We Built Microsoft Teams in 23 Minutes (And You Can Use It) & GPT 5.4 Impressions - EP99.37

So Chris, this week we've been having a little bit too much fun with the AI again that we almost just didn't bother recording the show or doing anything. Like life tasks are just out the window. It's a very addictive feeling. Right now we do have a new model to talk about. Open AI released 5.4 and 5.4 Pro. But really you need Billy's in the bank to use it. Just a reminder of that track. What a track. So we'll, we'll get to our initial impressions in a moment of OpenAI's GPT 5.4 but before we get going because I mentioned it at the very end and I don't think anyone would be stupid enough to listen to this whole thing. So below is a link that will work this time to our still relevant tour. I actually gave the wrong link in the description because I like stupidly vibe coded it previously. So the correct link is below. Also, if you're not interested in our like model discussion, please at least skip ahead or stick around because we have a really cool demo for you that we spent a lot of tokens on and we'd love you to participate in so you can skip to that. There'll be like chapter markers if you don't want to hear our ranting about a model at the front.

0:02

So just I do like starting the episode with we screwed up our link and also please skip part of this episode. Classic opening.

1:18

I mean you gotta be honest, right? So we have a new model here. It is up the screen now. Introducing GPT 5.4, our best model yet. And you know we like just. I think it really is helpful to just quickly go through some like reality checks on this stuff. So a lot has changed with this model. So with GPT 5.4 there's a few variants like always. There's 5.4 Pro came out at the same time in the API. I do enjoy a zero day API release here. Like a huge fan of it and full props to OpenAI for doing it. This new model is really interesting. So it's got just over a million token context window 128k out. It's got the highest level of reasoning for, for a model. They say it's medium speed, but we've got some thoughts on that. The price $2. 50 per input million and $15 per output million. So it's a little bit more pricey. It's, it's priced higher than GPT 5.2 which was A$75 per million input. But I would say arguably from initial impressions, a much better model.

1:26

And I think if you used to paying anthropic prices, the 250 is a breath of fresh air. I saw and I felt a sense of relief. I was like, oh, that's actually pretty reasonable.

2:41

Yeah. Like finally we can actually afford to run even crazier experiments with this stuff. So I think the price is good. Context windows great. There's a bunch of different reasoning level that you have support for. So low, medium high, extra high and none if you don't want it to like reason at all. What are your initial impressions? We've been using it for a couple of hours now.

2:49

Yeah. So I've given it, I gave it two separate tasks that, that have been sort of eluding me and just seeing how it it worked on those and one of them it spent a long, a long time on like 40 minutes to get to a conclusion. But it seems to have actually solved the issue which is amazing. So I think there's a reason the OpenAI kind of fans always brag about how long it takes because it's quite slow. I find like in the agentic loop it's quite slow. However, it did get there in the end. I gave it another task on a project I'm working on it. I was able to complete that. The. It's interesting the way it works because it really is a no nonsense model. There isn't a lot of chatter between tool calls. It is just, I mean I'm talking purely in an agency mode here. It comes up with its plan and then it just methodically goes through running tasks until it's complete. It gives quite a nice chipper, upbeat summary at the end of it completing the task. So I haven't had a lot of time with it. Obviously it only came out this morning for us, but yeah, it's a plotter. It's just plotting along, getting the task done. The thing I'm experimenting with now as we speak because like, like you, I'm an addict and actually propose we just never even release our update and we just keep it for ourselves because I have so much fun all day using it and I have like, I literally have five tabs open now running agent loops, doing stuff for me. What I've done is I've actually lowered 5.4 to its lowest level of thinking just to get that extra juice on the speed because I actually find it's performing all right and I don't like sort of these long periods of indeterminate thinking in models. I think I'd prefer it just get on with the job.

