AI Solves an 80-Year Mystery, Microsoft Agents Take Over, & Anthropic’s Claude Mythos vs. Fable | [Sidecar Sync Episode 138]

Summary:

 In this episode of the  Sidecar Sync, Amith and Mallory explore three major developments shaping the future of AI. First, they unpack how an OpenAI model independently solved a decades-old mathematical problem—challenging long-held assumptions and proving AI can generate truly novel insights. Then, they break down Microsoft Build 2026 and the shift toward an “agent-first” world, where AI systems take on real work across your organization. Finally, they dive into Anthropic’s powerful new Claude Fable model, discussing its capabilities, real-world applications, and the growing tension between performance and transparency. Along the way, they connect these breakthroughs back to what matters most for associations: data strategy, flexibility, and staying ahead in a rapidly evolving AI landscape.

Timestamps:

👥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:

Claude Fable & Mythos ➔ https://www.anthropic.com/news/claude-fable-5-mythos-5

Microsoft Build 2026 ➔ https://news.microsoft.com/build-2026/

OpenAi Solves Math Problem ➔ https://openai.com/index/model-disproves-discrete-geometry-conjecture/

Member Junction ➔ https://www.memberjunction.org

LangGraph ➔ https://www.langchain.com/langgraph

CrewAI ➔ https://www.crewai.com

n8n ➔ https://n8n.io

👍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 - 00:00:25:24]
Amith
 My name is Amith Nagarajan.

[00:00:25:24 - 00:00:27:16]
Mallory
 And my name is Mallory Mejias.

[00:00:27:16 - 00:00:38:11]
Amith
 And we are your hosts and hosting we will do today because we have a lot to cover. We've got a lot of craziness in the world of AI. I think we could probably put that part on repeat Mallory.

[00:00:38:11 - 00:00:40:23]
Mallory
 Yeah, maybe we just made that part of the intro.

[00:00:40:23 - 00:00:56:08]
Amith
 Yeah, maybe so. But it's kind of hard not to say because here we are on June 10th, a day after, you know, anthropic drops, mythos, fable, you know, I got to give them credit. They have more interesting names than open AI, but it's still a little bit hard to follow.

[00:00:56:08 - 00:01:18:10]
Mallory
 Yep. And then, I mean, I feel like today's episode as well, we have three topics lined up and they're all around AI, but all kind of in different directions. And also to the model naming point, we're going to talk about Microsoft's family of models, which we kind of touched on last week. And I think their model is my thinking one or something like that. I don't know. I prefer fable in my humble opinion.

[00:01:19:17 - 00:02:31:13]
Amith
 Yeah, I mean, I've said this before. I really think this is going to be true that at some point in time, the AI is going to be kind of like the engine in your car, where unless you're like really a car person, you don't really know exactly what kind of engine you have. And if it's sufficiently powerful and all that and gets good mileage, you don't really care that much. And I think at some point, most AI use cases will become like that, which I'm actually, as much as I am an AI geek, I'm kind of looking forward to that because for most business leaders, actually a lot of this stuff is noise. The trend line is, is that AI is getting smarter and cheaper and smaller really, really, really fast, which means that you can do a lot with it. And that's the broader trend line. And we're going to keep following that obviously here in the sidecar sync and continue to inform all of you on the important things that are happening in AI and try to contextualize it in a way where you can take it home into your work at your association and do something interesting with it. So we'll be getting into all of that as we get going. So excited to do that. And I'm also pumped about what Argentina is doing, Mallory, I know we're not going to cover that today, but I shared an article with you yesterday about Argentina really leaning in hard to the world of AI. What did she think of that?

[00:02:31:13 - 00:02:58:02]
Mallory
 I think it's quite interesting. Amita, I know you have some travel, some upcoming travel plan to Argentina. Argentina is a place, a country that I've researched at length because for a while I thought I would honeymoon there, we ultimately did not, but it's a place that I just have on my radar always as somewhere I want to go. So hearing that they are going all in on artificial intelligence is exciting. And I'm curious to see if you will figure out any more while you're there.

[00:02:58:02 - 00:03:04:14]
Amith
 Yeah, I'm going to dig around a little bit. I'm going to go talk to people and see if, kind of what the attitude is towards AI.

[00:03:04:14 - 00:03:09:15]
Mallory
 And I'll see if-- Good luck with that. And you're going to go to a restaurant and say, hey, let me pick your brain really quick.

[00:03:09:15 - 00:03:30:18]
Amith
 I don't know if that's going to be like weird, because I'm happily married, but it might sound like a strange line or something like that if I do that at a bar. But I don't think so, but you never know. But I'm going to go around talking to random people when I'm there and it'll hopefully be some fun conversations. And I'll report back here on the sidecar sink and let all of our listeners know if I've learned anything interesting.

[00:03:30:18 - 00:04:21:16]
Mallory
 I am looking forward to that. Last week, I actually got to host two webinars, one with Association Latinos about the AAIP, or Association AI Professional Certification, which was a great time. And then the following day, I got to participate in Foresight Fridays, and it was a conversation around AI and member value, which I thought was a great time, good turnout for both of them. I also, of course, I shamelessly plugged the podcast only a few times in each one. So we may have some new listeners from either one of those webinars. If so, welcome. But it was really exciting to see so many Association professionals coming together and really excited about the conversation. You know, I guess you never know how it will go, but I felt like I was channeling my inner Amiith for the Foresight Fridays. I was like, what would Amiith say to this question? I think it went pretty well, though.

[00:04:21:16 - 00:05:43:16]
Amith
 That's awesome. Well, you know, over the last week or so, since we last recorded Mallory, I've had a number of very interesting conversations, some deep dives with some Association leaders. I had a group of executives from our large Association in Washington come down for a day of really just like, deep thinking and collaboration down here in New Orleans, where we spent the day together and had a nice dinner. And we talked through a whole bunch of different ideas of challenges that are facing their organization culturally, opportunities, ideas, obviously with AI. And then I've also had a whole bunch of conversations in the last week with people that are half an hour, type length, and really actually I find those to be really helpful. Hopefully they're very helpful for the folks on the other end of the line, but I find them very helpful for me because I'm learning in real time where people are struggling, where people are succeeding. And what I'm generally seeing is the same thing you just described. There's a tremendous amount of enthusiasm. There's certainly caution around security and safety, which I continue to urge associations to think very critically about this, but there's ways to do that with AI. There's ways to keep safe and also progress very rapidly with artificial intelligence. But generally speaking, there is this overwhelming enthusiasm and optimism about what AI is gonna do in their sectors and for their associations internal ops too.

