Summary:
In this jam-packed episode of Sidecar Sync, co-hosts Amith Nagarajan and Mallory Mejias tackle five cutting-edge AI developments, all (mostly) fitting into three big buckets. From the powerhouse that is Claude 4.5 Sonnet to OpenAI's stunning new Sora 2 video model, this episode dives into real-world use cases, experiments, and emerging tools that are reshaping how associations can work smarter. They also break down the promise of computer use models, the future of AI-native browsers, and what "boring AI" really means for legacy software. Amith shares highlights from the Blue Cypress Leadership Summit and explains why early adopters might just lap the competition. Packed with demos, practical takeaways, and some existential questions for creatives, this is one you don’t want to miss.
Timestamps:
👥Provide comprehensive AI education for your team
https://learn.sidecar.ai/teams
📅 Find out more digitalNow 2025 and register now:
https://digitalnow.sidecar.ai/
🤖 Join the AI Mastermind:
https://sidecar.ai/association-ai-mas...
🔎 Check out Sidecar's AI Learning Hub and get your Association AI Professional (AAiP) certification:
📕 Download ‘Ascend 2nd Edition: Unlocking the Power of AI for Associations’ for FREE
🛠 AI Tools and Resources Mentioned in This Episode:
Claude 4.5 Sonnet ➔ https://www.anthropic.com
Claude Code ➔ https://www.anthropic.com
Otter.ai ➔ https://otter.ai
MemberJunction ➔ https://www.memberjunction.com
ElevenLabs AI Voice Agents ➔ https://www.elevenlabs.io
Perplexity AI ➔ https://www.perplexity.ai
Sora 2 Demo ➔ https://www.youtube.com/watch?v=1PaoWKvcJP
Neon AI Browser Demo ➔ https://www.youtube.com/watch?v=MKQ98413RCI
Learning Content Agent Episode ➔ https://shorturl.at/nxEry
https://www.linkedin.com/company/sidecar-global
https://twitter.com/sidecarglobal
https://www.youtube.com/@SidecarSync
⚙️ Other Resources from Sidecar:
More about Your Hosts:
Amith Nagarajan is the Chairman of Blue Cypress 🔗 https://BlueCypress.io, a family of purpose-driven companies and proud practitioners of Conscious Capitalism. The Blue Cypress companies focus on helping associations, non-profits, and other purpose-driven organizations achieve long-term success. Amith is also an active early-stage investor in B2B SaaS companies. He’s had the good fortune of nearly three decades of success as an entrepreneur and enjoys helping others in their journey.
📣 Follow Amith on LinkedIn:
https://linkedin.com/amithnagarajan
Mallory Mejias is passionate about creating opportunities for association professionals to learn, grow, and better serve their members using artificial intelligence. She enjoys blending creativity and innovation to produce fresh, meaningful content for the association space.
📣 Follow Mallory on Linkedin:
https://linkedin.com/mallorymejias
🤖 Please note this transcript was generated using (you guessed it) AI, so please excuse any errors 🤖
[00:00:00] Amith: Welcome to the Sidecar Sync Podcast, your home for all things innovation, artificial intelligence and associations.
[00:00:14] Greetings, everyone, and welcome to the Sidecar Sync Your Home for content at the intersection of artificial intelligence and the world of associations. My name is Amith Nagarajan.
[00:00:25] Mallory: My name is Mallory Mejias,
[00:00:27] Amith: and together we are your hosts. And boy oh boy. Do we have a lot to cover today, don't we Mallory?
[00:00:32] Mallory: We do. I felt like this morning I was preparing for a marathon, but not a running marathon.
[00:00:37] A marathon of the sidecar sink. We've got a jam packed episode, but I encourage everybody to listen to it all the way through. 'cause you don't wanna miss any of these insights.
[00:00:46] Amith: Yeah. And I know that you like routine, Mallory, so I guess trying to fit all this into three topics, even though we're really, I think you were saying we really have five, but you kind of found a way to make it fit the three topic format.
[00:00:56] So
[00:00:57] Mallory: I think to make myself feel better, it's still [00:01:00] fits into three buckets, but yes, technically we kind of have five topics today, but we're gonna go through them. It's gonna be quick, it's gonna be hard hitting, and it's been really fun to prep for this episode because with so many releases and updates and demos to watch, I got pretty excited.
[00:01:16] Amith: Yeah, well, you know, if you've been bored with AI and the pace of AI this summer, it is now October 1st when we're recording this, and so we are officially in the fourth quarter of the calendar year. And I guess most people would consider it the fall here in New Orleans. It's still pretty warm, but uh, it is the fall and it is, uh, the fourth quarter and, uh, there's a lot to report on and a lot of exciting applications of these ideas to our friends in the association market.
[00:01:41] So can't wait to get into it.
[00:01:43] Mallory: Yep. Amith, before we hop into everything, I know you were at the Blue Cypress Leadership Summit last week. I was hoping you could share a little bit about your experience there. We had some association leaders attend this time, which is incredibly exciting. So how was it?
[00:01:57] Amith: It was fantastic.
[00:01:58] So it was the fifth annual Blue [00:02:00] Cypress Leadership Summit up in Park City, Utah. We hosted every year. That is a meeting of, uh, roughly 30, 35 of our, uh, senior leaders from across all of our companies come together once a year. We're a highly distributed organization, so we have our companies all over the place.
[00:02:16] And so we had some international attendees and we had people from all over the US come together up in the beautiful clear, uh, mountain air up in Park City, Utah. And it's really a combination of two things. We are trying to form closer relationships and form collaborative. Opportunities across our companies.
[00:02:34] And we are also, of course, focused on education and continually investing in ourselves as a team and learning and growing together. And so we focused on, uh, two primary themes, uh, education wise. One was the Blue Cypress Operating System, which is our way of running our companies. It's kind of a mixture and a mashup of scaling up from Verne Harnish, uh, the EOS system.
[00:02:55] And then really heavily leaning, leaning into, uh, OKRs, which is the framework for [00:03:00] and from Andy Grove, uh, at Intel originally, and then popularized by the Measure What Matters book by John dors. So we've taken all those components and over many years actually, we've put together the Blue Cypress Operating System, which is our way running the companies.
[00:03:13] So one of the days was dedicated to that decidedly not AI related. Quite interesting. And I had to mm-hmm not talk at all really that day because it wasn't AI related. And then day two, of course, was all about AI and agents, and we ran a workshop, uh, for everyone within the company, uh, all including most of our leaders are not developers, not technicians.
[00:03:31] And so, uh, we had everybody building AI agents by the end of the second day, which was super fun.
[00:03:38] Mallory: That's awesome. So that was
[00:03:39] Amith: the, yeah, it was super cool and, uh, a lot of people got really excited. Uh, Johanna Snyder, our group of companies, CEO, she led one of the groups and in fact, she had a really innovative idea.
[00:03:49] She had everyone get pretty close together and all speak, uh, you just have a collaborative conversation. She recorded that using Otter ai, which is a really great, uh, you know, audio note [00:04:00] taker on her phone. And she took the transcript of that and gave it to Claude. And then Claude kind of, you know, iterated on it a little bit with her and I, and a couple other people that she was working with.
[00:04:09] And then that in turn communicated back with Member Junction, which is our AI data platform that hosts our agent architecture and it automatically created the agent that they would brain that they were brainstorming. So all of that happened over the course of a couple of hours and they had a completely functioning agent, uh, for their use case, which was really cool to see.
[00:04:29] So the Blue Cypress Labs team, blue Cypress Labs is our innovation arm where, you know, it's kind of like our very small version of the idea of a skunkworks type of thing. Um, or, you know, uh, those kinds of advanced forward deployed, uh, research folks. And so we, we do all sorts of thinking there in terms of, uh, future disruptive innovations we can bring to the market.
[00:04:47] And those folks we're helping lead those conversations, but they weren't, they weren't the ones driving. It was the business leaders that were driving. So that was the leadership summit part of it. And then the evening of the second day, um, we had a really fun [00:05:00] reception at my place. And then we, uh, invited, uh, all of our sidecar MVPs.
[00:05:05] And if you're not familiar with the MVP program, it is an exclusive invitation only program for our most, uh, you know, our proudest supporters of sidecar. These are folks in the community association professionals who really believe in our mission, uh, which if you haven't heard us talk about it here in the sidecar sink, our mission at Sidecar is to educate 1 million association folks on AI by the end of the decade.
[00:05:29] So it's a big number. Um, we are well on our way to achieving that. You know, we think that we're, we're somewhere in the neighborhood of 50 to 75,000 people so far that we've touched either through webinars, the podcast, uh, through our writing, of course, through our A A IP certification program. The MVP program was inspired by, uh, early days of Microsoft, um, where they, you know, had community advocates come together to try to help, uh, spread the word about things they were doing.
[00:05:54] What we're trying to do with the MVP program is arm the MVPs. There's about 15 of them, [00:06:00] uh, and most of them were, were in Utah. Uh, really arm them with knowledge on AI, of course. And then about how sidecars various offerings, most of which, like this podcast are totally free, how we can get them into the hands of more and more association professionals around the world.
