Microsoft Pays Anthropic to Fix Copilot, Meta Returns to the Frontier, & Courts Put a Price on Your Content | [Sidecar Sync Episode 156]

Summary:

Microsoft spent years building Copilot around a single model maker, and the fix turned out to be paying for a competitor's. Amith Nagarajan and Mallory Mejias unpack three stories that matter for association leaders weighing their AI choices. They trace how Charles Lamanna convinced Satya Nadella to put Claude inside Copilot, explain what "model monogamy" cost Microsoft, and consider whether your association should keep its Copilot licenses or equip power users with more. They break down Meta's return to the frontier with Muse Wonder, why its new models are proprietary rather than open-weight, and how to choose between frontier labs when capabilities are this close. They also unpack the Third Circuit's ruling in Thomson Reuters v. Ross, the first appeals court decision on fair use for AI training. Along the way, Amith makes the case that protecting your content starts with knowing where it goes and whose incentives you're trusting.

Timestamps:

00:00 - Blue Cypress Summit Recap
03:33 - Digital Now and Mallory's Keynote
06:35 - How Claude Got Into Copilot
11:55 - Ending Model Monogamy
16:05 - Keeping Your Copilot Licenses
19:40 - Secret Cyborgs on Staff
23:49 - Meta Hits the Frontier With Muse Wonder
28:13 - Muse Pricing and Open Weights
31:02 - Testing Models, Trusting Vendors
36:14 - The Ross Ruling Prices Your Content
41:57 - Protecting Content Without Burying It
47:46 - Refresh Your AI Choices Quarterly

👥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

https://learn.sidecar.ai/

📕 Download ‘Ascend 3rd Edition: Unlocking the Power of AI for Associations’ for FREE

https://sidecar.ai/ai

🛠 AI Tools and Resources Mentioned in This Episode:

 Microsoft 365 Copilot ➔ https://www.microsoft.com/en-us/microsoft-365/copilot

GitHub Copilot ➔ https://github.com/features/copilot

Azure Maia 300 ➔ https://azure.microsoft.com/en-us/blog/deep-dive-into-azure-maia-300-architecture

Claude Opus 5.5 ➔ https://www.anthropic.com/claude/opus

Claude Haiku 5.5 ➔ https://www.anthropic.com/claude/haiku

Claude Code ➔ https://claude.com/product/claude-code

Claude Design ➔ https://claude.com/product/design

gpt-oss ➔ https://openai.com/index/introducing-gpt-oss

Gemini ➔ https://gemini.google.com

Gemma ➔ https://ai.google.dev/gemma

Vertex AI ➔ https://cloud.google.com/vertex-ai

Muse Wonder ➔ https://about.fb.com/news/2026/10/introducing-muse-wonder-models

Llama ➔ https://www.llama.com

DeepSeek ➔ https://www.deepseek.com

Kimi ➔ https://www.kimi.com

GLM ➔ https://z.ai

Fireworks AI ➔ https://fireworks.ai

Artificial Analysis ➔ https://artificialanalysis.ai

MemberJunction ➔ https://memberjunction.org 

👍Please Like & Subscribe!

https://www.linkedin.com/company/sidecar-global

https://twitter.com/sidecarglobal

https://www.youtube.com/@SidecarSync

Follow Sidecar on LinkedIn

⚙️ Other Resources from Sidecar: 

• Sidecar Blog
• Sidecar Community
• digitalNow Conference
• Upcoming Webinars and Events
• Association AI Mastermind Group

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

Read the Transcript

🤖 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 - 00:00:25:20]
Amith
 My name is Amith Nagarajan.

[00:00:25:20 - 00:00:28:05]
Mallory
 And my name is Mallory Mejias.

[00:00:28:05 - 00:01:14:03]
Amith
 And we are your hosts. And we have a lot of fun planned for you today, as always. We're gonna be talking about something that has been the butt of jokes for some time with Microsoft co-pilot Mallory. I'm looking forward to hearing what's new in the world of co-pilot. I'm not super up to speed on it, but I'm looking forward to hearing what's new and chatting about that. I have some thoughts on it and all the other things that are going on. But I have some breaking news here in New Orleans. It is not disgusting outside. It is actually really nice. So everyone's kind of looking around going, "New Orleans, are you messing with us? It's mid October. We got a couple of nice days and hopefully it sticks around." But I was in Utah last week for Blue Cypress's leadership summit, came back, got off the plane and I thought we might've landed in the wrong city.

[00:01:14:03 - 00:01:27:06]
Mallory
 Man, I'm happy to hear that the weather is so nice in New Orleans. For all of you that are just listening audio only, Amitha's here in a green sweatshirt. So I don't know if it's maybe a little crisp in the air, a little chilly for you and me.

[00:01:27:06 - 00:01:55:08]
Amith
 No, the issue is, so we're in 110 year old house here in New Orleans where our team is. And in order for my office to, actually, sorry, in order for the office overall to be comfortable for everyone else, my office is basically an icebox. Probably in the mid 60s here. And I tend to need to wear a hoodie anyway, even when it's like in the low 70s. I'm just outright cold, but I've got to keep all the programmers and the BC Labs team from overheating because we want them to be happy and productive. So we keep it on.

[00:01:55:08 - 00:01:57:20]
Mallory
 I think I'd rather be cold when I'm working than hot.