[00:05:43:16 - 00:08:31:00]
Mallory
 Yeah, very exciting. Having all these conversations with association leaders is very encouraging and inspiring, and we're gonna do more of that on today's episode. So today we're starting with AI that cracked a math problem that stumped people for 80 years on its own. Then we're gonna be talking about Microsoft Build, where the whole event was about handing your work over to agents, no surprise there. And we finish with Anthropic's Fable 5, a model so capable they had to make it safe before they could release it. Honestly, I feel like we saved the best for last year, so stay tuned for that conversation. But topic one, back in episode 84 of this podcast, we actually covered Google DeepMind's Alpha Evolve breaking a 56 year old record by finding a faster way to multiply four by four matrices, beating the mark set in 1969. This is kind of the next chapter of that story. So a model from OpenAI just resolved a famous problem that had stumped mathematicians for the better part of a century, and it did it autonomously, so without humans working through the equations. I'm gonna give a high level of this problem. I am not a mathematician, but the puzzle is one of Paul Erdosch's favorites, the unit distance problem. So in plain terms, if you scatter a bunch of dots on a page, how many pairs of those dots can be exactly one unit apart? Erdosch conjectured in 1946 that arranging them in a grid was about as good as it gets. That makes sense to me for sure. The model though didn't confirm this conjecture, it disproved it, finding an arrangement that does better than the grid. Most humans who'd attacked the problem were trying to prove Erdosch right, not prove him wrong. OpenAI had handed the problem to an internal model as a test of its abilities. The researchers initially didn't believe the results, so they searched for errors, had outsiders verify it, and checked the work with their own coding agent before accepting it. The Wall Street Journal offers three reasons the AI succeeded where most people didn't. The solution was deeply counterintuitive. The model synthesized across fields humans usually keep separate, pulling from both algebraic number theory and discrete geometry, and it simply had the patience to stick with an approach a human probably would have abandoned. One detail that lands well out loud, the model's chain of reasoning, even in abridged form, ran more than 75,000 words. About the length of the first Harry Potter book, a former OpenAI researcher estimated the whole thing took under 32 hours and roughly $1,000 in tokens. So we talked about Alpha Evolve and episode 84 as algorithm discovery. This feels like a step further. AI disproving a long-held belief on its own. What is exciting to you about this?

[00:08:31:00 - 00:08:48:16]
Amith
 I wanna highlight a couple things. Well, first of all, just listening to you describe that, I feel smarter already today, so it's exciting. Really? I'm not a math guy. I'm not a math guy. I'm pretty good at computers, but I'm not a deep math guy. I like math, but I'm not math at this level.

[00:08:49:20 - 00:11:17:15]
Amith
 Anyway, the thing that I wanna point out is the counterintuitive part. So you mentioned that it was deeply counterintuitive, also that it's synthesized across many fields. I think a lot of people assume that AI is good at this synthesis to be able to reason across all these different vast areas of knowledge and bring together insights, which I find stunningly exciting when I work with AI to have that level of knowledge available. But the counterintuitive part is, I wanna double click on that. So a lot of people have been saying that essentially AI thus far is essentially a stochastic parrot, AKA a probabilistic parrot, something that only repeats back to you, that which it has learned. So what they're saying, this is kind of an old argument now, but a lot of people, well, I hear AI speakers saying this all the time, the stochastic parrot line, this sounds kinda cool, sounds really smart actually when you say it like that. But it's been false for a while, especially like starting pretty much with strawberry, which became O1, that was the first time models were able to have the backspace key and correct their own work. They weren't simply predicting the next token, they were actually looking at their work, correcting it, hypothesizing, testing ideas. And it's actually been a false statement the entire time we've had agentic loops or agentic iterations, things like open-claw popularize, but member junctions agent system works the same way, the line graph works in a similar way where you can essentially have something loop and iterate in order to basically continue to test different ideas. The point I wanna make though, is that this model within the model itself was able to reason across these different, totally divergent areas of mathematics and form a counterintuitive hypothesis and then prove it. And that's very interesting because it basically undermines this whole idea of AI only parroting back to us, that which we have already created. So if AI is now creating novel contributions to the body of work that is collectively what humankind has produced, that's really interesting, right? And we're seeing this happen, we saw this happen with alpha fold in a very narrow domain. As you pointed out, Google had a very interesting thing earlier in the year. And, but this particular breakthrough I think is yet another stunning one that showcases that even in a fundamental field like mathematics or physics or biology, an area like that, that AI can have a novel contribution. That's the first thing that really sticks with me is that particular part of what you discussed.

[00:11:18:20 - 00:11:26:24]
Mallory
 Do you think for an association, there's a version of give the model a really hard question and let it run and see what happens? What would that look like?

[00:11:26:24 - 00:11:47:14]
Amith
 Yeah, a great example is to look for a paradox in your data. So one of the things we like to do is say, well, we make a lot of assumptions. I'll give you a classical one. Associations assume that people who come to their events are close to those events. They'll say, well, if we have the event in Nashville, we're probably gonna have people in the Midwest and the East Coast come.

[00:11:49:00 - 00:11:50:09]
Amith
 Maybe, maybe, maybe not.

[00:11:51:10 - 00:12:05:00]
Amith
 They often find that when they analyze the data after the fact, and this is usually done qualitatively, not quantitatively, they'll find that there's people from all over the place that came there like, oh, wow, we were so surprised that Nashville attracted people from Germany and from Australia and whatever.