3:12

So some other notable improvements. This release feels like a huge catch up or just like compete with Anthropic. I mean let's, let's be honest, it's a catch up. In fact, the blog post looks almost reminiscent of the Opus 4.5 launch where they showed improvements in Knowledge Works. So you can see up on the screen now for those that watch, they've got a comparison of like spreadsheets, document presentations between GPT 5.2 and GPT5.4 and you can see they've tuned this model so it makes like design friendly, like you know, nice spreadsheets and documents. It adheres to styling better. You're starting to see them implement skills. They have a demo of it building presentations in PowerPoint and they look stunning. If this really plays out in real life, I haven't had a chance to check yet. They also have released a ChatGPT for Excel add in. So now they're competing obviously with Anthropic along those lines. Now keep in mind, I was able to build in like a single day a plugin for SIM Theory for Excel and Word and PowerPoint. So I think all it's doing is just giving the model a harness and focus in, you know, in the, in the Microsoft Office suite. I mean I think everyone is just so jaded by how bad Copilot is. Like this is an opportunity obviously for OpenAI and Anthropic to push knowledge work. And I think a lot of knowledge workers have gone to Anthropic because of these plugins and various aspects lately because they are just like pushing ahead there. So to me, if you're just looking at it from like a market point of view, this is clearly a catch up. But I think also interestingly there's been a huge leapfrog according to them in computer use and vision.

4:51

That is big and we need to test that out because one thing that I've noticed lately in my agentic loop is the model's ability to control the computer by what I'd call non computer use techniques as in the command line. So just running software in a terminal and the other one is browser use. And through that combination, what we've going on with Symlink, I'm finding the agents doing way more complete end to end tasks. Like so it handles all my deployment now it handles testing. There's a whole bunch of stuff that I was doing manually before. Like we've gone from a world of copy, paste, reload, test it yourself to just asking for what you want and Then once it gets the idea or you get certain skills in there, it's completely able to do the entire process. So I was saying to you this morning, yeah, an updated computer use models, Cool. But I don't even feel like I need it as much as I used to. Like all of the use cases we talked about, like completing your compulsory work, security training, logging into proprietary systems to extract the data, operating systems for you, like, you know, running my software for me, so I can just run it from my AI console. All of those can be covered by the terminal usage and the browser usage. So other than, I guess, operating like Photoshop or some like desktop software, it's harder for me at least to come up with use cases where I'm truly excited by that, especially knowing how it works and how expensive the looping process is because you've got to do so much more on computer use.

6:38

Yeah, I've spent a significant period of time working on the agentic sort of harness for Chrome. And the learnings from it are that it's still sort of no one size fits all. Like you can't just let the model run wild, like it simply can't afford it. So you've really got to figure out, like, where are the optimization points? And I feel like we've gotten it to a good enough mix now that I don't even think about it, as you say, like, it's not something anymore where I'm like, oh, I can't do that. It just sort of goes off literally on the screen behind me. You can sort of see it if you watch and it'll just spin up a browser tab and do what it needs to do if it wants to like test something or as you say, like once it has come out of the console to get feedback for the actions it takes as well. It's pretty game changing. I mean, they've also showed off things like this theme park simulator game. They built RPG games. Like the vibe coding thing has just come like some, some like insanely long way. And we'll show you some of the things we've been able to build in a minute, which is just phenomenal. Like, it's just getting so unbelievably good right now that it's sort of hard to take in. And there's just this level of excitement I can't really even explain where you can really just do anything finally. And I think, you know, almost what like a year ago we were talking about, these models are already pretty much good enough. They just need context. And I think that Some of the breakthroughs really have been looping through smaller chunks of context so you don't get context drift. Clearing the context in this like agentic loop so that it just has more clarity around the task that it's doing at a given time. And some of these, these combinations coupled with the labs now focusing on, on models that are just really good at tool calling and agentic looping, I just think has brought this to a point where as we'll demonstrate in a minute, it's kind of scary. Like we have some products that we're going to release today, SaaS products. And I want to, like, I just want to clear this up. Like we built these before the show and deployed them. They are live. Like, you'll be able to sign up in a moment when I tell you the URLs. Actually, you probably just read in the description, but play along for a minute. You can sign up to these apps, there's one in particular where you can sign up and you'll be able to like interact with other people that have signed up to it. And like, what, how? Like, I think, I think the longest one was like 23 minutes it took to go through. So I don't know what this means and I do want to have a discussion about what it might mean. But I think to summarize my thoughts on GPT 5.4 only, we have a ball game, people. Like, we have some competition and I don't think, I really don't think OpenAI had a model. I know there's going to be the Codex sims out there that try and put their thing in the comments. Like, I use Codex all the time. Like, no, I'm sorry. Like there was no model until now that was comparable in my opinion, to opus 4.6, but now there certainly is. And I, I don't know, I think you got a competition and it's going to drive down price, which is just great for everyone.