[00:06:14] So we invited those folks to come to Utah and then on the Saturday we had an amazing session with them. Uh, and so it was super fun and uh, got a lot done. It was really a great listening opportunity for the sidecar team to hear input from the MVPs on things that were doing well, things that we need to improve.
[00:06:29] And, uh, that's pretty much, uh, what we were doing up in Utah last week. Uh, got outside a little bit, which was great. It was beautiful. Weathers,
[00:06:36] Mallory: well, at least you got a taste to fall. Amit, it sounds like it was a good few days.
[00:06:40] Amith: It was fantastic. Uh, I can't complain. Being up in the mountains anytime of the year is great.
[00:06:44] And for me, even though I'm a big skier, I, uh, really enjoy, you know, September October timeframe up there, probably the most out of any season.
[00:06:52] Mallory: Hmm. I love the fall too, here in Atlanta, we're getting a, a small taste of it right now, but. Alas, we've got a lot to cover. In [00:07:00] today's episode, I will let you know what our three slash five topics are.
[00:07:04] First, we're talking about Claude 4.5, which was released this week, and how AI interacts with our digital workspaces. Kind of within that Claude 4.5 bucket. We're gonna be talking about Claude computer use, and we're also going to be talking about AI browsers, which I don't think we've ever spoken about on the pod.
[00:07:22] Then we'll be covering the brand new release added in this morning to this podcast outline of SOA two from OpenAI. And then finally, we're gonna talk about the early mover advantage. AI adoption, so let's kick it off. Claude Sonnet 4.5 released by Anthropic on Monday of this week, represents another leap in AI capabilities and is widely considered the best coding model in the world.
[00:07:51] The standout feature is its ability to work autonomously on a single task for over 30 hours, far exceeding Claude Opus fours [00:08:00] seven-ish hours. It leads major coding benchmarks and reportedly achieves a 0% code editing error rate in some tests beyond coding, though Claude 4.5 excels in advanced context management processing book length or code-based length outputs without losing focus, and now offers an agent SDK giving developers access to infrastructure like virtual machines and memory tools for ENT workflows.
[00:08:27] The model also supports context editing and checkpointing, allowing developers to save progress and resume lengthy work sessions. Built-in file creation for spreadsheets, slides, and documents makes it practical for business Use new extensions for Chrome and vs. Code target power users and long horizon workflows.
[00:08:46] Pricing remains the same as Sonet four, and industry reception describes this shift toward AI as a teammate capable of full project ownership. Also, anthropic positions it as their most aligned model to [00:09:00] date a it well. It just came out Claude 4.5 sonnet. I got to use it a little bit this morning. I'm impressed so far.
[00:09:07] Uh, and I know you've been, you've been working with it a little bit, so what do you think?
[00:09:12] Amith: I think the people at Anthropic have just been kicking back and relaxing all summer long. Oh, really? Not, not a whole lot in that release. Uh, but in, in all seriousness though, it's, it's pretty impressive. I mean, they just released Opus 4.1.
[00:09:25] Uh, they mentioned when they released that, which was I think mid-July, it was the literally moments before OpenAI dropped GPT five. They did Opus 4.1 and they said, Hey, big things are around the corner. And sure enough, here we are and, uh, 4.5 sonet drops. Uh, I will say from my own experience, I, I started using it yesterday.
[00:09:42] Okay. So. I don't have a ton of experience with the model, but, uh, in, in the work that I've done with it, purely through cloud code, I haven't really used it much through cloud desktop or cloud web. But in cloud code it's a notable improvement. It is significantly smarter, uh, for a lot of the things that I threw at it.
[00:09:58] And, uh, also the [00:10:00] new cloud code interface, which is the command line interface that developers have. Really, it's become a phenomenon, um, that people are really excited about Claude Code as an agent coding tool. That, of course, uses the Claude model, uh, or models I should say, when using the newest cloud code interface with Claude 4.5 sonet, it's come almost unrecognizable from last week in terms of what you can do.
[00:10:22] You can give it very complex requests. To give you an example, uh, most of the work that I do that's technical in nature with code is with the Member Junction open source project. So I spend a good bit of my time working with our team on advancing that piece of software, uh, that is our totally free open source, um, AI data platform that we, you know, provide to the association community and, and anyone else who wants to use it for that matter.
[00:10:46] But in any event, that's a fairly complex code base. Uh, I believe there's, there's, I don't even know what the count is in terms of lines of code, but last time I looked it was well over a million lines of code and it's 121 distinct individual projects in a large [00:11:00] repository. So it's a very sophisticated piece of software and a complex code base, and so it's hard for AI models to get their, you know, get their hands around it, so to speak.
[00:11:08] Uh, quad 4 0 5. Has has done the best job I've seen so far, and that includes comparing it to, uh, what we've seen so far from GPT five, uh, and Codex specifically. Uh, so we've given it some fairly complex tasks. Like the task I gave it, um, yesterday morning was we had a completely new specification. We're preparing Member Junction.
[00:11:28] Three, which is this next major release. We're on the two X lifecycle. And a big part of that release is going to be a completely new conversations user interface. So, uh, thus far in Member Junction, in our standard, uh, in our standard application, you have, uh, multiple different chat experiences. You can run agents, for example, through a test chat experience, skip.
[00:11:50] Has his own, uh, chat experience and we have the idea for 3.0. We wanna have a new conversational UI that is multi-agent, so you can have [00:12:00] conversations with any number of agents at the same time, as well as with any number of people. And that's something that our team has prototyped and created a really nice, uh, interactive functioning prototype.
[00:12:09] Um, and so I said, well, I've got this prototype. It's just an HTML file that the, the team built, and I've got some documentation of what we're gonna do and why don't I see what Claude Sonnet 4.5 can do? And a couple hours that came back and it wasn't perfect, but it was, it was working. Um, so, and that's a very complex piece of software to give to an AI agent.
[00:12:28] Uh, it was impressive that it actually even compiled, frankly, but um. This morning I've given it a few pieces of feedback and it's been working, in fact, it's working right now as we speak on, on improving some stuff. And, uh, it's pretty exciting 'cause I, I've written literally zero lines of code. I think it's written, tens of thousands of lines of code.
[00:12:45] And I've reviewed quite a bit of it and it looks, it looks really good. It's followed our patterns from the repository. So what does this mean? For those of you that are listening that are not particularly technical, it means that the barrier has further decreased between you [00:13:00] and software that does your bidding.
[00:13:02] So software has become more and more accessible, less and less costly. Less scarce, more abundant. We've been talking about this here in the sidecar sink since the beginning of this podcast and well before that, in other venues, um, various forms of intelligence are decreasing the cost curves or pressing down on the cost curves in lots of scarce categories.
[00:13:23] Obviously, code is one that we're super focused on here, and for associations who are not coders or thinking about coding all the time, the reason it's so important is many of your problems could be solved if you had access to really high quality, abundant low cost code. Your website, for example, that is the bane of your existence, or I should say, of your member's existence and yours as well, because it's decreasing your members' engagement with your website and spend, instead of spending hundreds of thousands of dollars or millions of dollars with website vendors and taking a year or two to iterate on your website, what if you could have functionality that was specifically [00:14:00] tailored to your needs built by AI and also critically?
[00:14:03] Maintained by AI with high level of quality, surpassing that of human developers and doing exactly what you want, the workflow your members want with low friction and high quality, that would be amazing, right? Um, most people don't go down that path 'cause they're like, well, I don't wanna spend the money and I don't wanna have to maintain it.
[00:14:20] I don't want the tech debt. All of these issues, those issues kind of go away. They're not gone yet. But the trend line very clearly shows that. So if you are thinking, Hey, I need to move forward with a new member engagement experience, and you're thinking about the old way of thinking about it, which is, you know, I need to go do an RFP and I need to hire an expensive vendor and all this other stuff, maybe you still do some of that, but your expectations should be different.
[00:14:43] And the main advice I have for you is build incrementally. Don't do a big bang thing. Pick your biggest pain point. Like let's say the member application process or the abstract submission process is the biggest pain point. Go fix just that one thing and experience what I'm describing. [00:15:00] If you don't have a developer on your team or in your, uh, partner community that is fluent in this stuff, go find one.
[00:15:06] There's plenty of people out there who are really good at using these tools or ask your folks to learn this. Uh, just don't be bullied by developers who have the attitude of, oh, the AI's not good enough. And there's a lot of developer bullies out there who will tell you, oh, you association leader that are not technical as I am.
[00:15:23] Um, you do not understand why we can't do that. And there's all those kinds of, in my opinion, totally fabricated reasons to not embrace this technology. Um, the bottom line is your members don't care. They really don't. What they care about is getting their job done. And if you can help them with that, you will be their best friend.
[00:15:40] And if you don't, they will just move on and move on to the next thing. So long-winded way of saying quad sonnet four, five crushes it on code, more so than any model I've personally ever experienced. And that's so relevant to the association community, even though the association community doesn't tr typically think of themselves as software developers.
[00:15:58] Mallory: Mm-hmm. [00:16:00] For context to me, you founded an a MS company a while back. We're talking about the AI data platform member Junction having about a million lines of code, which might be abstract for some people. How many lines of code would you say an an A MS would have just to give people context?