[00:12:06:07 - 00:14:48:04]
Amith
 And then they'll hold an event in a place that has a lot of their members, whatever that particular nexus may be, and they might not have as good of an attendance. And they're like, well, this is really strange. We held the event there because that's close to where a lot of our, where our members are. And so these counterintuitive things that happen, they're like little mini blips on the radar that occur in our business. What we wanna do is have AI kind of automatically start to guide you down exploring those pathways. So an example of that is with Skip, our AI data analyst, we've trained Skip to look for these patterns. An example is recently we were looking at some data and we said, hey, Skip, do an analysis, actually this exact problem of, analyze the geographic proximity of attendees to all of our events over the course of time. And we wanted to do it with some degree of accuracy. So the geographic data in various databases like your event system and your AMS, usually it's sparse or non-existent, meaning you might have like a city and state or a country, but it's inconsistent. So it's very hard to analyze quantitatively and say, well, we have like four different bands, people under a hundred miles away, people between a hundred and 500 miles, people that are say 500 to 2000 miles and above that, then just making up these bands. But you can't really do that with just city, state, zip or country. You need to geocode the data. Geocoding is a process that you essentially use a form of an AI actually to take in your unstructured data and turn it into latitude and longitude from which you can then compute very easily the distance between any two points in the globe. So the point though, and that's fairly straightforward, but the data is non-existent. So first of all, you have to geocode your data. The good news is there's a way to do that actually. Member Junction has this built in, but there's lots of ways you can use data enrichment tools to get latitude and longitude for your data. And then once you have that, you can have the AI go to town and say, hey, like I'm gonna do all these different hypotheses and I'm gonna try to look to see if there is a correlation that supports our prevailing belief, right? Our understanding, our theory of operations or undermines it. And oftentimes you find that there's some counterintuitive results that say, well, actually this event, maybe for our annual meeting, that assumption is true. Maybe for some other meetings, it's the inverse to that because people look at that meeting as an opportunity to bring their families and therefore they actually kind of like it when it's destinationy. Whereas this other meeting we do is much more business focused. And so people don't really wanna be in Orlando or some other destination city like that because they'd rather be close to home. So it's a shorter trip and so forth. So there's all these interesting things that come up. Another classic one is engagement predictions. People are saying, well, the more people are engaged, the more likely they are to renew.

[00:14:49:12 - 00:15:58:11]
Amith
 Well, is that true? One would think that it is. The more services and activities and things that people do with the association, the more likely they are engaged. And I would say that in my experience looking at the data, it generally is true, but not universally true. There are nuances to that. For example, if someone has been to every single event you offer year after year after year, maybe after a while they get bored, right? Maybe they haven't done anything different with you. And so it's not just engagement, it's variety of engagement or it's engagement that they find emotionally satisfying and how do you get those signals? So these are the kinds of things AI can start to help you kind of tease out and ask harder questions. It's not even harder questions, Mallory. It's asking questions that we wouldn't think to ask. I've trained my brain to ask some of these questions, but in my own business, I don't ask a lot of the most important counterintuitive questions. So I find that very exciting. You can find something, it's like this gem that's sitting under a rock that's right in front of you, you just have to turn the rock over basically. So that's the thing I would point out, that's a very practical implication of what this particular tool has highlighted in the field of mathematics, bringing it back to the world of associations.

[00:15:59:23 - 00:16:17:11]
Mallory
 That's helpful. With the location example for the annual meeting though, it sounds like the precursor for solving this problem is having a good amount of data collected on who's attending. Are they bringing a guest with them or not? Making sure you kind of have a full picture of this member. Do you agree with that?

[00:16:17:11 - 00:16:37:24]
Amith
 Yeah, if you don't have your data unified, it's really hard to do some of the stuff that I'm talking about. If you have your data scattered about in 10, 15, 20 different systems, the good news is there's relief for that. There is a free open source data platform purpose built for this exact reason. It's called Member Junction. It costs $0. It's free for the entire association community

[00:16:39:01 - 00:16:51:19]
Amith
 and anybody can use it quite easily these days. So that's why we built that particular tool. But there's other ways to do it of course, you can use all sorts of different data tools. This is just one that was built for this community and the data sources it commonly deals with.

[00:16:52:20 - 00:17:25:15]
Amith
 But you do have to get your data house in order, otherwise it's hard for AI to do this. But more broadly speaking on the point of this particular piece, Mallory, I think when you work with AI, rather than kind of reconfirming your own biases and asking the AI to work in your swim lane, how about asking the AI to tell you what's wrong with your position or alternatives? Some of you probably do that, but I think AI as a thought partner, rather than just a worker bee, can be quite powerful because then it harnesses this breadth of knowledge and the synthesis that became evident in this particular mathematical breakthrough.

[00:17:25:15 - 00:19:54:10]
Mallory
 Mm-hmm, I agree with you. Sometimes it's hard to do that when you've been working so diligently on a project and then asking the AI basically rip it to shreds, but I practiced a version of that as I was prepping my outline for the Foresight Fridays conversation I mentioned. And I collaborated with Claude on this outline for weeks essentially. And then when I was done, asked it, "Okay, if you were attending this session and you heard what I had to say, could you walk away with something practical and then had it do a final pass of, okay, here's where you really need to make the tweaks?" And you could do that over and over and over and just think about how much better your work would be. So I agree with you. Moving to topic two for today, the agent first shift at Microsoft Build 2026. So Microsoft Build is their annual conference. It ran June 2nd and 3rd in San Francisco. And the whole event had one through line, a shift from software you operate to agents that do the work for you. Sounds very similar to what we heard coming out of Google I.O. recently. Microsoft is calling it agent first. The framing worth pulling out, I think, for our audience is this idea of ownership, not access. Microsoft's argument is that as AI models get more capable and more available, the thing that will set an organization apart is no longer access to intelligence. It's owning a system built on your own expertise, your own data, your own way of working, rather than one that funnels value back to a consultant or the model maker. The product behind that idea is a set of context layers called Microsoft IQ, now generally available across their developer tools. The most relevant piece is Work IQ, which captures how work actually happens across an organization, the people, emails, documents, meetings, and how they all connect. The most relatable demo was Microsoft Scout, an always on personal work agent that lives in the tools you already use, like Teams and Outlook, and proactively handles things like meeting prep and scheduling conflicts without being asked. On the model side, Microsoft released a family of seven new in-house models, all under the my name, MAI. Worth a quick word on why Microsoft is even shipping its own models now, because for years, its OpenAI partnership effectively kept it out of the foundation model game since Microsoft was leasing OpenAI's technology and had little room to build a direct competitor. Two rounds of restructuring that deal loosened those terms and Microsoft stood up its own team, AI team, under Mustafa Suleiman to go independent.

[00:19:55:12 - 00:20:53:00]
Mallory
 These seven models are the first real proof of that shift. They break into a few clusters, a reasoning model, my thinking one, two image models, a couple of voice models plus transcription, and a coding model. The headline one, which is my thinking one, is a mid-size reasoning model that Microsoft says independent raters preferred to Claude Sonnet 4.6 and that matches Claude Opus 4.6 on a coding benchmark. The single point that holds all seven together is the real takeaway. The era of one model for everything is giving way to a portfolio you pick from per task. Also a quick throwback for the long-time listeners. In episode 71 on this podcast, we covered Microsoft's Majorana Quantum Chip. And at this event, they announced Majorana 2, claiming a thousand times higher reliability than the previous generation and a path to a million qubits on a single chip with a scalable quantum machine targeted by 2029.

