Summary:
A drug company CEO just made the best case yet for why platform choice matters less than most associations think. Marking three years of Sidecar Sync, Amith Nagarajan and Mallory Mejias dig into Pfizer's new playbook for organizing around AI, Thomson Reuters' $40 million bet on building its own legal model instead of renting one, and the six-day mystery of Ox Alpha, an anonymous coding model that became OpenRouter's biggest launch ever. They trace Pfizer CEO Albert Bourla's three moves — structuring data, federating decisions, and building AI fluency — and translate each for associations of any size. They also unpack why Thomson Reuters trained its own model on Alibaba's open-weight Qwen instead of paying frontier labs, and how that differs from deploying a knowledge agent like Betty. Whether you're weighing a custom model or just trying to get your team AI-fluent, this episode lays out what actually moves the needle for associations.
Timestamps:
00:00 - 150 Episodes and Counting
04:53 - Pfizer's Blueprint: Adoption Isn't Transformation
13:24 - Federate AI Without Abdicating It
19:38 - Why AI Fluency Has to Come First
29:19 - Thomson Reuters Builds Its Own Legal AI Model
34:56 - Knowledge Agents vs. Custom Models
38:29 - Why Associations Need Domain-Specific AI
42:52 - Ox Alpha Unmasked
50:49 - Closing Takeaway: It's Never Been About the Tech
👥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:
Sidecar Sync Ep. 142 The Data Episode ➔ https://tinyurl.com/4vryx82a
Claude ➔ https://claude.ai
Claude Cowork ➔ https://claude.com/product/cowork
Claude Fable ➔ https://www.anthropic.com/claude/fable
ChatGPT ➔ https://chatgpt.com
ChatGPT Work ➔ https://openai.com/index/chatgpt-for-your-most-ambitious-work/
Microsoft Copilot ➔ https://copilot.microsoft.com
GLM (Z.ai) ➔ https://z.ai
Kimi ➔ https://kimi.com
DeepSeek ➔ https://www.deepseek.com
OpenRouter ➔ https://openrouter.ai
OpenCode ➔ https://opencode.ai
Qwen ➔ https://qwen.ai
Betty ➔ https://meetbetty.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 🤖
[00:00:00:14 - 00:00:09:17]
Amith
Welcome to the Sidecar Sync Podcast, your home for all things innovation, artificial intelligence and associations.
[00:00:09:17 - 00:00:24:04]
Amith
My name is Amith Nagarajan.
[00:00:24:04 - 00:00:26:05]
Mallory
And my name is Mallory Mejias.
[00:00:26:05 - 00:00:47:04]
Amith
And we are your hosts. And it's not the home for artificial associations, it is the home for artificial intelligence and associations. That combination, Mallory, I think is the key to why people tune in and hang out with us week after week. In fact, for 150 weeks in a row now, right? Is this 150?
[00:00:47:04 - 00:01:00:17]
Mallory
150 weeks in a row. It's crazy to think about it. We've come a long way from the beginning of the podcast. And I know we have some listeners that I believe have been there since episode one. So shout out to you for being with us on this long journey.
[00:01:01:19 - 00:02:21:15]
Amith
150 weeks is almost three years. In AI timescales, that's six doublings or more of AI power. It's pretty incredible. We appreciate everyone who has been part of the journey with us, whether you've been with us through from single digit episodes or double digit episodes, or if you just recently found us, thanks for being here. Our mission is to educate the world of the nonprofit community broadly and specifically focusing in on our good friends and colleagues in the association market, helping you think about how AI can drive business outcomes in your association. Yes, we get technical here a little bit sometimes, but the purpose of that is to educate you at every level, whether you're the CEO or the board chair or you just graduated from college and you're starting your first association job and you don't know which way is up yet. We are here for you. We're here to help you on your journey. The journey of life and the journey of an association through this transformative period, it's going to be interesting. I think it already has been interesting now, but the next three years between now and kind of the wrap up to the decade, we're at that midpoint of that six year window that we started talking about in terms of end of the decade when we started the pod. Our goal is to educate a million people in the market by 2030.
[00:02:22:16 - 00:02:32:08]
Amith
We're over 10% of that goal now in terms of the total touches we've had and accelerating rapidly. I'm quite optimistic that based on exponentials, we'll achieve or beat that number at this point.
[00:02:32:08 - 00:02:50:04]
Mallory
Wow, Amith. Thinking about the end of the decade, starting this decade off with the pandemic, the advent of generative AI, it will have been a crazy 10 years. Did you think when we started this podcast we would get to 150 episodes or were you kind of just shooting in the dark saying, "We'll see what happens."
[00:02:50:04 - 00:02:54:01]
Amith
I always just seem everything's going to go perfectly and last forever. That's fair.
[00:02:54:01 - 00:02:54:19]
Mallory
That's fair.
[00:02:54:19 - 00:03:03:20]
Amith
That's just, I mean, entrepreneurs tend to have like a little bit too much of that type of Kool-Aid in their systems because I think otherwise entrepreneurs would never start anything. I
[00:03:04:20 - 00:03:11:20]
Amith
wouldn't call it blind optimism, but it's kind of endless optimism or a well of optimism. I figured we'd be doing this for a while.
[00:03:13:15 - 00:04:12:06]
Amith
What's rewarding about this to me is, first of all, it's a lot of fun doing this each week, catching up with you and going through all this cool AI stuff. It's always good fun. Then when I hear from our listeners, whether it's at a conference like ASAE, which just wrapped up, or additional now, which is coming up October 25th through 28th, by the way, if you haven't registered and put it on your calendar, it's going to be awesome. It's in BC, October 25th through 28th. But regardless of where I run into people, I consistently hear from people that Sidecar and specifically the Sidecar Sync pod have helped people, have helped people move from zero to one a lot of times, like going from having done nothing with AI to getting started. We've also helped people go from one to 100, where they've already moved, they've already started going, but they've been trying to figure out what the use cases are, how to balance the risk of AI relative to its potential upside, how to talk to their board about AI, how to educate people who may be reluctant, or perhaps are just flat out anti-AI
[00:04:13:09 - 00:04:27:24]
Amith
people. So we've helped people with so many different backgrounds. That's been super fun. So if you're enjoying the pod, the best way you can help us is by sharing it with people that you know who we may be helpful to as well. And we'll keep growing this community and go transform the association landscape.
