EP. 03 - Personal AI Agents, Hacking, Flirty Voice Bots & the Skill Issue.

JJ: for me this week, I I feel like it's been all about AI agents, personal agents. I don't know about you guys, but I've just been spending a ton of time between Grockbot, Muse, Instinct, and now Claude Duet.

Jack Zoppa: Just upset that you have Claude Duet and none of us do.

Trevor: I know. I know. haven't tried it yet.

Michael: Yeah.

Trevor: I've been using Instinct a lot more, I think over the past couple of weeks. I I don't think I've opened Muse in over a week, honestly. I've been just using Instinct for everything. Yeah. And then I have like for Grockbot, it's connected to T three using the MCP and then it monitors like anywhere a bug would be and then it will just auto fix any bugs that like the team reports or comes in.

JJ: what's your reasoning, Trevor, for using instinct over the others?

Trevor: I'm always an iMessage. So it's like it's just there and it's a lot easier. And I think the group chats are awesome. I've added it to a handful of different group chats. I kinda like that like it kind of just does things like Jack was saying last week. Like it does things and you like don't even really know what it's doing, but it just does it. So it feels like it

JJ: Yeah, it is weird. There's not a lot of transparency behind

Trevor: Yeah.

JJ: Instinct compared to some of the other ones, but it is interesting. And for the listeners, we connected Instinct for our group chat for this podcast. And so now it's doing the podcast prep for us as well, just in our iMessage group chat. you had a reaction on when Trevor said he wasn't using Muse much anymore.

Jack Zoppa: No, I just think it's I I literally had this thought last night. I was like, gotta start dialing in like one of these things more primarily

Trevor: Mm-hmm.

Jack Zoppa: than the other or two. Like I really like to split like personal, call it executive assistant type stuff versus

Trevor: Mm-hmm.

Jack Zoppa: heavier workload stuff. but yeah, I mean I've added instinct to yeah, like a lot of group group messages at this point. I really like it. I think it's like pretty proactive. It's nice that it's a nine message. I I had some Instinct built us a prep doc for our pod today based on our group chat. One of the things that I was gonna hit was so I'm I'm buying a building right now, like a commercial building. And in order to find it, like Grokbot literally emailed back and forth with like forty or fifty different sellers that I was like of what I was looking for.

and then set up

Michael: how many dashes?

Jack Zoppa: the set up the whatever. I went and saw it. And then I mean and like one of my operations guys, we put Instinct in a group chat. It's literally called like the move in group chat with one of my other assistants. It's us three and Instinct. Instinct has planned the fire marshal inspection that I already went to, it planned the building inspection, went back and forth with both of their offices to do it. And it has a sixteen point like move in checklist that it's like continuously following up on like appraisal scheduling or like y whatever we're getting done.

And I don't know, it's pretty crazy just to have this iMessage group chat that this is all happening through and all of us can hit it the whole time

Trevor: Mm-hmm.

Jack Zoppa: and it actually acts right in in a group chat as

Trevor: Yeah.

JJ: Mm-hmm.

Jack Zoppa: well, which I think is big.

Trevor: Yeah.

JJ: Not like some of the Slack agents that we have that they sit there and stream their whole thoughts.

Jack Zoppa: Dude, my Hermes early on, you remember like

JJ: So bad.

Jack Zoppa: in one of our slacks was like literally just bombarding everyone. Yeah, yeah.

JJ: Seventeen messages per one response.

Trevor: Yeah.

Jack Zoppa: are currently raising at thirty billion instinct. Yeah.

Trevor: Instinct, which makes

JJ: Instinct, gotcha.

Trevor: no sense to me because it's completely free and I'm kinda confused what I mean, it's probably the same angle that Muse is going at, but Muse just has so much more like it has meta behind it. It has Facebook and Instagram and all that other stuff behind it. So I'm a little like I don't know what their plan is or like what they're doing for like monetizing. Yeah.

JJ: Selling data. It's all about or you know, eventually too. I think that once these agents start pr making purchases on their own, once it gets to be like commonplace, then they're they're gonna be monetizing through that,

Jack Zoppa: Yeah.

JJ: either it's you know, through the processing fee or through like sponsorship, so that if you say, I need to grocery shop, it's gonna direct you towards Whole Foods versus Kroger or or whatever, and then Whole Foods is gonna pay for that. placement. That's that's my theory on where these free

Trevor: Mm-hmm. Yeah. Yeah.

JJ: ones are going.

Jack Zoppa: I agree.

JJ: Michael, what agents are are you excited about currently?

Michael: I've got my Grokbot team pretty dialed in. I even just went through just laying in bed last night, scrolling through all the new connections

Jack Zoppa: Mm.

Michael: that are in there, seeing if I've got any that shoot, I've used this app. I should probably get this one connected or some random ones. I'm like, I think we had an org chart on Miro at one point. Let me connect that just so it has context. I feel like every once in a while I'll just randomly try to make sure that everything's got context. The hard part is then doing that across Muse, across Instinct, across T three with Claude and Codex and you're like, wait, I thought you had connections to this. And you're always like reconnecting things.

JJ: Yeah. I Jack brought up a good point of like you need to like go deep on one because

Trevor: Mm.

JJ: that you know, that context is hardened. I I actually so this week I started using use per your guest's recommendation and I actually really like it. I was I was pretty

Trevor: Yeah.

JJ: hesitant at first, but I do really like it. So I am I moved everything off of my grok pot for for personal purposes and moved it all into Muse. So I exported all of my skills that Grokbot had and I exported my memory and everything. And now I'm doubling down on Muse as as my personal assistant. And then I'm gonna sunset Grokbot as a personal assistant, but I think I want to start Grokbot's through the company. And that's something I wanted

Trevor: Mm.

JJ: to talk to you guys about is what's you know if you want your your employees to be using an agent, what do you think is the best route right now? Do we go, do you do a self-hosted Hermes Open Claw? Do you use Grockbot? Do you use Instinct? w what are what are your takes on that?

