The Claude Mythos Leak & the Reality of Computer-Using Agents | [Sidecar Sync Episode 128]

Summary:

This week on the Sidecar Sync, Amith Nagarajan and Mallory Mejias unpack a whirlwind of Claude news, from Anthropic’s accidental “Mythos” leak and what it could signal about the next major leap in AI capability, to the increasingly practical reality of Claude using your computer to complete real tasks on your behalf. They also dig into Mistral’s new open-weight text-to-speech model, the growing cybersecurity implications of more powerful frontier models, and what all of this means for associations trying to balance experimentation, security, and real-world productivity. Along the way, they explore where AI is already replacing repetitive work, where caution is still warranted, and why now is the time to start testing these tools for yourself.

Timestamps:

  🙋‍♀️ Chat with Grace on the Sidecar website

https://sidecar.ai/ 

👥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 ➔ https://claude.ai

Claude Code ➔ https://code.claude.com/docs/en/overview

Claude Cowork ➔ https://support.claude.com/en/articles/14128542-let-claude-use-your-computer-in-cowork

Claude Desktop App ➔ https://claude.ai/download

Mistral Voxtral TTS ➔ https://mistral.ai

ElevenLabs ➔ https://elevenlabs.io

MemberJunction ➔ https://www.memberjunction.com

👍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 - 00:00:25:06]
Amith
My name is Amith Nagarajan.


[00:00:25:06 - 00:00:27:22]
Mallory
 and my name is Mallory Mejias


[00:00:27:22 - 00:00:39:01]
Amith
And we are your hosts and We have quite an episode today for you. We've got a lot of interesting things coming from the land of Claude Are you excited to talk about all that Mallory?


[00:00:39:01 - 00:00:57:12]
Mallory
 I am I did some experimenting yesterday in preparation for this podcast I'm excited to share that you all know that we're a big Claude shop over here So some very exciting things coming out of anthropic, but I'm Heath I hear you just got a new member of your family and so you're a little tired today. How are you doing?


[00:00:57:12 - 00:01:52:03]
Amith
 Yeah, we're a little bit on the exhausted side. So we have another puppy that has joined our family We have a week old. Yeah eight week old golden retriever pup His name is Winston. He's awesome and he actually last night he slept for almost four hours in one stretch That was pretty awesome so Compared to the preceding week or so or it'd be an average of an hour and a half or two So my wife and I take turns on this and I tend to do more of the early morning shift and she does more late night shift, but We're both pretty tired. So but but he's awesome. The family loves him the two older dogs We have we have 11 month old Lucy who's actually been on this podcast before she she made a surprise appearance unbeknownst to me and jumped it jumped in but Winston will probably show up at some point in the not too distant future as well and And even our cat seems to like Winston Well,


[00:01:52:03 - 00:02:08:20]
Mallory
 you sent me a picture he's adorable I certainly have been there with the waking up every Hour and a half two hours. It does get better pretty quickly though. So I'm happy to hear that that was a four-hour stretch Eight weeks old is pretty young. So maybe you know who get to six hours and then maybe you can get to seven eight


[00:02:08:20 - 00:03:45:23]
Amith
 Yeah, I'll look forward to that. It's it's it's not too far off Well, I've had puppies before and they're they're they're awesome. Yeah, like human kids they tend to learn very quickly and they tend to Get to a certain level at least dog level of maturity within a handful of months and there's certainly within a year So but you got it. You got to kind of savor the savor the time. So they're puppies once You know Mallory talking about Claude and the big news from them, you know, they've they've They've just been doing a good job like putting stuff out into the wild unintentionally or maybe maybe unintentionally in quotes, I know we're gonna talk about one of those but Since we prepared this pod a day or so go in terms of the outline another thing happened Which is all the source code for Claude code the most popular AI coding assistant Was leaked on and perhaps unintentionally perhaps there's some more detailed subplot here that is unfolding but I think the people in entropic are very likely extremely smart people They have a lot going on. So they probably are moving really quickly and that may be the cause of the leaks But the entire code base for the Claude code product So Claude is a proprietary set of models the Claude code Software which you can use in the terminal as well as in places like Visual Studio is also not open source. It's proprietary code And there was something called a source map, which was unintentionally published To a place where you distribute the software from and as a result of that the source code is in the wild So it's kind of interesting


[00:03:46:23 - 00:03:53:07]
Amith
 And again, I don't know if this was maybe one of those leaks that was designed to get some press who knows


[00:03:53:07 - 00:03:58:01]
Mallory
 Okay, I was gonna say what would be the intentional reason for such a leak the press.