[00:16:15] Amith: You know, it's, it's hard for me to remember.
[00:16:17] I don't even remember what, um, what the a MS that I was involved with had in terms of the lines of code, probably quite a bit larger, I would imagine. Okay. Um, but at the same time, you know, part of the reason the MJ code base is as substantial as it is, is because we've had the benefit of AI for the last two and a half, three years since we started building it.
[00:16:33] So lines of code, many of them have written by developers, including myself, and we're very proud of the code base. But AI has written an amazing amount of the code. I, I'll give you one example and then we'll come back to the, the, the question. Um, you know, with, with integrating, uh, member Junction's all about integrating different data sources, right?
[00:16:49] So integrating different kinds of database systems. Files all sorts of things for unstructured data. And so we have this thing called file providers in mj, and that connects to things like [00:17:00] SharePoint. And initially we built support for SharePoint and I think, um, Amazon's S3 storage were the two providers we had.
[00:17:07] Uh, and then we said, well, we wanna implement support for a bunch of other things like Google and Box and Dropbox. And so AI built all of that, right? So that's a multiplication of the original code base that was based upon the human code. But, um, I would say that, you know, member Junction is probably comparable in its code, uh, sophistication to any, any a MS, perhaps more so in some ways.
[00:17:29] Um, but yeah, I mean, a million lines of code is no joke. That's a substantial software project. You know, most custom apps that I see out there are somewhere between, you know, 30,000 and 50,000 lines of code is typical for a small, small to mid-sized app. Maybe a hundred thousand, 200,000 lines of code would be a, a bigger custom app.
[00:17:45] So it's, it's a pretty large code base. Mm-hmm.
[00:17:48] Mallory: I was doing some writing with Claude 4.5 sonnet this morning and it's a little bit less tangible I feel like, with the coding example because you can clearly see, ah, this is something I got right this time, that it didn't last time, but it [00:18:00] feels better again.
[00:18:01] Just played with it for about an hour for some blogs I was doing for sidecar, but I am liking what I'm seeing so far on the writing front. Amit Anthropic is calling this their most aligned model. Can you talk a little bit about that?
[00:18:15] Amith: Yeah, so what Anthropic means by aligned is around safety. So they're talking about alignment with their constitutional AI process.
[00:18:22] So we've talked about that previously in the pod, and actually the surprising benefit of constitutional AI and an alignment more generally, uh, actually providing more capability where a lot of people have thought, oh, well AI alignment and safety and research along those lines will actually, uh, taper what a model is capable of doing because it's preventing the model from being able to do certain things.
[00:18:42] In fact, actually the alignment reinforces the behaviors. Um, it's kinda like a value system for people, right? So what they're essentially saying is, is that in their benchmarking internally, um, they have found this model to be the smartest in terms of keeping in alignment with their constitution, uh, [00:19:00] which is exciting.
[00:19:00] I like the fact that philanthropic is continuing to focus on that. You don't hear that much about alignment or safety research from the other major labs. I know they're doing work in those areas as well, but philanthropic, at least publicly in terms of what they're sharing, seems to be most focused on this.
[00:19:16] And it's one, one of the reasons I like those folks is that they're really highlighting in a very critically important part of the AI race is to do it in a, in a safe and responsible way. Now, alignment is not the end all be all. Just because this is the most aligned model doesn't mean it's a perfect model or it's safe, and, and they have not said so.
[00:19:31] Uh, but I just want our listeners to understand that alignment doesn't mean it's a guarantee of perfection or perfection or a guarantee of safety. And that's true for, for any model. It's also true for any piece of software. You can use Microsoft Word to do a lot of damage.
[00:19:44] Mallory: I wanna move on to computer use, which is our second topic slash subtopic under Claude 4.5.
[00:19:51] So Claude 4.5 achieves a about a 61% success rate on the benchmark for computer use up from about [00:20:00] 42% with its predecessor. This measures how well Claude can actually control a computer interface, moving cursors, clicking buttons, typing and navigating between applications just like a human would. The key insight here is that Claude learns to use our existing software rather than requiring special integrations or APIs.
[00:20:19] This opens up possibilities for automating workflows and legacy systems that lack modern APIs, including perhaps older a MS platforms, proprietary tools, and enterprise software that associations may be locked into. Right now, it's AI adapting to work with the tools we already have rather than forcing us to change our tools.
[00:20:38] As a note, right now it's only available via API and there's a wait list currently for the computer use Chrome extension, which I'm very excited about. Amis, I looked back, we talked about computer use back in episode 53, and it was still early days. Do you think we are approaching more mainstream use for this?
[00:20:57] Amith: I think we're gonna get there in the next [00:21:00] 12 months. So I'm excited about it. This opens up a lot of doors. Let me draw a parallel for a moment. We've spent, I think, maybe two or three episodes of, of the last hundred plus pods now, um, talking about robotics and in the world of physical AI or robotics, that human form is often the target of what people are trying to do or human-like form, right?
[00:21:22] Or, and, and the reason for that, um, is not to creep you out. Uh, that is not the reason for it, although I think human-like robots definitely can achieve that. Uh, we'll get used to them eventually, I think at least, uh, for those future robots that are listening, we're already used to and we love you.
[00:21:36] Mallory: Yeah, we love you.
[00:21:38] Amith: But, uh, in, in, in practicality, what the reason people are doing human form robots is that we've built an entire world that's based on us interfacing with stuff. Something as simple as a doorknob. It's at the height of the average human hand, something like a computer. The way we interact with a mouse and a keyboard, it's built for us.
[00:21:58] The way we drive a car, the way we walk down a [00:22:00] sidewalk. We have shaped the world to fit us. So Humanoid Robot is interesting because we have all these user interfaces, if you will, in the world that fit us. And so if we can plug in a humanoid robot that has our form approximately, um, and some form of artificial intelligence in it, that's quite interesting because then it just snaps right into the, the that which has already been built.
[00:22:24] So then come back to what you're describing here with computer use. It's the same kind of idea. There's a ton of software out there. There's of course all this web-based software, but people forget how many Windows applications that are out there. And even predating Windows, people who have terminal based applications, I know many associations who still use mini computers.
[00:22:44] And these mini computers are kind of like small mainframes that people run with terminal applications that you connect to. So, you know, for those of you that aren't watching us on YouTube, you can see Mallory kind of question that like, really? Yes. In fact,
[00:22:57] Mallory: okay. I didn't know that Amitha, I did not know that.
[00:22:59] Amith: [00:23:00] It's, it's really a thing. And these are like, you know, it's like going back in time to a time machine to like the 1960s, seventies, eighties and, you know, predating windows in the Mac when there was no graphical interface. And, and there are tools out there. People are still using, in fact, a lot of our modern, modern financial system and uh, things like airplane logistics, there's a lot of old school software that's out there that if it stopped working, that would be a very large problem for everyone on the planet.
[00:23:24] Um, and so, uh, these kinds of old software tools that don't have. Classical, you know what I call now classical APIs or more modern versions of that, which people call MCP servers and a to a servers for, uh, agent interaction. Um, there's a lot of software out there that we can't access for AI agents until we build a computer use model, which is capable of interacting with software just like a person.
[00:23:49] And so that's what Anthropic is talking about is the 61% success rate. That's still a D, but it's not an F anymore. So it's not perfect, but it's like going from a 2-year-old using a computer where they [00:24:00] can click around and do stuff to, or maybe a 5-year-old, right? That's what 40% probably is to, uh, maybe like a elementary school student, like a third grader or a fifth grader, or maybe maybe an eighth grader.
[00:24:11] So. I find it incredibly exciting because there's a whole world of legacy software. Um, also there's this other use case, which is to try to test software like a human. So when we build software, we wanna be able to test it, just like our users will test it. And there's ways of simulating that. In fact, there's a whole genre of software around automated testing, automated qa, uh, which has historically required QA engineers to write these really, really involved scripts, which are basically, it's a form of programming to control a computer and to control a browser and to control a Windows interface, uh, to try to emulate what an end user would do.
[00:24:47] And they're very limited because these, these scripts are very, very specific, very What's happening in computer use essentially is this, you have a model, which is a multimodal model, which I always love saying that it's, it's a mouthful. Um, [00:25:00] this multimodal model in Claude four point five's case is essentially taking a lot of screenshots.
[00:25:04] So just like frames in a video, the model will take a screenshot, see what's on the screen. We'll then compare it to what the user's instructions were. So Mallory, let's say for example, you said, Hey Claude, in my browser I have Amazon. I'd like you to find me a really great, uh, you know, auxiliary battery for my cell phone so that the next time I go on a long trip I have extra power and I have an iPhone.
[00:25:28] So you tell Claude that and you give it Amazon. You say, I'd like you to use Amazon because I have an account there and it's easy to use. So Claude will say, okay, I'm on Amazon. And based on that instruction will say, what should I do? Well. Claude has the intelligence to know that probably what Claude should do is search.