[00:01:57:20 - 00:02:13:18]
Amith
 Yes. For sure. I can add more clothing. So it's no problem. But yeah, it's good stuff. But New Orleans being in a comfortable temperature range and my San Francisco 49ers being five and oh, it's a good October so far.

[00:02:13:18 - 00:02:22:09]
Mallory
 And the Blue Cypress Leadership Summit, fresh off of that of me, how did that go? Yeah, that was pretty good too. Our listeners might not know exactly what that is. So can you give them an overview?

[00:02:22:09 - 00:03:32:17]
Amith
 Yeah, so up in the mountains of Utah, where some of our listeners know I like to spend a bit of time, we once per year get together a group of about 35 or so of our senior leadership across different companies and the Blue Cypress family. This year we actually invited a couple of our clients to join us for a portion of it. It was really great. And the whole point of it, first and foremost, like to remind people that it is an opportunity for us to engage and think and have a little bit of fun together. You know, we do things like going for a hike as a team, doing some of the things, dine arounds around Park City. And the goal of it is to certainly connect the team, build relationships. That's the fundamental goal. But beyond that, when you get yourself into an environment that's a little bit different than your day to day, you tend to think differently and sometimes think bigger, but sometimes just solve problems that may have been challenging to solve if you were just looking at the same screen, at the same desk, in the same office every day. So it's an exciting time for our team. We always run that either in late September, early October. And it's usually a few weeks before our Digital Now conference. So it's a great kind of lead up to the big event with our client community.

[00:03:32:17 - 00:03:42:10]
Mallory
 And speaking of Digital Now, I mean, that you just mentioned, it's coming up soon, right? What are the dates? Tell our listeners, if you haven't registered yet, you should, what are you waiting for?

[00:03:42:10 - 00:04:57:24]
Amith
 You really should. There are still a few spots remaining. We probably will sell out this year. It's the biggest Digital Now ever. It's really exciting. It is in the Washington DC area, specifically at the Hilton, the brand new beautiful Hilton property across the river from Georgetown in Roslyn, Virginia. It's called the Hilton, the Key in Roslyn. It is October 25th is the opening day, which is a Sunday afternoon. And we'll be going through Wednesday, October 28th, around mid-day. So it's basically two and a half days of nothing but amazing content and stories from deep within the association sector, all the great things that are happening with AI, as well as interesting speakers and folks from outside of the association market. We always bring in what we like to call Silicon Valley's latest thinking into the association space. Outstanding speakers coming to share the latest and greatest of what's happening in a variety of fields. So it's really exciting. And we'll have hundreds of people there this year from different associations. A lot of people are coming as a team. We have quite a number of organizations bringing 10 or more attendees, which is really exciting. When you come together and you learn together, you grow together. And so what better place to do that than Digital Now, Mallory?

[00:04:57:24 - 00:05:20:00]
Mallory
 I love it. More than 10 people on their teams. That's fantastic to hear from me. I agree with you. I think it's one thing to go to a conference like that with one or even two people. You have all these ideas and then you come back to work and then you realize your to-do list is 50 items long. And it's very easy to quickly kind of forget the momentum that you had leaving the event. But I think having a team of 10, that's gotta be really impactful.

[00:05:20:00 - 00:05:33:10]
Amith
 I should also mention, totally. And I should also mention to our listeners that Mallory, you've been part of the Digital Now story for quite a number of years. Now I know you had your first exposure to AI, I think back in 2021, was it? Or 2022.

[00:05:33:10 - 00:05:35:18]
Mallory
 I think it was 2022 in New Orleans.

[00:05:35:18 - 00:05:58:00]
Amith
 Yeah, it was the New Orleans event, right? 21 was Nashville, but you were in New Orleans in 2022 and had your first aha moment with what was going on with AI at the time. And you've been a big part of Digital Now ever since. And I'm excited to let our listeners know that this year you'll be returning as one of our keynote speakers. So excited to hear all the exciting things you're gonna share with our audience.

[00:05:58:00 - 00:06:35:02]
Mallory
 Yes, Amith, I'm very excited to be joining in that capacity, a little nervous too, as I'm still kind of working through the keynote. But I am just really grateful and I am going to try my best to fuse my two very different worlds that I sit in. One being AI and associations and the other being this more creative world of performing and acting and pursuing that. And then seeing how I can fuse them together. So if you are attending, be on the lookout for that Wednesday morning. I believe it's at 10 a.m. Very excited. And Amith, you'll be there, I'll be there. So if you're an avid listener of the Sidekarsync podcast, we would love to have you join us.

[00:06:36:21 - 00:08:22:06]
Mallory
 Moving to today's episode, Amith, one of your favorite topics. We're starting with Microsoft Co-Pilot and the executive who fixed it by convincing Microsoft to pay for a competitor's AI. Then we're gonna be talking about Meta, which we haven't covered in a minute, which just released a frontier model. And we're gonna close with a court ruling that says the market for licensing your content to AI companies is real and worth protecting. So starting off with Co-Pilot, this is the AI assistant many of you have heard of built into Microsoft Office. Word, Excel, PowerPoint, Teams, Outlook. Many associations that are listening to this may already be paying for it, whether or not anyone on your staff is using it. And for a few years, I would say Amith is nodding his head. The honest verdict was that it wasn't very good. By the summer of 2025, Microsoft was fielding complaints that it was buggy and struggled with basic work in Excel and PowerPoint. The information reported this week on how that got turned around and the story runs through someone named Charles Lamanna, the executive who led the fix and now runs Office. His pitch to Satya Nadella was that Microsoft should pay Anthropic to put Claude inside Co-Pilot. That's a little bit stranger than it sounds because Microsoft gets open AI's models at no extra cost through its partnership. So this meant spending real money on something it technically already had for free. Lamanna's word for how it landed internally was heresy. He won the argument with demos showing Claude was better at the math Excel needed and at making PowerPoints that looked decent. Amith, it's been a rocky few years, I would say with Co-Pilot. Can you help us understand why is it that Co-Pilot has seemed to lag behind other AI companies like OpenAI and Anthropic?