[00:03:58:01 - 00:04:29:03]
Amith
 I mean So the software can be decompiled and reverse engineered pretty easily See if you kind of understand the source code even if you don't have the source code having the actual source code certainly helpful but I don't know it's one of those things where I Don't know. I don't think there's necessarily an alternate storyline to this But it's just kind of fun to speculate because you know, we're AI land. There's not that much drama It's just we're talking about AGI and the replacement of all white collar labor at some point. That's not dramatic So let's have some leaks, you know, let's let's figure that out first So


[00:04:29:03 - 00:05:15:22]
Mallory
 well, we're gonna dive into another leak intentional unintentional from anthropic this week But before we do that and me I wanted to add this in as a topic But I feel like we're gonna already have a jam-packed episode with just focusing on Claude Mistral the French AI company released vox troll TTS their first text-to-speech model It's open weight just 4 billion parameters small enough to run on a laptop or a phone and Mistral says it matches or beats 11 labs and human evaluations of speech naturalness So it can clone a voice from as little as three seconds of audio and it supports nine languages right now We've covered 11 labs a ton on the show that seems to be a preferred tool across the family of blue cypress companies So I mean, what is your take on Mistral's version of a text-to-speech model?


[00:05:15:22 - 00:06:04:17]
Amith
 Yeah, so let's unpack that a little bit So audio has two major components. There's something called STT which is speech to text and it's what it sounds like It's when we're speaking on this podcast or if you're in a zoom call and you see the transcription happening in real time And you can see the words associated with what someone is saying that is speech to text and the inverse of that is Text to speech or the generative side where you're generating and they're both generative They're generating a different type of modality. But in the case of text to speech you're generating human like audio forms And so 11 labs certainly has been one of the stronger players. There are many others actually that do text to speech But having a open model that's very small that you can run potentially locally and there will be probably even smaller versions of this soon


[00:06:05:18 - 00:07:44:01]
Amith
 That's really exciting because it opens up optionality lowers cost increases speed All these things that are important particularly with audio. You have to have really fast low latency So you want to be able to convert, you know in this whole pipeline of text to speech Reasoning and then back to you know, the whole the whole earth Sorry speech to text reasoning and then text to speech in that pipeline. It's gotta be quick Otherwise it feels really strange You know if you've tried grace our audio AI agent on the sidecar.ai website that is powered in part by 11 labs and partly also by Claude haiku and the net effect is as you can get typically a response from grace is between 300 and 500 milliseconds or 0.3 or point to 0.5 seconds. This is really quick. It's almost human-like in terms of its speed And if you can make it even faster than it up opens up Opportunities for more like real-time back and forth more of a full duplex kind of a feel so it's pretty exciting I think in general these models are going to not only become better and more more commoditized But they're gonna become smaller and smaller and smaller In fact, I know we're gonna be covering this in the near future But there's been a lot of things happening in the world of model compression Which is where you make models smaller and smaller and smaller yet They're as powerful or in some cases more powerful than very large models of a prior generation and that continues to be the case There are a lot of new techniques that are emerging that the computer science and mathematics community are putting into the world and they're quite Exciting so we're gonna see more of this as the short take on it and it's really great in terms of making Making all this stuff more and more ubiquitous


[00:07:44:01 - 00:07:53:04]
Mallory
 Is the next iteration of this? I don't know what the technical term is Is it speech to speech models or audio to audio models where we skip the text portion?


[00:07:53:04 - 00:08:04:03]
Amith
 Yeah, that's a great question and actually there are models like that out there that are natively Natively doing the whole pipeline in one model. So you both get better


[00:08:05:03 - 00:08:59:17]
Amith
 latency time so faster lower latency time is better, but you also don't lose anything in the translation So when you go from speech to text you are compressing the information you're losing a lot of nuance But if the model directly is hearing what you're saying and then reasoning over those audio tokens rather than translating a text first It potentially can not only give you faster Times but also richer experiences and more detailed understanding Right now the way we approximate for this is that when we go to a speech the text model to then feed into a language model We asked the speech the text model and we've trained speech to text models to do this to give us additional information Beyond the words themselves about speed volume Potential sarcastic tone things like that that we try to pick up on but it's it's essentially still very much The cliff notes as opposed to the whole novel


[00:08:59:17 - 00:12:24:23]
Mallory
 Mm-hmm. Okay interesting enough We will certainly be keeping an eye on models like that in the future on the pod But we've got a lot to unpack here at me The first topic we're talking about is the Claude mythos leak and accidental quote-unquote Maybe data exposure that revealed anthropic's most powerful AI model to date one The company says represents a step change and capabilities and raises major cybersecurity questions And then we are shifting gears to something more practical Claude using your computer opening your apps clicking buttons completing tasks on your behalf while you're away from your desk One story is about what's coming and then the other is about what's right here right now So the Claude mythos leak on March 26th fortune reported that a misconfigured content management system ad entropic left nearly 3,000 unpublished assets publicly accessible including a draft blog post announcing a new model called Claude mythos After fortune contacted anthropic the company locked down access and attributed the exposure to human error The draft described mythos as the most powerful model anthropic has ever built and introduced a new tier called capybara Positioned above opus which was previously anthropic's highest capability tier the leaked blog reportedly said the capybara dramatically outperforms Claude opus 4.6 on coding academic reasoning and cybersecurity benchmarks and Anthropic confirmed it is testing the model with early access customers and called it a step change in performance Company also noted the model is very expensive to run and not yet ready for general release What moved markets was the cyber security language? so anthropic's draft reportedly described the model as currently far ahead of any other AI model in cyber capabilities and Warrant of an upcoming wave of models that can exploit vulnerabilities in ways that far outpaced the efforts of defenders Which is pretty scary? Crowdstrike dropped about 7% Palo Alto Network spell 6% the iShare cyber security ETF lost about four and a half percent Analyst Adam Borg wrote that the model has potential to elevate any ordinary hacker into a nation-state adversary And anthropic's own rollout strategy reportedly focuses on giving cyber defenders early access to strengthen their code bases before the model reaches broader availability In November 2025. I think we covered this on the pod But anthropic disclosed the first documented large-scale cyber attack executed predominantly by AI a Chinese state sponsored group had jailbroken Claude code disguising malicious commands as legitimate security testing and used it to autonomously target roughly 30 organizations including tech companies financial institutions and government agencies and Shockingly enough the AI executed about 80 to 90 percent of the operation independently discovering vulnerabilities writing exploit code harvesting credentials and exfiltrating data Small number of attacks succeeded an anthropic ultimately banned the accounts and notified authorities But the incident demonstrated that AI power cyber attacks at scale are happening right now So a lot to unpack here me But the leaked documents described mythos as a new tier above opus for listeners who don't follow Model names as closely as we do what does that actually mean to you?


