Summary:
Meta and NVIDIA are turning up the heat in the open-weight AI race, while Anthropic is taking a very different approach by watermarking Claude-generated content. Amith Nagarajan and Mallory Mejias unpack what these moves mean for associations, from choosing models and inference providers to avoiding lock-in at the agent harness layer. They also explore why AI models are becoming increasingly commoditized, where proprietary models may still have an edge, and why Claude’s new watermarking strategy probably isn’t a foolproof solution for detecting AI-generated work. Plus, Amith shares lessons from Blue Cypress’s latest AI hackathon, including how associations can use focused offsites to accelerate adoption and generate new ideas, and the hosts discuss a more productive approach to AI-generated submissions: using AI to evaluate quality, relevance, and originality rather than trying to ban it outright.
Timestamps:
00:00 - AI Dog Training?
02:51 - The Blue Cypress Hackathon
08:06 - How Associations Can Run Their Own Hackathons
12:52 - Meta Returns to the Open-Weight AI Race
18:35 - NVIDIA Pushes Deeper Into Open AI Models
24:04 - Why Agent Harnesses Matter More Than Models
27:14 - Open Versus Closed Models and Preserving Choice
32:29 - Anthropic Watermarks Claude-Generated Content
36:38 - Why AI Detection Remains Unreliable
40:10 - Using AI to Vet Association Content and Submissions
43:52 - Closing Thoughts
👥Provide comprehensive AI education for your team
https://learn.sidecar.ai/teams
📅 Register for digitalNow 2026:
https://digitalnow.sidecar.ai/digitalnow
🤖 Join the AI Mastermind:
https://sidecar.ai/association-ai-mas...
🎀 Use code AIPOD50 for $50 off your Association AI Professional (AAiP) certification
📕 Download ‘Ascend 3rd Edition: Unlocking the Power of AI for Associations’ for FREE
🛠 AI Tools and Resources Mentioned in This Episode:
Claude ➔ https://claude.ai
ChatGPT ➔ https://chatgpt.com
Gemini ➔ https://gemini.google.com
Grok ➔ https://grok.com
MemberJunction ➔ https://memberjunction.org
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 - 00:00:25:07]
Amith
My name is Amith Nagarajan.
[00:00:25:07 - 00:00:27:14]
Mallory
And my name is Mallory Mejias.
[00:00:27:14 - 00:01:09:14]
Amith
And we are your hosts along with Winston today. Winston is not on camera and he probably won't hear him, but he is my six month old golden retriever. And Mallory and I were just chatting before we started recording the show that he is so relaxed that it's kind of amazing at six months. So, and it's not an AI thing. I mean, AI did help me Mallory to find an awesome dog trainer here in the New Orleans area. And they did a fantastic job with both Winston and Lucy, who's our one year and three month old, four month old Cavalier. But they're just different dogs now. So it's quite amazing. We've had to do a lot of work since we got them back. But
[00:01:10:21 - 00:01:15:02]
Amith
AI did help me find a great trainer, but since then AI has not been part of my dog's journey.
[00:01:15:02 - 00:01:27:09]
Mallory
You said you told me you used Claude Cowork to find a good place to send him to learn some manners. Did you just give it a simple prompt and say, you know, research the best dog training places? Do you tell it to look at reviews? How do you do that?
[00:01:27:09 - 00:02:20:13]
Amith
I had it go and dig pretty deep into Google reviews and do a whole bunch of other stuff and it's spent, I don't know, Claude Cowork probably spent maybe on the order of 45 minutes or something doing research. And this is before Fable was available to the rest of us. I think maybe, maybe Anthropic had Fable at the time, but it was, it was Opus 4.8, I believe, that did all the work and got back a great report and clicked on all the links to see if these were real dog training companies, which, which they were and found one that just appealed to me. And I made a few phone calls, talked to a few different people and found one that I thought was kind of philosophically aligned with the way I think about dogs and what I was willing to do. Because, you know, different dog trainers, it's, it's like education for kids. There's lots of different philosophies from Montessori to more structured curriculum to everything else. So, dog trainers also have a wide diversity.
[00:02:20:13 - 00:02:26:18]
Mallory
Yeah. Well, I mean, co-work, Claude Cowork, good for your association, but also good for training your dog potentially.
[00:02:26:18 - 00:02:51:17]
Amith
Yeah, seriously. And, and actually, Mallory Winston had his first airplane ride and he ended up in Utah last week for our hackathon. And so he was incredibly well-behaved. He just was basically, you know, a format for the whole thing. So pretty cool. Um, yeah. So, and the hackathon is AI related, so we should probably maybe talk about that a little bit, but it was, it was good fun. We got a lot of, a lot of great things done there.
[00:02:51:17 - 00:03:00:08]
Mallory
Okay. Can you tell us anything new that might have come out of the hackathon or any problems that were resolved that you had been thinking about for the past few months?