[00:20:54:04 - 00:21:04:09]
Mallory
 So Amith, I summed it up as fast as I could. There's a lot coming out of Microsoft Build. What do you think is most exciting or most relevant for associations out of all of that?

[00:21:04:09 - 00:21:16:23]
Amith
 Yeah, the folks up in Redmond have not just been sitting around. They've been moving fast, doing cool things. I think the Quantum Chip is super exciting. We did cover that probably about a year ago, I think it was.

[00:21:16:23 - 00:21:17:24]
Mallory
 Yeah, I think so.

[00:21:17:24 - 00:22:48:24]
Amith
 I find that fundamentally just very exciting. That opens up a whole new dimension basically in computing generally. So that's exciting to follow. We'll keep following along with that as Microsoft and others pursue quantum at scale. To me, the most practical and useful thing is the fact that Microsoft's getting into the model game. So they're obviously a major tech company. They haven't had models of their own other than they're very small models, the Microsoft Fi models, PHI. And I haven't seen an update for those since the series four of those about this time last year. They're very good small models, but they didn't seem to be particularly interested in commercializing them. And there was an agreement, as you mentioned, with OpenAI where Microsoft was precluded from building models above a certain size, and that is no longer the case. And so Mustafa Saliman, and for those that aren't familiar with his history, he was actually one of the co-founders of DeepMind back in the days of the early 2010s, worked with Demisa Sabis and others to found that company. So very, very deep AI guy, and he's the head of Microsoft's AI efforts. So that's exciting to have someone of that caliber leading that effort and driving forward a major organization. So the seven models they have right now, I think are good. Microsoft does want their GitHub unit with their co-pilot coding product to be back in the gate. GitHub co-pilot was the coding assistant that started it all, so to speak, back in the early, early days, going back years.

[00:22:50:03 - 00:23:50:17]
Amith
 And it started off as just smart auto complete for the code writer, and then it's evolved. They do actually have a CLI version of co-pilot, which is much like Cloud Code, but it just hasn't been that good. And it's been dependent upon other people's models. Far and away right now, the most profitable use case in AI is around coding. And so Microsoft's trying to position GitHub to work on in-house tech, so that their gross margins essentially are palatable to scale that business and to be competitive there. And I actually think they're gonna be a competitive player in the game. Microsoft is deeply embedded in the developer stack with Visual Studio Code, with GitHub as the dominant source code repository in the world. They have an opportunity there to really leverage that, to build better coding tools and to give Cloud Code and Cursor and others a run for their money. So I find that exciting. And then just in general, the more models right now for consumers, the more competition, the better it gets. And competition not only means lower cost, it also means more innovation. So I find that exciting as kind of the headline coming out of build.

[00:23:52:02 - 00:24:32:12]
Mallory
 And to that point of me, I wanna talk about model fatigue because we cover a lot of models on this podcast, which is great. And we always talk about how important it is to build flexibility into your infrastructure so that you can use the best model for your specific use case. However, I feel like with so many models at this point and the fact that most of the major AI companies are creating, they're matching each other within at least a couple of months in terms of model capability. What do you think about the idea of an association just saying, we're gonna commit to this AI company or to Microsoft or to Anthropic because we believe in it and this is what we've always used and we're just gonna stick with the family of models they come out with. What do you think about that?

[00:24:32:12 - 00:24:34:17]
Amith
 I think that's a deeply flawed approach.

[00:24:35:18 - 00:26:52:10]
Amith
 And the reason is, is I think right now more than ever, people need choice in their life, they need optionality. They need the ability to be vendor independent. So as much as I'm excited about Microsoft's advances and their agent tooling and the IQ framework they have, I'm also excited about what Google is doing. I'm also excited what Anthropic and OpenAI are doing and countless other labs that are producing amazing technology, both at the model level and at the application level. And no one vendor will fit all of your needs is the most likely outcome. And the best vendor for your needs will likely change even month by month, but certainly year over year. And so if you anchor yourself to just one vendor, that's going to create a dependency that you really don't wanna have. There's no need to have that dependency. It's not like an AMS where you need to be closely coupled to a single vendor. You can play the field a little bit here. The way you do that is you have to have a framework of your own that you own where your data, your expertise is in fact owned as you pointed out earlier. That's a really key concept. I just don't think Microsoft should own it. I don't think Google should own it either. I don't think anyone should own it other than you. You're the association, it's your data. It's high time that you have ownership over one of your systems. And this is going to be the most important system ever. This is gonna be the system that you run your AI agents from. This is gonna be the system that you run your AI workloads from. You need to have control over that. So there's many ways to do this. I already mentioned Member Junction once in this episode, so I don't wanna belabor that point. That's one way to do it. You can also use tools like Langraph. You can use tools like Crew AI. You can use N8N. You can use a number of independent agent frameworks that allow you to orchestrate agent workloads independent of a particular vendor and have optionality. So I think it's exciting, but I would still urge associations to look to maintain a unified architecture, but one that allows them to plug in any vendor they want, not just a particular vendor. Now we'll say Microsoft offers a lot of convenience. When you're the incumbent and you control the office workspace for many, many clients, associations predominantly are office clients. There's some that are Google workspace people, but most are office Microsoft 365 folks. All your documents, all your email, all of your stuff is with Microsoft. So does it make sense to use Copilot? Maybe,

[00:26:53:12 - 00:26:56:20]
Amith
 that's kind of with you inside Word. It's with you inside Teams,

[00:26:57:20 - 00:27:36:15]
Amith
 but you can also plug in external agents into those frameworks. But even if you use Copilot, because it's useful in those productivity tools, that doesn't mean that Microsoft is the only way to go when it comes to enterprise level agents and your agent architecture. So to me, this is not actually about model fatigue. It's more about thinking ahead about what's likely to happen, which is even you might be tired of hearing about new stuff, but I kind of think it's too bad. I kind of get some model fatigue myself at times, but it doesn't matter. It's just gonna keep changing. And if you want to be positioned to take advantage of something that's contemporary and works well, you've got to have optionality.

[00:27:36:15 - 00:27:55:11]
Mallory
 I hear you. I just think for the average association leader, keeping up with the capabilities of all the companies and the models would be challenging. So what do you think about an association building an independent agent framework, having that plug and play flexibility, but just opting to use, to plug in models? Totally. Okay.

[00:27:55:11 - 00:28:02:04]
Amith
 Yeah, and totally. So like another way to think of it is, have you guys bought a TV recently? Like when you moved, did you get a new TV?