[00:04:27:24 - 00:04:51:14]
Mallory
Absolutely. And you're right, Amith, it is fun to do this every week. We have never skipped a week of the podcast. I'm trying to think back. I think one Thanksgiving, we did an AI episode that just had an AI avatar. So you might say we skipped that week. But other than that, we have put out an episode every week for the past 150 weeks. So as Amith said, thank you all for being here and we look forward to the next 150.
[00:04:53:08 - 00:06:34:15]
Mallory
Moving to today's episode, we've got three things that we are covering. One is Pfizer CEO publishing an essay on how he's actually reorganizing the company around artificial intelligence. And it's more about org charts than technology. Thomson Reuters launched its own AI model built on top of someone else's for less than half a million dollars. And a model showed up online with no name on it and no lab claiming it until this morning, right before we recorded this episode, I did a check and realized the name is out there. So we will be reporting on that. But first, how is Pfizer thinking about AI? Earlier this month, Albert Borla, the CEO of Pfizer, large American multinational pharmaceutical and biotechnology corporation, published an essay called How Pfizer Thinks About AI. And plot twist, it's barely about the technology itself and it's about how you reorganize a company. His opening claim is that adoption and transformation are two different things that nearly every company has done the first adoption and that what holds companies back is inertia rather than the technology choice. And he's pretty blunt about it. He thinks leaders spend far too much energy picking a platform because today's leading model already does more than most companies ask of it. And the rankings change every few months anyway. He lays out three moves that he is using for Pfizer. One is structure the data. Pfizer has been running experiments since 1849 and his line is that failure is data, not a setback because a molecule that doesn't work teaches a model as much as one that does. My question for you here, Amith, is what do you think this move, if we want to call it that, looks like for associations?
[00:06:35:17 - 00:06:44:07]
Mallory
Associations are not quite pharmaceutical companies. So is it important, you think, for associations to have a log of their previous failures, quote unquote?
[00:06:44:07 - 00:09:00:05]
Amith
I think that keeping track of all the experiments that you run and knowing what worked and what didn't, as well as some insights about what may have caused the failure would be really helpful. If you tell me that you tried a certain project previously and it didn't work, it would be really helpful to know why. So for example, if you tried to do live streaming of your content from a conference 20 years ago, it was technically feasible already in 2006, but it was very, very technically challenging. It was extraordinarily expensive. So you might have said, well, it was great, but it was low bandwidth and it was glitchy and it was super low resolution. It was way too expensive for the value and not that many people tuned into it. But if you just had a data point that said experiment failed, I might not have tried it again five years later or 10 years later when the circumstances have changed and things have passed both externally and internally that the data on experiments in terms of the why, not just the outcome, is really valuable. And I think that's a lot of what Pfizer is talking about here is that a molecule may have not worked for a particular application, but maybe it actually could be useful for something else or maybe something else changed where that molecule could work in a way that we didn't think about it before. So it is directly applicable in my mind to what associations and all of us do in our businesses is to have good data on what we do. It's one of the reasons that culturally it's so important to get out in front of the idea that failing is bad. And so many associations have a deeply, deeply inculcated mindset that failure is unacceptable, that we only succeed here. And sometimes that's driven by board mandates. Sometimes it's driven by just years of time, not even the people that are there. But people need to buck that trend. If that's your culture, you need to basically create an environment. And if you're the top leader in the organization, this is definitely your job. But even if you're not, there's ways you can approach this. We talk about this actually in the Ascend book in terms of bottom-up transformation, being able to take small bets that are within your domain. But the essence of what I'm saying is to make failure an expected outcome, a good percentage of the time.
[00:09:01:17 - 00:09:07:11]
Amith
Roger Federer gave the keynote at, I think it was Dartmouth, maybe two years ago. And his
[00:09:08:24 - 00:10:00:09]
Amith
commencement speech is what I'm trying to refer to. His commencement was essentially saying that he failed, I think, on 49% of the points, or 47% of the points that he played. He lost them. He lost Roger Federer, one of the all-time greatest tennis players ever. And incredible. He's like, whoa, he's invisible. He has absolutely no emotions when he's playing. He's ice cold and keeps his composure. But in his head, he's not too happy about losing points. He's super competitive. And he loses almost every other one. So how does he come out on top? And a lot of it is looking at the data with clinical precision and trying to understand what went wrong, but not getting hung up on it. So that's a big part of it. And if you consider most successes, they weren't overnight and they weren't the first shot. They weren't the first at bat. So you have to look at experiments as an opportunity to learn.
[00:10:00:09 - 00:10:13:01]
Mallory
Now Albert uses the term structure the data, but we've talked about on this podcast before how AI is very good at handling unstructured data. Do you think that's just semantics? Do associations need to think about structuring their data? Or what do you think on that?