[00:03:00:08 - 00:05:14:18]
Amith
So those of you listeners who know me know that I am pretty well known for kind of coming up with all sorts of different ideas. Um, but I tend to be somewhat of a, an agent of distraction within my own companies, at least my team has told me that for decades. And so I think a lot of entrepreneurs can kind of nod their head in agreement if, if they're kind of reflecting well enough on themselves to realize that's usually what entrepreneurs are pretty good at. Um, and so what we realized over here at Blue Cypress over the last couple months is, uh, we've been making a ton of great progress with clients. We've been making progress on core technology, really innovative stuff around real time agents, um, really, really amazing. Yet at the same time, we weren't investing enough in ourselves. And so, uh, we decided that we no longer would want to be the cobblers kids with no shoes, as they say. And I've, I've kind of done that to my companies over a long period of time where I invest far more in our client relationships and far more in, um, technology that the world uses, but then don't put enough effort into implementing it for ourselves. And Mallory, as you know really well, uh, AI transformation, it's part technology, but it's really way, way, way more about all of us people adopting the technology and using the technology and, and a lot of other pieces. So we spent the entire week with a group of 11 folks doing nothing but focusing on Blue Cypress's internal tech and our internal AI agent adoption. We made a ton of great progress, really, really exciting stuff. So nothing, you know, like interesting probably from the external lens. Uh, but you know, it's just really AI agent adoption within Blue Cypress, pushing really hard, working across different kinds of functional areas, not just technical, and we made a ton of progress. So we have a whole quarterly focus this quarter to go live on a new transaction processing system for our back office, uh, a new marketing and go-to-market agent suite that is going live in the next few weeks, um, a brand new learning experience platform for our sidecar, uh, AI learning hub. Uh, community, which is, you know, about 10,000 people now that are, have active access to the sidecar learning hub. It's, it's come a long way. It's as in it Malice, the days when you and I recorded those first few videos to get a kick started in what was it early 2024 maybe.
[00:05:14:18 - 00:05:23:05]
Mallory
Um, 10,000. I didn't know that number that it has certainly come a long way. Very exciting to see so many people in our community, um, wanting to learn more.
[00:05:23:05 - 00:08:06:23]
Amith
Totally. And it's so cool. And so we've got, you know, we've actually, I think it's over 10,000 active learners there. And we've had a couple thousand people complete the AIP certification, which is, you're not familiar with it. It's the association AI professional certification, uh, that we, um, you know, have had in the market now for about 18 months, roughly a little bit longer than that, and it's just taken off. And, uh, we've got all those folks that are in the community are actively pursuing it. So we expect that number to grow immensely between now and the end of the year. Uh, but we, we felt we needed a new learning platform and we couldn't find one that we liked and we looked at, we looked far and wide. We looked inside and outside of this vertical and couldn't find a learning platform that was truly AI native, that was truly focused on a one-to-one relationship with the learner with extraordinarily deep personalization, with a focus on a genteck interaction, where essentially the learning platform becomes your one-on-one tutor, Mallory. And so that's what we're deploying, um, a little bit later this year. Uh, we're going to deploy a new learning experience platform. It's going to replace our current learning management system, uh, completely to end, and it's going to be totally personalized to the learners. So the learner will onboard just by having the conversation either in audio or they'll, uh, they'll type whatever their preferences. Uh, and then the learning platform will essentially adopt the individual and, and kind of take them on a journey. We of course do have structured learning paths for the AIP and so forth, but an individual can kind of create their own adventure. And then as they go through this process while they're watching videos and, you know, the interactive things, uh, to learn, um, Sid, which is the interactive, uh, AI agent we have, um, that, that agent is always available to be your conversational buddy and also to tutor you. In fact, Sid can pop up a whiteboard and start illustrating concepts that you don't understand and do a lot of other cool stuff. So that was a big part of the hackathon too, was the, the emphasis on pushing that over the finish line. We're really, really close on that launch. We're not going to announce the date just yet, but it's, it's fairly soon. So, uh, but that, that's kind of what we worked on. We worked on all these pieces and parts that are about us doing a better job supporting our clients, just like we encourage our association clients to better serve their members. So that's what the entire week was. And we have a whole team of people that continue to stay focused on this that we've allocated internally now for, for pretty much the foreseeable future. But this quarter is going to be particularly exciting. So yeah, that was, that was last week for me. Didn't get a lot of sleep. Um, I did have to kind of force all of these programmers to get out of the house a couple of times. Um, one time I took them out and, you know, took them up to 10,000 feet and, uh, walked around for a little bit and showed them that there's, there's nature out there. And, um, you know, actually most of them were pretty into that, but there's a couple of folks that were just so into their, their work that we were like, listen, get out of your chair. Let's, let's go for a little drive.
[00:08:06:23 - 00:08:30:14]
Mallory
So, and even when you're in a flow state, sometimes it's good to get some fresh air, get some creativity from the nature around you, and then bring that back. I mean, it sounds like it was a great success. I'm curious for our listeners who are hearing about this hackathon and think, huh, that's something we would like to replicate, but maybe aren't sure of the different pieces that you need. Do you feel like everyone involved in the hackathon must be a software developer or could you pull in any?
[00:08:30:14 - 00:08:30:22]
Amith
Not at all.
[00:08:30:22 - 00:08:31:11]
Mallory
Okay.