Trevor: I think both. So I think instinct is so good. Like most of the people on our team do have iPhones or iOS, so then we you can get a group chat going and then it being able to work within your work with your own instinct, AJ, I think is really cool. and you can connect an MCP to it. So like our operations system that that we've built. has its own MCP that has access to absolutely everything. So if you're on the fly and you just want to do something really quick and instinct, I think it's totally okay to ask it to go do something. I think Grokbot is just really good if you're gonna f really utilize the sub Grokbots where you have one managing all the other ones.

and you could really like Jack was saying, like have it do some crazy ongoing tasks and like finding a building for you to buy and going back and forth and whatnot. I feel like both are useful and important. And I think Grockbot's more work focused. And then Instinct or Muse is more like productivity focused, where it's like just on the go. that's my take on it.

Jack Zoppa: so for the team, we have like a self hosted Hermes agent that 'cause we do so much in Slack. So like the Slack bot is very nice to be able to like interact with everything. I really like the idea of going in to like XAI's team plan potentially and having these grok bots in Slack and out of Slack for the team to interact with. My only like Worry there is usage pools. I I don't know so with with a Hermes agent everyone interacting with, I know the back end usage, right? I can monitor that, I can switch subs, I can increase it, an open router, whatever.

But with Grockbot, I'm like, what happens if our really important team agent just hits weekly usage and it's just done? And you can't Yeah, I I don't know. I'm I'm still kind of going back and forth on that. That's my literal only hesitation. I have a GrockBot in Slack as a team agent that I've been testing and it works really, really well. honestly, like I don't know, most of the listeners of this call are probably like self-hosting a digital ocean droplet of a Hermes agent and getting it really, really right for Slack is not the easiest thing in the world for everyone.

Yeah.

JJ: Which is why I haven't set one up a hundred percent. Now that now that I'm seeing Grockbot and it's like, I just add a couple

Trevor: Yeah.

Michael: So

JJ: of C Ps and throw it in Slack, that's so easy compared to Hermes

Jack Zoppa: Yeah.

JJ: and Open Call where it's like I spent ninety percent of the time just troubleshooting while it wasn't working. you still have some Hermes agents, right? Are you thinking of switching those out or wh where is your mind

Michael: We haven't. yeah, we've got a couple on a on a mini computer in my cabinet that's running a maturity of the company, which is a little sketch. we've moved some of it to DigitalOcean, but like Jack said it's a little tricky and there's a lot of nuance there. Thankfully I can use my Grokbot and like GPT work with full computer access to go and get all those little settings dialed in right. but it's really nice knowing that hey, we've got these agents for each department. we kinda nerded out they're all Lord of the Rings themed. So like Gimli's our maintenance guy, we've got Gandalf, he's like the chief of everything.

He can code everything, he can change settings on the back end, he can remove users. And then some of the other ones are a lot more scoped down and we set this up before Rockbot and Muse and stuff even came out. And so it it took a while to get the settings dialed in, but I didn't want my maintenance tech to be able to like send something to in Slack to Gimli about, hey, what's the furnace size at this house, furnace filter, or hey, is this one on weekly mowing or bi weekly? And it accidentally have enough power to go email the owner. wait, hey, we need to make sure owner

JJ: Right.

Michael: relations is completely separate from like the maintenance guy that's like, hey, you're stupid, you're not answering the right question and like commingle those things.

JJ: Mm-hmm.

Michael: And so I think what the new ones have done, they've gotten really good at those permissions, without you even really setting them up. But yeah, like what Jack said, it is mysterious. You're like, I'm not sure where it's working or what it's doing, but hopefully it does the right thing.

JJ: Yeah, and that's something I I'm planning on setting up a team Grokbot. I've got a couple of use cases. One is for our sales guy to use so he can update the CRM as he's kind of going and that sort of thing. And then the other one is for our owner relations team to use to be able to pull in metrics from like the various sources so that before they get on an owner call they can know like, here's you know, here's the rating, here's what all the tasks that were completed. here's the revenue numbers, all that kind of stuff. my personal grok bot, if I connect an email, that email is connected across all of the sub agent grokbots.

And so for like you said, Michael, I don't want the sales guy asking Grokbot a question, and then somehow it's like s emailing an owner from the owner relations grokbot. So I I I assume there's some permission settings on the team Grokbots that you can only set certain permissions on, but that's something I'm absolutely curious about and also a little bit hesitant because we you definitely wanna go like lowest possible scope on on some of those.

Jack Zoppa: Michael's sales guy, I helped him set up like a grok bot with an MCP into the into like the CRM and it's crazy how useful that is. I feel like there's a couple like quick little unlocks for team members when they're like, I can never have to get on a software again. Like I can voice to my phone

JJ: Yeah, yeah.

Jack Zoppa: for everything now. It's pretty wild.

JJ: That's that's my plan for the sales guy, is like let him call and talk to prospects and then he can just voice to text Grok and say, Send out this contract, move this stage, add this note, all of that stuff where he doesn't even have to get in the CRM at all, but now the C RM's up to date. Michael.

Michael: Yeah, Jack really helped getting Brian dialed in. Brian is awesome. I wish we could clone him five times. Our sales guy's just crushing it. But he's like the classic. He came out of tech sales previously and he's very like dialed into like what's gonna help me close the most? How do I like be as successful as possible? And he's he is very, I guess we used to say claw pilled, but he wants the Tesla to just drive him to every house while he's like voice texting his Muse agent,

Trevor: Yeah.

Michael: sending contracts and he's like s showing up at their door and knocking on it, closing a deal. that's what he's trying to be like as efficient as possible and it's like AI agents for salespeople is like you're supercharging one person a ton. Whereas it's a little harder to do that with like your housekeeping team or your maintenance team. Like that's dispersed,

Jack Zoppa: So I do have an agent and it's a Trellis agent. but for lot like if you wanted to do this yourself with a tool, you could literally just set up a Grok bot to do what I'm about to say. And I might actually end up doing that. but it's in a Slack group or in a Slack channel, and all of the cleaners report things or communicate through this one Slack channel. And so what's cool is like my COO that runs a cleaning company, he has the entire channel. double translated into English and Spanish, which is really neat. So every single message is like a double part English and Spanish so everyone can read everything.

And then the agent in there takes all of the information they send of whatever's wrong with the property or whatever, and it like triages it back to the team members that need to get it fixed or schedules the work order for the next day or that day or lets the guest know you know this thing's not gonna be working or whatever. the cleaners are tagging it now. And you you can go through and yeah, they're like they'd rather talk to this thing that can just like get it done for right now. It's kinda Yeah

JJ: Instant answers. Yeah.

Jack Zoppa: yeah.

JJ: I love it.