[00:08:22:06 - 00:11:30:07]
Amith
 I have so many conflicting feelings about this one, Mallory, Co-Pilot was one of the coolest launch events and launch videos I think that I've ever seen from any vendor and Microsoft did an outstanding job. This is, if I remember correctly, spring of 2023 when they were rolling this out, this is just after the chat GPT moment, after GPT four came out, Microsoft said, "Hey, we're gonna have a Co-Pilot for everything." At this point in time, Co-Pilot as a broadly understood term in AI wasn't really out there. In the developer world, GitHub Co-Pilot was already a pretty powerful developer tool at the time. It was like really smart auto complete was all it was back then. This is way before a Claude code. And so when Microsoft announced they were gonna bring Co-Pilot to Word and Excel and PowerPoint with the theoretical idea that it has the full context of all the files you have in your OneDrive or SharePoint, I was really excited. We signed up quickly for early access and paid for all of our employees and very quickly turned it off because we found that it underwhelmed at least in its first few incarnations. We wanted to empower our team to have really frontier capabilities. And back then that mattered a lot because the models were so limited and so really kind of infantile in their abilities compared to today's models. So my early experiences with it were negative. I should disclose that I've a long time Microsoft stockholder and generally a fan of what they're doing with enterprise AI. I think they have the right mindset around things like governance and stuff like that, which is by the way, one of the reasons associations default to picking them because it's kind of the easy button for AI to say, "Hey, I already trust Microsoft with a bunch of my data. We'll just turn on Co-Pilot." But unfortunately for the last few years it has been pretty poorly executed. So I'll give you a couple of reasons why I think it has lagged behind other AI companies. One is that Microsoft's a behemoth and they were not built in the AI era. They were built long before that. And while Nadella who's been now at the helm for about 10 years has done a remarkable job rebooting the culture, it's still a big, fairly slow moving organization compared to the frontier labs. So that's one factor is culture. And a lot of association listeners can empathize with that because changing an organization that's been around for a long time is tough sometimes, especially a very large one with tens of thousands of people. The other factor is the scale of what Microsoft is doing means that the cost implications and the resources required to properly power Co-Pilot with really great models would have been pretty tough up until fairly recently. So back when they first rolled it out, there were reports that were really, I think largely based on the rumor mill that Microsoft had used lower powered quantized models which are essentially models that are not as powerful as the headlines within Co-Pilot in order to make it scale up. So that was part of the problem. And then finally the issues we ran into, and I should also say I haven't used Co-Pilot in probably about six months. So I have not personally experienced the latest. I'm gonna go back and check it out soon to see what I think. But

[00:11:31:08 - 00:11:54:00]
Amith
 it did not really have great access or retrieval to the files in SharePoint and in other places, which is part of what you'd assume that your agent would be able to do. So I believe a lot of those things have improved, but that was the reason that they, I think they lagged initially. Microsoft certainly has the horsepower to be able to do much better. So it sounds like they're taking it more seriously now.

[00:11:56:00 - 00:13:21:21]
Mallory
 The phrase that I wanna point out too with all of this is model monogamy. So Co-Pilot had been built on the assumption that it would only ever use one model maker. Lamanna's argument was that if you bake in that much dependence on one model's behaviors, you can never swap in anything better later, which is something that might sound familiar because we say it all the time on this podcast. So Microsoft rebuilt Co-Pilot to route across models and the product you use today is different in a few concrete ways. So underneath it now sits a small in-house AI whose only job is choosing which model handles a given task. Most people never see that happen as it's occurring behind the scenes. And then the other change is Agent Mode, which makes the model slow down and reason through a task answering it immediately. Customers used to have to switch it on themselves. Lamanna found it worked better for nearly every task, overhauled Co-Pilot to make it the default by this past spring and Microsoft saw a sharp increase in internal customer satisfaction data right after. Microsoft says conversations per user nearly doubled this year and weekly engagement now matches Outlook and Teams. Independent surveys still put weekly usage at well under half of license seats. So the product improving and people using it are not exactly the same thing. Amitha I kind of touched on this briefly because it is something we say a lot on the podcast but what do you think associations can take away from this idea of model monogamy with Microsoft?