[00:08:31:11 - 00:10:45:18]
Amith
Not at all. And the hackathons can be used for anything, you know, in, in, in Blue Cypress and also in prior companies, we've had hackathons for marketing teams and hackathons for, uh, teams well outside of anything related to engineering. The idea essentially is to get people outside of their normal day to day. And these days, especially with people working a lot remotely, it's really important, but if you're in the same environment, whatever it is, home office or, you know, an office with your colleagues, you can get stuck in a rut pretty easily in terms of your creative process. So if you want to be able to think differently, think broadly, think creatively, it's great to get to a different environment. You don't need to spend a whole week on it. Like that's our, our approach to it because we have people widely distributed all over the United States and beyond. And so, um, we make a pretty big investment for at least twice a year. We're actually ramping that up to quarterly going forward into 2027, where we get a group of roughly 10 people together. I think that there's a couple, there's a couple of learnings I would share. One is once you get the group size above 10, it starts to really break down. Uh, at least for our format we've been using, I know of hackathons that are massive, but they're a little bit different animal than what we do. Um, and the other thing is, is to have a clear theme, to have alignment. Cause if you have people working on radically different things, there's less cohesion there, uh, but it's not about software development necessarily. Um, last thing I'd say is every, everybody's a software developer now, right? Everybody can talk to an AI in English and get software out of it. So, um, you know, we've had people at hackathons that are decidedly non-technical who've contributed enormously because actually because of that, because they aren't technical, they don't think the way a developer thinks and they come up with different ideas. So it's very, very powerful. And I encourage every association and anyone else listening to try this. If you haven't, and maybe don't start off with a full week, you know, take a team offsite for two days and, um, you know, one of the things you can do is go rent an Airbnb somewhere that's very large, you know, so you can get the whole team together, uh, when you divide the cost by the number of people you have attending to go rent, you know, an eight bedroom or a 10 bedroom house somewhere, it sounds expensive, but compared to hotels, it's actually usually a lot cheaper. So there's a lot of economical ways to do this and it's really, really fun. It's a great culture building team bonding thing. Um, but it's, it's unbelievable in terms of the new, new ideas that you come up with. I've been doing this for decades now.
[00:10:45:18 - 00:11:01:22]
Mallory
Yeah. So it sounds like you've got some good takeaways. So a central theme is important. I mean, do you find it's essential to have the day broken up into portions like this morning we're working on this altogether or you go work on this, you go work on this and then we'll come back together. What do you think about that structure?
[00:11:03:00 - 00:12:07:09]
Amith
Well, um, the only structure we really have is, um, every night at, uh, 9 PM, uh, when I was a little bit younger, it used to be 10 PM. And we were joking around that it used to be actually, I think a little bit later, like 11 or midnight, we used to do these, do this daily meeting and it's a demo. And so, um, you know, nowadays it's, it's nine o'clock and we'll get everybody together and everybody demos what they've been working during the day. And the demo isn't necessarily live software. A lot of times it is, but it can be even concepts. So everybody connects, you know, their computer to the TV screen talks for five ish minutes and that usually takes an hour or two. And then, um, some people are night owls and that's kind of the beginning of their day, not necessarily beginning, but it's the middle of their day. Let's say, um, for me, it's not, it's usually after that I start to wind down and then start back up again early the next morning. But, um, we, we try to be very, very open in terms of the approach. Um, people have tons of little sidebar meetings and depending on the project they're working on, but that that's the only structure we impose is to have demos and the idea is, is by the end of the hackathon, the last demo, um, you got to show something cool. Expectation.
[00:12:07:09 - 00:12:20:08]
Mallory
Okay. So show something cool. Probably 10 or less people. If you want to keep things focused, having some centralized clear theme. And as an association, you could start off with maybe one day, two days and see what happens.
[00:12:20:08 - 00:12:50:19]
Amith
Yeah. And associations might even consider inviting a couple of their close and key members to join. Like, so let's say that your goal is to brainstorm new member engagement models or new ways of delivering value to members. You know, there could be hackathon type events for that, uh, where you bring in outside people now, you know, you may or may not want to do that, but there's lots of different ways to do this. We've done hackathon style events with our clients in the past and it's very powerful to, to get that external feedback and then to collaborate with people you don't normally collaborate, whether it's internal or external.
[00:12:52:05 - 00:14:16:20]
Mallory
Well, moving on to today's episode, two of the biggest names in tech, meta, and Nvidia, both made major open source moves within a day of each other. And both said out loud that this is now a race against China to build the best free open AI on the planet. And then kind of on the flip side, anthropic is starting to watermark the texts that Claude writes. So you can tell where it came from, or at least allegedly. So starting with the open weight race heating up quick term to land early, just to remind you all open weight or open source here means the company gives away the model itself so you can download it and run it on your own machines versus a closed model that you can only rent through their website or API. So on August 10th, meta released a new open weight model called Muse Glimmer. Meta says it's small enough to run locally on a decent laptop or desktop. And it's a slimmed down version of Meta's flagship Muse Spark. Alongside it, Meta also said it plans to open up the weights of that full flagship Muse Spark 1.2 in the coming weeks. Mark Zuckerberg framed all of this in a long essay arguing in his words that American open source models should be the best in the world. Amit, I feel like this is a bit of a reversal for Meta. They were the open source champion with llama back in the day. Then over the last year, they kind of pulled back and even started charging for their flagship models. Why do you think they would swing back to open? What are your thoughts on that?