8:07

A few other interesting notes about it. It brings a new feature which is called tool search. And this is something we do anyway in our agentic loop where instead of loading all of the tool calls into the model with their full definitions and then the model deciding which one to use and therefore knowing the parameters. A better technique is sort of give it short descriptions of what all the tools are and then if it wants to, it can load those tools in and get the full definitions to run them. We were doing this anyway because we found that with intense looping, you just use way too many tokens. If you're presenting Like a hundred tools every time you do anything. So now it's, it's much more tight. But OpenAI has actually put this as part of their official API, so they've got a built in tool search. So you just provide the tools, it handles that for you and you're getting the same benefit. The second thing that I found really interesting is we've built our own skill runner. So like Anthropic for example, has a skills thing that'll run it in some random cloud container kind of thing, but it doesn't have network access. There's limitations to what it can do. So we have our own skill runner that we run through simlink and also in the cloud to actually execute skills. And the OpenAI 5.4 just dropped in immediately and was able to start working with the skills just fine. So I think that's another advantage too because in the mix of all of the things you're talking about in terms of context management and all that, being able to competently load and execute skills at the right time is just such a bonus because there's just so much benefit in having these tight, well defined skills that can run in a sub agent, get a really, really efficient part of your process done in a reliable way. And so I think that early days with it, it's got all the check marks of what you would want for a model like this. I think the only thing I'm really noticing as a downside is the slow speed like I'm used to now with all my loops being able to get updates fairly regularly on, on what's going on. And I just feel like this one's churning away for a bit longer than I'm comfortable with.

11:19

Yeah, I think this is the challenge with it, right Is like if you run it on Max versus Opus 4.6 and I say max, we call it Max, but it's really just the like extra high thinking or whatever the highest level of thinking is. And so if you run it on those settings it's incredibly slow and painfully slow. But I think it's like for the cost benefit, like because it's cheaper maybe it doesn't matter, like you just set it off and just accept that it will take longer. But you're not, you're not just burning dollars on fire. So it, you know, it could be in that regard a good thing. I think it's still not as strong, I believe in sort of like a turn by turn chat mode at MCP calling it's, it's a lot better, but it's it's definitely not on par yet, in my opinion. So far, very little time with. With like an opus or a sonnet or even a haiku. But, man, it's getting close. And I think that's so healthy for. For competition right now because it's like, if you go all the way back to, like, the sort of GPT four days where there was just zero competition, and then Claude had this sort of like, you know, safety sex cult model out where it was just incapable of doing anything, we were really concerned back then. Like, oh, there's no competition. They're just like, flogging everyone. And it's. It's how the tides have turned, you know, so it's like now it's sort of like, oh, there's really no competition for this model in reality. I know some people disagree, don't care, but I think now there is, and it's a good sign. And to me, the next one to two iterations of OpenAI, if they stay on this path as predicted by me at the end of last year, I think they could have the best agency by the end of the year. What do I win? Do I win anything? I'm not sure.

13:20

If you go on Polymarket and bet on it, you can.

15:12

All right, well, I'll take the bet. I'll take the bet. So there's also the pro model, which we should speak of, and I don't

15:15

know which, like, organ in my body it is that has a physical reaction to the price of this model. But one of them is reacting. Maybe spleen, maybe liver. I don't know. But I have a physical reaction to the price of this model. So let.

15:24

Let's talk for a minute about, ooh, or why you would pay three $30 per million input tokens and $180 per million out. Like, I cannot fathom a use case.

15:39

You must have. You must have experienced it right when you're working on an agency process. And it'll sit there chewing up all of its output tokens in thinking. Imagine doing that, knowing, like, every little chunk of thinking tokens coming through is costing you, like, $5 or something.

15:54

So, like, I'm curious if there's anyone in the audience who says, I'm willing to pay, think $30 per million tokens input because this model is superior.