Trevor: Yeah, it's the equivalent of Jarvis, the but that's more like customize what what we built. It's not like on a Grokbot or anything.

JJ: I set up T three code. I've got my Claude and Codex subs on there. It has been actually really nice being able to toggle. I I was like I I don't think it's that important to switch between them, but I'm seeing a lot more now with code stuff that Claude is getting a lot more like, I'm not gonna do this, you know, and scraping websites and things like that. So it's been useful being able to just toggle over to open AI and have open AI do it.

Trevor: Yeah, there's a couple skills that I use. One of I saw this on Twitter. It's called Soul Loop. So on T three it's really nice because basically, you know, I'm trying to like prompt less and like let the agent go and work and do its thing once I've like determined that it we're on the same page of what I'm wanting it to do. So I trust and want Opus five point five managing and doing everything and doing most of the coding as well. But then being able for it to like call codex GPT 6.1 soul sub agents to review all of all of its work that it's doing.

I think it it it's just different. It's better at different things. Like it's more thorough and it's really good at like going deep and finding bugs. Like it will just like work nonstop. So I think that it's cleaned up a lot of the stuff that that we've been shipping, which is why I like using T3 because everything's still on one thread. Like it's very easy for it to orchestrate sub agents. you could add Muse Wonderful on there, you could add literally any other you could add Groc if you want to. and just getting more opinions 'cause the different models are good at different things,

JJ: Let's dig in on that a little bit because I'm looking for ways to improve my coding flow. So right now I'm using linear and it's that I feel like alone has been a great hack where I will scope out an issue or you know a feature or whatever, and then Opus will go through, write this really nice long linear ticket, definition of done so that you know it'll ship it once it has completed those features. But I would love to learn more about the second sweep that you're talking about because we I don't think I'm doing that right now. I do have a pretty cool setup where it's going through linear overnight and picking up tickets that are agent ready.

And if they're not agent ready, it will like ask me a question on more information to finish scoping the ticket. And then another

Trevor: Mm-hmm.

JJ: agent will come in and pick up the ticket and work. And then a final agent will come in and make sure that that ticket got done and close it out and make sure that it got pushed through GitHub. but other than and Michael, you can tell me because I know you set this up, I I believe, but other than like the GitHub CI checks, I don't think there's any like additional checks going on. So I I think it might be s wise if I'm coding with Opus to then have Codex go in and double check it. So is that is that what you're talking about, Trevor?

Trevor: Yeah, so that's exactly what I'm talking about because CI is also expensive. So the more you do it, yeah, yeah.

JJ: We learned that this week. We maxed

Michael: Ha ha ha.

JJ: out the plan we had to upgrade, so

Trevor: Yeah

Michael: But I think it I think at one point we even had it like where it was a little overkill where it was like Claude would ride it and then it would go to GPT and then cursor was double checking it and then your agent

JJ: wow.

Michael: would do something and you're like, Wait, how did I hit usage on everything? That's like, Well, you're double checking everything five times.

Trevor: Yeah, yeah.

JJ: Gotcha.

Trevor: I've dialed it in a lot this week. So that was one of the main things I worked on this week. because our bill was also getting super expensive and I think there was a lot of unnecessary stuff going on. So actually Yash at conduit, I don't know if already like posted this, but on their blog, he basically wrote a blog post about their whole like CI workflow and what they do. and I had switched last week to Blacksmith instead of using GitHub because Blacksmith just uses better and faster and more like stronger machines. so I'm using Blacksmith now and it's going faster, but then also like dialing in for your agents, like A, I think doing like the codex check on everything that you're more ready to go code so you're not wasting time.

running CI and everything and then getting it back and then having to rerun it over and over and over again. and then also kind of like breaking down more what you're actually checking. Like you don't have to do a full check every single time you're shipping like one small feature or one one thing that you're doing. And most of this you could just dial in with in within your like agent MD settings. But yeah, and I another another skill that I'm using, it's called full send. So where basically I'm like give a task I'm like, okay, go execute this, use soul loop and full send.

So then that's basically go through, finish this out, do your testing, merge it, and then deploy it, like all the way to the end. So you're not having to like go back and check it once I feel confident. it feels weird because I feel like I'm like you're more trusting of your agents and they're doing more work on their own. But I feel like I'm not working as much. So then I'm like, are they doing what I want them to do? So there's a little bit of trust involved with that. But I think I've been able to manage more going on. And then because the orchestration on T three and being able to use all those different sub agents and check their work and you know, it's almost like a council that you're building out of of agents to to check everything.

our workflow has been a lot better.

JJ: And where are you putting those skills? Are those skills in T three or are they in the agent that's doing the code?

Trevor: So you just get I strictly speak through T three, but it like T three's running the codex and Cloud Code CLI. It's just like through proxy. So it's basically s I mean guess it's like technically saving on T three, but those agents should have access to those skills no matter what. so yeah, everything I'm doing right now is through T three.

JJ: put the skills in the in the show notes because that that sounds super useful.

Jack Zoppa: Yeah, so everybody can download. And I can I can run those JJ on my on my new daybreak model.

JJ: Yeah, t Jack, tell us about daybreak.

Trevor: Yeah.

Jack Zoppa: Well, I think Trevor got it too, but Tre I mean

Trevor: Yeah.

Jack Zoppa: on the last pod, I was like, I gotta have it. So

Trevor: Yeah.

Jack Zoppa: on the pod, I ordered it for next day delivery that then got I paid extra for next day delivery and it got Delayed, which is insane. Unbelievable.

JJ: Ha ha ha.

Trevor: Yeah.

Jack Zoppa: but yeah, you get like a UB key and it's like a specific open AI Ubi key. and basically you have to like go into your codex account and be like, I'll give you my firstborn if you let me use Project Daybreak and like

Trevor: Yeah. Yeah.

Jack Zoppa: Like on your codex, you basically have to be like, yeah, like I might not be able to do X, Y, or Z anymore.

Trevor: Yeah.

Jack Zoppa: I will no longer be able to sign in on my phone easily, all this crazy

Trevor: Yeah.

Jack Zoppa: stuff. which is fine. And then every time you have to like tap the Ubi key, enter the password, do the recognition, and then retap the UB key to get in. Once you're in, you're in.

Trevor: Yeah.

Jack Zoppa: that was the setup. Trevor texted me, he's like, How the heck do you set this up? But yeah, that'll do it.