[00:14:16:20 - 00:16:39:04]
Amith
They're swinging back in that direction simply because they want to have the absence of an open weights model for Meta would mean that I think other than people that are captive to their platforms using Meta AI within Facebook, Instagram, WhatsApp, etc., probably they'd have a hard time getting adoption outside of that. There's so much competition. There's strong leadership. Google's Gemini app just crossed a billion monthly active users. That doesn't include people who use AI mode in search. That's a much larger number. Obviously, chat GPT has been hovering around there for a while. I don't know what Claude's monthly active users are now, but I know that they've been growing like crazy. So for Meta to compete against these other tools that are closed models would be really tough, both because of adoption, but also frankly, they're just behind at this point quite a bit. So I think coming up with open source or open weight models, putting them out there is actually a good move for them because it puts them back in a conversation. Back in the days of Llama 2 and then Llama 3, they were very much at the center of conversation. Llama 3.3 in particular was a great model. And then Llama 4 just completely flopped. It just didn't work. And so they kind of threw out the whole name, which I'm not sure if that made sense or not. But the new model series they're coming out with, I have not personally tried it yet, but I'm hearing good things about it. And I am very open to welcoming anybody and everybody who wants to put an open weights model out there because again, you can inference these models anywhere. So, you know, inferencing your model isn't the same thing as where did you who built the model, right? It's kind of like the roads you drive your car on versus the manufacturer of the vehicle is one way to think of it. So or, you know, you run your own software on your own computers or on a cloud based server. Who is the cloud provider? Right. Is it Meta? Is it somebody else? So with open weights model, you have lots of choices and most likely Meta's models will become available on a variety of inference provider clouds, including the major ones as well as, you know, some of the so-called Neo clouds like fireworks that have hundreds of models available. So more choice is good. And I actually agree with Zuckerberg, which is not common for me. But I do think that American companies should be very active in the world of open source. And we should absolutely compete to have the best open source models in the world.
[00:16:40:11 - 00:16:59:04]
Mallory
You touched on this a bit, but I feel like at the inception of this podcast, we spent a lot of time talking about Meta and the llama family of models. You said llama four was kind of a flop. Now with these new models, I mean, are you open to using them? Do you still feel like Meta is behind in the greater landscape? What are your thoughts on that?
[00:16:59:04 - 00:17:52:20]
Amith
When it comes to open weights, I'm open to using models developed by anybody because you can put them on an inference provider that you trust or even your own hardware is becoming increasingly possible. So it's not an issue of do you trust the company that made the model. The model, once it's open weights, is capable of being run anywhere. So it's who do you trust to run the model is the really big question. I think there's a lot that you have to consider when you think about your inference workloads. Who do you run the model with? That's a really important question. Who made the model is important in terms of its reliability, the value system that went into the model, things like that. For example, some models are very limited in terms of what they'll do for you. Other models are very wide open. You know, if you use, for example, Grok with the K, the Grok Think Fast model or the real time models, they're excellent. But if you ask them to, they will swear at you,
[00:17:53:21 - 00:18:34:06]
Amith
which open AIs and anthropic models will not do. I experimented with this actually the hackathon one late night just for fun. And it was very strange, actually. I've never heard an AI kind of yell at me and swear at me before. Maybe it's hopefully not a vision of the future. But you have to be thoughtful about the model developer in the context of what the model can do. Certainly the guardrails on the model in terms of what it'll do and not do. But the really key question on top of that is the inference provider. So to answer your question, I would absolutely be open to considering models from Meta or anyone else once they've been tested widely and depending on the inference provider that they are that they're running with.
[00:18:35:21 - 00:19:43:07]
Mallory
So one day later after this on August 11th, the information reported that NVIDIA is building its own top tier open model reportedly called Nemetron for interesting names, aiming to be one of the best open models in the world. The scale is being reported right now around a trillion parameters, which is frontier territory. NVIDIA did separately and officially ship a smaller open model the same day called Nemetron 3.5 Lightning. NVIDIA says that one is built for the repetitive execution work AI agents do all day, like calling tools, checking results and handling long running tasks. And that's only because it activates a small slice of itself per request and it can run locally on your device rather than needing a whole data center. NVIDIA's angle is a little different from Meta's and worth sitting with for a bit, which is that NVIDIA makes the chips. Its CEO Jensen Huang has basically said that free AI is great for NVIDIA because free models still have to run on something and that something is NVIDIA hardware. I would ask Amith if you're surprised to see NVIDIA entering the open source AI model game. But I imagine based on your previous answer about Meta, you are not surprised.