[00:12:24:23 - 00:13:00:14]
Amith
 Well, I think anthropic is a lot more creative in their model naming than some of their contemporaries, which is it's kind of fun I actually really have always liked the haiku sonnet opus kind of trio cuz they're I think pretty clear in terms of their Relative sizes, which is kind of cool. I think they've you know, they've run out of names and that you know right in that group, but you know mythos or mythos or Callebara or whatever. It's called. I think there's just some interesting things to think about but whatever like The extra large model is what I'm gonna call it. So it's like short tall grande and venti So this is like the Starbucks.


[00:13:00:14 - 00:13:01:02]
Mallory
 Okay.


[00:13:01:02 - 00:15:10:08]
Amith
 Yeah, we'll do some we'll do some Starbucks I could use a cup of coffee right now actually not that this is an interesting conversation. It's just time of the day that time of the day, so anyway, so That extra large version of the model which is the extra powerful version. Here's the bottom line. We all know this is coming We know that AI is on a doubling path every roughly six months in terms of power Relative to price and that's continued to be the case in the last six months It will continue to be the case in the next six months. In fact, it might even accelerate further It looks like that's actually what's going on So you won't feel that in terms of it doubling every month or something But it'll probably be like five months and four months and the speed at which we're doubling power is pretty incredible That's happening on multiple different vectors. One is that we're getting access to more hardware Particularly as the year continues to go along there's more capacity coming online. There's bigger and bigger models And we're getting smarter at it, which is another vector people are getting better at model training There's a lot more capacity in terms of upside with RL. There's a lot more going on with you know model quantization And so all of these things ultimately come together to make models both faster And in many cases actually we're just talking about smaller models here I'm quite confident that this is a massive massive model, but it's incredibly smart. So You know my general thought process here is that I would expect in the next 30 to 60 days We'll see some actual announcement from anthropic talking about what this thing is We already think of Opus 4.6 is the best coding model on the planet And if this truly is a step change, you know, this is even more than like a 5.0 release, right? This is a big big deal because a step change actually sounds like a small thing, but it's it basically means a nonlinear Rate of increase relative to where you just were so you're like literally going like straight vertical You know in the and stepping over in terms of the chart, so I don't know I think it's I think it's really exciting I also think all the other labs probably are in similar places They all have stuff that they're cooking up that they obviously they're not out there telling us and they probably have two versions ahead of that But they're that they're working on planning right now.


[00:15:10:08 - 00:15:32:16]
Mallory
 I think what's hard for me is I guess since we're mere humans You know Opus 4.6 most powerful coding model in the world It's hard to fathom what a step change in that even means so where where does your mind go? I mean for associations in terms we don't even know everything that Opus 4.6 can do now So how are you framing your mind to think about? Okay another step change in model power?