16:08

And there'll be some dickhead. Hang on. Be someone who's like, oh, well, I'm solving the very hardest physics problems, and for us, this is par for the course. You know, something like that.

16:18

But they never provide evidence like, there's never proof. It's always just vague posts. To me, the way to consume it would probably be to just have like ChatGPT's highest plan, right, and try it through that paradigm where. But even then, I think they rate limit you so hard on it that, you know, these, these sort of like, like Pro. And then Google has their Ultra models. It's all sort of untouchable stuff. And I think because you don't, like, you're so scared to play around with it, even as a developer, you, you're like, what can I build with this? Because I can't even afford to, like, test it.

16:28

And I think that's, that's part of it, right? Because then they can brag that they're smashing all the benchmarks, like it's the best model in existence, but you can prove them wrong because you can't afford to. Like, it's just absolutely wild. Like, I added it into our system and then I just, I panicked and deleted it immediately because I'm just like, I don't want to accidentally have the ability to trigger this thing off and spend like $1,000 in a few minutes. Like, it's just, it's nuts. I just, I can't think of any industry where it's just going to be so profoundly valuable that you can afford that. I mean, I'm sure there is some, I know there probably is some, but I just don't think it's going to be that much better that it justifies it. Certainly the last Pro model wasn't.

17:06

Yeah, I, I. Anyway, I'm curious to hear. It's just at a point, I think, where, like, it's even hard for us on the show to justify testing it. We did come up with some ideas and then we're like, it's going to cost us like, two grand. So we, it's just like, we're too poor to be able to test this model. I guess we're like, very token poor.

17:50

You need to make a song about that, I reckon.

18:12

Yeah. Well, I think we have, I think we have. I think Billy's in the bank. Wasn't that the whole plot line of that, of that song? All right, so I think we should, we should move on. But we are pretty excited about this model and we think it's, it's gonna be a hit. One thing I did want to call out is someone over on X posted this. They were able to like, fully recreate a fully working version of Minecraft for the first time. Now I've tried this with models before and you know, it just hasn't been that good, right? It would trip up like you got the 3D environment or whatever. But not only is this thing humming, but it has like all the different block types he said or she said, I don't know. It said it took around 24 minutes to make Minecraft. And like, I know people will naysay things like this and say, well, it's just replicating things that have already been done. And sure, like that's fine, but if you think about it, most things people do have already been done. So, you know, I think it's quite impressive and it just shows how far these models have come. And he also pointed out that with GPT 5.3 Codex, it was able to do it as well. And. But Opus 4.6 for some reason got stuck at some point when trying to build it. So I don't think that's like some definitive, like this model is better than that model. But I think it just shows for the first time that this is a super competitive model. And it also sort of now shows how disgustingly bad Gemini 3.1's release is, that they're just not even on our radar anymore. And I'm just sad, I'm sad to

18:14

say it, like it's. It's actually lowered my own self esteem because I'm like, why can't I get this thing to do even the basics that the others are able to do? Like, honestly, it makes me feel like our system's broken when I work with it and test on it. And I've just kept trying to prove myself wrong and be like, oh, it's us, it's us. And then it just isn't. That model just isn't good. I just don't understand what they did to it. And admittedly I am using it in a very, very tight use case, as in agentic loops. I'm not even trying it for other stuff, I'm sure, like analyzing documents or videos or audio. It's great at. And we still extensively use Gemini models on single shot analysis of things and stuff like that, which it's great at. But in terms of this new way of working, and it is a new way of working, it is so far off, it's not even close. Like GLM5 smokes it, Kimmy K2 smokes it in terms of this kind of work. And this is an expensive large model like that is their flagship model. Admittedly, again, they call it a preview. But what's our alternative? We don't really have Like a Google model right now that you can use at the front level of this kind of stuff. And I think the demonstrations you're showing, yeah, it's replicating Minecraft yet doing whatever, but the change in the way we work is just so profound that like the way I work now is I issue commands, like I set goals and the system goes off and does it and then I go set another goal or I queue up more goals. It's just, it's so different that the models that power that are incredibly essential now and for a major provider out there not to have something that competes in that space is a real concern.