[00:19:43:07 - 00:20:13:12]
Amith
Not at all. And actually a clarification is this is their fourth generation model series that they're talking about. So Nemetron has been around for a while and they've consistently been investing more and more and more hiring stronger and stronger researchers to grow their team. They're a heavyweight, not just in terms of their overall size, but in terms of their model development capabilities. So I'm really excited about this because they have both the resources intellectually and financially to produce world service.
[00:20:14:13 - 00:22:31:19]
Amith
And I don't know if Meta does or doesn't. They've gone on this big hiring spree to try to get more talent in, but they seem to have a revolving door from all the departures they've had in terms of top researchers that come in and then pretty much very quickly leave. NVIDIA is not that at all. NVIDIA is very much a desirable place for top people, both on the hardware side and now on the software or the model side to be part of. So and yes, they're motivated exactly by what you're describing. They're motivated because they are right now very much in control of the supply that is needed to run all models. That's changing. There's a lot of other opportunities and options coming out there. Everybody's in the game to build their own ships. But at the moment, NVIDIA has a very strong control on that. So growth in AI means growth for NVIDIA. I would say too that the motivation for some of these other companies that are putting open source models out there, it really is about being able to go higher up in the stack. So if you think about open source and say, well, why would a company invest potentially billions of dollars and then give away a product? There are the altruistic sides of this, of course, that that's a component of what some people believe in. But really, it's a business model that's about giving away one piece so that you can monetize something else. And the something else in the case of a lot of these companies is the harness. It's the piece of software that the model lives within that actually allows you to do much more work. So Claude Cowork, for example, is an example, is a harness. Codex or what's called now GPT work is a harness that's very similar to Claude Cowork. Gemini is doing similar things because if you're in their ecosystem through the app layer, you're not, I wouldn't necessarily say stuck, but you're more likely to stick with a particular app like Claude Desktop than with a model. Because, you know, if there's a new model like Kimmy K4 comes out and let's say it's even better than Claude Fable six or whatever's next, you'll switch to that model instantly. It's like super commoditized. It's very easy to switch to the model level. But the app level is a lot harder. Like just imagine yourself, Mallory, would you rip out Claude from all of your personal workflows and replace it with something else? Even if the model was slightly better, it might take you a minute to consider that because you're you're used to using the model. Sorry, not the model, but you used to using the application.
[00:22:31:19 - 00:22:40:12]
Mallory
Right. So you think that Nvidia, well, they have the chips, but do you think they have their own or do they have their own harness that they are monetizing?
[00:22:40:12 - 00:24:04:07]
Amith
Sort of. I mean, Nvidia has a harness that's based on the Open Claw. It's called Nemo Claw, I think. And they they put that out there because they wanted people. I think their corporate customers were using Open Claw, which was which was really scary because Open Claw is this thing that came out. It was a big sensation a few months ago. And basically, it's a harness that has like zero governance and zero car rails. So there was all these problems happening. So they created a version of that that they forked it and they created a version that's a little bit safer. Probably still not a great thing from an enterprise governance perspective. But they have a different motivation, right? Because if they own the vast majority of inference and training market for chips, AI growth means Nvidia's growth, at least for now, until they have meaningful competition as a percentage of the market. AMD is going after a lot of other people are building custom chips. All the major hyperscaler clouds like AWS, GCP and Azure all have their own custom chips for inference specifically for for running the models either in development or are deployed. Right. And Google is like the best, most notable one with their their TPUs. So there is competition for Nvidia, but at the moment, Nvidia controls the lion's share of the market. So any growth in AI means growth for Nvidia, at least for now. And probably for a reasonable foreseeable future. Whereas if you're a model company, you have to have a game plan to go up the stack because the model itself is pretty clearly a commodity.
[00:24:04:07 - 00:24:22:10]
Mallory
Mm hmm. I guess I'm also thinking through this from a business perspective. If the harness is something that you can monetize and the model is available and open to anyone, then couldn't a company just use someone else's open model and then create their own harness around it? I guess what is the benefit of as the company creating the model and the harness?
[00:24:22:10 - 00:24:58:17]
Amith
Well, I mean, you can create tighter integration. You can make sure your harness works extremely well with your models, which is what the sensation behind cloud code is partly that cloud code works really, really well with cloud as the model. And you can create a tight ecosystem. Right. So what entropic has done brilliantly is they've executed this game plan where there's an artifact ecosystem, right, where you build artifacts and you share them with colleagues and it takes them right back to cloud. Right. So it's got these kind of viral loops built in for growth. They're building an application stack. Right. And that's that's the big thing that the major labs are doing and open eyes doing the exact same thing.
[00:24:59:20 - 00:25:56:09]
Amith
So I would say that this is actually also really worth highlighting for our association listeners to be very thoughtful about the harness that you use and harness again is it's the app that you use day to day as a user to some extent. But it's really what you build your enterprise, a genteck AI on top of. So if you're thinking, hey, I need to deploy a member services assistant or I need to deploy an AI agent for marketing or whatever your agentic use cases are, that's where you need an agent harness. And so this is the thing that basically runs the model to form more complex tasks for you than what you do as an end user to do it on a repetitive basis, either on demand or on a schedule. And to be able to govern it, meaning giving it controlled access to limited data, controlling where it can go on the Web, recording what it does. Right. So observability. These are concepts that a more enterprise level harness would do.