[00:13:21:21 - 00:16:04:03]
Amith
 Yeah, my thought is that choice is really helpful in most things when it comes to the technology it creates a lot of optionality, flexibility and different models are good at different things. So the idea of a small in-house model being able to make a decision about which larger model should be handling a task makes a ton of sense and we do this all over the place in systems we build. You don't necessarily need the most complex model making that initial choice and in some cases though you do need the most sophisticated model handling a problem for you. So an Opus level problem should be given a model that's capable of handling the user's needs whereas many requests that users have are very basic. So for example, if you're in PowerPoint and you want to change the font of the heading of your slide probably doesn't require a ton of horsepower really simple you say change the font to a bigger font and a very simple model should be able to do that. Whereas if you ask it to reason across, hey look at the last 10 PowerPoints I've done make sure I'm not being redundant and think deeply about the best way to position this statement for this audience that's gonna require some research, some reasoning probably a lot more intellectual ability on the models part and so the router is probably gonna take you to something like an Opus 5.5 capable model. And Microsoft by the way is also investing heavily in larger models in house. So long-term I would strongly suspect that the plan is for Microsoft AI models to power a large portion of co-pilot. In fact also running probably on their Maya chip architecture which they had a big announcement on yesterday talking about the Maya 300 which is their latest chip that's going into production next year where reportedly they are going to be inferencing most of their AI inference on the Maya 300 and successor chips down the road as opposed to Nvidia GPUs which is of course the same playbook that a lot of other large scale companies are trying to do in order to obtain a cost savings. So there's so much margin in Nvidia's product and for a lot of the high throughput workload you don't necessarily need it. But coming back to the essence of what you're asking I think that choice is critical especially in a world where we don't know what's gonna happen next and I don't know about you Mallory but I've never felt more like every day really is a surprise right? In terms of what's gonna happen next. Sometimes it's exciting, sometimes it's a little bit nerve wracking but when we don't know what's gonna happen next locking ourselves into any one thing is really I think almost guaranteed to be the wrong choice because there's the probability of picking the right model or the right model family right away is very, very low.

[00:16:04:03 - 00:16:18:02]
Mallory
 So you said you haven't tried out Microsoft CoPilot in about the past six months. Let's say you go back and try it out and it's better but still in your opinion not quite as good as like Claude or chat GPT.

[00:16:19:05 - 00:16:31:03]
Mallory
 Let's say we have an association listening to this saying well we have CoPilot licenses, do you think we should keep them or do you think it's a disservice to our staff giving them tools that maybe aren't quite to the level of what's out there?

[00:16:31:03 - 00:19:23:04]
Amith
 I think that like many things the short answer is it depends. I mean a year or so ago I would have said CoPilot really isn't a great idea just because it was so limited and so buggy that it made sense to give your users a better option but now it seems as though CoPilot is a legitimate choice, it does have a lot of powerful capabilities and if you personally find comfort in being in that walled garden that Microsoft provides you could do that and you could say hey CoPilot's our default and we're now comfortable with its capability for most users. I would also say that there are capabilities that are moving very very fast and products like Claude and also chat GPT, Gemini as well, that you probably want to give some of your power users access to. Some of the capabilities that are announced literally every week by some of these little bit more forward looking organizations or at least faster moving I should say forward looking, more forward looking may not be fair to Microsoft but they're able to just put features in really rapidly like Claude design for example that you don't see that coming to CoPilot anytime soon probably. So there's features and capabilities that you should evaluate and maybe have a couple of choices that you vetted and then of course for enterprise workloads what I would call the really critical things to choose when it comes to your corporate data, your business data going to the model that's where I think you have to look at the overall architecture really closely. So CoPilot is gonna be a great place to work with your unstructured data, your Word documents, your Excel documents, things like that. But if you want to have AI agents work with structured data that came out of various other systems like your AMS or your LMS or other systems you have to have a data layer as we've talked about on this pod a bunch of times that you own and that's really a very strong advice that we give to people in our community because the absence of owning your data layer essentially is almost guaranteeing that you're going to end up in essentially someone's tenant meaning someone's gonna control you and limit what you can do or monetize you in a way that you won't really like. So I think that it also points back to this idea of provenance, sovereignty and the idea of really being able to be a destiny for your agents. Remember that the agents that you use will often use 50 to 1,000 times as much underlying AI as you use when you're talking to Claude or chat GPT. Because you're asking for one thing at a time, agents are capable of doing things in parallel, 50, 100, 1,000 times as much as what you're doing. So if you actually look at consumption of AI tokens, agents are already the vast majority of that consumption. And so for your association, if your primary use of AI is just your staff talking to an AI like chat GPT,

[00:19:24:05 - 00:19:39:09]
Amith
 what we're talking about here is directly relevant. But if you're thinking about it from the viewpoint of building agents that do customer service, member service, hit your database, all this kind of stuff that's more enterprise level, that's where you need to step back and look at your data platform layer first too.

[00:19:40:22 - 00:20:13:19]
Mallory
 I also wanna double click on what you said, Amith, about power users and perhaps giving them access to tools like Claude or chat GPT because I know I've had conversations with association professionals who have co-pilot sanctioned by their organization and then have admitted that they keep a computer off to the side, that they have Claude running on, that they use for work for the association anyway. So it's just something to consider. You might have some secret cyborgs on your team, as Ethan Malik likes to call them, but basically power users who are going to find the best tools to do their work, kind of whether you're aware of it or not.