Trevor: Yeah. Yeah. Well I figured it out. It's 'cause I was on Safari. Like the UB key it it kept pulling up like the the Mac OS keychain instead of the one on on

Jack Zoppa: Yeah. Okay.

Trevor: OpenAI, but

Jack Zoppa: But it's cool. so I have like a personal app. Me and Michael have like similar maybe personal apps that we've built. And I had it scan like security vulnerabilities across it, found a ton of like little things here and there, especially on the superbase side of things, and I was like, yeah. Should probably get my finances locked down better.

Trevor: Mm-hmm.

Jack Zoppa: So and then

Trevor: Yeah.

JJ: So tell what's what's the difference, right? If you're if you're using Daybreak Crib something, another model. why why is this special?

Jack Zoppa: So yeah. So I from my understanding, we don't necessarily know the underlying model. I believe it's like five point six, but it's probably now like six point one soul. Okay, yeah, so it's probably

Trevor: I it's six. Yeah.

Jack Zoppa: six point one soul or something like that. but it's essentially just a fully unrestricted version of codex, like somewhat fully, to where like it no longer will tell you I won't do this, so I won't look for vulnerabilities, I won't try to like hack your own system. and all that thing. it's it's just unrestricted. And you access it obviously through your plugged in UB key and go through the facial recognition, ID verification. you have to do the whole flow to get it. It's a lot, I just sent my ID. But they have Daybreak Red, which is like Fortune 500 only, I think, or something

Trevor: Yeah, yeah.

Jack Zoppa: like that. but yeah it's it just seems like extremely persistent. So like I put it through like StateOS, one of our big projects and I obviously sent you all the vulnerabilities. Like a lot. and it but it was like two hours it was going for. I I I I was watching like what it was doing and I was like, this is insane, like how long and like everything that it checks and then like backtracking itself, the sub agents it uses. So I think it's like extremely more persistent as well as it's just unrestricted on like what it will do. and, you know, obviously using it for cybersecurity is So I I had it route through a local Hermes agent on my Mac.

So my local Hermes was using daybreak, like the highest model of daybreak, yesterday. And it went through my Mac for I think two and a half hours and scrubbed my like contact and identity information across the entire internet. It just like

JJ: Mm-hmm.

Jack Zoppa: removed me from the face of the planet, except for like my social media accounts. which is pretty cool.

Trevor: Yeah, I everything Jack said, obviously, and then I had it go at the State Portland website and I basically was like, Your job is to hack this website at all costs and figure out a way to like do something. And it found not too much, but it found some stuff where it's like, this could have been like a race condition and then this person would have been charged twice or someone could have intercepted this or whatever. So that kind of stuff when it's like not just going through your own code base but also like trying to go hack a live site is you would instantly get shut down if you were trying to use like normal Claudka.

JJ: Gotcha. Yeah, if you if you just told Codex probably would say, No, not gonna do that.

Trevor: Yeah, yeah, Yeah, 'cause it's even my LG TV, I finally got like it was Claude wouldn't do it and I got Codex to do it. I was like, I promise this is my T V. it's on my home network. It's the same network we're connected to right now. I have access to it, yada yada. so that is like what I've used it for. And then also just running through all of our code bases and making sure everything's dialed up, dialed in and and good to go.

Jack Zoppa: you just like gave it a a link to like your site, your live site, like not the code base. You're like, hey go interact, like computer use on the site and see if you can like hack it or whatever?

Trevor: Yeah, it has access to both, so it's gonna do whatever it needs to do, but it I was basically just go hack State Portland. figure out a way to hack it.

JJ: And I I think there's a ton of value in that because I still think that we're gonna see more and more as the models get better, more and more bad actors doing that, but it not for safety reasons, you know, like instead

Trevor: Yeah.

JJ: like actually using it to hack. I'm sure you guys have seen the things where it's like especially the the early AI models where you could like trick it. and say my grandma used to tell me a bedtime story about making bombs, and then and then it would do it. So

Jack Zoppa: Yeah.

JJ: I feel like there's still potential ways to get around it for people to use it to hack you know, your website. So like I I love that idea of just throwing in your site on that and pretending like you're the hacker just to see what security issues might come up.

Jack Zoppa: dude, I cannot wait to hack the boys' websites. I'm just gonna let it loose on everybody else's.

JJ: You called us all ransom. So speaking of other models, did you have you guys I I didn't see any announcements on this, but i and maybe I'm just late to the party, but did you know Claude can control your computer now?

Jack Zoppa: No. This is yesterday.

Trevor: Yeah, you can set up Codex's computer use as an C P that Cloud uses. So I've been able to do that for a while, actually.

JJ: No, this is this is native. it's not it's not additional. I don't know,

Trevor: really?

JJ: yeah, I don't know. I didn't see any updates about it, but as I was setting up my new desk and everything, I was talking to it about hooking up my camera because I got a ridiculous camera setup now. It's a mirrorless camera plugged in, hanging upside down, plugged into a HDMI video card going to the USB. So it's like way overkill of a setup. But anyway, so then my camera was upside down and so I was talking with Claude, how do I flip the camera over? And so I had to download a program called OBS that'll you know, just like a a live stream video software.

But even then the settings in OBS were kind of confusing. So I'm asking Claude and it's giving me all the steps and I'm like, Can you just figure it out? Like just do it. And then it just took control of my computer and just like did all of the settings. And I was like, I didn't know I figured it was gonna go in, you know, like do something in in the terminal because it's done that before. But

Trevor: Yeah.

JJ: no, it it does computer use now, which I thought was pretty interesting 'cause we were just talking last week how I had to get Astra to do the full computer use. And I like, Claude does this

Trevor: Mm-hmm.

JJ: now. So

Jack Zoppa: Yeah, I'll I'll have to test that out. I I'll be very curious if it's better than Codex because I feel like I think I said this on the last pod, but Astra Codex computer use I feel like was a GI for me. I was like, This is unbelievable. I

JJ: Mm-hmm.

Jack Zoppa: don't do anything. So I'd be so curious how the clause works.

Trevor: Yeah, yeah, I agree. Yeah, I I think Codex is is still better. 'cause I I I think I had the same experience as you. Like it it was able to do stuff, but it's it's more limited on certain things. Like I don't think have you tried using it in a browser?