[00:15:32:16 - 00:20:16:15]
Amith
 Yeah, look I mean, I think that's a great point I mean, you know I think most associations haven't even gotten to the level of fully taking advantage of GPT for era capabilities, right? GPT for turbo which was the model that everybody was trying to chase since you know Chat GPT was late 2022 and then I think it was April May something like that in 2023 And for like a year GPT for turbo reigns supreme and people were trying to figure out how to get anywhere close to it including anthropic right they were they were a fair bit behind and Now all of a sudden everyone's leapfrogging each other every every single day, but you may you raise a good point So even if you go back to what GPT for turbo could do A lot of associations haven't taken advantage of even those capabilities, right? So some of the basic business processes that associations run in terms of like classifying content There's people that are manually tagging content to put on their website There are people out there that are manually reviewing abstract submissions for speakers and for for articles for journals and stuff Without using any AI pre-processing to filter out garbage There are people that are manually doing a whole ton of other things that AI Could have done for them three years ago. So you raise a really good point I think the flip side of that is the more capable the frontier gets the cheaper Everything else gets and the more available it is to run even on your local machine Which actually does open up some use cases, you know it kind of ties into that earlier kind of mini topic that you introduced with with vox drill, but You know ultimately if you can run models and your own hardware it opens up Opportunities for associations that are ultra concerned about safety and security to run these things in their own dedicated environment so we know some clients that are doing that where they have let's say HIPAA type data or Financially sensitive data that they just do not want to share ever in any format with any inference provider So those are all positive things but in terms of a step change I'll give you a couple things not from the world of associations with the world of software development on what? Something like a mythos could potentially do that. I think is exciting if you have a step change in reasoning power the horizon of Length how long you can run an agent independently for will grow. That's actually how I like to measure power of these things It's not so much, you know, what are the benchmarks say and those are all a bunch of numbers ultimately, right? But I like to look at a different number that's very simple that I can process with my brain Which is how long can an agent productively run on its own and with like Opus for it was like an hour I think maybe a little bit over that then Opus 4.5 It was something on the order of two three hours and Opus 4.6 It's commonly, you know, it's possible to get Opus 4.6 to run for six seven eight hours In fact, there have been some experiments that are double that so what if it's it's possible For this mythos edition of Claude to run for a week or a month on its own And again, that has to be productive obviously where it's doing useful things You can definitely create an infinite loop pretty easily without any AI at all and let it run for as long as you want But you want the AI to be doing productive things We're working on a whole bunch of really exciting things for the next generation of our AI data platform at member Junction and In that new edition that we're working on we're adding a bunch of functionality around vectorization, which is this capability to basically, you know go in and You know take any form of structure unstructured data and convert it to a format But the AI can really understand natively and that sounds like a really technical thing which it is But the the benefit of that is to be able to do true Universal semantic search across literally everything the association has all the structured data all the files that you have all of your websites And then to be able to reason over those things and that that Synthesis of both structured non-structured data is something of a holy grail that hasn't been reachable until basically recently And so this is a very complex piece of software We're putting together and so we've had Claude Opus 4.6 through a number of different harnesses including Claude code and some other AI As well working hard on this and it'll typically run for four six eight hours and it will get pretty far But then it'll fall over itself because it's just too big of a project even as smart as it is So a step change would be welcomed and I think you know We're certainly not unique in the software development world looking for even more autonomy even more power to build bigger and better things So and there I think there's parallels to that in every other industry Like you know, how far can the model go with like a legal case? How far can the model go and diagnosing a patient's, you know medical condition, etc?


[00:20:16:15 - 00:20:34:10]
Mallory
 Yep, I Filmed another podcast episode today earlier me which I don't want to tell our audience too much about because it's a surprise But it's a really good one and we talked a lot about cybersecurity so I'm curious where your mind goes to with a model of this power and associations and cybersecurity


[00:20:34:10 - 00:22:23:02]
Amith
 Well, we have to realize that open source models that are freely available and are capable of doing literally anything that the user requests are Six months maybe twelve months behind generally at the most right and so the ability to do cyber attacks With even the current AI that's out there open source is Already there and we've been talking about this on the pod for a while It depends on the strength of the target and so some cybersecurity targets are extremely weak They fall over from a light breeze Others are really robust and it takes a category 5 hurricane to even you know, budge the door open slightly, right? And so Claude mythos is probably capable of doing things that you know We really don't want to have happen like hacking into you know grid scale utilities hacking into government institutions really bad stuff Whereas current open source AI is perfectly good at hacking into like individuals probably and maybe even small businesses which associations fall into And so I think there's there's a lot to be said around just being thoughtful about this in general My view is it's really great that we have frontier AI That's good at the defense side because in order to be good at defense. You have to know what the offense is gonna do You can't play defense, you know in a fully isolated area. So if you have a model that's extraordinary at cyber attacks Then you can use that intelligence to build models that are extraordinary at the defense side and share that broadly with the cyber industry And coordinate how do we protect against these problems because the whether these problems are coming now or in six or twelve months It does make a big difference because it's time to prepare But it basically means, you know Claude maybe at the forefront of this but not for long And so the only good defense ultimately against AI cyber threats is really good AI for cyber defense


[00:22:23:02 - 00:22:40:13]
Mallory
 Mm-hmm. This is more of a question for you and me like more of a personal question But what are your thoughts on the company? That's so big on safety and security having multiple leaks within the past few weeks Whether intentional or unintentional I do you see those as like things we should think about as related or or not


[00:22:40:13 - 00:24:03:17]
Amith
 I mean they're their core fundamental Value system revolves as I understand it revolves around the whole idea of safe AI And so they're hyper focused on this concept of alignment. They have this concept called constitutional AI where they're deeply focused on ensuring their models are Safe and are smart enough to be safe as well because to be safe You have to be really smart in order to have nuance and understand what's actually a good use and what's not, right? And you know at the same time, they're just a bunch of people over there just like us and Ultimately, that means they make mistakes and so not everything's automated and even obviously automated systems make mistakes, too so these leaks I think are Probably a function of them moving ridiculously fast I have a ton of admiration for those folks on multiple different levels one of which is their willingness to Just really rapidly iterate and put stuff out there to see what their community is interested in Building things that may not be desired and then seeing what works right and that kind of experimentation in public takes a good bit of courage It also results in a speed of execution that sometimes, you know has these outcomes So you kind of take your lumps along with your wins I think in when you when you decide to execute that quickly So that's my strong suspicion is that's really what it is. It's probably that simple But as I was saying earlier, it's kind of fun to speculate