[00:10:13:01 - 00:12:22:22]
Amith
I think what he's really getting at is getting your data house in order is the term we like to use, which is to say that you need to get your data organized so that you have access to it. So Pfizer had theoretically had 150 plus years of experiments, but didn't know where some of it was or some of it was in people's handwritten notebooks and some of it was in SharePoint and some of it was somewhere else and they weren't quite sure. That would not be very helpful. They may actually have truly structured data in the sense of experiments and outcomes and structured data on molecules and things like that. I'm sure they have a lot of that. But a lot of the data is inherently unstructured. People's observations, which is just text or even images, right? Things that would be considered unstructured. So I think the move has more to do with organizing your data in the sense of having access to it. My advice to you as a nonprofit leader is to make sure that you have your data at hand. What I mean by that is yes, you technically own your data regardless of where it resides. Almost certainly that's true. You know, you have data in Salesforce, you have data in Microsoft, or you have data in Google in these different places. You do own the data, technically it is yours by contract. But as a practical matter, operationally, can you do what the Pfizer CEO is talking about and access it ad hoc at any time you want across all these different places? Usually the answer is no. Usually it's like, well, I have it in this silo and I can get this piece of information to get that. I have to export it. Then I have to go to this other place and get this other piece of data. And to get that, I also have to export it or maybe figure out how to do backflips enough times in the right order inside the platform to get it. And so yes, you have technical or legal ownership of the data, but you don't have operational control. I think that's a lot of what this move is about, is getting your data house in order. How do you align your data? How do you get it into a surface area that you have direct access to whenever you want ad hoc access to all of your data at any time? And then you can think about what the applications are and whether the data needs to be truly structured or if unstructured or semi-structured is fine, which to your point, Mallory, many times actually AI thrives on unstructured data.
[00:12:24:06 - 00:12:39:06]
Mallory
If you're listening to this and thinking, "Ah, Amith and Mallory, we have some ancient systems in our association," I would say be thankful you don't have notebooks from 1849 with handwritten notes. And maybe you do actually have those, but AI could probably, you could take some pictures, scan them, and AI can handle that.
[00:12:39:06 - 00:13:23:11]
Amith
You can go back in time to the mid 1800s and you will find associations thriving. They've been going back even a couple hundred years before that. But many associations have pre-digital, it might be other formats, but they have microfiche. They have things that remind people of days gone by in terms of access. And so yes, digitization of some of these things is something worth considering. So maybe you have stuff literally in boxes at a storage unit and you don't have these things digitized. And historically, maybe the cost of scanning and doing OCR to get those into a format was prohibitive. Perhaps that's not the case anymore. Perhaps there are services that you can use now that would make that a lot more palatable.
[00:13:24:21 - 00:14:02:08]
Mallory
So the first step was structure the data. The second move that Pfizer is making is federate. So instead of concentrating AI expertise in one central team of specialists, he pushed it out to the people doing the work with the center, owning the platforms, the data and computing power while leaders across the business own accountability. I mean, several associations have told us, some on this podcast, that they have a group or a committee of AI enthusiasts that spends more time with this technology than the rest of staff. Do you think this is not the right move? What could the idea of federate look like within a smaller organization, like an association?
[00:14:02:08 - 00:15:30:13]
Amith
I think that's a great point to zoom in on, Mallory, that Pfizer is obviously a very large organization. Most associations are a fraction of that size. Yet many associations have a hard time decentralizing decision making. Like, for example, one of the things he mentioned in the opening part of the piece is this idea of spending so much time on platform selection. Associations are used to saying, "Oh, there's a new technology. Let me do an RFP. Let me hire a consultant. Let me spend a year figuring out what I'm going to use and then go use it for 10 years or longer." And this stuff doesn't work that way. You can't make those kinds of bets. You have to move in a lot more agile way. And part of that is decentralization and giving more autonomy to teams. Possibly one tool for one department, like let's say your events department. Maybe they can really use Claude to its optimum level and they're getting incredible value, let's say, out of Claude Co-Work. The Claude Co-Work tool, if you're not familiar with it, is kind of the every person's agent where you can go into the desktop app or the web app and you can have Claude Co-Work actually take actions on your behalf. It can use your computer and it can access websites. And by the way, chat GPT has something similar to this called chat GPT work, which is basically the exact same thing. But the point is, is maybe Claude for some reason is fantastic for the events department. Maybe there's something specific that the membership department really needs to use and they prefer co-pilot for some reason. And there are people who prefer co-pilot.
[00:15:31:16 - 00:18:22:04]
Amith
They are out there. I have met them. So, but the point is, is that that tool may actually also be okay. And a lot of people try to simplify it and say, "Oh, we only allow one AI tool." The other thing that they do is they tend to do the usual play of saying, "Well, anything that you want to do that's more than just talking to a chatbot, you have to get central approval for." So I think the question is, is what is your policy around AI experimentation? What requires the sign off of the board or the CEO or some central committee that you mentioned and what can individuals play with on their own? The other comment I'd share is, and in my particular belief, I should also cap that off with is, I think it's really important to have a lot of autonomy at every level, particularly the department level. But even if the individual level, and that kind of aligns with this idea of federation. But the other thing about federation is that it's not the same thing as abdication. So what I like to say is that delegation, or in this case, federation, is the granting of power that's decentralized. But it comes with accountability. Whereas abdication is when you essentially say, "You know what? I don't care, whatever. I'm just going to let everyone do what they want. I'm not going to bother checking in with them. I'm not going to hold them accountable." You might be saying, "Well, that's ridiculous. Why are you even talking about abdicating?" Well, I call it that because that's what it is in many organizations where people say, "Well, you know, we're going to let our teams do what they do." It's like, "Okay, well, what are they getting done?" And the answer is usually, "Well, I don't really know." Or, "It's not a whole lot," or, "Not what I hope they'd get done." It's like, "Well, how are you holding them accountable? How are you creating clear alignment so that they have clear business goals that they're supposed to be pursuing?" So the operational decisions to use tool AB or C, or the operational decisions to implement an agent or not, those are fine to delegate if you have alignment with what the objectives are for the overall organization. And many organizations have, I would say, a lack of clarity or a lack of specificity. They might have a strategic plan that has five pillars and they'll say, "Oh, we have a high degree of alignment because we have these five pillars." But each of those five pillars might have 35 bullet points underneath them. And so people really get lost in that sea of potential priorities. And oftentimes those types of plans aren't even looked at other than the board meetings to say, "Well, actually, how do we do? Oh, well, let's go back and look at the strategic plan and let's fit our work back into the categories that the plan called for." That's an ineffective way of running an organization, in my opinion. And it ultimately leads to the inability to advance AI or any other type of change initiative. AI is just a human thing much more than it is a technology thing, just like the Pfizer CEO said. So I think that's the essence of Federation, in my mind, is it is the ability to push down that responsibility, but to hold people accountable at the same time.