19:53

Yeah, it's just, it's just shocking to me how they went from being seen, at least by the public market, which all that seems matter to most of these companies from being like, oh, wow, Gemini 3 knocked it out of the park and then it just all fell to bits and I'm not entirely sure why, but everything's moved on. The world has changed with these agentic models now to the point where they need to compete. So I think the positive is they will and we'll get something out of them. And I think if they can get like a flash level model running in an agentic loop, it's like super cheap, super fast and just wipes the floor with the other guys. They're still in the race. I mean, this is not over. It's. It's so, so early.

21:35

Yeah. And look, this happened before, right, when they had their preview models and they were hit and miss. Like one day to be great, the next day to be bad. Then Gemini 2.5 came out and it was the best model for a very long time. So, yeah, I'm just saying, Google, come on, guys, like, get us something. Get us something we can use. I want it, like, yeah, I'm happy to eat my words, I'm happy to retract, I'm happy to contradict myself, whatever it takes. I just want the models.

22:13

So I thought this part of the show, we would do something a little different. We. So a number of episodes ago we talked about, I talked about I'm just going to rebuild Trello because they didn't have a good MCP at the time. I've since built my own off their API, so it, it works better now. But I did say I was just going to rebuild it to try and prove. And this was quite a while ago, probably like a week ago, like three months ago. But you know, it could be longer, I'm not sure. So we said we'd do it. And this was at a time when I don't think it was that possible. Like, I would have had the vibe code quite a lot to get this done, right. So I set this off last night overnight. It didn't take that long. And I was able to build Trello. Now, I know, I know the play here, right? You're thinking this is, you know, it's going to be missing features and like, it's really the finer details and. Well, there's all these other elements you don't really understand. Right. But I'm about to freak you out. So. So it was able to deliver me Trello, which is my Trello knockoff called Trello Please Don't Suicide last year. This is a joke. So it's trello.simtheoryapp.com and you can just sign up and use it. You can actually, like, assign tasks. It has all the features, unlimited boards, and it just completely works.

22:36

It's great.

24:12

And so don't actually use this because I may not keep it up forever. But if you do like it, you can comment below. So here's a board, right? Looks pretty similar, right? It's beautiful and fluid, exactly like the real thing. I can assign members, I can do labels. Everything works. I can have a due date, I can have attachments. I've got like checklists. I mean, it's unbelievable. I can like move cards and they animate exactly like they do. I can invite to my board if I want. If he has an account.

24:13

I do. I do.

24:51

So, like, I get it. Trello is a pretty simple app, but this is like an absolute full clone. So now.

24:52

And it's. And it's fully deployed to production. Yeah. You didn't do any of that.

24:59

No, it's. I just said deploy plus. I'm not even kidding. Like, so we have this in the new Sim Theory. We have this ability to. When you create stuff now, you can just deploy it to our servers, sim theory app.com and then you can use it. It'll have authentication. We're going to have like user roles and permissions. Like all that stuff just baked in for you. And I know a number of services already do this, but it's a little bit different because it's just. You're dealing with raw metal. Like, it's just a raw metal service. You can build anything. And so anyway, pretty cool. So we've like, has AGI been achieved? Maybe. But I think it opens a broader discussion about just software in general. Right. Because for us, like, we. We occasionally use Trello, right? Before the end of a project, right. To get it over the line. And we have hit our, like, max board. So they want us to pay $99 or something a month, usually to create new boards. And so it's like, all right, I'll just recreate the entire app and share it with Chris in, like, what, 33 minutes? And doing nothing. Like, I just said, copy Trello plus and call it Trello for our use case.

25:03

It does the job. We actually don't need anything else. Like, that's good enough.

26:16

Yeah, it's just, like, compostable, sort of like throwaway software if we really want. Look, I'm not saying that this is, like, going to take over, like, Salesforce or something immediately, but what I am saying is any of these apps that are just, like, tools that are you using, like, 10% of the features, it's over. Like, it is completely and utterly over. And I know because I've said it before, he's like, come these don't want to maintain these things. But here's my counterpoint is they don't need to, because we can soon have a community of people working on different pieces of software that like Trello and put that in a store, and then you can just go and install these apps and you don't have to maintain them. And they're far. Like, you know, they cost nothing. So.