[00:24:05:18 - 00:24:23:15]
Amith
 Maybe they're maybe they're trying to put this out there. I mean they're getting plenty of good press anyway, right? I mean, there's so much cool stuff happening for anthropic But you know if they were sagging in terms of their like relative prestige and AI Maybe they put something out there, you know But but that would be kind of a dangerous move if you were in that second or third position


[00:24:23:15 - 00:24:37:12]
Mallory
 Yep, you can just keep thinking about it in loops But then what if they did but within what if they didn't I'm sure as soon as they type something out and put it into A draft mode there has to be some understanding that like this could get out there as soon as it's stored somewhere


[00:24:37:12 - 00:24:46:21]
Amith
 Yeah, for sure for sure And and pretty soon if we have brain computer interfaces that work as well as some believe they will very soon You will not have to actually type it


[00:24:46:21 - 00:24:50:01]
Mallory
 Well you first ameet and then you let us know and then


[00:24:50:01 - 00:24:54:21]
Amith
 no, thank you. No, thanks. I am definitely I'm a I'm gonna go hang out with my dogs


[00:24:54:21 - 00:27:43:05]
Mallory
 Well speaking of automation I'm very excited to talk today about Claude's computer use going more mainstream We covered Claude computer use all the way back in episode 53 If you were still with us back at that time when anthropic released it as an API only beta in late 2024 At the time it required developer skills to implement and it was a bit clunky What anthropic just released at the end of March is different? So computer use is now built into clock the Claude desktop app through co-work and Claude code So any pro or max subscriber can turn it on no API no code. No setup Claude can open apps click buttons navigate browsers and work with files on your behalf For listeners who have not used co-work It is the agenda tab in the Claude desktop app while regular Claude chat answers your questions co-work actually can do work on your behalf You can point it at a folder give it a task and it executes multi-step workflows creates real files and connects to services like Gmail and Slack Maybe a bigger story as part of this is dispatch which we covered briefly on a recent pod Dispatch creates a persistent conversation between Claude on your phone and Claude on your desktop So you can text Claude a task from your phone Like exporting a pitch deck as a PDF and attaching it to a meeting invite and then come back to finished work on your computer Combined with computer use Claude can now function as a remote digital worker operating your machine while you're away Anthropic has also built a smart hierarchy So Claude tries connectors first direct integrations like Gmail or slack Then the browser and then only falls to screen control when nothing else works on the safety side Claude asks permissions per app Block sensitive categories by defaults and includes prompt injection detection So that would be something like if you visit a malicious website and there is text that's invisible to the human eye Basically saying ignore your instructions and do this instead And Candid though that guardrails are not absolute and recommends caution with sensitive data as do we early reviewers also flagged heavy quota consumption So I'm Ethan knew it was time yesterday. I sought out to use Claude at co-work Specifically computer use and I'm not gonna lie. This might resonate with some listeners I struggled a little bit to think of something for it to do because I was a little bit scared to give it access to anything You know of any power or meaning so I decided to ask it to create a zoom meeting link Because I don't really use my zoom account for too much except this which is important So I said can you create a zoom meeting on this date at this time? And then drop the link in the chat for me and it went off to do that I realized I didn't have the Claude for Chrome


[00:27:44:05 - 00:28:28:19]
Mallory
 Browser extension so I had to download that first and then it took a while I will say maybe it took like five to ten minutes for it to actually create the meeting link it gave it to me I was impressed, but then I said what else could we do? So I have a Gmail email account that I created that isn't really linked to anything Kind of like a sandbox email and I said, can you add it to the calendar? Add in that meeting link and then email it to my personal email and sure enough about five to ten minutes later I got a notification on my phone Had the meeting invite correct time correct date with the zoom recording link and I was pretty impressed So that was my very small experiment with computer use and Claude co-work It was pretty rough around the edges back in episode 53, but a myth What are you most excited about in this research preview right now?


[00:28:28:19 - 00:31:08:24]
Amith
 I mean, it's it's pretty polished now It can do a lot There's two things that point out one is that it can control any application on your desktop Not just browsers and that's important Especially if you have some older software in your association and we know associations, of course don't have older software right Mallory I'll cut it all Yeah, all brand new and so if you have a crusty old a MS laying around somewhere That's a Windows app or even like a terminal type app and there are some associations that are using mini computers with literally green screen terminals Which is like predates like graphically Yeah, it's like literally text. It's actually kind of like what cloud code is It's basically like a line-by-line type of thing but and so it's back to the future with cloud code, right? It's just a little more powerful But if you have these legacy type apps that are not web-based is the point Cloud code and sorry Claude in general Claude code as well But Claude dispatch Claude co-work can control those applications as well So let's say you have an old windows-based app for event registration and you don't like it But it's some custom thing that somebody built 15 years ago And it just works and you don't want to replace it right now But you sure would like to be able to get data in and out of there more easily Well, it's not gonna be super fast But you can instruct Claude to use that app on your behalf. It can learn the app You can actually save a lot of the information that would be important for it to know how to use the app in a skill Which allows Claude to Claude co-work to reuse a lot of the context about you know What does this app do and how do I navigate around it and so forth? Kind of like a human the first time Claude might go and use this old app It might have to click around a lot more to figure it out But then it'll learn from that you can ask it to save its notes on what it learned from that session into a skill And it'll be much faster the next time So that's really exciting I think because it forms a bridge between the world of AI and the world of older software that requires Literally a human touch because it's only available through older interfaces not a web-based interface So I think that's interesting a worth noting I would encourage people to experiment with that and try it out on a small scale The other thing is right now if you used it, you'll find that Claude is very smart Generally makes the right decisions, but it's also very slow in using your computer So clicking through a website making a choice clicking on the next thing typing something in all those actions that it takes Oftentimes take five seconds seven seconds even ten seconds in between actions I don't know if you had that experience if it was a little bit snappier for you