[00:18:22:04 - 00:18:36:11]
Mallory
Okay. So having a committee on your association of your most enthusiastic folks about AI is not necessarily a bad move, but if that's the only place within your organization that conversations around AI are happening, that's probably not ideal.
[00:18:36:11 - 00:19:37:03]
Amith
Yeah, because part of what it's signaling, it depends on the nature of the committee. If the committee is there to help move things faster than a single person might, that can be helpful. It can also be a step in the right direction compared to just a single all-powerful CIO or someone like that who historically held the keys to all technology decisions. So I think it can be helpful. It can be really helpful if the charge of the committee is to educate everyone on the team or to help educate everyone on the team. So they're not sitting high up in the sky as the committee on AI that hands down decisions, but rather they're actually actively working with the rest of the organization. And each person on the committee has a responsibility and is held accountable for activating certain people or certain groups of people in the organization. That would be a really fantastic way to use a committee. But I think if the committee is the sole source of authority in terms of choosing tools or how the tools are used, that is a choke point.
[00:19:38:15 - 00:20:13:11]
Mallory
A good segue into their last move. So we talked about structuring the data, the idea of federation, and finally fluency. So CEO says Pfizer launched an AI certification program in June covering every role and function worldwide, including a course developed with Indiana University and that 98 percent of eligible colleagues have now taken the first course. And self included. I mean, do you think the move of fluency, if we're going to put that in a category, needs to be done first before the other moves, before structuring your data, before federation, or can they all kind of go hand in hand?
[00:20:14:19 - 00:24:49:14]
Amith
I think that it needs to be at a minimum in parallel with the other activities. But I actually think that it's one of the first things you should do, especially if you're, you know, contemplating, well, how should I organize this? Should I do federated model? Should I do a little more central control? What do I need in terms of board approvals? All these other things associations have to grapple with, which you definitely can get agreement to in basically every organization is that education is a good thing. Most associations believe that deeply because they themselves are educating bodies in their professions. And so being focused on educating your team, gaining AI fluency and fluency isn't a term to just throw around, right? You can study, for example, a foreign language for decades and not be fluent. I'm a great example of that with Spanish. I can barely speak it even though I spent 10 years in school learning Spanish. Can't speak it. So I'm not fluent. I'd love to be fluent, but fluency requires a lot of effort, a lot of time, and that requires resources of the individual to spend time ultimately. But it also, in the case of AI training for your organization, requires fluency skills that come from either a resource that you buy or something you develop. There's lots of ways to approach it. But fluency is so incredibly critical. Put another way, if the goal is for everyone on the team to be able to speak in this other language, if you will, the absence of fluency means people can't participate in the conversation. So if you are not fluent in this new language of AI of AI land, right, and you want to participate in this transformed world of AI land, how do you expect to be an active, robust, fully involved participant if you are not fluent? And if you are the king or the queen of that new AI land, and you are the one person who doesn't know anything about AI, how do you expect to be useful? How do you expect to lead with vision if you yourself don't understand the technology? That's the beauty of what the Pfizer CEO also pointed out is not only did 98% of people take the course, but the CEO was one of the first people to take the course himself. I get a lot of calls from CEOs of associations Mallory and I've talked about this in the pot a bunch of times. And they're always asking me essentially the same two or three questions. One of the questions is always about how to talk to their board about being ready for AI and making investments. That's actually a lesser concern now because most boards are actually saying, "Hey, we need to do stuff with AI." But that was a conversation a couple of years ago, much more frequently. The next one is how do I get myself and my team more further along in our understanding and awareness of this? And the third one is how do I get actual business value out of the technology? Like how do I go from a bunch of experiments to actually seeing an impact? I'll come back to the third one. But the middle one is really all about this topic of fluency. If you're not focused on fluency, like if you're just doing a checkbox exercise of saying we're doing a simple AI training, bam, everyone's been trained on AI. You're really harming yourself. It's not just about completing one course or reading one book or listening to one podcast. It's a journey. You have to commit to it. And like with a language, a human language, if you don't keep up with it, you lose the skill. The skill decreases over time. And unlike a human language, which changes fairly glacially, AI is changing so fast that you have to keep up with it to have a chance of understanding the next dialect, which will emerge in weeks or months. So I do think there's a really key comparison to language fluency when we're talking about working with AI. So to me, it's a critically important thing. When people ask me the question, what should I do first, which is the other thing I usually these calls usually end with or start with, I usually tell them education. And granted, I'm obviously biased as one of the leaders of Sidecar. And I think deeply that Sidecar's AI education is great for associations, et cetera. I always tell people, look, Sidecar has all these great free resources like the pod. We have hundreds and hundreds of blog posts. We have free webinars we do every single month on intro to AI. And yes, we have some paid offerings like the AIP certification and so forth. But it doesn't matter to me what you use, ultimately. My mission and my message are the same, which is you have to become fluent. Obviously, we're here to help with that. But the point is go figure it out and become fluent in AI, because that is, in my mind, the most important single thing that you do. Because the absence of fluency almost is a blocker to the other two things that we just talked about.