[00:28:02:04 - 00:28:05:16]
Mallory
 We didn't, we want to get an OLED, but we did not buy one yet.

[00:28:05:16 - 00:30:59:12]
Amith
 Well, if you go to Best Buy, or I don't know who else is a retailer these days for electronics, but. Costco. Oh yeah, of course. Costco. So if you go to one of these places that has a lot of TVs, I don't know about you, like I love TVs. I think big TVs are super cool, but like I honestly cannot tell the difference between the vast majority of these things. Like I walked into Best Buy a couple years ago to buy a TV and I looked around and I'm like, well, here's an $8,000 thing and here's a $800 thing. And they both, to me, to my eyes, they both look really good. Now, a lot of our listeners who are video files or audio files and are really into this stuff will tell me, no, no, no, it's not true blacks and the display resolution isn't as good and all that. But you know, once you get to a certain point for like an average consumer like me, it's good and it's good enough. And AI models are much the same way where there might be a Best Buy full of great AI models. Some are crazy expensive and some are just fine. And if you have a high, high precision need, maybe you need the $8,000 TV, which would right now be like the Claude Fable 5 or the latest Gemini 3.5 or whatever. But for the average use case, you don't need the OLED with the blackest blacks and the richest, brightest colors and the biggest screen or whatever. You just need something that works. So to your point, just finding models that work well for your use case and largely sticking with them is fine. So long as you're sticking with them because they're the right cost and the right capability, not because you're stuck and that's where optionality comes in. So I would say at a minimum once a year, to look at the models your agents are using and say, hey, like, okay, we're using GPT-40 on stuff. Well, why are we still using that? That's like a two year old model. It works, but it's more expensive than GPT-55 Mini and GPT-55 Mini is three times as fast and half the cost and twice as intelligent. And there's a lot of people who have systems like that, that get stuck, just like, you know, you get stuck in an old AMS. The difference is this is much, much easier to stay up to date with. My example there was OpenAI to OpenAI, but what if you were on, you know, GPT-40 from OpenAI and you're like, whoa, I hear Gemini 3 Flashlight is a really fast, really capable small model and super cheap and is way smarter than GPT-40 was in its heyday. So let's switch to that. We'll save a bunch of money. We'll be able to process our stuff faster. That sounds good. But if you built on OpenAI's agent framework, guess what? You're out of luck. So that's the point is that I'm not suggesting that people churn models. I mean, we try out models literally every single day here, but that's our job. We're an AI lab focused on applied research. So we have to do that. Our job is to make that easier for our clients, but ultimately associations, my whole point is that you just need choice. You don't know what the best model is going to be for you. Most of the time, the TV you had last year is probably fine, but if you're paying three times as much for it now, and it's a TV that's just as good, why not upgrade, right? If it's gonna save you money or it's just better, right?

[00:30:59:12 - 00:31:11:24]
Mallory
 Yeah, okay, I think that makes sense. So building with that flexibility in mind and then holding yourself to maybe once every six months, once a year, looking at the models that you're using and making better choices if there are better models out there.

[00:31:11:24 - 00:31:37:04]
Amith
 Totally, and just one more thing on the TV metaphor is that like, I was in college, I thought I was really cool because I had a color TV, so that dates me quite a bit, but it was a color TV, it was a CRT, it was like this big. Sony, it wasn't actually very big. I mean, the box was big, but it was this plastic, giant plastic case around a Sony TV. I think the screen size was maybe 18 inches or 20 inches. It was actually, by today's standards, microscopic.

[00:31:38:12 - 00:31:41:12]
Amith
 Now that was the coolest thing in the world. And I kept that TV for like 10 years.

[00:31:42:14 - 00:32:10:11]
Amith
 And it was cool, but after a while, I'm like, wait a second, I can get something like this thing, just moving it, like my medical bills for hurting my back are starting to be more than the cost of a new TV. So I guess that's the kind of thing that happens to associations a lot with technology. Don't let that happen to you. And there's no need for that to happen. With AMS's, I empathize, because those are hard to change. But with this technology, as long as you set yourself upright to begin with, you don't need to have that problem in the future.

[00:32:10:11 - 00:32:21:01]
Mallory
 I'm laughing about the TVs. I don't remember a time where TVs weren't in color meat, but I did have TVs where they had the built-in VHS player. So, you know, I can understand a little throwback. That's a pretty cool feature. Yeah, I enjoy that.

[00:32:21:01 - 00:32:21:21]
Amith
 That would still be useful.

[00:32:21:21 - 00:32:38:12]
Mallory
 Oh, in terms of quantum computing, I just have, I'm always interested on this topic because I have a feeling over the years, we'll probably be talking about it a lot more on this podcast. Do you think that's something our listeners need to be stopping and focusing on right now, or just something to keep kind of in the back of their mind?

[00:32:38:12 - 00:33:32:14]
Amith
 The way to think of quantum is the way to think about energy breakthroughs that they're right around the corner. They haven't yet directly impacted our day-to-day, but they will. Like, not only in our lifetimes, that's what we used to say about major scientific advances, that, hey, hopefully in our lifetimes, there'll be a cure for cancer or something like that. But now what we're saying is, like, look, in the decade, we're gonna have major energy breakthroughs where the cost of energy is gonna plummet. We're gonna have major computing advances. We're gonna have like a million-fold advance in capabilities, right? We talk about, Nvidia has been setting a crazy pace of compute advance every single year. They're advancing by like 10X, 20X, which makes the Moore's law curve from prior decades seem quaint in comparison. But this acceleration, when we get to a step change like quantum, it's probably like a multiple step change, right? Because it's a completely different paradigm from computing, and it's gonna make certain problems that right now are still out of reach with all the compute we've got.

[00:33:33:14 - 00:34:28:13]
Amith
 Totally within our hands. A lot in the scientific research realm, but also just generally AI will be a really great candidate for many aspects of quantum. So that's one thing. I think these exponentials, we think about energy curves and compute curve. Quantum is just another tool in the tool belt, is the way to think of it, but it's this like mega tool that makes our current compute infrastructure look like hand tools, as powerful as they are. So it's gonna be like another supercharger on top of what we've already got for AI. It's gonna be kinda nuts. So keep it on your radar. I don't think anyone needs to go study it unless you find it interesting, but it's kinda like studying what's happening with fusion or what's even happening a little bit nearer term to that. In my opinion, it's like small modular reactors for fission, which is a new type of nuclear reactor that's actually being deployed right now, that's gonna change the energy game in a very meaningful way. So these are things that are like right around the corner on the horizon in terms of the next several years, but don't affect you just yet, but they're extremely exciting.