[00:25:44] So Claude will go figure out where the search box is. It's not hard coded to say the search box is at this particular location. It's looking at the screenshot, just like our eyeballs would do, right? To say, okay there, oh, there's, that's what seems like a search box. Lemme click there. So the computer usage agent issues [00:26:00] a command, which is to click somewhere.
[00:26:02] Um, and then after that, issues commands to enter keystrokes, right? So it's essentially emulating how you and I would use a computer with the mouse, the click the mouse movement, et cetera. And then based on that, we'll take another screenshot and keep taking screenshots until something's changed and see what's happened.
[00:26:17] And then that process continues until it's achieved its objective. So it's a non-trivial, it's a fairly sophisticated and complex multimodal exercise. And so instruction following in context, carry forward is really important. And Quad 4.5 sonnet is the best in class. Uh, I haven't personally used it yet, but the videos I've seen are really impressive in terms of its both, its accuracy, but also how fast it's, it's much, much quicker.
[00:26:41] Um, Amazon used to model called NOVA Act earlier this year, which was actually quite good. It was probably somewhat comparable to the earlier Claude, the 40 ish percent. Um. Which is actually usable for certain use cases, right? It's, it's, it can be 95% in, in simple use cases, but the benchmark we're testing against is more, you know, more [00:27:00] sophisticated.
[00:27:00] Uh, but with, with this level of success rate, you can actually use it in production. A lot of production use cases, but the speed is really good. Um, so that's really exciting. I think that this opens up a lot of doors. Um, it's not just the, the case of testing or integration with legacy software. It's anywhere where you might want to do some work and you think you can't automate it because there's some step you have to do that requires a person to click a button somewhere thinking about something and then clicking a button.
[00:27:26] You can now figure out a way to build that into your workflow flow.
[00:27:29] Mallory: Mm. I will say it sounds quite resource intensive to be taking screenshots over and over again. Do you think that's something they will be solving for?
[00:27:39] Amith: Well, ultimately, I mean, the stimuli that describes the state of the application, which is the, the visual, you know, status of the screen, right?
[00:27:47] Probably will have to be, uh, you know, it, it'll have to be somehow ingested into the model. Now, it might be smarter over time where instead of taking, you know, screenshots, maybe it's taking somehow like a diff between the screenshots, like [00:28:00] what's changed between the two. Okay. Who knows? Maybe Claude's already doing that in order to make its inference more efficient.
[00:28:06] Uh, but ultimately this is one of the reasons why people are projecting inference is growing not at like a thousand fold or even a million fold. You know, Jensen Wong, CEO and founder of Nvidia is talking about a billion plus fold increase in inference. Um, of course that's driving, you know, bullish, uh, perspective on Nvidia and, and all the other inference, uh, providers that are out there, or people who provide the, the picks and shovels, so to speak, for all of this ai uh, growth.
[00:28:31] But the bottom line is, is that we're probably gonna have multimodality. Deeply baked into this because the more sensory input you give a model, the more capable the model is. Just like our brains, if you closed your eyes, you'd be a lot less effective at using a computer. Or if I said, Hey Mallory, you can use your computer, but guess what?
[00:28:47] I want you to open your eyes for a split second every 10 seconds, and then close your eyes and try to do some stuff and then open your eyes again. That's kind of what the models were doing now. They're, you know, of course the frame rate is increasing. They trying to [00:29:00] take more pictures. Right. So that's how these things are evolving.
[00:29:03] Mallory: Hmm. Yeah. This is really interesting to think about. I'm curious for you, amis someone who is very AI friendly, I would say, I think we can all agree, uh, when this is more mainstream and maybe we're getting better benchmark scores, maybe like in the 80% range. Would you feel comfortable with employees across the Blue Cypress family using computer use kind of all the time?
[00:29:25] What would be your parameters for maybe think twice before doing.
[00:29:30] Amith: I think, I think for personal use cases, it's good to utilize these tools in a way that is kind of supervised by the employees. So maybe have it, you know, populate your basket at Amazon or your shopping cart at Amazon, but don't have a checkout.
[00:29:43] Mm-hmm. Um, of course, checking out isn't necessarily, you know, something you can't undo these days pretty easily anyway. But, um, I think there might be a human in the loop kind of mindset for those kinds of arbitrary tasks that a, a person might wanna perform. Um, but at the same time, uh, I think [00:30:00] there's a lot of specific things where the use case is narrow enough that I can get the AI to repeat it.
[00:30:05] Even with the, you know, the 61% on the benchmark might be 0% for some use cases, but it might be close to a hundred percent for others. So part of it might be to figure out, oh, okay, you know, there's this one use case where we have an employee. That uses this really old meeting software in order to upload, um, you know, information about our exhibitors or something that's like repetitive and mind numbing and time consuming.
[00:30:30] And the software is really antiquated and you know, it works this certain way and maybe for whatever reason, when you test out that use case with cloud four five sonnet or, or uh, open AI's operator or something else, it works really well. Well if you can codify that into a set of prompts and instructions essentially that teach the AI to do that, and if you can get 90, 95% accuracy, it might be better than what most people would get.
[00:30:53] And so you can still have a review process in most, in most processes you design. So you might find, even with [00:31:00] this stage where it's 60% on a benchmark, you might find use cases where it's 95 to a hundred percent for you. And you might also quickly find use cases where it's close to zero. So I think it depends, I think for agen use, which is really more of, uh, autopilot, right.
[00:31:14] What I'm looking for are places where, you know, for example, you have a sales rep says, oh, well I need to go use LinkedIn every time I research a prospect, I'm gonna go to LinkedIn, I'm gonna send a message, I'm gonna find more information, I'm gonna do all these steps. And then that can lead to an automated process using a more contemporary software like a HubSpot or some other agentic flow or whatever.
[00:31:35] It's like, well those steps are interesting to think about 'cause could we use an operator or computer use agent for that? And there's tons of stuff like that in the association world. Lots of manual steps that people think like, oh, I go to my a MS and I click the export button. And then that export button creates a file and that file gets saved to this particular folder.
[00:31:53] And then from that folder, I take that file and I open it in Google Sheets or Excel. And from there I do this. And from there I like, it's like [00:32:00] this se sequence of manual steps in order to kind of like the human being, the integration. I know tons of association staff who do stuff like that. Highly repetitively.
[00:32:09] Um, and there's opportunities for those kinds of processes to be fully automated. Mm-hmm. So I find it very exciting, but I, I do think there caution is important. You should, like everything else we talk about here, you should test it out. Take a small dose and then see what happens. And then if it works, great, keep rolling with it.
[00:32:26] If not, don't kill it. Just put it on pause and say, Hey, you know what? That didn't really work, but this is an experiment I want in my inventory just sitting there waiting for cloud five or GPT six or Gemini three, which is, uh, supposedly right around the corner. So keep those experiments, the failed experiments in an inventory that you just constantly, you know, keep right in front of you in a folder that's ready to go so you can test them again.
[00:32:50] Mallory: As you were talking, I was thinking one thing I can use Claude, uh, computer use for is checking in for my Southwest flights. I think we've talked about that before. Meet I'm, we do fine at that. And [00:33:00] another thing, as you were talking about the event example is I know in the past when we've used our event app for digital now, it was a pretty manual process for us to load in the sessions and the titles and the speakers.
[00:33:11] Like, I don't think we could do that in bulk at that time. That would be a great use case for something like this.
[00:33:16] Amith: Totally. Yeah. And that's an example of an event software that at the time, at least the one we were using didn't have that feature. And so it would be great to have the ability for the AI to do many of those things.
[00:33:26] And what I always tell people to do is don't necessarily try to find the most complex thing, but find a thing that annoys you the most that you have to repetitively do. If you have to do something like once a year, that takes you an hour, I don't know if I'd spend a lot of time automating that. Um, but if you have to do something like, you know, six times a week, that is five minutes, that's actually probably worth looking at.
[00:33:46] And maybe you don't do that as well as you could or as consistently as you should. And this is an opportunity for AI to improve your execution consistency, improve your quality, and of course, save you some pain and suffering too. By the way, one other quick thing that I [00:34:00] wanna share on this is that audio is not part of our agenda today, specific to this topic, but is worth mentioning because in the context of these kinds of agentic flows, that's a very natural thing to think about, right?
[00:34:10] So if you say, oh, okay, well, audio's getting really good. Um, in fact, for our listeners who have not yet played with the 11 Labs agent framework, I would highly encourage you to do that. You can go to 11 labs and their website allows you as just a business user to go on there or an end user and define an audio agent that you can call.
[00:34:29] You can give it a phone number and it can talk to you, you can give it context, you can give it documents to have knowledge. You can give it the ability to call tools and you can connect these things with other ai. So if you have a knowledge agent or if you have some other systems out there, you can connect it.
[00:34:42] Um, but the idea though is well, what if you can connect that world into this world, right? And these things are all interchangeable. So you connect audio modalities with this kind of computer use and then connect that of course to enterprise agent flows. And it really opens up the door to a lot of interactions, like more and more of the surface [00:35:00] areas that previously required us to make a phone call, for example, to like book an appointment somewhere.