[00:24:50:24 - 00:25:44:07]
Mallory
Yeah, I feel like I just had a light bulb moment, Amith, with comparing AI fluency to foreign language, which feels very obvious because of the term fluency itself. But thinking the experience you shared is not unique, which is people taking years and years of foreign language class in high school or through college and not actually feeling fluent in the language. And I was thinking through why is that as someone who spent a large portion of my life trying to learn Spanish? And I think the missing piece there is when you're learning a foreign language, you have to at some point throw yourself to the wolves. You have to go to the country. You have to be uncomfortable. You have to be so confused for so long. And then it's like all of a sudden things start to click together. And I feel like AI is no different. You can take all the courses you want, but until you throw yourself to the wolves, in theory, whether that's an experiment or a pilot you're scared about. Or confused about, I feel like on the other side of that is when you can achieve fluency. Do you agree?
[00:25:44:07 - 00:27:28:18]
Amith
Totally. And yeah, so that like in a way, part of what we're talking about is, you know, you're dipping your toe in the water by taking a course. But to jump into the deep end is what you're describing of going to the foreign country and immersing yourself. And that truly is the way that you have to learn artificial intelligence. And the way you do that is by using it in your day to day work and not just for toy examples, but to actually use it. A lot of people who listen to this podcast are saying when they hear this, yeah, I'm way past that. Like I use AI every day. I use Claude and chat GPT. They're my best buddies. I'm in those tools all the time. And to them, I would say that's awesome. And now the next question is, is how can you get the AI to do more of your work for you without you asking for it? So a lot of users are in this, you know, ask and ask and, you know, get response kind of mode where they're going to the chat bot and they're getting a response. And it's an increasingly good response, right? You go to Claude and you say, give me an email I can send to my board. Give me a pitch deck for this meeting. Give me whatever. And the magical AI gives you something pretty damn good, which is awesome. But then the next thing is, is okay. Well, how can I stop having to ask, right? What can I do to drive automation to have it just do the tasks, either that I review and approve or maybe ultimately for certain things. I mean, I have to review and approve certain activities, right? If we get to that confidence level, you can't get to that point if you haven't done the first thing. If you haven't had lots of those conversations, you're not going to have the confidence in the language, so to speak, to do the semi-automated or fully automated approach, which is really where all this stuff you hear about, agentic AI. That's really what I'm describing is agentic AI is AI that can take action on your behalf. People are not going to be comfortable or even understand how to use AI agents if they don't understand what the AI can do from a capabilities perspective.
[00:27:30:05 - 00:28:43:22]
Mallory
In that essay from the CEO of Pfizer, he said that anyone can license a model, but nobody else has Pfizer's data, which is a good segue into the second topic for today's episode, which is that Thomson Reuters just spent $40 million acting on that exact idea. On August 24th, they launched their own large language model called Thomson 1.0. Thomson Reuters owns Westlaw and Practical Law, the legal research databases most American lawyers effectively live inside, and they didn't build it from scratch. They started with Alibaba's Quinn 3.5, an open-weight model, meaning the maker publishes the actual parameters so anyone can download it and keep training on top of it. Then Thomson Reuters specialized it on their own legal content with their own experts reviewing the output. The final training run cost about $450,000 inside that $40 million two-year total, counting the acquired team and the compute as well. Frontier Labs as a note spent hundreds of millions on a single training run, so pretty impressive, but $40 a meath is not chump change with the audience that we're talking to. Is there a version of what Thomson Reuters did that could be feasible for associations that's not, you know, training your own model?
[00:28:43:22 - 00:29:47:07]
Amith
Yeah, I think what's happening is so we go back in time a little bit to something that a lot of people have beat up, you know, pretty extensively, which is the Bloomberg custom model that was around the GPT-4 era. So it's between GPT-35 and 4. Bloomberg had Bloomberg GPT where they came out with their own model. I don't believe they started with an open source layer at the time. I don't think there was anything suitable for that. So they built their own model from scratch and it by the time they had it done, GPT-4 or GPT-4 turbo came out and it was better. And it was better even though Bloomberg had all this proprietary data. So this is a great example of something that didn't work three years ago, maybe works now or maybe doesn't work still, right? And there's different reasons. Like why didn't it work? Well, Bloomberg certainly has a treasure trove of proprietary data that is truly unique to them. Just like Thomson Reuters has in the legal space and Pfizer is one of several firms that has that scale of data in the world of drug discovery.
[00:29:48:10 - 00:33:06:03]
Amith
So what is it that we're talking about here? Ultimately, let's first zoom out and say like, why do people want to do this? Like, what's the point? Ultimately, they're trying to gain competitive advantage in some way by having an AI tool that's smarter in their domain, again, smarter in their domain, not in everything than anything else in the world. So their domain being in the world of Thomson Reuters, legal research, you know, Thomson and West, you know, they have one of the best legal databases that most lawyers use. They have the capability to really answer just about any legal question through their classical like software tools. And the data also includes behavioral insights. What are people searching on? What, which articles are good responses to certain queries, things like that, which all of that feeds back into how can they make a tool that's even better? How can they compete with their competitors in that space and be an indispensable aid to the legal sector? So the outcome is to provide a competitive advantage. And so the idea would be then they have to believe that this effort will produce something better than Claude Fable, right? Or a GPT 5.6 sole or whatever comes next, which is a pretty high bar. So you have to believe that your data is this unique ingredient that will produce, you know, out of band performance for your domain, a.k.a. something significantly better than what the generic models can do. And so I think this is a possibility now with the idea of reinforcement learning post-training or RL post-training, which is a lot of what they're doing in this type of experiment. And essentially what that involves in is taking, you know, giving a model post-training where you're essentially saying, hey, I have the base model, whether it's QWEN or something else. And you're feeding it essentially questions and responses and giving it good feedback and bad feedback. And through that process, you're giving it lots and lots of domain specific examples that essentially tune the model to be better at your type of content, your type of questions. And it's effective. It's one of the reasons that you see the large labs going out there and spending billions of dollars getting experts in a variety of fields to answer questions in those domains. And so Thomson Reuters, you know, these other firms, they have this data, associations have this data as well. So coming back to your question, Mallory, the way I look at it is how do you create some kind of a durable, competitive differentiator in your domain using your data as one of the potential cornered resources? And so I do think this is something worth pursuing. I don't know that training a model or post-training a model is necessarily the right tactic to use. It could be. It depends on the data that you have. You could actually just use an agentic layer on top of a standard model and get really good results potentially. But the short answer is it depends. And it also depends on the type of workloads that you want to throw at the tool once you build it. So for Thomson's purposes, it was probably a combination of both the data and the type of data they have, but also the use cases and the volume of use. They probably thought that they would be able to save a ton of money on inference. Aside from quality, they need to be able to do this at enormous scale and do it cost effectively.