[00:34:28:13 - 00:34:32:03]
Mallory
 We won't be launching a quantum learning hub anytime soon to me, right?

[00:34:32:03 - 00:34:52:09]
Amith
 Not quite yet. And I'm hoping that by the time we see that come to commercial reality, that our world of associations is so well-versed in AI, that they're gonna be able to just jump on the quantum wave when it comes. And it'll probably be kind of automatic because it's gonna just affect certain things that we do kind of seamlessly in the background is the way I see it happening.

[00:34:53:06 - 00:35:16:05]
Speaker 9
 Digital Now 2026 is the AI and digital strategy conference built specifically for association leaders. Join us October 25th through 28th in Washington, D.C. for four days of world-class keynotes, hands-on sessions, and a plan your team can implement immediately. Register now at Sidecar.ai backslash Digital Now and use code "DNOWPOD" for $100 off your registration.

[00:35:16:21 - 00:35:19:23]
Speaker 5
 Are you ready to become the AI leader your association needs?

[00:35:21:11 - 00:35:47:13]
Speaker 5
 Sidecar's Association Artificial Intelligence Professional Certification, or AAIP, is your fast track to mastering AI skills built specifically for the association world. Whether you're just getting started or you're already experimenting with AI, getting your AAIP will give you the practical tools, real-world examples, and hands-on training to boost productivity, enhance member value, and confidently guide your organization into the future.

[00:35:47:13 - 00:35:52:14]
Speaker 6
 And right now, listeners of this podcast get an exclusive discount.

[00:35:53:18 - 00:36:06:07]
Speaker 6
 Visit sidecar.ai.ai.p and use code AIPOD50 today for $50 off your year-long pro-level subscription. That's AIPOD50.

[00:36:07:22 - 00:36:13:01]
Speaker 6
 Become the AI savvy leader your association deserves with Sidecar's AAIP certification.

[00:36:13:01 - 00:36:39:06]
Mallory
 Our mission at Sidecar is to educate one million people in the association community around the world on artificial intelligence. And the good news is you can help. If you're enjoying this podcast, please take on the challenge to share it with one friend or colleague each time you listen. This will help us spread the word and hopefully bring one million association leaders and volunteers into the age of AI.

[00:36:39:06 - 00:38:44:22]
Mallory
 well, it's time for the topic everybody's been waiting for. This one is hot off the press, and Tropic announced it just yesterday, the day before recording this episode. They released two models at once with the relationship between them being kind of the whole story. So first, let's talk about Claude Fable 5. It is the public model, a frontier tier model. Anthropic says they made safe for general use. We also have Claude Methos 5, and it's the same underlying model with the safeguards lifted in certain areas. And it's restricted to a small group of cyber defenders through their project, Glasswing Program in collaboration with the US government. So the same brain, the guardrails are the only difference, which is why they get two names. Anthropic says that without safeguards, Fables abilities and areas like cybersecurity could be misused to cause serious damage. So when the model detects a question touching cybersecurity, biology and chemistry or model distillation, it quietly routes that response to their previous model, Opus 4.8, and it does tell the user when that happens. On capability, the pattern Anthropic emphasizes is that the longer and more complex the task, the bigger Fables lead over their other models. Their example, this was during early testing, Stripe used it to do a code migration on a 50 million line code base in a single day, work that would have taken a team over two months by hand. And the benchmark story is also striking. Anthropic says Fable is state of the art on nearly all tested measures of AI capability, and the gap widens the harder the task gets. A few that land out loud on a tough software engineering test called the SWE bench pro, it jumped to around 80% from roughly 69% with Opus 4.8, so that's an 11 point leap. One Frontier Physics research group reported it matched a leading competitor's four day result in about 36 hours while burning a third of the reasoning tokens. And on vision, it can rebuild a web app source code from a single screenshot.

[00:38:45:23 - 00:39:56:01]
Mallory
 Ethan Mollick, someone we talk about often on the podcast, he's an AI researcher and professor at Wharton. He had early access and tested the model on everything except security, and his central reframe, last year he described using AI as working with a wizard, you cast a spell and something happens. With Fable, he says it's closer to a patron who commissions work. So he briefs the model, pays for it, and then judges the result, but the actual work happens somewhere he can't watch. In his words, he no longer steers, he commissions. Mollick backs that up with what he watched it do. It worked up to a dozen hours on multi-page specs. One project ran nine and a half hours straight. On another, he watched it spin up its own cheaper sub-agents to research over 2,200 flights, plus rail and road data, then launch more agents to test its own code. The tension that Mollick raises, and I think one that is most relevant to discuss on the pod, is the flip side of that power, how little he did, and how little he could see. The model made hundreds of small judgment calls he never got a vote on, and he frames the black box as possibly being the price of capability.

[00:39:57:08 - 00:40:40:23]
Mallory
 Speaking of price, Anthropic lists both models at $10 per million input tokens and $50 per million output tokens, which is twice the per token rate of standard Opus 4.8. On access, the subscription rollout is staged, Fable is included free on Pro, Max, Team, and Seat-based Enterprise plans from launch through June 22nd, and then it will move to usage credits with the aim of folding it back into plans later. So folks on those plans can try it out for free for about two weeks. And on a privacy note, Anthropic is requiring 30-day data retention on all traffic for these top-tier models. It says it won't use the data for training or anything outside safety, and it deletes it after 30 days in almost all cases.

[00:40:42:02 - 00:40:51:17]
Mallory
 So Amith, you said you got to test it out. I have not tested it out yet. I told you, I don't even know what to give a model like this, but what are your initial thoughts on Cloud Fable?