[00:35:05] Like I, I don't know about you, I just love calling my doctor's office to book an appointment. It's so, so awesome. So I would love to have an AI robot do that. And then there actually are some AI that already do that. But I'd love to tie that into my workflow and then have it automatically add a calendar entry in my calendar and then remind me the day before and whatever, you know, it'd be great if I could send a humanoid robot to the doctor on my behalf.
[00:35:28] Mallory: We're probably not there yet, Amit. No, I was just thinking, we've talked about on the pod many times how Amit, uh, used to and still does a lot of work while he goes on walks and is talking to ai. But imagine how much work you're gonna be doing, Amit, when you can like run your life on a walk. You're gonna be super fit and all your businesses will be thriving.
[00:35:48] It'll be great.
[00:35:48] Amith: Sounds fun.
[00:35:50] Mallory: All right, let's shift to AI browsers. So while computer use represents AI adapting to existing tools, AI browsers take the opposite approach. So rebuilding [00:36:00] tools to be AI native from the ground up. These are browsers with AI designed at the core, featuring realtime contextual understanding, built in chat and co-pilots and AgTech capabilities for multi-step autonomous tasks.
[00:36:13] The key distinction from traditional browsers with AI plugins bolted on is that AI browsers have deep context tracking across tabs and sessions, allowing them to understand your work Holistically. Leading AI browsers right now include Microsoft Edge copilot perplexity comment. Browser and brave ai. The thing that caught my attention this week was that opera just launched a Neon, which is another AI browser at 1999 a month, positioning itself as a power user browser with cards for creating reusable automated prompts and agentic features that can complete multi-step tasks across browsing content.
[00:36:53] These tools consolidate search, writing, workflow and automation into a single unified workspace, removing the need [00:37:00] for multiple apps and extensions. So I'm kind of surprised to me, I don't think we've ever talked about AI browsers. I personally do not use one to you.
[00:37:09] Amith: I have tried Perplexity comment product and uh, which is, I think it's a fork of the same underlying browser technology that Chrome uses.
[00:37:17] Uh, one thing is that there, there are a couple of different code bases for open source browsers that makes it very easy for people to create forks of these things. And it's kinda like actually Cursor, which was a fork of visual Studio code in the, in the developer world. Open Source is perfectly, uh, positioned to support all this explosive growth in ai.
[00:37:34] But in any event, um, yes, I have used perplexity comment. I wouldn't say I've used it extensively. I probably spent a total of a couple of hours in it. It's really good. Uh, it's really good for basic tasks. It has its own form of computer use type functionality as well as the stuff you mentioned. And, uh, I think it's the future.
[00:37:50] I think that's where people are gonna go. Now, I don't know if it's an AI browser as much as it is. Maybe cloud four, five extension plugs into Chrome where you already are comfortable working. [00:38:00] Uh, and by the way, um, last time I checked Google was pushing pretty hard to, you know, do some cool stuff with AI as well.
[00:38:06] And, uh, they haven't forgotten that they own the Chrome browser and that that's probably a good place to throw some additional ai. They just actually added a really cool, um, uh, Gemini extension that's like native to the Chrome browser that I haven't really used in, in a lot of detail, but it has some, it has some cool stuff in terms of integrating search, multi tab type of context, all that kind of stuff that you just described.
[00:38:27] So pretty much this is just gonna, uh, uh, this is what I say about AI in general. As much as we're massive proponents of ai, one of the reasons we want to educate a million people by the end of the decade in this sector on AI is I think the term AI might even just go away by the end of the decade. 'cause the word browser is really the operating word.
[00:38:44] And AI is like, yeah, like if it's an internet browser, you don't say internet browser or web browser anymore. Of course you're browsing the web. Of course you're connected online and of course you're gonna have ai. So, but I do think that that's not where most people are now. So they should play around with this stuff and see if there's some use [00:39:00] cases that are really interesting for them.
[00:39:02] Um, you know, I think it's also notable to look at the way search. 'cause most people's experience in the browser starts with search. Uh, and most people search starts with Google. Uh, and so Google is, you know, their AI answers is very much like what perplexity burst on the scene doing really well. Uh, their answers have gotten increasingly good and really, really fast.
[00:39:24] Um, so I'm, I'm impressed with it. I think, I think you're gonna see all these things blend together into kind of, you know, ultimately a singular experience almost.
[00:39:32] Mallory: Okay. That was my question is do you think it's an either or with computer use and AI browsers or it's just all gonna be like a, a conglomeration.
[00:39:41] Amith: I think ultimately this stuff blends together in some interesting ways. The browser and the operating system have become more and more unified in some ways. Where, you know, is the browser, the operating system is the operating system, have some browsing capabilities, you know, built into it. And there's kind of ebbs and flows in both directions where you see like web widgets.
[00:39:59] In the [00:40:00] Mac operating system or Windows, and you see the inverse happening where, you know, the Chromebook phenomenon that happened a number of years ago. And, and that's still continuing. So you have kind of this technology boundaries being, you know, uh, permeable, let's say. And so ultimately what I think is gonna happen is, is that, you know, I, I don't know which category necessarily will win or not win, but I do think a lot of software that we've been using in these different boxes, whether it's Word and then Excel, and then PowerPoint, and then the browser, and then the file system, some of those boundaries might go away.
[00:40:31] And you might end up in a situation where in Chat JPT or Clot or Gemini, you do most of your work. And then those that Super Smart Agent is capable of operating across all these applications for you through CPS and through other forms of integration. Uh, or you might go into Excel, and Excel might have a really great future version of copilot.
[00:40:50] Microsoft just upgraded the copilot in Excel. I haven't used it, but I've heard it's considerably better than the last iteration. You might have specialist tools in specific apps. I think the same [00:41:00] thing is gonna happen between browser versus, you know, AI extension versus AI being inbuilt into the browser and, and search, et cetera.
[00:41:07] These are all, you know, in, in the mind of the user. They're trying to achieve a task, right? They're trying to get their job done. Mm-hmm. So I think it's gonna be who provides the most value with the least friction.
[00:41:19] Mallory: Yep. We talk about that all the time. Totally. I think I'm gonna try out perhaps, uh, neon or I'll do some more research into the ones I just listed to see which one is quote unquote best.
[00:41:27] Because in their demo, which we know is, you know, putting their absolute best foot forward, it seemed pretty neat. Like you have, you're watching a video, you've got the chat interface, you can ask, you know, what is this woman in the video saying that's different from this article I recently read, and then it kind of summarizes those, or mm-hmm.
[00:41:43] Uh, one is a cooking video and you say, Hey, can you actually take note of the recipe as I watch this and the ingredients, the AI does it, and then you say, okay, go ahead and order these on Instacart for me directly to my door, if that works. That's pretty cool.
[00:41:56] Amith: Totally. And I think the lesson to be taken away, aside from the [00:42:00] benefit you can get as an association staff member or a volunteer, uh, is to try this stuff out to see if you can get benefit from it.
[00:42:07] But the, the main message is, is that this is what consumers are growing to expect. And unlike the world of automobiles where people do have a feeling in their hearts about classic vehicles, I don't think people are gonna feel that way about classic software. In fact, I don't feel that way about Siri or the old Alexa.
[00:42:23] Uh, I don't know anybody who does either. So I don't think people are gonna have a warm place in their hearts for, you know, aging software, uh, the way they do for classic cars. I'm a big classic car, you know, enthusiast by the way. But, uh, I don't think that's gonna apply to software. Uh, so I think you need to be thoughtful about this and make your experience low friction and find a way to end up being where the user is working.
[00:42:43] So if people are working in this new. Realm of whatever the AI browser is, you need to find a way to plug in your association services to be seamless into that. Maybe that's a browser extension, maybe that's something else. It adds complexity to your job, but in a way it [00:43:00] actually reduces complexity because perhaps what you're doing is surfacing some of your features, some of your knowledge, some of your agents, perhaps in an environment like perplexity comment, um, or elsewhere.
[00:43:10] And so if I'm an association leader, I'm gonna be thinking about, well, what's the trend line in terms of consumer experience? And what kind of extensibility do these future browsers and other technologies have, and how can I plug in? That's the key thing. How can I plug my value so that it feels like a native experience in that environment rather than being an afterthought.
[00:43:30] Mallory: Hmm. Shifting gears a bit, I wanna talk about SOA two, which is open. AI's latest video and audio generation model launched alongside the new SOA app for creating and sharing AI generated videos. SOA two stands out through its highly realistic physics world modeling and direct avatar insertion features, letting users generate lifelike clips complete with synchronized dialogue and sound effects simply from text prompts.
[00:43:59] SOA [00:44:00] two can simulate complex physically plausible scenarios like Olympic gymnastics and sports routines with improved fidelity to real world physics. It supports realistic soundscapes dialogue and contextual audio tightly synchronized to the generated visuals. The model demonstrates advanced instruction following, which enables multi-shot sequences, character consistency, and sophisticated world state persistence.
[00:44:24] Cameos allows users to create AI avatars by uploading short videos and audio, then insert themselves or designated people into generated scenes with facial likeness and voice accurately rendered. SOA two excels in cinematic, photorealistic and anime style renderings. Generating clips up to around 10 seconds per prompt.