[00:20:13:19 - 00:21:49:13]
Amith
 Yeah, totally. I mean, that's gonna happen. That happens all the time. You've seen this happen in your associations for years, where for example, you tried to mandate that people do certain work in the AMS, for example, like you might say, hey, you have to use the AMS for event planning or something like that, where the AMS kind of sort of does it, but isn't great. And then you have your events department saying, well, yeah, no, we're not gonna do that. And then they have a spreadsheet for their event planning or they use some specialty tool that may or may not even be something that IT or anybody else centrally is aware of. These kinds of things happen because people are trying to get their job done. It's not because they're intentionally being disrespectful or they don't want to comply with policy. They're just trying to get their job done. And if people see a better way, the more creative, the more entrepreneurial minded people will go and do exactly what you just said, now they're gonna solve the problem. They're gonna bring in their own equipment. They're gonna do things like that. And you're much better off acknowledging that and having a program internally where people can maybe apply to be part of a forward deployed kind of forward looking group that maybe gets additional training to help them understand cyber risk better, help them make sure they really do understand what it means to use an AI tool like a cloud outside of CoPilot. And then once they've had that training, maybe they even signed some kind of acknowledgement that they understand it and what they'll put in versus what they won't put in. Then they have a mechanism through a provisioned tool by a cloud through teams, for example, as opposed to just individually signing up to a free account, which is blending the concerns a little bit better, I think.

[00:21:51:00 - 00:23:27:23]
Mallory
 Well, everyone, we will certainly circle back on this CoPilot topic. If you are somebody that is an avid user of CoPilot and loves it, let us know on LinkedIn and we'll cover it again. Moving to topic two for today, Meta is catching up. Meta is the company most people had maybe written off in this race. Llama, its open weights model family was free to download and quite popular for a while. And then Llama 4 landed badly in 2025. The frontier moved on without them and Meta spent a year as the cautionary tale. But Zuckerberg's response was to spend an enormous amount of money, hire Alexander Wang, the founder of Scale AI as Meta's first chief AI officer running a new group called Super Intelligence Labs. That bet perhaps paid off on October 12th when Meta released Muse Wonder. You may recall Muse Spark as the model family Meta launched in April of this year. Muse Wonder is a bigger class sitting above it. And Zuckerberg's claim is that this is the one that brings Meta to the frontier, competing with Claude and GPT rather than trailing them. Artificial analysis, which test models independently scores Wonder at 52 against 48 for the latest Spark with coding work improving fast. But Wonder costs about three times as much per task and Meta still recommends Spark as the default for new work. Claude Opus 5.5 as a note scores 58 ahead of both of them. So Meta is closing the gap but hasn't quite passed anyone just yet. Amith, what are your thoughts on Meta coming back into the game and competing with a frontier proprietary model?

[00:23:27:23 - 00:26:23:13]
Amith
 So in the past with the Llama series of models, Meta had embraced open weights. And now they're not doing that. They still have some open weights and open source stuff that they do but not in their core LLMs. So the Muse models are proprietary and the only place you can run these models is through Meta's data centers, through Meta themselves. So that means that you are signing up to use Meta either through an API or through their Muse app. So that's the first thing to note is that you don't have choice here. That used to be the case with Meta's models and it would be easy to confuse the new models with the old when Meta was well known and widely applauded for their contributions to open weights but they've chosen with this new model series to reverse that into go down the proprietary path. Much more like Claude or OpenAI chat GPT which is not necessarily a criticism. It's just an observation that this is not an open weights model which means that you have to be a lot more thoughtful about who Meta is as a company, who runs it, what their values are, what their priorities are and therefore how comfortable you are as a user of those apps regardless of how powerful they are in regards to sharing sensitive data or even having conversations of the models that you don't necessarily want to be incorporated into future Meta products possibly used for targeting you on Instagram or WhatsApp and so on. Now Meta does say that you can opt out of using your data in these conversations for any other Meta products but just like with all other companies you have to trust the people behind that that that is true and that will be true over time. So again, this is not a judgment. I'm simply highlighting that first before I talk about the models power and capabilities because it's really important for our listeners to understand the difference between what Meta is now providing versus what it had been providing over the last several years. So Mallory with that said, I am excited that there is more choice, more choice, more competition that I generally think is always good. And this model is on the level of probably like an Opus 5, maybe an Opus 4.8. So it's a good model that the newest Muse model is quite strong. There are a lot of people saying it's quite capable but it's just yet another model that's quite capable. So it didn't break any records. It didn't really introduce any new capabilities. For Meta as a major technology company, as a major player or desire to be a major player in AI of course it's important, but I would just say we'll see what happens. I don't see anything compelling enough to go and even try it. I have not tried any of the Muse products partly because I personally am not comfortable taking any of my data and sharing it with them. But that's my personal opinion not what I'm recommending for anyone. But I think that it's good that they are aggressively going after this because the more investment and the more choice there is the better it is for consumers generally.

[00:26:23:13 - 00:27:24:00]
Mallory
 Meta technically shipped two versions of that model and then on October 12th, Muse Wonder is on the API at $4 per million input tokens and $20 per million output which is about the same as Opus 5.5 from Anthropik. Muse Wonderfull, got to shout out the name. That's pretty good. Muse Wonderfull, the stronger one has no per token price at all. It isn't sold through the API in any form and the only way to touch it is a subscription to Muse code. Meta's coding agent at $50 or $200 per month. And then on the proprietary note, Amith, Spark and Wonder are both closed models meaning you essentially rent access and the model stays on Meta servers. In August, Meta did release Muse Glimmer which is open and free to download and use commercially. I think it's a smaller model and Meta has also promised to open up Spark. Zuckerberg has said that in August and again in September but that has not happened yet. So it seems like they might be leaning proprietary but at least still toying perhaps with the open models.