[00:33:07:14 - 00:33:10:12]
Mallory
We have a company in our family of companies called Betty. And
[00:33:11:18 - 00:33:21:06]
Mallory
I'm sure there are some other companies out there in the association space at large. But can you explain how how a knowledge assistant is different from what Thomson Reuters did?
[00:33:21:06 - 00:36:40:08]
Amith
Sure. Well, a knowledge agent can have a custom model as part of its architecture. So if you have a custom model, a knowledge agent can leverage it as one of its tools. But what a knowledge agent does is it takes typically multiple different models, not just one. And it uses this multitude of models to both search for content that's relevant to what the user is requesting and to then reason over that content to form hypotheses, to test the possible answers and to ultimately come back with reasoned and grounded responses. That is what Betty does specifically. Not all knowledge agents or so-called knowledge agents do all of this. Some of them do very simple retrieval and what's called RAG. Some people are familiar with the term R-A-G or RAG. It stands for retrieval augmented generation. And it's a fancy tech term that simply means that before answering question, the AI is essentially there's a system that searches for content that's similar to what the user is asking and then injects that content into the model's context window or gives it that gives it like chunks of content. And then the model answers better. And that is actually how Betty worked three years ago. The very simple early versions of these RAG tools worked quite well, but very, very quickly you run into a performance ceiling with that because you can't guarantee that the results are truly grounded. So it makes a really good knowledge agent really good. And to get it to like 99.99 plus percent accuracy, there's a lot more engineering involved. You have to do a lot more systems level thinking. You have to have the ability to do things like even write code and to test the outcomes that the agent's providing. And so a tool like Betty doesn't provide you answers instantly. It's more like chat JPT research where it might take 20, 30 seconds or sometimes minutes to give you a great answer. But it sits on top of the association's knowledge as the core fuel for making all of that possible. So knowledge agents have evolved dramatically, but they're also incredibly accurate now. And they're also very capable like Betty's new release that was just unveiled at ASAE conference earlier this month in August the 26th. Showcase full multimodal reasoning. So the ability for you to provide images and get images back. Also the ability for you to have it generate images as well as to reason over audio and video. So you can talk to Betty, you can get audio responses, you can also do all sorts of different things like that. So the idea here isn't so much this being a Betty commercial, but the point is is that systems like this are getting increasingly advanced. And if you have a reason to think that a custom model under the hood would improve a tool like Betty, you can plug it in. If the architecture is sufficiently flexible, which Betty is, you can plug in any model or any set of models for the way the agent architecture works. So they're not necessarily mutually exclusive decisions. You may choose to build ultimately an agentic knowledge worker like this that's available for your members. But you might also decide that you want to have some kind of fine tune or post train model like Thomson Reuters did.
[00:36:40:08 - 00:36:59:23]
Mallory
Okay, so not necessarily an either or. I felt like that was a helpful explanation. As a whole, how do you think association should be thinking about domain specific AI? Is it something they need to be a part of? Do legal associations need to come out and vet this Thomson Reuters model? Or how do you think association should be thinking about that?
[00:36:59:23 - 00:39:09:03]
Amith
In my mind, associations are the business of connecting people that have similar interests, helping enrich their journey, whether it's personal or professional. Oftentimes, people who are listening to this are in some kind of professional society or industry association. And then to ultimately improve the quality of service in that sector. So providing education to upskill people in a given space, be it a field of medicine or law or accounting or architecture or whatever the field may be. And so all of these endeavors require really good knowledge. And the association often has been unique in providing hyper specific content in their domain, which may be like a very narrow subfield of medicine or a particular geographically scoped area of real estate where they have the best content in the world. In a particular county, in a particular state, right? Something like that. And so when you have this hyper specialization, the value you provide to people in your very narrow niche is extraordinary. That's why you exist as a business, right? That's why people come to you. So AI enabling that resource is extremely important because if you don't, you have a non-AI resource. You might have the best content in the world for your county in the state of Arkansas for real estate. But if you don't make that really, really powerful from an AI perspective while also protecting your IP, by the way, which is a whole other element of this conversation we can touch on. But if you don't do that, then chat, GPT or Claude or Gemini will probably be a better tool for most of your members. Because even though it doesn't have your hyper specialized content, it's way easier to use and it can reason over its more generalized content. And even if it doesn't have your exact data, it can do a lot of things that you cannot do. So you have to be part of that conversation or you really will. I think you will find your relevance fading extremely quickly if you don't provide AI enabled services to your members. And a lot of that is about leveraging your domain expertise.
[00:39:10:13 - 00:39:16:05]
Mallory
You said you could briefly touch on protecting your IP. What is the short and sweet on that, Amith?