[00:44:47] Another interesting fact, the new SOA app is going to be structured like TikTok with vertical videos, algorithmic feed, remixing tools, and engagement features like comments and likes. It incorporates identity protection [00:45:00] requiring users to verify their likeness before their AI cameo can appear, and users are notified if their appearance is used in others' content.
[00:45:07] OpenAI is developing a SOA two API for third party developers to integrate the video generation engine into their own tools. SOA two is seen as a big leap forward in world simulation, aiming toward models that accurately understand and interact with reality beyond just video creation. For now, this is invite code only through the SOA app, and I think the SOA app is only available for on iOS, on Apple iPhones for now.
[00:45:34] So in me, you sent me a video, a demo of this, this morning. What was your initial reaction?
[00:45:40] Amith: It, it surpassed my expectations. I've been hearing the rumor mill talk about SOA two for, I don't know, a couple weeks or something. And, and honestly, like video is, uh, even though I find it incredibly interesting, it's, it's a little bit less on my radar than some of the other things that we talk about.
[00:45:53] So I don't follow it quite as closely. But SOA two, if you haven't checked out the, um, the demo yet, we'll link to it in the show notes [00:46:00] and on the YouTube, we'll, we'll have a link to it as well directly. Um, but it's, it's pretty impressive. It's a mixture of all the things that you mentioned. It is, uh, really hard to distinguish between the portions.
[00:46:11] Like there's portions where Sam Altman, the CEO of AI is presenting and talking about stuff, and you really can't tell unless you look real closely that it's an AI generated version of his face in various contexts talking and lots of other cool things happening. The physics simulation, the world model capabilities, clearly there's been some, it's a step change, it's a major advancement in capability compared to Sora one.
[00:46:34] Uh, and you know, I know recently we did a episode with Thomas Saltman and Mallory Yu and he had a, a great conversation about world models. Um, and, you know, the, the, the work that Google is doing in that area is pretty stunning. There's companies out there like fefe Lee's new, uh, world, uh, I think it's World Labs or World, uh, something like that.
[00:46:53] She's done some amazing work in that space that I'm really excited about as well at the model research level. And so, uh, there's gonna be [00:47:00] a true explosion in intelligence in terms of physics and world models and how that ties to all of their modalities. It's, it's really fascinating. I think there's some really cool applications for associations as well.
[00:47:12] Mallory: You led me into my next question. What are some of those applications you think potentially am meet for, uh, for SOA two?
[00:47:19] Amith: Well, one of them is learning. And so what I'm excited about for sidecar, which effectively is an association style business model, right? I think that, you know, a lot of what we do is very similar to what our colleagues in, in the association market itself are doing.
[00:47:32] We are really focused on delivering the best possible education we can, uh, for association professionals on the topic of ai, obviously. And so, um, for our active students that are in the Learning Hub already, you know that we utilize extensive AI to. Generate videos that include AI avatars that include AI generated voices.
[00:47:53] The content itself is written by us and generated based on our knowledge and based on our corpus of content, which is a mixture of different [00:48:00] things. But then we generate all sorts of different great content from that, and we do that. Uh, obviously there's an efficiency play, but the bigger reason we do that is to keep it up to date to constantly evolve the learning content.
[00:48:10] We've even built our own, uh, learning content agent around it, and we've had prior. Spot episodes that we'll link to as well on that topic. So what I could see us doing with sidecar is incorporating SOA two or, or a similar type of technology to create way more dynamic learning experiences, possibly even personalized down to the individual learner, right?
[00:48:29] Thematically including, uh, a lot of things that learner finds interesting building content up that is specific to where they are in their learning journey, to really tailor an experience to ultimately optimize not the sizzle, I mean the sizzle's cool, uh, but to optimize the learning outcomes so we can help people really take away tangible, actionable skills.
[00:48:50] And I think this kind of video has the potential to really reshape the way learning is done. So asynchronous learning. Uh, I also think right now, you know, I'm, I'm assuming that SOA two [00:49:00] videos take a good bit of time to generate. I don't know the exactly the, the type of time involved from prompt to completion, uh, but it's definitely not a real time technology yet.
[00:49:09] But we also thought that to be true about video translation about two years ago, and now we're approaching the point where we have real time video translation at the level of quality we had for batch processing only a couple years ago. So as this stuff continues to compress and become smarter and faster, if you imagine what's happening with SOA two type videos, or even like soro one type videos, if you could have real time interactivity with that as a learner, that could be really interesting in terms of simulations and so many other categories of learning that I think would be fascinating.
[00:49:39] So to me, that's the number one thing that popped up. Obviously entertainment value is incredibly high. Uh, that's not where my mind goes, but what are your thoughts on that? Because you, you know, that's only, you know, a tremendous amount more than I do.
[00:49:50] Mallory: Yeah. This is, I've been doing a lot of thinking on this, Amit, because I don't know if you've seen the headline recently, I think it was this week or last week, about Tilly [00:50:00] Norwood.
[00:50:00] Have you seen anything about that? She is, uh, perhaps the first AI actor, quote unquote, people are really not liking that, that title and talent agencies were looking to sign her. Uh, she's completely AI generated and performed in some, you know, short film or something like that. Anyway, the whole creative community is really against Tele Norwood, uh, sag aftra.
[00:50:23] The Union for Actors and other creative professionals has come out saying like, we don't support this. We don't recognize her as an actor. Uh, so anyway, it's been heavy on my mind recently, especially given my role in the pod and with Sidecar and as an actor too. But, uh. You know, I might have to circle back to this, Amit, I'm really struggling with it, which I'll talk to you about on the pod and I'll share with you all.
[00:50:46] It's just, I think there's so much good in it, and I think of SOA two in terms of capabilities for storytellers who didn't have access to anything even close to this for a lifetime. Right? So many creatives who just couldn't create because they [00:51:00] didn't have the resources to do so. That's incredible. The reverse right is not using actors, not using writers, not using, you know, video editors and things like that, which sucks to be frank.
[00:51:11] So I don't, I don't know what the answer is except that I believe we should have more creatives educated on this stuff and involved in the conversations as opposed to creative saying, you know what? I don't want anything to do with that because I don't, I just don't think that's the smart way to handle it.
[00:51:27] But I go back and forth.
[00:51:30] Amith: I think it's a really interesting dynamic and the technology doing what it does, which is relentless progression at this geometric pace, which is, you know, it's hard to really even contemplate much less, you certainly can't control it. And so the idea of a union, uh, saying, Hey, we're not going to recognize this.
[00:51:46] We're not gonna support it. I think it's counterproductive in the sense that, uh, ultimately it, it does exist. People are going to use it. Um, and so actually even saying what they said somewhat. Actually adds fuel to the fire for the use of the [00:52:00] technology in kind of an inverse way to what they, they desire at the same time.
[00:52:03] I totally get it. I totally get where they're coming from and I empathize with that. And I think the, the creative community, which adds so much to the human experience and part of what we all seek to enjoy, not just for entertainment, but just kind of in, in our journeys, um, is impaired by this. And I would think that it's, it's going to be, it's kinda like coders, you know, it's a totally different thing, but with software developers, um, there are going to be elite software developers who continue to thrive in the profession.
[00:52:29] Even with the cloud 4.5 sonnets that we talked about in Quad 10 or whatever comes in the very future, you know, that is so much better than a human. There'll be people in the profession. And there will be an explosion of output in terms of the amount of, of AI generated software that is serving the world.
[00:52:45] But will there be as many software developers? I don't know. I mean, I think that there will be demand for a dramatically larger amount of software. Um, but I don't know what the, the, the, the story is in terms of the vast majority of the software developers that are out there. Certainly people who are doing like routine maintenance [00:53:00] on, on legacy software where, you know, there's really not any creativity involved necessarily.
[00:53:04] It's more of just keeping stuff running, which is by a lot of estimations. Somewhere between 50 and 80% of the develop software development workforce is people who just keep the lights on, which is really, really important by the way. Hmm. You'd hate it. If those people stopped showing up to work, because you wouldn't be able to use your banking app, you wouldn't be able to turn the likes on in your home.
[00:53:21] Uh, but a lot of that stuff is, is absolute bullseye for what even current state AI can automate. So what happens to all those folks? I don't have the answer to that, but it's, it's concerning. And I think the same thing is, is the case for, uh, kind of the middle of the curve for people in the creative world because people who are at the kind of elite end of the profession of, of any profession, probably there'll be demand for them, but they're already doing fine.
[00:53:43] Right. You know, it's the people that are in the middle of the curve who are doing kind of just the fairly routine stuff that people aren't thinking about as much, but they make their livings from it. I think a lot of that is a target for automation because, you know, if, if you're a local retail store in New Orleans, you don't necessarily need the ultimate creative expression to get [00:54:00] your ad out there on, on radio or cv.
[00:54:02] Mallory: Mm-hmm.
[00:54:03] Amith: What happens to all those gigs? You know, I don't, I don't know what happens ultimately the people who make a living in that. That's a good
[00:54:08] Mallory: point. And I mean, I think too, it's very different, but I think too, in the sense of. Films and television shows being created by major production companies, really for their commercial benefit.