[00:40:51:17 - 00:41:18:22]
Amith
 It's as good as people are saying it is. That's the short version, and it's worth checking out for your most complex workloads. So I threw something at it yesterday that Opus 4.8 was working on, but struggling with, and a very complex software plan for building out a capability. We're right now adding something called a co-agent capability to the Member Junction framework. The way that works is basically any agent inside Member Junction,

[00:41:19:24 - 00:42:32:18]
Amith
 a lot of times these agents do long-running tasks. They'll work for seconds or minutes or even hours to do complex things for you. We want the ability to do real-time co-agency, which essentially means being able to have live audio voice conversations with any agent in MJ, as well as potentially down the road, having video conversations, being able to sketch out ideas on interactive two-way whiteboards that the agent can draw on, and you can draw on, stuff like that. That's how we see the future of this going, but it is something that requires a different category of models, something called a real-time model from either OpenAI or the live model from Gemini, and other models of that class will come from other vendors. Those are the two that right now are best. And so this is a big, big project to add that into a generalized agent framework like MJ, and make it work automatically for all existing agents, whether those agents are things like from our agent catalog, like Betty or Skip or Izzy, but also agents that customers or users have built on top of MJ automatically getting a voice. So it's a non-trivial thing. So we had a super detailed plan, and Opus was working and getting some stuff done, but then I gave the whole project to Fable, and it worked for, I think, four or five hours while I was in a bunch of meetings,

[00:42:33:21 - 00:43:47:17]
Amith
 and it came back and it had a working version. And I went and looked at the code. It was very well architected. It stuck with the plan. Opus, a lot of times, will veer off a little bit here and there, and you have to kind of give it some steering. Fable really nailed it. So it's definitely an improvement from a coding perspective, and that tends to be some of the more complex work we can throw at these things. I would say at the same time, Google's new Gemini 3.5 Flash, which is dramatically less capable than the Fable model, and actually somewhat less capable even than Opus, is good enough to do most of the things we do at our team for coding. So I guess it's this paradox of, yes, for the most complex things, you end up with a choice of using the most powerful model, the best TV at Best Buy, right, versus most of the things you're doing, you don't need that. You might get by just fine. Using a car metaphor, it's like, I don't need the fanciest, fastest car for everything. I might be totally fine in a Corolla for most of the stuff that I do. So I think it's just an interesting time to be thinking about this. I think people should get familiar with it. I love their release strategy, by the way. It's like, hey, it's free for two weeks, you can try it out. It's kind of like, in the software industry, you call that the drug dealer strategy, right?

[00:43:47:17 - 00:43:52:04]
Mallory
 I was just thinking about drugs, like here, have a taste, and then you're like, and now you pay for it.

[00:43:52:04 - 00:44:29:19]
Amith
 Here you go, do you like it? It's like, all right, now it costs a whole bunch of money. But I think if you kind of look at Fable as it's another tool in your tool belt, it is going to be matched very soon by Google and OpenAI, undoubtedly, I'm sure. In the next week or two, you will not hear from one or both of those labs having something cool to release. But the bottom line is, is that this stuff is advancing really fast. The capabilities that you have tomorrow should make you rethink what you believe you have to do. So going back to Malik's comment of going from a wizard to going to a, what did he say, a commissioner of work, a patron.

[00:44:31:00 - 00:44:42:11]
Amith
 Ultimately, that is really what it is. I would say there is, the downside to it is that, that opaqueness of how the model got to the answer. Now with code, there's less of a problem because if you're a coder, you're inspecting the output.

[00:44:43:13 - 00:45:14:02]
Amith
 If you're someone who's just like looking at the result of a website, you may or may not know what you're getting under the hood. So there are some potential risks with that. I think observability of the model is an area that's going to get better over time. You can inspect the thinking logs, but that would require you to invest an enormous amount of work. So I think as a practical matter, you kind of have to accept that. But that's true. I think if you had a team of people working for you for six months on a project, you're not gonna know about every decision they made. And that's the same thing essentially here.

[00:45:14:02 - 00:45:15:12]
Mallory
 That's actually, that's a good argument.

[00:45:16:14 - 00:45:46:21]
Mallory
 Yeah, and I think anthropic in the past has been very forward thinking in terms of observability and how their AIs work. I just think maybe even they are like, "Hmm, we don't know exactly how Fable works." But my mind was going to, because in my scenario, I'm like, "What could I possibly use this for? Do associations have use for a model so powerful?" But I think a good example is one that you already touched on, maybe a paradox in your data that might be something very difficult to figure out. You could throw a Fable type model at it. Anything else you can think of with me besides coding?

[00:45:46:21 - 00:47:19:16]
Amith
 I think so, but Pehid to what Mallory mentioned earlier that in the first 30 days of use, they are retaining the logs of your use, which means that they say they're not gonna use it for training, so that's reassuring. But I would be careful about that. I advise clients regularly to always use for their most sensitive AI workloads to not use the consumer interface of chat, GPT or cloud, because those always retain your data, but to use APIs and to use a different harness of any variety you want, because that utilizes APIs, which generally are zero data retention versus what you had just said about the next 30 days. So be thoughtful about that. So I wouldn't take your whole membership database and say, "Hey, Fable, here it is." But I do think you should start thinking about throwing problems at Fable that you can't solve right now. So if you do have a dilemma in your member engagement or you're having a tough time attracting younger members or something like that, these are the classical problems associations have been talking about for years now. Get Fable to think about that with you, right? You all of a sudden have this unbelievably brilliant AI that can work with you. Talk to about some of the problems you're having. Don't go to it like specific work output saying, "Hey, I need a blog post that talks about these three." Of course it's gonna be great, but it was great two years ago, with Regit with Opus 4 or whatever it was and models before that as well were pretty good. So go to it with the things you think it can't do is the short version of what I'm planning to test with Fable. I do that whenever new models come out, I always throw things at them where I don't have a specific output in mind.

[00:47:19:16 - 00:47:39:11]
Mallory
 I wanna hear your thoughts Amithan how concerned you are about the safety element. I was very excited to see this drop yesterday, but I also had a feeling of heaviness in terms of, okay, this is out there, that means the other companies are gonna match it soon, that means we're gonna have open source versions soon, and thinking about what that means, how concerned are you?

[00:47:40:18 - 00:47:51:20]
Amith
 I'm quite concerned about cybersecurity just in general. I think this particular model, yes, it's at the moment the most advanced in many areas, and if it wasn't guarded the way they have guarded it with Fable,