[00:27:24:00 - 00:29:12:11]
Amith
 Yeah, for sure. I'm glad you clarified that and the idea that Glimmer is a tool that people could use as an open weights model is great. And again, more contributions to open weights is good. It's just such a small model and relative to all the things they're focused on is not the same as what you're using. So I think it's just really important that people distinguish between them. It's kind of like with Google, Gemini is their proprietary closed model set that you can only get through Google or through their enterprise Vertex service. And in comparison, they have the Gemma models and the Gemma models are their open source open weights models that you can freely download, install and run anywhere. So they're pursuing probably most consistently out of any of the major labs, both in parallel. OpenAI briefly toyed with the idea of an open weights model with the GPT-OSS series back around the same time as when GPT-5 was announced last August but hasn't updated those models since. And to my knowledge, Enthropic has never released anything open weights. So I do think it's good that they're contributing something back to open weights. Now, when you're talking about open weights model, just as a reminder for all of our viewers and listeners, that is where you can take that model and you can run it anywhere. You can run it with any inference provider, you can run it on your own hardware. And in that situation, you're not sharing data with Meta or anyone else. And that's true for deep seek, it's true for Kimmy, it's true for GLM, it's true for all of the open weights models. You can run them at an inference provider of your choice and that completely changes the game in terms of privacy, which is why we here at Blue Cypress are totally comfortable running a wide array of different models within MemberJunction in our open code data platform because those are run at inference providers that we trust and are in locales that we're comfortable with. So that is a really important distinction. I'm glad you brought that up.

[00:29:13:22 - 00:29:42:12]
Mallory
 So, Amith, if someone listening to this is in the market for a Frontier model to use for a project that they're working on, you kind of hinted at this kind of your personal preferences with Meta, but I'm wondering, I guess from a professional perspective, when all of these Frontier models are hovering around the same capability level, how do you recommend that a leader thinks about distinguishing between companies like OpenAI or Anthropic or Meta or Google? How do they decide essentially which to go with?

[00:29:42:12 - 00:32:38:05]
Amith
 The first choice is to make sure you have choice. So instead of building something, if you're building an agent or building some kind of app, make sure that you're not coding directly for one of these companies' APIs. So there's ways of creating a layer of indirection or abstraction, as we like to call it, that allows you to switch easily between companies and models. That's the first choice to make, is to say, "Hey, I want to have choice." Then I would say, "Test them out." So if you think Google and OpenAI and Anthropic might be the best choices or Meta, and you're comfortable with those companies, then test your app, test your agents with a variety of models. That's what we do here at BC. We are constantly testing every major model that comes out, and we test them against a whole suite of benchmarks across all of our agents in our family, including Betty and Skip and Izzy and all these other products, to see how they perform. So every time a new agent comes out, we have this whole set of evals or benchmarks that we run, which tell us how effective the model is, how performance it is, how costly it is, and then we make choices. And for us, it isn't necessarily all alone. We say, "Look, we might support multiple different choices of models at any given point in time." So for example, with Skip, which is our AI analyst agent, we are primarily focused on the Google Gemini model series because we found that for Skip, Gemini just outperforms everything, including stuff from OpenAI and Anthropic that are bigger models and theoretically more powerful. But for what Skip does, Gemini 3.8 Flash is the best model there is right now, and we are constantly testing new models that come out. However, some people want to use Anthropic's Cloud or OpenAI's GPT-6 with Skip, and they're welcome to do so. We have parallel configurations of these agents that are supported, but we test all these configurations to say, "Hey, we know that it works well with these different models from these companies." So I would encourage association leaders or anyone listening to test a variety of different models and you don't need to do this every day and every week the way we do here, but maybe make sure that once a quarter, you rerun some tests with whatever the latest and greatest are. And not necessarily the latest and greatest, by the way, Mallory, as we talk about a lot here on the pod, doesn't necessarily mean biggest. There might be a new model coming from someone that's, let's say, Haiku 5.5, just was released by Anthropic, and it's the new, very small version of Cloud. It's not nearly as smart as Opus or even Sonnet, but it's dramatically less expensive and much faster. And so for certain workloads, it might be great. And it's actually quite a bit smarter than even the last generation of Sonnet, which was kind of the middle-sized model. So as new models come out, including the smaller ones, you should test them out and see, can a small, fast, and an expensive model take the place of a slow, big, and expensive model from a prior generation? Oftentimes you'll find the answer is yes.

[00:32:38:05 - 00:33:11:00]
Mallory
 Hmm, so test out the model against your specific use case. I think that's a really good example with 3.8 Flash. You might think, oh, well, I need the strongest fable out there, but maybe a Gemini 3.5, or 3.8 Flash works just fine. But I think there's also a degree of, I don't know if you want to call it trust of the company that you're working with, but you made a comment, Amith, of we're probably not gonna plug in the latest meta models into our workflow. So kind of on top of that, some level of do I trust that this company is handling my data as they say they are.

[00:33:11:00 - 00:34:25:09]
Amith
 Everyone should make a choice with the vendors that they trust in life and in business. So where do you choose to have your personal email? Presumably if you're using Gmail, which billions of people are, you have some level of trust in Google. If you're using Outlook.com, you have some level of trust in Microsoft. And similarly in business, companies and associations make choices all the time. When you pick an AMS vendor or an LMS vendor, there's a strong element of who are these people and do I trust them as part of the decision, not just features and functionality and cost. The same thing should apply to when you think about both inference providers and if you're using a model provider that also does their own inference like Anthropic or OpenAI to consider those factors as part of your evaluation. And again, also realize that things are moving so fast, you can't do the same type of thing you've done for procurement with AMS and LMSs in the past, you have to be a lot more dynamic, but you can still have the general thoughts asking yourself the question, well, Muse is amazing, let's say in your tests, is it an organization where you're comfortable sending your content and your data? Maybe it is, maybe it isn't, but that's a choice you should make as objectively as you can, but also as clearly as you can, it should be a step that you at least think about.