[00:54:19] I mean, like the shows that we see on TV and the big blockbusters, yes, are created for the audiences, but are also very much created to make a ton of money for the people who've invested in them. But we still have a thriving indie market for films and, and web series and things like that. So I, I'm a firm believer that we'll probably see a path split where we have, you know, AI actors and some people will be totally fine consuming that and it will be entertaining and that's what it's for.
[00:54:43] And then there will be others who say, well, no, I wanna see that. That human element, that emotional reaction come out in this actor that brings out something in me as I wash them. So that's where I'm at. But I will say it's been, it's been an interesting week to be in both worlds because it's very [00:55:00] much, we don't want anything to do with ai.
[00:55:02] And I'm like, ah, I don't know if that's the answer either.
[00:55:05] Amith: Yeah, you live in two polar, polar opposite worlds between the, the things you spend your time on. So that, that's super interesting. I think therefore, your perspective is also, uh, distinct and it's really interesting to hear it. I would say for our association listeners, or if you're a, um, a, a vendor helping associations or a volunteer that, uh, works.
[00:55:23] To help us lead your industry association or your professional association? Um, the common theme that we're discussing, which we keep coming back to in the pod, uh, week after week, and is this issue. And I think that the, the number one thing that I would recommend people do is learn as much as they can about ai, uh, which is of course in the association market.
[00:55:41] You know, I'm, I'm, uh, pretty well known for saying that there's gonna be two kinds of association staff members in the not too distant future, probably 18 to 24 months. There are going to be the association staffers who are versatile and very well trained and fluent in ai, AI forward folks, if you will.
[00:55:56] And they're gonna be the rest of the association staffers who are no longer [00:56:00] employable. So therefore, it is the responsibility, uh, the moral imperative, if you will, for the leaders of these associations to train their staff. Now, even if that means they're dragging and kicking and screaming, most association leaders are more accustomed to consensus.
[00:56:13] They're more accustomed to opt in, volunteerism, like volunteering, their employees volunteering to learn something new. I think you gotta stand up and put a stake in the ground and say, yes, we're gonna give you a carrot to complete learning of, of whatever kind you wanna do in ai. We're also gonna give you a bit of a stick, which is to say, you, you, you need to do this.
[00:56:30] This is not optional. Uh, and I think this is true for your, for your, uh, members as well. And so that's really the last point I wanna make, which is as an association for whatever it is that you represent. It could be an industry, uh, in construction. It could be, uh, serving a professional like architects or engineers or a branch of medicine.
[00:56:47] Um, I also think it's the association's role to stand up and ensure their, their members are well prepared for ai, uh, providing AI learning, providing resources. And a lot of associations aren't doing that. They're doing it kind of at a very [00:57:00] superficial level, if at all. I think associations need to go deep and figuring out all aspects of how AI will affect their sectors and help prepare, you know, their communities.
[00:57:10] That's a big part of what associations need to do. To do that, you have to first become pretty knowledgeable about AI yourself and get your volunteers knowledgeable about ai. But that's a big lift. Um, but I don't think that's an option if you don't do that. Pretty much know it's coming for you. It might be a few years, but it's coming.
[00:57:25] If you do that, I think some interesting and really cool possibilities are in front of you as well. So I think it's about preparedness and preparedness comes through learning.
[00:57:33] Mallory: Hmm, preparedness learning and moving early, which is our last topic for today. Recent insights from the information AI agenda live.
[00:57:42] By the way, the information is a great technology news source that Amit and I love. But at their AI agenda live conference, they revealed a massive opportunity for early movers. Sarah Go, founder and partner at Conviction VC notes that AI adoption will mirror cloud computing's 20 year trajectory. [00:58:00] Large PE firms are working on five year transformation timelines, meaning associations that start now can build years of competitive advantage while others wait.
[00:58:09] The twist here is that boring, quote unquote, AI gets adopted fastest. Enterprise software. Giant SAP reports 34,000 organizations using their AI capabilities, not for flashy demos, but for automatic document processing, receipt scanning, and mundane automation. The successful implementations are ones that easily allow human behavior to adapt quickly.
[00:58:32] Failed pilots shouldn't concern anyone. This experimentation phase is where early movers learn what works. Meanwhile, a critical memory crisis is limiting AI infrastructure. While chip shortages for training have eased, GPUs now lack adequate memory for inference workloads, AKA, running the model restricting simultaneous users and explaining the server crunches.
[00:58:54] Companies like OpenAI and Cursor have been complaining about this bottleneck must be solved before [00:59:00] advanced AI agents become viable. Industry experts predict major leaps in AI coding capabilities, which we've talked about within 18 months though we're really seeing them right now and advanced computer use within three to five years.
[00:59:12] So Amit, I feel like we'd spend a lot of time talking about cutting edge AI applications. That's the fun stuff. What, what do you think about these kind of boring AI applications that are easy for humans to adopt and adjust to?
[00:59:25] Amith: I think that's the cool stuff. I mean, it's, it's boring. It lacks the, it lacks the sizzle, but that's where the stake is.
[00:59:30] So you gotta go and look for this stuff in your organization. You know, we mentioned last week we were, uh, a bunch of us were in Utah for the Blue Cypress Leadership Summit, and we had this agent workshop. And my advice to the team when I was speaking about agents is to find the workflows that have the most pain.
[00:59:47] Don't look for the sizzle, don't look for the coolest use case of, you know, multimodal audio agents with this and that, and like things we can't do yet. Those are awesome. Like, I love talking about disruptive ideas and all that innovation, [01:00:00] but let's talk about the thing that's a giant pain. You know, what is stopping you from enjoying your day?
[01:00:04] Right? And look for the stuff that's just crushing your staff or your volunteers. Uh, good examples of that is when you find yourself shuffling documents around, right? So if you have a committee that's a mixture of some staff, but a lot of volunteers that's responsible for reviewing, let's say, proposals for your upcoming conference or for your journal.
[01:00:21] That is a nightmare of a process to just route the documents around. A lot of you who are listening know exactly what I'm talking about, uh, or just trying to connect people at your conference. And this is super manual if, if you do it at all. Um. These pain points are opportunities and they're actually pretty simple things.
[01:00:38] So an SAP example, like scanning receipts, you know, that's super easy for ai. I mean, AI's been able to do that for quite a while and it's now like even the cheapest, uh, cheapest and easy to use models you can run like on a phone or on your computer can do that, uh, incredibly well. So we should use this stuff, right?
[01:00:55] We should take advantage of this stuff. Like we're talking about the frontier and a lot of what this pod talks about is the [01:01:00] frontier. We just got done talking about Claude four, five sonnet talking about computer use. We're talking about all this cool sizzle stuff and there's a lot of stake in those models as well, just to be clear.
[01:01:08] But you can use a generation older model, like you could go and say what's equivalent to like a GPT-4. Caliber model and there's a lot of great open source models that check that box. They're basically free to run. You can run them locally or you can run them on a fast inference provider like Grok or Cereus.
[01:01:26] Um, and you can get unbelievable performance for next to no money to do some of your most boring workloads. So I say find the pouring, find those, those pain points that can open up time and give you more creative, uh, cycles. Because if you're spending your time like, you know, just doing the basic, you know, things in your business or a number of your employees are doing that, um, it's eating up their energy.
[01:01:51] Uh, and that's resulting in them having fewer reps on creativity.
[01:01:55] Mallory: Hmm. Okay. I wanna take your example myth of routing documents, which is pretty [01:02:00] simple, but sounds like a huge pain for many association folks listening to the pod. So with that example, as someone who's deep in this, what would be like, what are the essentials to make something like that happen in the sense of.
[01:02:13] Dollars, how many developers you would need, uh, if you wanna run an open source model. And like anything else you can think of right now, what would you need listening to this podcast?
[01:02:22] Amith: Well, let's take a very typical association that runs a professional conference that might have, let's say a few thousand attendees.
[01:02:29] They might have several hundred sessions at an event like that, and they might receive a thousand, or even more than a thousand proposals to populate those several hundred sessions. Sometimes they receive, you know, 10 to 20 x the number of proposals that they have room for, uh, in some, in some fields. And so the, the first step is often like.
[01:02:47] Did the person who submitted the proposal meet the basic criteria, um, for our requirements? Like, are they proposing to speak on a topic that is even germane to the conference? You know, a lot of times conference organizers will [01:03:00] say, Hey, here's six pillars of content we want proposals on. And then someone submits something that doesn't hit on any of those six pillars of content.
[01:03:08] That's actually a fairly common problem where they fail to identify it. Uh, or perhaps there's a, a requirement that says, Hey, in order to speak at this event, uh, you have to be a member, or you have to have some other qualifications that are a little bit more qualitative, that are less on or off kind of things.
[01:03:25] Uh, and so they'll have a staff person that's literally reading through proposals to try to figure out, is this proposal aligned with any of the content areas? Does this proposal meet certain kind of, you know, standard check boxes in the rubric? And these are all basic things, but they require human labor in almost all associations today.
[01:03:45] And even a very basic, you know, multiple generations ago, LLM can do that incredibly well. And people actually realize this because if you wanna test this out, do your process. So let's say you have a rubric or some other guidelines document says, this is the set [01:04:00] of requirements for my association's conference.