[00:31:08:24 - 00:31:16:21]
Mallory
 It took a long time and I was watching it make mistakes and then a reason like this drop-down menu isn't working the way I thought it would which was just funny to see that


[00:31:16:21 - 00:33:25:24]
Amith
 Yeah, exactly. So it's doing a lot of heavy lifting relative to something we consider super simple, right? And of course, we've got a lot of more experiences in these these tools than Claude does The models right now have to go through a pretty inefficient process to use the tools But they are able to what they do essentially is they're taking screenshots and they're saying oh what's on the screenshot? Oh, this is what I see just like your eyes would inform your brain what you see and then oh I think I see a button over there that says something that seems to be related to what I should do next and then it'll click On the button and then it'll see what happens next It'll wait half a second or second take the next screenshot and then the LLM will say oh, this is the next screenshot My objective is to achieve this goal of booking Mallory's meeting What should I do next and so it's kind of figuring out the application as it goes so saving? Information from the prior session as a skill can kind of give it a bit of a recipe so that Instead of having to figure it out totally dynamically It has a little bit of insight of what to expect in the user interface And you can tell Claude to save insights on how it navigated the user interface into a skill and it'll be faster the next time It'll still be pretty slow But my point is is don't be discouraged by this actually because you know that AI has just continued to go faster and get more powerful You will see models go much much much faster very soon And I don't know if you can do this with co-work, but in the Chrome plug-in Mallory There's a way to change the model and you can downgrade yourself to sonnet or haiku Instead of using opus and if you try computer control or compute brow in this case just browser control with haiku You'll actually find that it's quite good because it doesn't take the most brilliant reasoning to figure out how to use Amazon calm thankfully But it haiku is way faster. It's the same model we use for grace for part of what grace does so In any event, I think it's really a notable thing It opens up the world to AI to explore and use apps that it previously couldn't touch and that that's exciting fundamentally It also of course as you pointed out is something you have to be thoughtful about in terms of security


[00:33:25:24 - 00:34:22:24]
Mallory
 Yeah, and if in my demo example as I described it My dogs are shaking off if you can hear them in the background, but they're very excited about this, too I was very excited So I did that was a big moment for me in terms of AI use just realizing what I could do and my mind started spinning Okay Well if it can create the zoom link and create this meeting invent invite on my behalf What else could I do with it? but where I kept getting stuck a myth and I feel like our listeners will resonate with this is What should I do with this and what should I not do with this? So thinking about my podcast workflow after this call, I'm gonna take the raw files I'm gonna put them into a spreadsheet thing that I use send them off to our editor I'm now realizing okay Claude co-worker could do all of that on my behalf But should I give it access to my the raw files and should I connect it to my outlook? I'm actually struggling a lot with that piece. So can you talk a meet to Mallory co-host to co-host? What should we not do with something like this?


[00:34:22:24 - 00:37:27:06]
Amith
 Well, the first thing I would not do is I would not be logged into a password manager tool So if you have a browser plug-in for one pass or any of these other Password management tools which are by the way password management tools are really good idea to use because they generate unique strong passwords for every website You visit as opposed to you know Using the same weak password that you can easily remember And if you're logged into that then through computer use Claude could click on your password manager and get access to all your passwords and then you know Do bad things with that or possibly use those passwords to get into any website? Similarly, I would not give access If you use a Mac you're able to connect your iPhone to it and you can do this thing where your iPhone apps become available On the Mac one of the apps that you probably have in your iPhone and I think you can do that with Android phones as well a similar concept well what you don't want to do is allow the multi-factor authentication program that you use in your phone like Google authenticator or Microsoft authenticator to be accessible to the AI Along with your passwords because now you have both factors and that's really bad because then you're talking about potentially getting into banking sites And doing wire transfers and all sorts of stuff So I'm probably painting a very scary picture for you by pointing those things out first, but they're worth noting So if you're gonna let Claude run loose, I would certainly think of it Almost the way I described open claw and if you haven't caught our open call episode that just dropped I think last week or the week before go check it out in that I say open claw is not that big of a deal for Associations and the reason I say that is because here we are a week later talking about very similar Capabilities that are available in a mainstream tool that does have some semblance of security in it some guardrails some tools that try to prevent Malicious use whereas of course open claws just the Wild West anything happens anything goes There are versions of it now that are somewhat safer like Nemo claw and stuff We talked about in the last pod, but the point is is that Claude code when you're sorry Claude Claude code and Claude co-work They're all basically one thing in my brain. So I keep reverting to saying Claude coke. That's that's the version I use all the time They're very very powerful. I definitely encourage you to work with them, but I'd give them a sandbox So if you happen to have like an old computer laying around a lot of us do you know from three four five years ago? Something like that before we last upgraded Grab that thing Refresh the operating system or if you have it and it's working it's fine, but start off with like a clean slate So give it just access to the web give it access to a very limited number of things But don't give it your whole computer. So I wouldn't recommend like letting co-work completely loose on my computer With all of my logins and other stuff that that to me is a bit nerve-wracking And there's ways to also do this where you can set up like virtual machines on the same physical machine, right? so you can set up like parallels with A different operating system on your Mac you can set up virtual machines on Windows PCs as well That may be a good idea for this use case as well