[00:39:16:05 - 00:39:54:09]
Amith
Short version is, do not load up all of your data and back up that truck to Sam Altman or Dari Amade's loading dock and say, "Here you go." Don't do that. That would be the equivalent of what a lot of people unfortunately are doing, which is saying, "Oh, I have a custom GPT and I just loaded up a thousand PDFs in there and I have a custom GPT. That's awesome." You literally just gave all your content to the firm like OpenAI or Entropic. Now you might say, "Oh, but Amith, I have an agreement with them or enterprise agreement says they can't do training with my content." To which my response is, "Okay, but what if they change their minds and decide to use your content? What enforceability do you have?"
[00:39:56:03 - 00:40:58:18]
Amith
There may be situations like that unfolding right now. I'm not accusing anyone of anything, but I would rather have a clear separation of responsibility between the inference provider I use, which is the person that runs the AI system for me, and where my content is held. There's ways to architect systems so that you don't have to back up that truck full of your data directly to the hungry mouths of the AI model trainers. I just think it's a good idea to have that separation to protect yourself while still doing exactly what we talked about. Now, this is not as easy as saying, "Load up all my documents and a custom GPT." That's the easy button. It's also the button that potentially could have issues for you down the road. So there are ways to approach this. It's a scope far beyond this conversation. Maybe that's a podcast episode. If you think it would be interesting to have a discussion and take a deep dive into how to protect your intellectual property while using frontier AI capabilities, Mallory and I would be happy to do an episode on that. So ping us, drop in comments, and we'll consider that as a topic.
[00:40:59:19 - 00:41:02:16]
Mallory
Don't feed those hungry AI mouths. I like that analogy, Amy.
[00:41:03:17 - 00:42:11:18]
Mallory
Our last topic for today is OX Alpha. When a lab puts a model out without its name attached, that's called a stealth release and it's deliberate. They want to watch how developers actually use it before the branding and the benchmark charts shape everyone's expectations. On August 20th, a model called OX Alpha appeared on OpenRouter, a marketplace that lets developers reach a lot of different AI models through one connection, free, about a million tokens of context, which is roughly a 700,000 word document. And it was pitched at coding and long running agent work. Developers went straight at it. OpenCode reported 42 trillion tokens ran through it in six days. It went to number one on OpenRouter's leaderboard, more than doubling DeepSeek's usage, which we've covered on this podcast before, and it became the biggest launch in that marketplace's history. Stripe CEO Patrick Collison, whose company is acquiring OpenRouter, called it very impressive and for basically a whole week, nobody knew who made it. Now, as of this morning, we do, but before I reveal it, Amith, have you tried this out and do you think it's safe to test out stealth models? Is there any concern there?
[00:42:11:18 - 00:42:34:18]
Amith
I have not tried out this model. And as far as the general safety for very simple things and where you're not providing proprietary data, sure, go try them out. Go check them out on places. You could use OpenRouter, you could use a coding tool like OpenCode connected to OpenRouter. There's usually a number of ways to get at these early models.
[00:42:35:21 - 00:42:45:15]
Amith
But don't drop in content that you wouldn't want to share with the world. Basically, throw at it things that you would have no problem with anyone on planet Earth seeing.
[00:42:46:15 - 00:42:56:19]
Amith
You can still test its capabilities just fine that way. But yeah, definitely be thoughtful about where your data is going when you're not even sure who made the model. And more importantly, even where the model might be running.
[00:42:58:04 - 00:43:01:03]
Mallory
Well, Amith, do you know as of now who this model belongs to?
[00:43:01:03 - 00:43:03:16]
Amith
Well, I did read ahead a little bit on the show.
[00:43:03:16 - 00:43:04:11]
Mallory
You skipped ahead.
[00:43:04:11 - 00:43:11:15]
Amith
I did. I have a bad habit of doing that. But I mean, honestly, prior to about 30 seconds ago, I did not. No, I did not know.
[00:43:11:15 - 00:44:15:13]
Mallory
Well, I do forgive you because you're allowed to read the outline ahead of time. I don't want to prevent you from doing that. But Bloomberg broke the answer this morning and its ZAI, a Chinese lab and OX Alpha is a new version of their GLM model. They said they'd release the weights, which means anyone can download it and run it themselves. ZAI also shipped its flagship GLM 5.3 model earlier this month, which the company says rivals Claude Fable 5 on some benchmarks. None of this is new as a tactic. Just as a note, in April 2024, a model called GP2 Chatbot turned up in a public testing arena. The guessing went on for days, though now as I'm looking at the name for this, I feel like it's pretty obvious that it turned out to be OpenAI's model. What's different now is how little the name seemed to matter. Developers moving 42 trillion tokens through a model with no brand, no model card and no pricing on the strength of how it performed in their own tools. So, Amith, I don't think we've done a full topic on the GLM models. Maybe we have, but what are your thoughts on those as a whole?
[00:44:15:13 - 00:44:22:20]
Amith
Well, first of all, I have to say I love the idea of using animal names for some of these models. So OX, that's pretty cool.
[00:44:23:23 - 00:45:08:04]
Amith
It's a very powerful animal. In fact, a little side story. One time, about 15, maybe 20 years ago at this point, it's been a while, I was in India and my former company was launching an operation in India. And we had just hired this awesome team. I think it was maybe like eight or 10 people. It was just getting started. Our plans were to grow it. By the time I sold that company, we had over 150 people in that office. But we were starting off with just a small group of eight or so developers. And so I flew over there to meet this team. I had other people in the company had gone over there to set things up. So this is my first trip. And I go out, spend a week, had an awesome time with this team. I think on the middle of the week, like Wednesday, they're like, hey, we want to go do an activity with you. This evening, I'm like, this sounds awesome.