JJ: No, I've used cowork, I didn't even know that it was going to take over the computer until I just saw like the orange

Trevor: Yeah.

JJ: thing that that it doesn't cowork too, but this wasn't cowork because it was your whole desktop computer. So

Trevor: Yeah.

JJ: we'll definitely have to play with that some more this week. And then speaking of other Claude models, I got access to Claude Duet, which is their response to Grokbot and dots and all that stuff. interesting story. I saw a LinkedIn post that somebody posted that she said that she got off of the wait list because she asked Claude about Grokbot. And so I did the same thing to test it. I'm tired of having to set up new agents. Like I have Grokbot and Muse, but I really just wish that I could use Claude duets because Claude knows me so well. And I kid you

Trevor: Mm-hmm.

JJ: not like three hours later I got off the wait list. So I a little creepy because that means that Cloud's definitely reading all of your messages. But yeah, so it's it's been it's been interesting. I I don't I'm not super impressed with Duet from the get-go. And I think this is

Trevor: Mm-hmm.

JJ: kind of like on the same lines of like what we've been talking about with the anthropic stuff is that it just seems a little too safe. I was messing with it this morning and I I was playing with an idea of thought leadership curation on LinkedIn where like I don't want my LinkedIn posts being drafted by AI, but like, you know, if you could find me some relevant news articles or industry topics, that kind of stuff, give me a daily brief of what's going on so then I can add some commentary on it. Muse connected, logged into my LinkedIn, no problem to look at my past LinkedIn post and all that stuff.

And do it refused. It was like I can't log into LinkedIn. I was like, okay. So it's just stuff

Jack Zoppa: It's unbelievable.

JJ: like that. But what was really interesting is I've already I've got all it it takes all the connectors that you already had from Claude, so there's no setup. So if you're like a heavy cloud user, that's a benefit because it's got all of your connectors

Trevor: Mm.

JJ: and all of your chat history. already there so you can just start working with it. I do like it it's still the same cloud experience. So I had it going through creating some documents for us and it was able to like pull pull stuff from my email and drive and then put stuff into a Google Drive and tag it in in a Google Sheet. So that would and it's a way better experience than trying to do that same stuff just through Claude natively. So I will say that is good. So I it might change the way that I use Claude on that type of thing I think is where that's gonna become useful.

Jack Zoppa: I really want access to it. I saw your LinkedIn comment. The second I saw it, I opened Claude and complained about Claude sucking and how Grocbot was amazing to see and I didn't get access. So I was pretty upset.

JJ: Well see, I talk I said that I wish that I could use Claude. I didn't say Claude sucked. I said that I wish that I could use Claude Duet because the other ones are frustrating. So

Jack Zoppa: You're right. That's that's on me for breaking the I broke the anthropic terms of service on being nice to the models, my bad. but I'm curious about it because I was texting you guys last night. and like I mentioned at the beginning of the show, I we have a team plan at our company. So like for anthropic we've had it for a while now, and like I would say there's five or six users who are like use it every day to the max usage, like a ton, which is awesome. I like love that. But I really want them to experience like the personal agent side of things.

It's just like I think duet is probably the way that I'm gonna go if if public access actually gets out because it's the same subscription. They can just stick stick on what they're doing. What I've been tossing back and forth is like should we just switch to XAI and go all in and have grockbots for everyone? But I think what they do a lot of is like these long workflows of like my jiggle manager, hey, go into our task system, communication system in this and kick off this hour long. I wanna know all of this stuff. And it's like it's kinda like hard for a personal agent to like do something like that.

That's like a super, super long workflow. And so it'd be nice to have something that does both. So I'm pretty excited to use it. I I wanna get it so bad today.

JJ: you don't want to have to pay for a Claude subscription for your whole team and then also pay for a Grok subscription and whatever else. maybe we're getting rid of some of our tech stack in some places, right? Like I I've got a goal

Trevor: Mm-hmm.

JJ: to get off of Airtable and like I've got off my phone system because it moved over to Conduit and like so I have less expenses in terms of software, but then At the same time we're adding more expenses from the different AI subscriptions. So it it is funny how it's just kind of a scale that's tipping in a different direction.

Jack Zoppa: Yeah. Did you see their so like Anthropics S one came out I feel like you're gonna have a crazy take on this, but like is it half the Fortune five hundred spends a hundred thousand a year with them or something like that and their S one and I I saw a bunch of people being like first of all, only a hundred thousand a year seems low for Fortune five hundred companies. Second of all What is the other half of the Fortune 500 doing? Like if they're not us are they Ye Yeah, that's

JJ: Right. Yeah, if you're if you're not using an AI tool, that's

Jack Zoppa: insane. Michael's on mute, he's just over there making faces.

Trevor: I think Grokbot though, I don't know if you guys have seen this, but Grokbot also will use Opus for certain stuff. Like it's not

Jack Zoppa: Yeah.

Trevor: tied just to to Grok.

Michael: Yeah and and Elon came out and said

Jack Zoppa: That's true.

JJ: Yeah, I I think that's super smart on their part too, because now it's like if you think that Anthropic has the better model and that was keeping you from using Grokbot, now that's not a a barrier to entry. And I think eventually they'll be their models will be just as good as anybody else's too.

Michael: Yeah, and and Elon came out and said that he wanted to make it headless, like what Salesforce has done and stuff like that, where it might be controversial in the moment and you could see like, shoot, how do we actually monetize this if we're like routing to other people's models and stuff? But if maybe if you see a longer vision, I don't know.

Jack Zoppa: I'm not gonna bet against Elon, I'll tell you that.

JJ: For sure. And plus their

Michael: Yeah.

JJ: the his model is gonna be the the compute in space, data centers

Trevor: Mm.

JJ: and all that stuff in space. So he doesn't even care

Jack Zoppa: Yeah.

JJ: about the AI model. He wants everybody using his data centers.

Michael: Yeah, Anthropic's still paying a lot of money for Colossus One.

Jack Zoppa: So Michael, are you you're buying the IPO, right?

Michael: I it's so hard to bet against them. I mean, either one, right? Anthropic or open AI. 'Cause there's just there's so much power, but it also just seems so fragile, it just scares me. 'Cause like open source stuff is like right behind it it seems. And

Jack Zoppa: Yeah.

JJ: Yeah, open source is crazy too. Spe speaking of open source, I have stopped using Whisper this week and started using an open source tool called Handy. And

Trevor: Mm.