[00:37:27:06 - 00:37:35:01]
Mallory
 So would you connect Claude co-worker Claude code to your crusty ams? Would that be overstepping?


[00:37:35:01 - 00:38:21:10]
Amith
 You think no, I would do that, but I would I would do it very carefully at first The first way I would do is I'd probably connected to like a test environment Most of these deployments have like the production environment But also a test environment and I've let it kind of run with that first And if I gave it access to my ams maybe in that same environment I wouldn't give it access to other tools like the web. I wouldn't give it access to like arbitrary websites I'm not so worried about like the the issue of Prompt hijacking or the issue of like, you know, the system just having some malicious underlying intentions that those are things that are much More far-fetched for most associations, but they but they are real I just think of it as a separation of concerns, right? You you partition the information that you're giving out until you increase your level of confidence in the system


[00:38:21:10 - 00:38:50:03]
Mallory
 Mm-hmm. I mean it sounds to me like with Claude co-work everyone listening should go out and test this but It could probably do the bulk slowly of a lot of people's jobs, which is concerning Definitely for first and foremost But maybe there's also some opportunity there in terms of the repetitive work that we've always talked about Automating eventually with AI and getting rid of quote-unquote with AI but that seems like today it is possible with Claude co-work That's no longer a future state


[00:38:50:03 - 00:39:06:02]
Amith
 I mean if you think about like an assembly line and kind of I always go back in my mind to the I love Lucy episode where Lucy was at the chocolate factory and trying to keep up with this conveyor belt of ever increasing speed of chocolates going by and all sorts of if you haven't seen that one go check it out.


[00:39:06:02 - 00:39:08:02]
Mallory
 I like I love Lucy. I don't know if I've seen that one


[00:39:08:02 - 00:39:56:13]
Amith
 Yeah, if you go to YouTube, it's probably on there, but I love Lucy and the the chocolate factory episode I don't know what it's actually called But that's my vision of assembly lines And I don't think that most people really like doing monotonous repetitive work over and over and over and over and over It just gets boring really fast and a lot of white-collar labor fits into that bucket There's a tiny bit of reasoning skill that's required in a lot of tasks that people do where in order to process any membership Application maybe most of its monotonous moving data from you know This the system to that system or this field to that field There's a tiny bit of reasoning where you have to look at it and say well I got to make a judgment call is this the right member category for this person? Do they fill in all the fields in a way that makes sense? Did they give us enough information right something that's somewhat subjective or qualitative?


[00:39:57:13 - 00:41:47:13]
Amith
 There's so many of those tasks and processes in so many jobs that yes Those things are going to go away, and I think that's actually cost for celebration. It is long term I should say that the the the thing that should put an asterisks on that statement for is that near term if all those changes Happen all at the same time there's gonna be a lot of unemployment And that's a problem that nobody's figured out what we should do about yet But the reality is is you can automate most of the tasks that most of us do myself included And so if that is true, which I firmly believe it is it's just a matter of time on how quickly we can diffuse the technology Into those processes now the same time there's a lot more work that needs to be done I don't think any association or any business for that matter would say we have reached the pinnacle The ultimate level of member service the ultimate level of content production the ultimate level of whatever I think there's always a higher bar to shoot for it a higher opportunity level a better way to serve your members your customers a Better piece of software that can be created and so if you think of this with the abundance mindset and say okay Well, I don't have to process membership applications manually anymore because the AI is doing it even though I've got the crusty old legacy system they I know how to use it now cool But what can I do with my time right and and not everyone's gonna have that attitude some people are gonna go well I don't want to learn how to do anything new. I'm not willing to change that that's a problem That's a really challenging place for some people and I have empathy for that But at the same time the exciting side of it is what can we do with our human minds and our human creativity? That we haven't figured out yet, right? It takes time to figure these things out So to me there's way more enthusiasm and excitement around it But I do realize there there is this gap between what the technology can do the time of diffusion of the tech and then ultimately What we do to reskill people and help help them in that journey.