[00:34:25:09 - 00:36:43:19]
Mallory
 I wanna move to topic three for today, a court has just said that your content has a price officially. So this one needs a little bit of setup, so stay with me, Thomson Reuters owns Westlaw, the standard research tool for American lawyers, you search it to find cases, statutes and precedent, but Westlaw isn't just a pile of court opinions because those are public and free to anyone, its value is what Thomson Reuters adds on top and the main thing is headnotes, so a headnote is a short summary and editor writes of a single point of law inside a case, so a lawyer can find the relevant piece without reading 80 pages, there are millions of them and people wrote every one, at least before AI. Ross Intelligence was a startup building an AI legal research tool to compete with Westlaw and it used about 2200 Westlaw headnotes as training data to do it, Thomson Reuters sued, Ross's defense was fair use, which is the rule and US copyright that lets you use someone's copyrighted work without permission in certain situations like commentary, criticism, parody or research, it's not a clean yes or no, courts weigh those four factors and ultimately decide and as a note, fair use is the argument every AI company facing a copyright lawsuit has made, which is why this case matters far beyond legal research software, so at the end of September, September 29th, the Third Circuit, a federal appeals court ruled against Ross, it's the first time an appeals court anywhere in the US has answered the fair use question for AI training and here is some of their reasoning, Ross's use was highly commercial and barely transformative because it built the same kind of product Westlaw already sells, transformative meaning you did something genuinely new with the material, Ross copied entire headnotes, not pieces of them and Ross ultimately harmed a market for licensing content as AI training data and the court called that market rapidly developing, so I think the third point is the one for us, a federal appeals court has now said there's a real market for licensing your content to AI companies and undercutting it counts as actual harm, Amith, what do you think is kind of the signal from this ruling for associations?

[00:36:43:19 - 00:37:10:11]
Amith
 Couple things, I mean, first of all, I think this is going to get litigated and litigated and litigated for a long time, I also think if you compare the relative size of the parties and their resources between Westlaw and Ross, which I don't know a ton about Ross, but it's a much, much smaller organization with far fewer resources, you're talking about an ant versus a large boot, so the boot stepped on the ant and I think Ross no longer is in business if I remember correctly.

[00:37:10:11 - 00:37:10:24]
Mallory
 That's correct, yeah.

[00:37:10:24 - 00:38:56:11]
Amith
 And so the thing is that independent, and that happened before the verdict came down, so the thing to remember about this is that the law sometimes is used as a tool as part of a business strategy, whether or not ultimately they ultimately prevail. In this case, Westlaw did. My view is that this is actually a really good ruling for associations because it says that the content has to be more narrowly interpreted in terms of what fair use would mean. Fair use as an interpretable kind of idea could allow people to say, hey, we're going to take all the journal articles your association has ever written and use them for other things, but in the case of AI providing essentially a substitute good relative to your association service, I do think it's actually very similar to the West case. So my view is that this precedent adds to the strength of the case associations may have against AI companies. The problem is is that in the case of AI companies versus associations, it's the opposite in terms of who's the ant and who's the boot. So the reality is is that very few associations have the resources to fight a case like this and you're going up against anthropic, open AI, Google, people like that. And so I think that you have to look at this as a factor in your strategy, but not certainly something you can just say, well, the case law now supports our position, we're good. Which of course, most people wouldn't think of it that way, but I think it's really important to have a very practical view of how legal opinion and legal outcomes like this would affect what you should and shouldn't do. To me, the best defense ultimately is protecting your content. That isn't to say that you should take all your content, lock it up in a box and bury it, a thousand feet below

[00:38:57:14 - 00:40:07:10]
Amith
 the earth. My thought is is that you have to be very smart about how you partition content, keeping enough available above surface so that you can attract people into your realm, but then being very careful about where you're allowing your content to go. So you don't allow people to, for example, like I like to say, back up a semi truck and unload it at Mark Zuckerberg's doorstep and say, hey, Mark, here's all our content, we're using the Muse agent. Help us analyze this content and tag it, please. And by virtue of doing that, you've given Zuckerberg all of your content. Do you feel comfortable that you will be able to defeat Meta if they were to improperly use that content later after you've used their service just to tag your content? You have to be very thoughtful about stuff like that, right? So you have to be very, very careful where you use AI services to do things with your content and with your data, because it's kind of like a one-way door. Once it goes out, it's very, very hard to say you're gonna be able to control it. And while the law may be a comforting factor here, I don't think it provides you really the level of defense most executive directors I know would really like to have.

[00:40:08:20 - 00:40:32:01]
Mallory
 And to that point, Amith, I also wanna point out, Ross was not a generative AI system. It didn't write anything. It was a search tool that pointed lawyers toward relevant cases. So it does not settle whether training chat GPT on books and articles is fair use. And several law firms have said exactly that in their write-ups. Ross also is saying to your point, it will ask the Supreme Court to review it. So litigation will continue. But Amith,

[00:40:33:03 - 00:40:53:09]
Mallory
 to your point about the big truck and backing it up to Mark Zuckerberg's place of business, that would make me think as an association leader, you know what, we should just lock everything up. We should put it all behind a member wall. That makes me feel the most comfortable. But there could also be an argument that maybe you shouldn't do that. So can you talk a little bit to that balance?