JJ: it's it's bit so we were talking about use cases for the local AI the local model for Apple intelligence. Remember I I texted you guys saying, like, did you know that Apple Intelligence has a local model on on the new Mac system, but it's like I was like, I have no idea what I'm ever gonna use this for. And

Trevor: Mm-hmm.

JJ: handy allows you to plug into Apple Intelligence to do the kind of like transcription cleanup that that Whisper does. So it'll take out the ums and the ahs, and you can even write your own custom prompt to say, I I made it say like if you correct yourself. Right. So if you're like, let's meet Tuesday, actually never mind, let's meet Wednesday, it'll just write let's meet Wednesday and take out that middle part. And that's all happening locally and with no subscription with an open source software. So not that Whisper

Jack Zoppa: So this is like

JJ: is like an expensive thing, but it is pretty cool that, you know, that's an open source software.

Jack Zoppa: So it's it's

Trevor: Mm-hmm.

Jack Zoppa: fully local on your computer. I'm s I'm assuming there's no iOS, right? No phone. Yeah.

JJ: No yeah, no, I don't think so. No, I think it's ju and I think it's only on Mac. I don't think it's on Windows either.

Trevor: The most annoying thing

Jack Zoppa: Yeah.

Trevor: about the iOS stuff though, 'cause I use Aqua Voice, which is the same thing as as Whisper, but it completely messes up your keyboard. And then when you wanna

Jack Zoppa: Mm-hmm.

Trevor: type, it just it creates the key it makes the keyboard experience so horrible. so I don't use it on my on my phone.

JJ: Yeah, me neither. I use the Apple one and I know that

Michael: It can't.

JJ: it's not very good on my phone, but

Trevor: Yeah.

JJ: it's also

Michael: No.

JJ: like if I'm just too lazy to type on on my phone, I'll

Trevor: Yeah.

JJ: just deal with it. Yeah, I wish the 'cause the Whisper keyboard sucks on on Apple too. It just And then it's slow too.

Trevor: Yeah, yeah. The horrible. Yeah.

Michael: Yeah.

JJ: You try to open it and then it opens up the Whisper app and then you have to go back. It annoying.

Trevor: Yep. Yep. Yeah. I'm sure Apple's next year will be great. I think that they're

JJ: Yeah.

Trevor: I I if I had to bet iOS twenty eight will have basically a Siri will be essentially a Grokbot or a duet and have that level of capability and then the voice message and things like that. So they kinda

Jack Zoppa: Yeah.

Trevor: have to. If not, they're in a bad spot. Yeah.

Michael: Better be. Yeah.

Jack Zoppa: I I think

JJ: Yeah, historically

Jack Zoppa: I

JJ: Apple's always been a little slow to release stuff 'cause they wanna make sure it's good. So I yeah.

Trevor: Yeah. Like f folding phones.

JJ: Finally.

Jack Zoppa: bro's got the duo. I mean, I think I told you guys, like, I didn't bring this up on the pod, but this is like two, three months ago. I built my own Whisper mod like I built my own on my Mac. And I built it on my phone. So I think I used like Whisper C C from OpenAI as like the first catch model, like catching the speech transcription, and then I did NVIDIA parakeet. as the arrangement model, which is basically like what you were saying, hey, I messed up. And I it corrects it in real time. it actually worked pretty well. it's like the iPhone is super annoying to like hack into or like get that downloading a local NVIDIA LLM or onto your phone is like not that easy.

Like not through like an official source.

Trevor: Yeah. Yeah.

Jack Zoppa: But I'm I'm really curious to try this one out. I probably will try it out because There was things I ran into that was in interestingly kinda difficult to overcome and so I'll be curious how they how they do it. I like Whisper a lot. I don't like paying for it.

JJ: They're you can choose a ton of different models too. Like I'm on the parakeet model as as the main one and then the as the main transcription and then Apple silicone for the kind of like redo.

Jack Zoppa: So that's that's what I would use Project Daybreak for. Like pre-downloading that, I would be super all these like open source projects.

JJ: yeah, yeah, yeah. So make sure that there's nothing sketchy in it.

Jack Zoppa: 'Cause what I've done in the past with open source is I did something else, I forgot the big open source thing I did, but I rebuilt the repo instead of downloading it. And s like I was like, let's kinda not take random code from the internet right now.

JJ: That's fair.

Jack Zoppa: So I'll probably try that.

Trevor: Have you guys seen Dessa?

JJ: Yeah, report back.

Jack Zoppa: Dude, that looks crazy. Yeah.

JJ: What was that?

Jack Zoppa: It looks wild.

Trevor: DESA. It's from Jack, the founder.

JJ: Mm mm. What is that?

Jack Zoppa: From Jack Jack Dorsey.

Trevor: it's insane. You watch you gotta watch a video of it. It's a voice model. It's not like publicly released yet, but there's gonna be so much scams going on over the phone once that's released. It is indistinguishable from a normal human. It is like its cadence and how it speaks and responds and all those things. It's like it's insane.

Jack Zoppa: Yeah. Go look at it on Twitter.

Michael: But is it isn't that exactly what we want to finally be able to talk to our guests?

Trevor: We do, but like all things,

JJ: Yeah.

Trevor: it's also gonna y be used for bad and yeah, yeah.

JJ: There's gonna be bad actors using it too, for sure.

Jack Zoppa: Yeah.

Michael: For sure, for sure. But in our world, that's a massive pain point.

Trevor: No, it's insane. Yeah. I bet within a year or two, at least us in this in this group will all be using some model of that sorts for our phone calls, you know.

Jack Zoppa: Yeah, let's get conduit needs to hop on this ASAP. Open call.

Trevor: Yeah.

Jack Zoppa: Get early access to DESA. Let us test it live on the pod.

Trevor: Yeah, for real.

Michael: I rem I I I remember when

Trevor: Let me let me text button.

Michael: when Conduit first

JJ: Mm-hmm.

Michael: came out with voice AI, I was testing it and maybe it was kinda wrong, but I was like, Yeah, booking dot com guests are kind of difficult anyway, I'll just test on them. And so like that's where the old people call in from, right? but I don't know. I don't know what I did. I created for our company for Renjoy, we created our voice agent called Joy. And she'd be like, Hi, it's Joy, let me know and she'd talk and then she got super flirty. And it was just kinda weird. And I was like, I don't think this is ready yet for for launching across all of our

JJ: Not ready for the big leagues yet. it's funny you mentioned that because right now we are still using dial pad, but the only part of dial pad that we're using is just for the call routing functionality, like the call tree. And Conduit

Jack Zoppa: Mm-hmm. Same, same.