[00:25:57:15 - 00:27:13:03]
Amith
But you also have to be very thoughtful about not getting into this trap of being stuck in one ecosystem. You know, picking a model is not really that important right now. It's going to change in 30 days, probably. But picking how you want to structure your deployment of AI is really, really important because you can end up with another AMS like situation where you pick something and you build around it and then you realize you're stuck with, you know, a company's harness and ecosystem that you really can't escape. People who are building directly on top of the agent SDKs from anthropic Google or open AI are falling exactly into this trap, even though those are very, very popular and they're great. You will not be able to use other models if you build on top of open AI or anthropic agent ecosystems, you will be using their models. So the way they're monetizing it is by giving you higher level tools, which of course are stickier, have higher switching costs, which by the way, is exactly the reason we built Member Junction, which is our free open source platform that is an agentic architecture and harness. And it's totally free and always will be and you can deploy it anywhere you want. And that gives you this leverage to control that destiny a lot more so than using these commercial harnesses from a variety of these companies.
[00:27:14:12 - 00:27:27:07]
Mallory
Seeing more and more open weights models pop up from various companies. Do you see a world of meath where the AI space in general is moving away from closed models or will we always have the choice of open and closed?
[00:27:27:07 - 00:27:46:21]
Amith
I suspect there will always be some closed models that are, you know, the leading, bleeding edge kind of things or maybe in certain domains where somebody has a really tight lock on a particular cornered resource like content. Who would have really unique novel content in a particular vertical?
[00:27:46:21 - 00:27:48:24]
Mallory
Maybe we might know some people. Yes.
[00:27:48:24 - 00:29:16:19]
Amith
Maybe some associations. So maybe associations maybe not train a model itself, but build a system where they do retain close proprietary loops around the mixture of a particular level of intelligence coupled with a particular domain of knowledge. And that is, of course, something people can do. You can do that at the model level. You could train an actual AI model on its own to be an expert in law or accounting or whatever the field is. But oftentimes you want to separate that as well. You want to have the knowledge layer be part of what we call kind of the agentic process as opposed to the model itself. But the point I would make is I do think there's cases where proprietary models could make sense. And I have no problem paying for services for proprietary models if they add value. I just like choice. I like the idea of being able to say I'm going to use a little bit of that proprietary model and then I'm going to use a whole lot of free stuff because the free stuff is really, really good at most of the day-to-day tasks. But I might use the really high end model like Fable 5 on Ultra mode or whatever it's called to do planning and to do review. But then I might use an army of really small or fairly small and really cheap and expensive models to do all the legwork and then have Fable review all that work. Right. That's actually a pattern I use every single day and it's both faster and way cheaper. But you have to have flexibility and choice to do stuff like that.
[00:29:16:19 - 00:29:30:11]
Mallory
So it sounds like it doesn't really matter which company or country is creating the open source model your association uses and we get to kind of sit back watch the battle play out while selecting the models that best fit our use cases.
[00:29:30:11 - 00:30:39:13]
Amith
Yeah, largely that's the case. I think the biggest thing you have to think about is the model developers value system. We've talked about this with anthropic and their so-called constitutional AI, which essentially is an embodied value system that they train deeply into every model that they build, which means that's how the model is going to behave, how it will act. It's the same thing as the value system that a child is raised in and how they will interact with humans in the world after they leave the house or leave their hometown. How do they interact with people? Models interact with us and do or don't do things based on how they were trained. So it is important that the model developers kind of overarching value system at least has some similarity to your own because the guardrails or the lack of guardrails may or may not suit your use case or your philosophy. So you do have to pay attention to the model developer. But I do think that models are largely commodities. So long as you do pay attention to the nuance I just described. And then to your point, Mallory, what matters a whole lot beyond that, if you assume these models are very, very similar, which most of them are, is the inference provider, as you said.
[00:30:40:17 - 00:31:32:02]
Mallory
I want to move to our next topic for today. On August 11th, anthropic said that Claude models launched on or after August 2nd now embed a watermark in the text that they generate. Anthropic describes it as invisible, something that doesn't change the meaning or the quality of the writing woven in at the levels of the words themselves. So it travels along when you copy and paste. For image files, anthropic says it's also attaching signed metadata using an open industry standard for content provenance. The driver per anthropic is the European Union's new AI Act, which starting this month requires companies to make AI generated content identifiable. I mean, for a while, and we've talked about this on the podcast, it seemed very difficult to detect AI writing. This seems to be some sort of solution to that. But I'm curious, we haven't talked about it in a long time. What do you think about anthropic's new strategy?
