment a success.
[01:38:39:08 - 01:38:50:08]
Amith
Point out that that part is the most important part of experimentation. Lots of people run experiments in the association market. Very few of them are willing to take a successful experiment and step on the gas.
[01:38:51:09 - 01:39:04:01]
Amith
So I hear about people that have even gone and built incubators, and they have these ideas of like, hey, we have an experiment fund, we allocated this money, and I totally think that's awesome. But then a lot of these same organizations, when they do have a success, they're like, oh, this is great.
[01:39:05:22 - 01:39:18:04]
Amith
And that's kind of where they stop. Or they'll say, oh, this is great. We gotta go present this to the board and try to get funding so that in fiscal year 2027 to 2028, we can do something, which is like not until March of next year, and it's July.
[01:39:19:05 - 01:39:43:07]
Amith
So there is, and that's not because the people don't want to do it, it's because there's this kind of impotence mismatch, and you have this problem, ultimately, where you don't have people aligned, and that's organizational, it's cultural, it's some governance. So point would be that the piece that you need to find a way to inject is the ability to experiment and then to actually do something at the end of the experiment with the successes so that you can scale it.
[01:39:45:22 - 01:39:51:22]
Mallory
That's a good point to me. Well, two model launches, a government force shut down, and come back for Anthropics' most
[01:39:51:22 - 01:39:57:14]
(Music Playing)
[01:40:08:07 - 01:40:25:06]
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.
[01:40:25:06 - 01:40:28:12]
(Music Playing)