[00:47:52:23 - 00:50:17:01]
Amith
 and undoubtedly people are gonna try to jailbreak that, right? And the people are gonna try to get unauthorized access to Methos and blah, blah, blah, but there will be plenty of other models coming right behind it from the other labs, some will perhaps not be as well safeguarded. There will be open source models on the heels of these proprietary models that will be available in months, certainly years, but probably just a handful of short months, and this will put really powerful AI in the hands of people who intend to do bad things. And we've been talking about this for years now on the pod. The reality is that every general purpose technology always has both sides of this. There's this duality of both good use and bad use. What you think is good or bad obviously can vary based on your value system and your culture, but generally speaking, I would say cyber threats or cyber attacks are bad uses of AI. And how do you protect against that? The comment I've made consistently here is still the same. You need a lot of AI because the threats are far greater than both our intellectual capacity collectively or our throughput is capable of keeping up with. The number of threats, the volume of threats, the styles of threats, the variations and threats, these are going to be unprecedented coming at our systems. And so you have to be prepared for these things. You've got to look at your infrastructure. You've got to make sure you're partnered with the right security vendors. I think auditing your security is a good idea with AI in mind. Cyber security firms are all over this. Any decent cyber security firm is very well versed in these risks and is working hard to ensure things like firewalls are updated and that their policies in place are updated. I believe though that the biggest risk of all will ultimately always be us. And that's been true with cyber security since the beginning of time. People who've intended to do harm have preyed on the weaknesses of our emotions, not so much our brains, but our emotions far more so than they've tried to attack the hard exterior, whether that's the physical exterior of a building or it's the hardened exterior of a firewall. It's much easier to get someone internally to mess something up and to leave a laptop unlocked at Starbucks and go to the bathroom and you're gone for two minutes and someone could connect to USB drive to your computer and steal all the data on it. I see people do this all the time.

[00:50:18:05 - 00:51:34:08]
Amith
 People leave their machines unlocked at the office. That's bad. I always get on people's case when I see them like, you have sensitive data, lock that stuff. We have a corporate policy that says you have to lock your computer anytime you're not physically in attendance. It doesn't matter if you have a timeout in three minutes. That's plenty of time to do damage. But like, the example at Starbucks is, I know it sounds ridiculous, but there are tons of people. I go to Starbucks and other cafes all the time. And first of all, I never even leave my laptop in there. I'd put it back on my backpack and go to the bathroom with the backpack as I was in New Orleans. So I see my computer will be gone when I get back. But even if it wasn't gone, I don't know who touched it or whatever. So a little bit of paranoia is a good idea. Just trusting that the world isn't interested in your stuff is a really bad idea. Even if you're just, you know, some average Jane or average Joe and you just, you know, you have your Mac book and you're sitting there just to your Word documents or whatever you think anyone's not interested. But you're an attack vector because you work at some company and if your machine is infected and you've been authorized to access their network, you carry the biggest threats to that organization. So I guess my point in all that is human training. There's a lot of cybersecurity training available out there. Companies like Ninjio, there's others that are out there that provide kind of drip training that keep this top of mind. But I think it's a really important thing for every organization to be thinking about.

[00:51:35:14 - 00:52:10:21]
Mallory
 I think if you're an association leader who hasn't thought about cybersecurity in a year, two years, hasn't revamped anything, now is probably a good time to do that. And also I wanna plug the great episode we did with Eric O'Neill, former FBI super spy. That one, it was just an incredibly fun episode. But two, I asked him, you know, what are practical things associations can do right now for cybersecurity? And basically the number one thing he said was turn on two-factor authentication, which is so simple. And I thought he was gonna have all these crazy ideas, but Ameth, you're right. Like humans are the vulnerability, keep that in mind.

[00:52:10:21 - 00:53:17:19]
Amith
 Totally, and if you want more of him and his sage advice in cybersecurity, I'm gonna plug Digital Now for a second, coming up in October in Washington, D.C. at the beautiful new Hilton, the key, at Roslyn right across the river from Georgetown. We are holding Digital Now October 25th through 28th. And Eric is one of our keynotes. So we're very excited to have him. I think cybersecurity is as important of a topic as anything in the realm directly about AI productivity, things like that. Now I'll add a couple of quick tips of my own. One is have verbal passcodes with your team and with your family. So talk to people occasionally about this, maybe once a year or once every couple of months even, and write down on a piece of paper and put it in your wallet or your purse, these codes. And these can be like passcodes of some sort, you can be creative and fun with it, pick different kinds of words that are very unlikely in combination to be guessed and use that as a passcode when people call each other to ask for authorization. And if you're not in the habit of asking for authorization for important things, like the classic one is wire transfers,

[00:53:19:04 - 00:53:55:01]
Amith
 make a phone call, verify the information and use the passcode system. Say, "Hey, what's your passcode?" It'll feel a little bit weird when you first do that, but like you're playing some game in middle school or something, but it's really important and it's easy because AI is really, really good at impersonating voice and even video. There's lots of stories about people getting ripped off by an AI that called them or an AI that wasn't a Zoom call with them. So protect yourself. And then beyond that, I'd say lock your damn computer, don't leave your computer sitting around. It goes with the MFA thing, even if you have MFA, but you leave your computer logged in and you're already logged into a website,

[00:53:56:17 - 00:53:58:19]
Amith
 you're kind of toast, so don't do that to yourself.

[00:53:58:19 - 00:54:20:18]
Mallory
 No, I'm proud to say, partially because of you and me, that conversation with Eric, we have come up with our own verbal passcodes and I encourage all of you just take the moment and do it, because it's something you think, "Yeah, yeah, I'll do that." But after you listen to this pod, sign off, go create your verbal passcodes, because you never know. It's so easy to clone voice with AI. I mean, there's just a lot of damage that could potentially be done.

[00:54:20:18 - 00:54:39:01]
Amith
 And do not do that on a recorded phone call, recorded Zoom call, do it in person or do it on like an old school like phone call and do not put this in a computer somewhere. Don't even put it in a password manager. Write it down on a piece of paper that you hold on your physical person at all times, like your wallet.

[00:54:40:08 - 00:54:52:08]
Amith
 Do it the old school way, because if it's digital, that means it's hackable. And if it's hackable, that means the AI is probably gonna get to it if it's got malicious intent and then it's gonna be real good at telling you your passcode verbally when you make a phone call.

[00:54:52:08 - 00:55:05:09]
Mallory
 We became super paranoid. We did it in person, but even in our own home, we were like, "We should whisper, right?" I don't know. Be in a safe place, make sure, as Amit said, you're not recording it digitally and maybe whisper too.

[00:55:06:23 - 00:55:12:21]
Mallory
 Amit, what do you think is the big takeaway here for our listeners for a release as big as Claude Fable?

[00:55:12:21 - 00:55:14:05]
Amith
 Go try it.

[00:55:14:05 - 00:55:27:23]
Mallory
 Easy takeaway, go try it. Three stories, one thread, AI, we can see is doing longer, harder, more self-directed work than it could even a few months ago. A map proof no human had found, agents that run an

[00:55:27:23 - 00:55:33:09]
 (Music Playing)

[00:55:44:02 - 00:56:01:01]
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.

[00:56:01:01 - 00:56:04:07]
 (Music Playing)