[00:31:32:02 - 00:33:05:03]
Amith
I still am suspicious as to how effective it'll ultimately be because it's move, counter, move type conversation again, whether it's cyber security, in this case, provenance tracking of which model generated the item, whether it's text, image, video, etc. As smart as your watermarking strategy may be, there will be models that are as smart or smarter that can try to undo that. And in fact, when it comes to text, if you simply take the output of Claude and you take it over to chat GPT and you say, hey, I have this draft, can you edit it for me? The watermark is essentially gone because the words and so much they're going to get scrambled. But a lot of the watermarks potential strength is ruined. What they're saying essentially is that the pattern of words somehow they've been able to mathically mathematically encode something that tells them that it's coming from Claude. I don't think they've disclosed what the algorithm is. I don't think they will. That would actually undermine the whole point of it. But it's clearly not. There's nothing there's no like hidden data in the text itself. It's just text. So if you copy and paste the text, how does the watermark come with you? It's the pattern of the text. So in that it's kind of like a hidden cipher, essentially. So I'm sure they use like cryptography techniques and very interesting. But also you could probably run every output from Claude through fairly low end models. Say, hey, this is a draft. Give me a revision. Take it to another model and very quickly effectively have no watermark from a text perspective. So that I'm fairly confident of even without knowing the specifics of their algorithm.
[00:33:06:05 - 00:34:48:12]
Amith
Now, if will people do that or is the goal to basically catch at scale distillation, which I suspect actually is really their motivation. So Claude and anthropic, they have been the victim of mass scale hacking to, you know, a lot of providers have come after them to use their models to generate outputs that can then be fed into the training process for newer smaller models, which in turn essentially picks up the knowledge from the bigger model. It's a process called distillation. Anthropic claims this, I should say. It shouldn't be assumed to be a foregone conclusion that they're right. Although, you know, the evidence that they've shared with the US government that is public is fairly strong in terms of the hacking and the things that have occurred, as well as just like, you know, fake accounts, armies of fake accounts that have logged into Claude to use it for this. So my suspicion is that their motivation largely is to try to catch these patterns, these watermarks in the distilled outputs from these derivative models or so-called derivative models and then to be able to go after those folks. I think that's probably their main goal because at the scale that I'm talking about, you probably wouldn't run it through a process like I'm describing. As an individual user, if you cared about this, you could. When it comes to images, there are also workarounds for that. There's ways to potentially change the image. There's ways to copy and paste the image without actually copying the file itself. There's a lot of different techniques. For every move, there's a counter move. So I wouldn't necessarily put a lot of stock in this being the solution to, hey, we know this was AI generated. I'm not suggesting that I think I know that there isn't a solution or there is a specific solution to this. I'm just a little bit skeptical that this is it.
[00:34:49:22 - 00:35:14:20]
Mallory
Well, I think that's helpful because my thought with this topic on the episode was, you know, credentialing bodies that are worried about AI and exams and certifications. We've got publishers and journals worry about AI written submissions and associations in general that vet content thinking, hmm, perhaps this is a solution we should keep an eye on in the coming years. But it sounds like it's move counter move. So this may not be a reliable thing to depend on as an association.
[00:35:14:20 - 00:36:13:03]
Amith
Totally. And I think we all think that we know a lot more about this than we really do myself included. For example, I kind of feel like certain things I read are probably clod when I get emails from people that sound clod like in terms of the just the rhythm of the way the thing's talking. I use it so much that kind of gotten used to its way of talking. It's probably changed the way that I write and talk. Right. If you that's true. If like, you know, you find yourself in a room with people who speak a certain way or use certain phrases and over time, you basically become a product of the people you spend the most time with. And if one of those so-called people is clod, you will very soon be talking like clod and thinking like clod. So I mean, that's the nature of the way our brains work, right, which is actually both amazing and also something just to be aware of and to think about. And I found myself doing that over time, right, where you go work in the UK for a few weeks with some clients and you adopt some of their phrases and some of their ways of thinking about things, too. So,
[00:36:14:05 - 00:36:16:21]
Amith
you know, I still call things French fries, though, when I go over there.
[00:36:16:21 - 00:36:33:09]
Mallory
Oh, yeah, that would take at least a few years to lose right now. But I used to when I was little, I, you know, I'm from central Louisiana. I had an accent. I don't think I do now. But when I'm around my family, I certainly sound a lot more Southern and people will point that out. So we are a product of our environment. Certainly.
[00:36:33:09 - 00:36:38:18]
Amith
I, for one, think the Southern accent is awesome. And you should kick that out on the on the pod sometimes.
[00:36:38:18 - 00:36:50:19]
Mallory
All right. I mean, I wasn't that bad. It wasn't ever that bad. But I hope we don't all start sounding like AIs. I don't know. I think that would be a little sad. Hopefully we can retain our humanity. Right.
[00:36:50:19 - 00:37:38:08]
Amith
Totally. Yeah. And I think, look, and this is the process of adaptation that has made our species, obviously, the apex of planet Earth, at least. And it's our ability to adapt. It's not because of our physical might. And it is certainly we are an intelligent species, but it's our adaptability. It's our continual learning. I don't think that's ever going to end. I think we we continue to do that at a pace that's really remarkable. If you think about how quickly we're adopting these tools, you know, I'm I'm always going out there kind of hammering on the drum beat of train your people, train everyone, make sure everyone has access to quality AI learning. And then and then, frankly, after you give them access, demand that they do it because it's not only necessary for your association or your nonprofit to exist in a few years, but it's also critical for those people's future employability.