[01:04:02] Right? Take that document and then you take a proposal that you got. You go to your favorite tool, it could be chat, CPT, Claude Gemini, really any of these tools can do it. And you upload both documents and you say, Hey, um, ai, please review the guidelines or the rubric and please review this proposal and give me this feedback.
[01:04:21] The feedback I am asking for are these, I wanna answer these five questions. The first one is, is like, does it qualify on this basis? And, and on and on and on, right? And maybe they answer the questions are like, just fulfilling, you know, what the guidelines say, like for each guideline, is it yes or no? Or on a rating scale of one to five or whatever.
[01:04:38] And then, you know, that's probably a process that your human reviewers are doing. Try to get the AI to do that just for one proposal, right? So if you can do the thing with a manual step by uploading a couple documents into chat PT and asking a question, well guess what? You can chain that together into a process where, let's say you wanna have a little [01:05:00] form on your website where hopefully that's how you're getting these, these submissions.
[01:05:03] Some people are still getting them an email. Some people may be getting them in the mail. I don't know, but like, let's just say, no, I'm not, I'm kidding. I, I know people who receive literal cash in envelopes to pay for memberships still in the, it's, it's not a joke. There are organizations that are that, so.
[01:05:20] But notwithstanding the, the kind of like the analog world, which has to be digitized and there are things called scanners where you can, you know, get that stuff digitized. But regardless, let's say you have some form of submission. It could be email, it could be off the website. Um, let's just say that ultimately these documents all land in a SharePoint folder.
[01:05:35] Or let's say they land in a box.com folder or something like that. Um, there are ways to automatically run prompts against documents when they land in a folder. So box.com is, it's actually one of my favorite tools, uh, for file sharing. They have a really cool set of, uh, agentic AI capabilities built right into box.com that a business user can get into.
[01:05:56] So you can literally set up rules that say something like this, you can say, whenever a [01:06:00] new document is added to this folder, I want to use this AI model. It could be Claude, it could be open ai, and I wanna run this prompt. So in that prompt, you can also feed it an a link to another document, which is your guidelines.
[01:06:12] And then based upon the answer from the prompt, I either wanna move the document to the. Accepted folder or the rejected folder, and then if it goes to the rejected folder, I wanna send an email back to the person who submitted it saying, thanks so much for the submission. Uh, but it wasn't accepted. And oh, by the way, not just, it wasn't accepted, but here's why.
[01:06:30] It is actually often omitted from the process of rejecting a proposal. You can do that right now with box.com. You can do that with Zapier connected to your SharePoint. You can do that with Member Junction, which is the idea of the platform i I described earlier. You can do that with a ton of different tools.
[01:06:46] The tools are there. The first thing you have to do is to see if you can do it once manually with the underlying tool. If you can do that, then you can replicate it at scale and it's not that hard to plug it into your workflow. Um, so that's just one simple example, but that [01:07:00] example fires me up. Malory. I don't know about you, but I, I can hear about that.
[01:07:03] And I think about the number of thoughtful, intelligent workers that exist in associations who are doing this crap manually. And it just pains me to think about that because that is not fun work. It's not really difficult work either, but it takes a lot of energy.
[01:07:17] Mallory: Yeah, I think it's a great example. I was, I've been writing down the essentials just to really, to really challenge our listeners.
[01:07:22] So it sounds like you need a rubric, you need some sort of file storage system, um, to hold all of your proposals. An accepted folder. Rejected folder. You need maybe one developer, because it sounds like there's some setting up rules, prompts, things like that. I don't even think that, I mean,
[01:07:37] Amith: that might've been true a year ago, but like, you know, box.com, you can go, you could go do that yourself, Mallory, without any development skill or experience.
[01:07:44] You could go do that with Zapier. I think right now you use, you've used Zapier, I believe in the past, and Yep. Zapier has a really cool agents platform. It's very consumer oriented. Um, there's all sorts of tools within the Microsoft ecosystem for copilot that are business user oriented. Member Junctions AI agents platform is [01:08:00] designed for business users to talk to an agent, to create another agent.
[01:08:03] These tools are increasingly accessible to business folks. Um, developers have a role to play in more complex workflows. I'm definitely not trying to obvious escape that statement. It's, it's important for developers to be in the mix, but, but I would just go try this stuff, right? I mean, if you can, maybe this isn't your production workflow.
[01:08:18] Maybe it's not the authorized, approved, sanctioned way to do it forever. Mm-hmm. But why not just go try it out? Maybe try it out for like a small event you have coming up instead of the big one and see what happens. Certainly get professional help, uh, to wire it all up and make it great and all those things.
[01:08:33] But, um, this is a time of empowerment. It's opportunity for everyone and it's really democratized. So, um, I think that if you're just a little bit curious and just a tiny bit technology savvy and, and you'll become technology savvy if you go do these things, um, you can do a lot. And that's what I find so incredibly exciting about this.
[01:08:51] Mallory: Mm-hmm. There's also platforms like I think make.com, I don't dunno if you've heard of that one, a Me through Innate N that are like Zapier, um, built for non-technical [01:09:00] folks to do tons of things. So it sounds like you don't even need a developer, at least for this basic example. And then in terms of dollars A me, I mean, I know I feel like this would be pretty affordable, but what would be your estimate?
[01:09:11] Amith: I mean, I don't even know, but I think like the basic box.com account's like 20 bucks a month or something. Wow. And Zapier, I think Zapier has a free tier, maybe. Yeah. And a couple Zaps. And there's like a couple bucks, like the consumer grade tools, which, which again, they, they're not the end all be all.
[01:09:24] They won't get you to the finish line for like enterprise wide automation. Your IT group might have a fit because they're saying, Hey, like, you know, you're not within like our secure environment, but that doesn't mean you shouldn't experiment with this stuff. With like, especially with proposals and stuff, like I'm talking about, these aren't like the most sensitive documents in the world typically.
[01:09:40] I'm not suggesting you take your employee payroll information and do this with it or whatever you consider, like really, you know, uh, you know, sensitive data. So, uh, I think that the, the consumer grade tools may be the beginning and the end of your story, and they may be perfectly fine, and particularly for smaller associations, they might be totally awesome in everything they ever need.[01:10:00]
[01:10:00] Um, but for, you know, associations that I say like. I would think maybe like more than 20 staff, more than 30 staff, maybe that's where you prototype stuff. Mm-hmm. And then you figure out like, where in your business infrastructure do you really do this? That's where like an AI data platform or maybe like a future generation of an A MS or something like that would have the facilities to have this workflow within it in a secure and, and like kind of managed environment, right?
[01:10:24] 'cause part of what you want for AI agents is observability. You don't just want the agent to just be kind of running amok and to not know what happened. You have to log this stuff. You have to have like really good instrumentation around it to keep it secure and to also learn from it. Uh, so there's a lot more to the conversation with the problem is, is that people think of this stuff as being such a big lift that they don't start.
[01:10:43] Mm-hmm. And so all I'm trying to do is encourage people to experiment, take things that are not sensitive information and go experiment with it, and then report that back. Like you could be someone that has no official authority over technology in your organization at all. Go and figure this out and then go demo it to people [01:11:00] who are in charge and have the ability to make change happen and say, listen, like I spent three hours of my time to build this solution on Zapier or box.com or whatever, and here's how it works.
[01:11:12] It's not perfect, but I can use this to save 80% of the time in doing this one step, right? So we found a pain point and we've used off the shelf consumer grade tools to solve a big chunk of it. Um, and then you can think about like what happens downstream from there. And once you get going, I think what happens is you've really started to, you know, you've got those mental reps and using these tools and you realize the possibilities to keep growing.
[01:11:36] It's not just this abstract idea of AI is really powerful and AI is really cool what I do with it, but you figure that out by playing in, in the sandbox.
[01:11:44] Mallory: And I did that exact thing. I talk about this stuff all the time, but until I wrote that little program that transcribed those audio files for episode 100, I, I just think of it in a whole different way because now I've actually done the thing.
[01:11:57] So if you're listening and you have a rubric, [01:12:00] a file storage system of some sort, 20 bucks, no developers, you can do that example that Amis just talked about right now. So what's stopping you?
[01:12:10] Amith: You know, in fact, Mallory, I think we'll go ahead and create a video showing how to do that with box.com and throw that out there in on the innerwebs and let people see it.
[01:12:17] That probably won't take more than a few minutes to set up.
[01:12:20] Mallory: For sure. Well, amis, it's been a, it's been a marathon of an episode, but a really good one. Do you have any key takeaways for this one?
[01:12:31] Amith: Go do stuff.
[01:12:34] Mallory: Alright, you heard it? Do you have 20 bucks? Do you have a rubric? Go do stuff everyone. Thank you for tuning in to today's three, maybe five topics.
[01:12:43] We will see you all next week.
[01:12:47] Amith: Thanks for tuning into the Sidecar Sync Podcast. If you want to dive deeper into anything mentioned in this episode, please check out the links in our show notes. And if you're looking for more in depth AI education for you, [01:13:00] your entire team, or your members, head to sidecar.ai.