26:20

And, and the other. The other thing to point out is this same technique works with way more complicated pro processes, like projects. I mean, so for the first time in my life, every side project I've ever wanted to do, I'm doing and getting fully online, you know, same day. And then I'm like, wouldn't it be nice if it had this feature? Add it. Wouldn't it be nice if it had this feature? Add it. Like, I'm literally sitting there just queuing up all of my wish list for what I want in these applications, checking in 20 minutes later and they're all done. And, like, and then we've actually made a cool thing that, like, we'll actually use the computer voice to let you know when it's finished. So it's like, mine is like, chris, I love you. I'm finished, and I can go check in, check on the new feature. And so the actual maintenance part becomes a joy because you're just like, oh, well, I need this. Oh, I need an administration command to reset someone's password, or, I need to be able to do this. And it's just like, I've added that for you and deployed it. And tested it, you're ready to go. Like, it is just a wildly different way of working. Even though we've had a lot of this technology for a while now, I just feel like something's happened. We've hit this point where it's just able to get so much further down the road, all the way down the road. Whereas before it was always pieces of it and you had to be in the middle and you had that mental overload of doing things. Now it's simply well specified goals and you get there.

27:08

I also think that you can imagine just building very customized like software for like project management and, and all these different things. And so I, you know, we, we got this Trello app out because that was like a goal from many months ago. Like could we just recreate the whole thing where people could log in, sign up was completely, completely and utterly free and they could do whatever they wanted. But then I, I was just joking around with you. I'm like, I wonder if it could like reproduce Microsoft Teams, like if it's at that level where you could have like almost every feature from Teams done. And so that's what we did. It's called Macrosoft Teams. Microsoft Teams. It's now available. You can sign up now. We haven't made it so you can create your own like team accounts yet because we thought it would be truly hilarious if Microsoft Teams was available to all of our listeners. So you can go now to teams teams.simtheoryapp.com and you can sign up and join the Microsoft Teams deployment. There is video chat so you, you can actually start a teams call if you click. I think it's like meet now at the top. We've tested this. It is incredible.

28:32

We tested it so much we forgot we were using it. Just having our morning catch up.

29:52

Yeah, we were doing our morning catch up in Microsoft Teams. So I don't, I don't know if I'm going to be able to join this without like sending out audio while. But we should try. Let's just try and see what happens.

29:56

Okay, I'll join your meeting.

30:10

I did update it before this so yeah, it's going to look a bit weird on my end because I've got all these like weird configuration things going on. I also was able to add background effects. I'm not sure if they work. They're probably not going to work for me because I'm in this like weird setup. But you can see like Chris is in the meeting and yeah, it like fully works. So we've, we've been doing meetings through this. Now when you join teams.simtheoryapp.com please do. I mean it could get bad in there because we, we, you know, we're not going to administer it. So apologies in advance if some toxic stuff happens in there. But anyway, so it's like fully working this. There's like different channels. You can go in there and chat. You can video chat with other listeners of the show. So you can, you could, you could start a meeting and other people can join. It supports I think up to 150 simultaneous video chats already. We can up that if, if need be. So anyway, check it out. Go and go and see this for yourself. Like see how far it's come. Teams.simtheory app.com Now I want to be honest here. I literally said please copy Microsoft Teams, call it Microsoft Teams and deploy it plus like that's the prompt. Then I had one follow up prompt which was like can you add effects and some other things to the video chat? Then one more prompt which is hey, can you make me some admin. Admin options in case this gets out of control. So I think if I spent like maybe a day on this, it, it would be close to fully featured, right? And then I honestly think like people going about like user roles and permissions and stuff. I honestly think in a single day I could connect this to Microsoft Entra and adhere to all the permissions.

30:12

I know you could. I know you could for sure like absolutely. The, the ability for it to do things. And this is because it has access. It can, it can run the code it need, it can use skills to do things, it knows how to do it and it can get the whole way through the process. And you think about how many companies are stuck in contracts for things like oh, you know, leave system time, recording systems, like all these like little pieces of software, they're paying these massive fees for every month that they could replicate. Like, I'm not saying that'll necessarily happen, but when you start to realize not only can you make any software you want, you can completely customize it. You can actually deploy these things in secure, maintainable ways. You can have individual deployments if you want. You can, you can like the, the Possibilities. It's my Endless Possibilities song. I feel like that should go down in all time, that song. But it's, it's just a totally remarkable thing to be able to say in this piece of software. Add this thing. I listen to that all the time when I'm working. I love it.