[00:37:39:09 - 00:38:19:10]
Amith
But at the same time, I'm also really impressed with the degree of adoption that we've already seen. Maybe not everywhere, but, you know, just seeing people adopt these these really strange, somewhat alien tools, right? Like you and I talking about our day to day workflow being radically different. Only a small number of years after these tools became widely available is pretty remarkable. And there's there's, you know, billions of people out there that are using these tools. There's many millions who have dramatically changed the way they work. So I think it's largely a good thing, our adaptability. But it's good to have, you know, this is also why hackathons are great, because you kind of take a deep breath. And you get away from your your normal environment. You're like, wait a second. I'm doing this thing that I didn't realize I was doing.
[00:38:21:05 - 00:38:52:24]
Mallory
To wrap up this topic, I mean, if you were a leader of an association that vets content within its field or within its industry, I guess I don't know if it's more of a philosophical question or strategic. But it sounds like the best path forward is not block all AI content, right? Because it seems like that would be very difficult to do. To associations that are afraid of the amount of AI slop out there, how can they think about vetting content that does include AI, but, you know, in a meaningful way?
[00:38:52:24 - 00:40:13:13]
Amith
And you're talking about content submitted to them for publication or for consideration at an event. I think what you do is you need to evaluate it on the quality. If you're looking for how to detect plagiarism, that's a really, really big challenge. If you're looking at it from the viewpoint of, is this thing correct? Is it technically sound? Is it aligned with what we want to be publishing or considering for a conference? There's a lot of opportunity around that. In fact, you know, we're actively working on solutions with clients right now that help them or helping them streamline exactly this process where you have an agent that's doing intake, taking in content like submissions for journals or submissions for conferences, doing an initial pass to see is there alignment between the topics that are proposed in this content relative to what we're looking for, which by itself normally takes an enormous amount of preliminary time from either staff or a lot of times from your volunteer leadership, your committees. And then from there, you know, continue to compare it, perhaps doing a comparison between a proposed article or talk against your body of knowledge. If you have access to it in such a way where you can very easily search it and synthesize from it. And if you don't, there are solutions for that as well. But the idea would be to use the AI to help you maximize quality and relevance.
[00:40:14:21 - 00:41:21:08]
Amith
The other thing I think that's really interesting about AI, it's not exactly on topic for your question, but it's related to this process. I'll throw it in there is kind of widening your thinking. So we all get stuck in a rut like we've been talking about. And so when we evaluate proposals for talks at our conferences and for our publications, we have enormous, enormous biases we have to overcome. And so when we're thinking about like, should we allow Amith to come speak at your conference from his proposal, you're going to be biased in evaluating it based on the person, their name, based on your experience with the individual, based upon the content. That's just how human reviewers are. Now, AI has its own biases based on the training data we put into it. But if you use multiple different AI systems together, you can actually not only mitigate some of those training biases, but you can kind of identify the human biases that are out there and identify potentially content that you would have said no to. Not because it's not sound and not because it's not potentially very interesting to your audience, but because your initial reaction was, no, this isn't a fit.
[00:41:22:10 - 00:42:02:23]
Amith
So I do think AI can potentially be very, very interesting for this. And I've talked about this before in our hiring process. We always have a human review every single application, every single interview that comes in. But we do have an AI suggest people to us based on its analysis of the AI interviews that we do. And we have found candidates that I would suspect we probably would not have spoken to just because they might have been a little bit out of band in terms of our mental distribution for what we think is the right fit. So AI can be an extremely good thought partner in broadening your way of looking at the world in the area that you're describing. And it can help you determine whether or not content is sound as well, like fact checking and stuff like that.
[00:42:02:23 - 00:42:19:19]
Mallory
Yeah, I love the idea of comparing submissions for conferences to your entire repository of content. And maybe even like you said, I mean, surfacing topics or talks, things that you wouldn't have considered maybe because of your own biases, but having an AI assist you in that process. I think that's pretty powerful.
[00:42:19:19 - 00:43:07:14]
Amith
Yeah, I think if you've ever been to a conference, you're like, man, this content is so similar to what it's been for the last few years. Can we talk about something else? And I know a lot of people have been to conferences and felt that way. It's a really hard problem for conference organizers. Let's say your event has 5,000 people coming to it and you have 150 different sessions and you've been doing this for a long time and you have a whole bunch of other conferences too. And your community is such that you have a lot of speakers that do overlap, which is potentially really good. How do you do this at scale where you can compare everything against everything else and do a good job of both allowing certain topics to continue in certain themes and certain speakers to kind of build on the success they've had before, but to avoid a situation where people feel like they're watching the exact same movie over and over again?
[00:43:07:14 - 00:43:17:00]
Mallory
Well, we started with a hackathon, what it actually is and how you can run a small version yourself. Then Meta and Nvidia both making big open model moves in the same
[00:43:17:00 - 00:43:22:14]
(Music Playing)
[00:43:33:07 - 00:43:50:06]
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.
[00:43:50:06 - 00:43:53:12]
(Music Playing)