JJ: just released that feature last week where the AI model can also do call trees now. So like it'll answer but do a call tree in the AI model. Have the call tree, it routes, let's say you press one for customer service, and then it routes to customer service, and then it'll ring to a real per like the real group of people, but if they don't answer, then the AI will answer and take over. I think that's the game plan that I want to go with. And and I do think I wanna make it super clear that it's AI and not like try to pass it off as a human. Cause I also think that that inauthenticity really bothers people.

I've had that experience like Talking to an insurance

Jack Zoppa: I agree.

JJ: company and it's like it just throws you off. You're like, hold on, is this a real person? And and it's like trained to say yes that it's a real person, but you're like, I don't think that this is a real person. Like it that grosses me out. I'd much rather say, Hey, this is Beach Side VR's AI assistant. Sorry, human can't get to you right now. Like I can help you with things like your door code or you know, booking a reservation, or you can leave a voicemail and then give them the option. on if they want to or not.

Jack Zoppa: Yeah. I would

Trevor: Yeah.

Jack Zoppa: I would love to hear Trevor, you like sent something a couple nights ago in our group about like basically you should be at like super high levels of automation rate or you should be able to do what JJ's talking about on the phone and if you're not it's kind of like a skill issue. I feel like that was super interesting. you should like read I feel like that's really in line with what we're talking about. 'Cause like my people doing phone calls they're probably gonna have similar struggles that the AI would and their capability to solve the situations, depending on what those are.

and I feel like you had a pretty good take on this.

Trevor: Yeah, I think in general, so right now I'd say that applies for everything other than voice, because voice is still for us, right? we're not trying to expose that. I don't I want most of our guests to come and stay and not even question or think that they're speaking to AI, which happens, you know, quite like it they rarely mention or or say anything. and I went back and forth early on of like, do I disclose this, do I not disclose it? And I've I've I've landed on like, yeah, if they say, Is this AI? just yeah, yes, it's AI, you're talking AI.

'cause most people, at least now in, you know, twenty twenty six, just kind of know like that's okay. And and I think you're you know, if you're worried about your reviews, which I'm always thinking about, it's like you probably are better off just telling them and then them being annoyed that it's AI rather than them getting pissed off that you lied to them and told them that it wasn't AI when it so clearly was. Again, super rare that this happens. But I think automation in general, as far as messaging goes, you know, the models and the technology right now is there to get to 99.

9% automation, in my opinion. I I think that there's so little if you have set up your system correctly, that a human actually does need to like step in and and respond to. Especially if you have the proper triage and stuff set up. So I'm excited for that to happen with phone calls too. one like if DESA was available on on conduit, I I definitely would try it and use it and and see how it is. I I ac I've never used a voice agent on conduit, so it's probably worth checking out and trying. I think people are just really scared and they, you know you you you limit yourself of what automation level you can get to.

And it's not just about cost savings. It's also about, in my opinion, quality and consistency. Because I've said before, I'm clearly bad at hiring VAs and finding good VAs and and training them. And I also hate training people. I'd rather train an agent. but our our quality of responses and consistency and all those things has increased dramatically since leaning into AI, you our automation rates 95%. So for me, I'm like, our views are better than ever. Our our experience that we're providing is better than ever. I want that to get as high as possible.

And I don't think that we should just like assume that a human has to be in the loop just 'cause that's always been the way it is. I think that it should be indistinguishable if you set it up correctly. And if it's not indistinguishable, then that is a skill issue. Like do a better job at setting up your system or what you're using. To where it isn't really noticeable. because the capabilities are there, the technology's there. And I know a lot of

Jack Zoppa: Mm-hmm.

Trevor: people like have had whether it's conduit users who left or are just not getting the experience they want or the responses they want. And then you like poke around in their workspace and it's like, well, yeah, obviously that makes complete sense to me why you're not g having the level of of results that you're wanting. which, you know. It should probably be easier to set it up and and other things like that. I don't know if that should fully fall on on the customer.

Jack Zoppa: You have you have like a lot of skills, right?

Trevor: Yeah, a ton. Yeah. Pro I think I have like sixty.

Michael: Do you do you think

JJ: I'm curious to see too. Trevor just just got into the Beachside Conduit account to to do some optimizing

Trevor: Yeah.

JJ: too. So I'm I'm curious to see, you know, after a couple of weeks how how that goes. 'Cause I think y you said like literally the first day you saw an uplift just by changing a couple of the skills and the way they

Trevor: Yeah.

JJ: were set up.

Jack Zoppa: We should this is crazy. We should do a head to head agent. So I use Trellis for for communications. We should like do a comparison of the chat agents between like basically how like the skills are set up, what they look like, the performance on you know, versus like a conduit. 'Cause I obviously know both of them super well. and the conduit skills I think are way easier to understand, like what it's calling in each situation and how it's set up. I'd be very curious to see how they run a test guest to the Haven site and the beachside site all the way up to the fake check in and

Trevor: Yeah.

Jack Zoppa: have the same issues be present before check-in and just like see what

JJ: Yeah.

Jack Zoppa: you know, kind of like what happens here. What is what is the difference between some of these softwares?

Trevor: Yeah. I think my workflow, for beach side for example, yeah, there's some low hanging fruit as soon as you get in where it's like, Okay, yeah, no duh, these things aren't gonna respond to, or this is the common response that's going. a big thing is your team jumps in and pauses AI and and messages.

JJ: They're very proactive to to their to their credit. But also like if you're trying to have the AI handle a lot of stuff, you're never gonna be able to train the AI

Trevor: you could potentially see.

JJ: if they're just answering.

Trevor: Yeah. Well, in a way you can because what I did is I went through the last thousand or so conversations that your teams responded to and I pulled everything from that. And then it's like, okay, what skills can we create just from this context? Like what knowledge gaps, what what times did like the agent respond incorrectly and then your team went in and responded? And instead of them teaching AI, they just left it and didn't do anything. And then getting the initial cleanup, restructuring the knowledge base itself, building out as many skills as possible based off of what I have.