[00:41:47:13 - 00:42:23:22]
Mallory
 Mm-hmm I feel like this computer use add-on to Claude co-work is has really made Personal AI assistants AI agents whatever you want to call them accessible to the masses because sure this was capable before But it seems like you had to go the Claude code route, which was intimidating for many people Claude co-work You use it just like regular Claude you just type out your request You approve some carefully approve access to certain websites or your browser And then it can do a lot of the work that you are doing today on its own. It's scary It's exciting, but I feel like today's a big day or this is a big announcement


[00:42:23:22 - 00:43:18:11]
Amith
 You know There's one thing that I'd flip my whole statement around if if we could add this to Claude co-work Which is an audio bridge the ability for it to make calls and receive calls I'd probably immediately suspend my concerns and just use it because I hate making phone calls and I really hate getting phone calls And not from like friends or something But like just just in terms of like having to make calls in order to like schedule appointment with some business They don't do internet they don't do email like, you know, like we'll always dogs We have in our house. We're trying to find a mobile groomer to come to our house occasionally None of these people do email they seemingly don't even have websites, but they have Google reviews and they have phone numbers Presumably they'd answer the phone if you called them Although that's not necessarily verifiable because I haven't called them But I would love for Claude to call the mobile groomers in the area of New Orleans I live in make a bunch of phone calls interview them and Get someone to come to my house to make my dog smell a little bit less. That'd be really cool


[00:43:19:21 - 00:43:28:24]
Amith
 That I would totally say hey, you can have access to all my passwords Not really, but like I would definitely not quite but I would definitely use I definitely use it tomorrow So


[00:43:28:24 - 00:43:41:08]
Mallory
 I thought what you were saying is Claude having a number you a meath could call and then you just tell it What were basically like, you know a voice in your face And then they're gonna suspend your account because I bet you would use that an awful lot if I had to guess


[00:43:41:08 - 00:44:00:07]
Amith
 well If you haven't used Claude's mobile app and their new audio capabilities, it is quite good You can't talk to Claude now and it's just as good as chat GPT chat GPT the only thing I was using chat GPT for like kind of the last six to twelve months was their mobile app on the iPhone to talk to chat GPT, but I think Claude has caught up on that as well


[00:44:00:07 - 00:44:14:09]
Mallory
 Mm-hmm. I well, it's been a couple months which in AI world is years, right? the only thing when I would use the voice mode on Claude is that sometimes my Chat history would disappear and it was really frustrating after I had had a conversation


[00:44:15:09 - 00:44:16:22]
Mallory
 But yeah, that would be my only note


[00:44:16:22 - 00:44:26:04]
Amith
 Yeah, they had some bugs in there and previously was more like walkie-talkie like where you had to press a button when you wanted To send your message, but that's all gone now. They have it. It's just a very natural conversation No,


[00:44:26:04 - 00:44:32:22]
Mallory
 well a meath lots of anthropic Claude news What is your one?


[00:44:32:22 - 00:45:01:02]
Amith
 takeaway, I Mean, I probably sound like a broken record to everyone listening to the pod But I think you should just go ahead and try this stuff, you know listening to us Listen to other people kind of watching videos, whatever It's a great way to be aware of what you can do But the light bulb doesn't turn on until you actually go in there and play with this stuff like your use case your killer app So to speak it's there. There's many of them there, but you have to go find them. So go go try this stuff out


[00:45:01:02 - 00:45:22:03]
Mallory
 I'm serious and I'm gonna double down that really go try out Claude co-work with computer use today Just do it today, you know, maybe by the end of the weekend by Sunday I'll give you a few days try it out if you haven't done it I like a meath just said I know a lot about it But that was actually my first time using it and whoa My mind is going in 20 different directions on everything I could do with it


[00:45:22:03 - 00:46:38:18]
Amith
 Yeah, and Mallory and that's kind of like if you think back to your first encounter with an AI where you Looked at that blank screen with that blinking cursor and you're like, what do I type into this thing? a lot of people have felt that almost anxiety from not knowing how to use it and and that quickly Changed if you did use it and you're like, oh it can do that. Hmm. Well, I can also get to do this Yesterday I got this phone call from a friend of mine who's decidedly non-technical But he was he was calling me because he said I just spent three hours with Claude code I said really he's like, yeah, I built this thing goes to my LinkedIn profile It pulls down this information then it goes to this thing that gets this other information Then it was this that that he was like talking like at a hundred miles an hour. He's so excited about it It's it's basically the same experience is called kill work that we're talking about here But using a genteck AI where it can perform real-world tasks for you You can't figure that out until you start using it that and that's kind of the conundrum But also the opportunity right because there is a time delay between the availability of the general purpose technology Which is where we are now and the point in time where most people are using it That's this diffusion delay and ultimately we're very much in that right now and you can take advantage of that That's an arbitrage opportunity for you to deliver Outsize gains to your mission


[00:46:38:18 - 00:46:44:08]
 (Music Playing)


[00:46:55:01 - 00:47:12:00]
Mallory
 Thanks for tuning into the Sidecar Sync podcast. If you want to dive deeper into anything mentioned in this episode, please check out the links in our show notes. And if you're looking for more in-depth AI education for you, your entire team, or your members, head to sidecar.ai.


[00:47:12:00 - 00:47:15:06]
 (Music Playing)