[00:40:53:09 - 00:41:42:16]
Amith
 Yeah, I think this is also where you have to think very carefully about where your data goes. So when your data leaves your organization, which vendors have access to it can touch it, right? And so if you use a single stack, which is to say, take MetaMuse, the proprietary versions, or if you take chat GPT, Google Gemini, any of these models where the model is provided to you through the same vendor that builds it, you have to think in terms of incentives. So why would Meta or Google or anyone else potentially want to utilize your content in a way that goes against the agreements that you will likely have with these companies? And that would be for one reason primarily, which is to improve their next generation model. So they think that using your proprietary content can somehow make their model smarter, better in some other way.

[00:41:43:16 - 00:45:55:23]
Amith
 I would suspect that there will be strong incentives economically to consider such action either by edict or perhaps by like, let's just say easing certain restrictions that may exist and hoping it happens. I'm not suggesting any one of these companies is actively doing this. I don't have such information. I'm not accusing any individual company of any specific thing. I'm simply saying that after the laws of physics, the laws of economics are perhaps the strongest in the universe. And if there's a sufficiently large prize out there to pursue and sufficiently motivated, hyper competitive people going after that, which I don't think we've ever seen something quite as competitive as what's happening in AI, stuff is likely to happen. And I'd rather be in a position where I haven't put my organization's data at risk, at least to the greatest extent I can. While Mallory at the same time, not walking away everything in a box where it can't be used. So how would I approach it? My suggestion would be to think about the economic incentives of the organization that actually has access to your data. So if you're working with, let's just say one of these Neo Clouds, that's an inference only shop, or really even some of the hyperscalers that aren't very much in the model game like Amazon, as an example, they do some model training, but they're not really in the model training game. Microsoft was in that same camp now as making models. But if you think about someone like a fireworks, right? Or someone out there who is just providing inference across hundreds of models. And if these companies do have a proper zero data retention or ZDR policy, you probably can put more like trust in that policy actually being true. Because since they're not building models, they don't directly have a reason to want your data. Your data is just actually a cost to them to try to log it if they weren't supposed to log it. Whereas the companies that do have an incentive to potentially log, improperly log, and then improperly use your data, generally the ones who would have downstream advantage in their business by doing these things, right? And also sufficiently resourced so that they might not care too much about lawsuits from an association coming their way. So again, this is not an accusation, it's simply looking ahead at what the economic incentives might be. And I think it would be prudent to think very, very carefully about where you use AI models. Again, I am an AI optimist, I also think there's ways to do this that are very safe. I just would think very carefully about the workload, the data involved in that workload, and then where that specific data set's gonna go. And this is also where training is really important. If you have an end user in your organization who can take a zip file of every journal article you've ever written and simply upload that zip file to a free version of Clod or ChatGPT, that has not turned off this switch, which you have to opt into to actually turn off training, you're actually saying to Clod or ChatGPT, you are allowed to use this content to improve future models. That's the default setting in both ChatGPT and Clod in a free account. You can turn it off, then you're in the trust game, but if you have it actually in the default setting, which is probably what most people do with the free account, you're actually saying, Clod and ChatGPT, you are allowed to do this. So the whole you have is, for one, if you are saying, hey, we're a co-pilot only shop, nobody can use Clod, and you have people who are using the free accounts and haven't turned this off and do what I'm saying, which is it's not that big of a zip file to upload all the articles, you can have a problem. So I think it's a mixture of things you have to think about critically and practically to protect your organization. My bottom line on it though, is you have to really be thoughtful about this from a practical viewpoint. If you just tell people they can't use these tools, they're gonna go use the tools. You're better off having a way of using these tools that are safe where you know those settings have been turned on correctly, and of course, to also consider the vendors that you're gonna allow in at all. So those are my high level thoughts on it. I really think this is a key area for associations, CEOs, executive directors, and boards to be thinking about and to set pragmatic policy on this right now.

[00:45:57:15 - 00:46:05:05]
Mallory
 I agree with you. Amith, what would you say out of everything that we've covered in this episode, what do you think is the key takeaway for our listeners?

[00:46:06:09 - 00:46:54:22]
Amith
 I think the key takeaway is that we've never lived in a world that's changing as fast, and we've also never lived in a world where there's as much choice. And so the good news about the fast changing, the crazy pace, is you have a lot of options. So if you look critically at what you're using and make decisions about what it is that you'd like to take advantage of and refresh that decision, I would say at least every quarter, you'll find there's a lot of good options out there. And there's a lot of companies out there who absolutely mean well, that are trying to do a great job and protecting you and your data, lots of choice. But you have to be willing to not do what you've done for decades, which is to pick one vendor and just say, hey, I've checked the box, I'm done with the AI choice, and I'm good for the next few years. That really will not serve you well.

[00:46:56:10 - 00:47:04:03]
Mallory
 I like that. Looking at everything, which is really what we've covered this whole episode with a critical lens as you're making these huge decisions for your organization.

[00:47:05:04 - 00:47:09:09]
Mallory
 Well, everybody, Microsoft fixed CoPilot maybe by admitting it needed

[00:47:09:09 - 00:47:14:22]
 (Music Playing)

[00:47:25:15 - 00:47:42:14]
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.

[00:47:42:14 - 00:47:45:20]
 (Music Playing)