32:02

So here is the real Discussion around this. Not to get too like carried away or anything, but I used to think all this stuff was fanciful. Only a couple of weeks ago I was still like, oh, you know, companies, yeah, the old me. I'm like, no one's going to actually do this in reality. But here's what I now predict will happen and I think this is becoming like the most clear vision of what will happen. Think about what happened with SaaS and applications before, right? So you would go to Google and you would pretty much start your interaction with the Internet with a search, right? So you would search like I need design software or chat software or whatever it is or like compare Microsoft Teams to Slack, you know, and then you would land on maybe an article or a review and then you click through to the product, you'd sign up and you try it, right? And then you would interface with that product when you were working on that task. But think about your day and I'm sure many of our listeners will relate to this. I right now was thinking and I, you know, I checked my browser history to verify this. I just don't go to any websites anymore. I just have my split view going in the new SIM theory and sorry to keep teasing it. We will eventually release it.

33:16

It's all a big tease.

34:33

It really is. This thing's a huge, just unsubscribe if

34:34

you feel like we're just, just quick teased you too much.

34:37

So anyway like I'm just sitting there all day in this thing and I mean look, this can be the same for if you're in like anthropic like it's the same relationship I think. So you're in it and you just don't feel the need to go anywhere else. And look there's, there's downstream consequences of that. For example, like why, why would anyone produce net new content? Because it's just not valued because no one goes to your site. So it breaks the economics of the Internet. But that's for another discussion. But the, I think the, the, the like importance of this issue is this you start and end your day in AI. AI is the operating system, your AI workspace, your AI chat thing, your Claude Co like whatever it is you using. That's where in my opinion you'll increasingly start and end your day if you're not already and you're doing everything in it and you're never leaving it like ever. And so what I think that does is at least for apps that are considered tools like say chat tools or trellos, all of A sudden the, the compression of them for you to go and evaluate and buy a solution like that, I think eventually goes to zero. Because you think, well, if I can start to build agentic, like we call them agent apps, right? This is like my, my pitch is this idea of an agent app where it's agentic in nature, but you can still have a ui, you can deploy it to a workspace, you've got roles and permissions. All of a sudden you can go around and a bunch of these SaaS subscriptions you're using. It's not like for most businesses, they care about the cost saving, it's just like more. Is anyone actually using half the stuff? Like, what are we really trying to do to improve productivity? Or like, what's the, what's the end result that we're looking for? And I think increasingly you're just going to manage agents and so you're going to be like, oh, just add that to the task list, notify Chris that it's done. Can you remind Chris's agent to do this? Like, the whole economics of how you communicate are going to change as well. So to me, like, more and more of what you do is going to be in the AI workspace. To the point. I think a lot of these companies are making mistakes because they're embracing like cowards. Like, they're like, oh, we have this McP and MCP UI in chat, GBT or whatever, right? And it's like, eventually those become tool calls to an agent and eventually you go, you know what, I don't even need the tool call anymore. Like, I can just have some dumb database on the end. Like, it's just so irrelevant all of a sudden. So I think the first to fall are going to be all these tools, in my opinion. And it's going to happen fast. I would predict in the next five years all of these things just go to absolute zero. And I guess what I'm saying is all the span, all the usage is going to be consumed by whoever owns the OS of AI and everyone else is pretty much doomed unless you've got entrenched enterprise software and processes. And I'm not suggesting all of this will be rewritten. I'm just suggesting the start and end page becomes the AI workspace. And the relationship you have where you're going to different software vendors and stuff, it's completely, it's over.