[00:45:09:07 - 00:45:52:08]
Amith
I am half Indian. And so in theory, I'm supposed to know something about India. Sadly, I did not learn much about India other than a little bit about the food from my dad when I was growing up. But this was my first trip to India. And so I went with this team out to this. I don't really know how to call it other than kind of like an amusement park of sorts. But, you know, it was in a rural area and the rides were considerably different than what you might expect at Disney. For example, there was a ferris wheel, but this ferris wheel was powered by a human who sat at the hub of the ferris wheel using his legs to push it and was incredibly good at it. In fact, everybody looked like they were having a really good time on it. I did not go on it myself. I checked out. I wanted to, but I did not go on it.
[00:45:53:10 - 00:46:12:11]
Amith
I have a video of this. It's just truly remarkable. But it made a lot of sense. It's just basic physics to get it going. And once they got it, it took, I think, two people to get it moving. And then after that, this guy like jumped between these moving like, you know, spokes of the wheel and jumped on the hub and used his legs to power the continual motion.
[00:46:12:11 - 00:46:13:10]
Mallory
Wads of steel.
[00:46:14:12 - 00:46:36:02]
Amith
Yeah, this he looked pretty skinny, but he was obviously extremely strong. But the thing that I saw that was really impressive there and it scared the hell out of me was we just hired this group of people and walk into this theme park and they're like, hey, first thing we're going to do is we're going to jump on this ride over here. And I'm like, okay, what is it? And they, they all jump in the back of this wooden cart.
[00:46:37:05 - 00:46:53:24]
Amith
And I didn't see it at the time, but there was an ox in the front of it that the guy who was in the front of the cart, he snapped a whip at it. And this ox just took off and there was these six programmers we had just hired in the back of this cart. I'm like, Oh wow, this is going to be really good. Productivity is going to be decreasing.
[00:46:55:12 - 00:47:01:15]
Amith
That's what I think about when I think about oxes. They're very powerful. And if I'm going to do an AI model named after that experience, I will remember.
[00:47:03:11 - 00:47:09:15]
Mallory
I did not know where that the whole time I was like, I really don't know where the story is going, but I got there right before the ending.
[00:47:09:15 - 00:48:50:23]
Amith
And hopefully marginally entertaining for our audience to be with me on that little little flashback. But you can call it AI hallucination. I have similar things occur in my brain. But on your question of GLM, I don't know what GLM stands for actually, but I'm very impressed with their models. GLM 5.2 was fantastic. 5.3 is even better. I guess this must be 5.4 or 6.0 or whatever it is. But it's at the caliber of the frontier of AI capabilities from the best closed source labs. So I'm a big fan of their advancing of the open source ball. They're right there with and probably ahead of deep seek. They're right there with the people from Moonshot, which is the makers of Kimmy. Some I'm definitely in favor of using their models. I would also say that any model that's not produced by a company in your home country, I would suggest that you find an inference provider for that model in your home country. So in the case of ZAI and the GLM models or the Kimmy models or the models from deep seek, they're totally safe to run. So long as you run them with an inference provider that is located somewhere and running a data center somewhere you trust. For me, that would be anywhere in the US or Canada. I'd be fine with a data center running workloads that I'm running because the privacy standards and security standards and so forth. I find compelling in those countries. Not to say that other countries aren't fantastic at this, but you do need to be thoughtful about where the model runs because that is where your data is going. Your data does not go back to the home of the developer of the model. It just goes to wherever the inference is happening. So short version of the story is double thumbs up on ZAI and the GLM model series. They're definitely worth considering.
[00:48:52:02 - 00:48:58:11]
Mallory
And Ox, Oxen are very strong and so models named after them are probably strong too, I think we can say.
[00:48:58:11 - 00:48:59:13]
Amith
I think so.
[00:49:00:16 - 00:49:14:07]
Mallory
Amith, what do you think is the takeaway from this episode? We talked about Pfizer. We talked about the moves that they're making within their organization. We talked about Thomson Reuters and their own custom model and then finally talked about Ox Alpha. What's the takeaway here?
[00:49:14:07 - 00:49:22:17]
Amith
Well, if you just invested a bunch of money in training a team on how to learn AI, don't load them up on an Ox cart because that might be a little bit risky.
[00:49:22:17 - 00:49:24:21]
Mallory
What's the liability Amith? I can't believe you did that.
[00:49:24:21 - 00:50:43:08]
Amith
I mean, I was not complicit in the planning or the consent. I was not given by me for this to occur. I just was simply witnessing it and pretty shocked. But retrospectively, it's a great story. I was a little bit nervous at the time. My takeaway Mallory in all seriousness, though, is this, is that when you think about the leaders of some of the best resourced and most powerful companies in our economy talking about the basic, simple things that they're focused on to make things work for them. When Pfizer CEO says it's not about the technology, it's about empowering people. It's about training people. It's about organizing your data. Those are the same fundamentals that apply to a two person association or a 200 person association. So while historically you might have said, well, Pfizer being able to do that, you know, they're big, they have tons of money. It's not about money. It's about priorities. And Pfizer CEO is simply laying out that they decided to, number one, make this a priority. And then once they've made it a priority, they narrowed it down to three very straightforward things that are simple conceptually, but hard to execute consistently. And so to me, the big takeaway is it's not about the tech. It's never been about the tech. It is about focusing on the things that will drive the outcomes.
[00:50:44:10 - 00:50:57:23]
Mallory
And to piggyback off that, I mean, as you said, I think it's really important to master the basics with this stuff when it comes to AI fluency. And then once you master those basics, that's when you can throw yourself out to the wolves.
[00:50:57:23 - 00:50:59:01]
Amith
Totally. And
[00:51:00:09 - 00:51:10:08]
Amith
let's just simply say the basics are not the same as they were six months ago. So checking the box, fine. But just remember, it gets unchecked on your behalf quite quickly.
[00:51:10:08 - 00:51:15:21]
(Music Playing)
[00:51:26:14 - 00:51:43:13]
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.
[00:51:43:13 - 00:51:46:19]
(Music Playing)