Riley sent me everything like SOPs and and your air table last night. So dump that in, see what we can configure with that. And then after that, I think it's just a self-learning loop. So I literally have a T3 tab open on my computer. Well, I guess it's running all the time on my phone too. And it has an hour timer. And then every hour it pulls. the last conversations from that hour and then it looks at them. What went well, what didn't go well. And then it will either create a skill, it will create a knowledge node, it'll remove a guardrail, add a guardrail.

And if it's like a knowledge gap that it just doesn't have and it can't find, then it will set that aside and be like, no, this is like something that we just don't know about this property. And we need to figure this out. And that runs every single hour. So that is I think the level that you need to get at. To get it to the point that you need. But after

JJ: Mm-hmm.

Trevor: a couple of weeks, you're gonna go from I don't know, 70 or 80% up to 90% just from the initial sweep. And then it's just that refining because the way I think about it is like, okay, you have your intention of what you want, and then you speak your intention to Cloud Code through the MCP. It builds what it thinks is gonna get that result, and then you have the actual outcome. And then you're using Cloud Code to then grade the outcome. against the intention. 'Cause it's an agent training an agent, if you

JJ: Mm-hmm. Well, and I think that's the thing that's been missing is up until this point, it's been we've had had to have a human go through and review all that stuff and it's like who's who's got the time? You know, who's gonna do that? But

Trevor: No, no.

JJ: now now you just had an agent review a thousand messages in an hour, you know, so that's where it is there's so much leverage there that you can just kinda I love the idea of having that training loop. So really excited to see where that goes.

Trevor: Yeah. And it's cool too, use the portals so I can get access to the portals. And there was some information wrong in the portals based off of things that your team had responded and corrected of a guest saying, hey, this says this, is this true? It's like, no, it's not. Well, I I get it. yeah, up update the portal. You're

JJ: Update the portal. Yeah, fix the issue.

Trevor: yeah. So then it's like fixing that as well. And then once you get into tasks, then you can really make some really cool skills that are our systems all internal, so all of our skills on conduit and tools are custom coded and built. which is still fully done by Claude, but it's extremely v extremely capable of of what it's able to do. If you you gotta just let it loose. You gotta you gotta trust trust your agents, you know.

JJ: So this is why it's so great to have smart friends 'cause then they just

Jack Zoppa: Yeah.

JJ: go and go and tweak your

Jack Zoppa: Wait, I feel like I have

JJ: operations.

Jack Zoppa: I feel like I have one last question. I know we have wrap up, but do you host those skills like on conduit specifically? Or do you like do you also have them backed up into like, you know, a simply repo where it's like, hey, here's all of our like guest skills or op skills. Is conduit pulling from the repo or are you like simultaneously updating the conduit knowledge base plus like the repo?

Trevor: Yeah, so what I've come down to is our knowledge base and our skills and everything is in Simply, and then I have like a a daily push or like an hourly push that pushes to both conduit and guesty and everywhere else. Because

Jack Zoppa: Hmm.

Trevor: what happens is like you go to Guesty, you change this, you have to go to Conduit, but now I change in one place and then whenever anything changes in that place, it will then push the information to the correct place for me.

Jack Zoppa: Are they living in a repo or are they living in a super base? Interesting. Okay. Okay, interesting.

Trevor: Stupid is. Yeah. Yeah.

JJ: Awesome. let's finish it up in just a quick fifteen, twenty seconds on either what you built last week that's super cool that we didn't have a chance to to hit on, or what you're looking forward to this next week of working on. Michael, you you start us off.

Michael: Ooh, all right. it's a kind of a smaller project, but we've built all of our own owner statements. And now that we've kind of switched back into guesty owner statements and we're switching all the stuff with Net Suite, it was kind of a good opportunity to re-assess like I was kind of going back and forth with JJ. The first one Claude spit out was like it was trying to prove how much context it had. So it was like narrating a whole story of like, you hosted Forty-two guests and two of them were on their honeymoon and I was like, Okay, well maybe that's a little too much.

Owners don't want that, but it is also cool to narrate how much property care you're doing and all those things rather than them just opening it up, seeing expenses, expenses, expenses, revenue, and their distribution. So I'm excited

Trevor: Mm.

Michael: to kinda take that one over the finish line.

JJ: Nice. Jack, what about you?

Jack Zoppa: I finished up my property onboarding project management system inside of our OS and built like a really cool MCP surrounding that for my team to interact with it through Claude and I think it's like kinda game changing everything for them this past like four days that they've been using it. yeah, it's like another one of those

JJ: I'm excited to see that one.

Jack Zoppa: yeah, yeah, it's pretty cool. I'm gonna take what Trevor just said about this like loop skill. I'm gonna build like a really intensive skill around like I think I'm gonna replicate it from like guest communications but also our operations agent and like how did we handle like developing these work orders issues? and I may

JJ: Yeah. That self learning is sweet.

Jack Zoppa: Yeah, I may open source it and put it in the show website.

Trevor: I ha yeah, I have a a a conduit skill that I built that's free and open source so I can add them.

Jack Zoppa: Yeah. We could put an open source skill library on our site for everybody to

JJ: Yeah, I love that.

Jack Zoppa: access the stuff that we're we're doing.

JJ: Love that. Trevor,

Trevor: Yeah.

JJ: what do you got?

Trevor: I think the main thing for me right now, Hayden and I are cranking away on getting this pricing stuff, you know, rolled out and figured out. Gonna test out Red Joy, see how that goes. I think so far we're seeing good good good results.

JJ: something I didn't get to talk about this show that maybe we'll dive in on another one, but I last week built a psychoanalytic personality test. So our recruiters can send this test or you can send it like internally and it'll show you it's it's kinda like if you've ever used culture index or predictive index. It's the same idea where you choose a bunch of words that describe you and then it'll come up with this matrix on where your personality sits. but the cool part is it pulls in based on my personality test, how would I manage this person based on their personality test?

Or how would I work with this person as a teammate and like your kind of strengths and weaknesses, what to avoid, how those people like praise, all of that kind of stuff. And then also how your whole team works together too. So everybody on your team can take this personality test and where where it falls. And that was something that was basically a one shot with Fable. So it's pretty insane what what you can build. But anyway, that that'll conclude our episode this week and great catching up with you guys.

Jack Zoppa: It's the part.

Michael: Bye.

Trevor: See ya.