34:40

And I think that even in the entrenched software scenario, I actually think a lot of this stuff can be done by force. So, for example, let's say you've got some existing software you just have to use, you're in a long term contract, but they won't add features you want or you can't get out the data you want and things like that. Here's my proposal. In Simlink we've got a browser plugin that can run code in the browser, right? Like, and the AI agent can connect to it and do what it likes, right? So you simply add features to their software anyway, whether they want it or not. You add a script that runs on top of that website when you're on it, that consumes their APIs via the authenticated browser, changes the buttons, changes the layout, does whatever you want. You can literally add features to other people's software insofar as your own or your team's use of it. That's one way you can do it. The second way is something we've talked about before. Again, you're in some entrenched software that you want to eventually move away from. What you do is you build your own version of that application in an agentic style. Again, whenever you're using that application, it's consuming the data directly from it and starting to replicate it into your own database. So more and more you're consuming it from your own interface with that as a data conduit to the point where you're like, hang on a sec, I haven't logged into that system for six months. I don't need that system anymore. Like we've talked about this as a sort of fantasy before, as theoretically possible, but I would argue right now, today, I could do this for any enterprise software system. Like the only exceptions to that would be like ERP systems where they're actually like, you know, like going into some like hardcore old school database. It might be a little bit more effort to replace those. But ultimately in terms of interface, in terms of the way you work with stuff, there is no, there is no idea that these entrenched software companies have to participate in this. If they don't want to do it, you can do it anyway.

37:57

But do you also think like you look today, so today Google, or maybe it was yesterday, released their own CLI tools for Google Workspace, right? We immediately like installed them on our agents and are using them already, right? And so I think that keeps all those, you know, Gmail and Calendar and all that stuff relevant because they, they, they handle like a lot of the protocols and all that stuff. Like you don't, no one wants to rebuild that, who cares? But now I'm never going to go to like I don't anymore. I haven't logged into my actual Gmail account in a long, long time now, probably months. And so I'm just consuming it all through my AI. And I guess like increasingly with these companies like as you start to consume them with cli, it's like around like the regulated industries as well. If you have a good data structure or data warehouse, there's new products out. Like databricks has this like real time interface now into a warehouse where you can basically like you could build a CRM for example, on top of the databricks data warehouse. Because it's now so fast it completely eliminates the need for like ETL stupid stuff that used to slow this relationship down. So if you think about some of these more expensive solutions in the enterprise, you could almost run them in parallel. So you could build like an AI agentic version of a piece of software that you're already paying for in a regulated market as well and run it side by side because they're both updating the same database as long as you had some sort of right permissions. And then when it's good enough you just roll the other one at the next renewal. Not. I'm not suggesting this is necessarily the way things will go. I think the small tools will fall first and then slowly over time people will start to just do more and more in their AI workspace and the, the AIs and agents are just going to get so much better that you're running agents. It's like none of this stuff is relevant anymore.

39:50

But even things like crafting tool like skills and MCPs, right? Like people I've been helping have had a system where they're like, I really want an MCP for this system. And I'm like, okay, all you need to describe is what you want to use it for. Let the agent go off and figure it out. Give it access to the things it needs. Maybe the computer it needs access to, maybe the network it needs access to and maybe like a login or something like that. Let it discover the best way to do it and develop it. And we're getting to the point where you're like single shotting or maybe like one or two updates to what you want and being able to create and publish this thing and then consume it. We did one where it was like 250 podcasts and all these documents and stuff and they wanted a database where, where it had the ability to query that and work with it as an mcp. It was like the thing figured out how to get the. The podcast downloaded via, like, you know, through a system that was trying to prevent it. Got them all in, built the indexer, built the search tools, and then it's available in the system, like, without any. No lines of code written. The system was just done. And we can do this for any system now. So, you know, we were all set to make videos on here's how to code your own MCP for your business. I'm like, that is now completely and utterly pointless. I would not recommend anyone learn how to code an mcp. It is utterly pointless. The models are fantastic at it. What you need is to basically ask for it. Like, what you need to get good at is asking for what you want. You need to know what you want and how to ask for it, and you need to guide the models towards building it for you. Like, that is the future of this kind of stuff. And I say this as a programmer, my skills are useless. I'm done.

41:49

The one, the one thing I would say is, is this, I think experienced engineers underestimate and probably, you know, completely underestimate how many things they know so that their prompting and guidance of the model is better. And also for projects that are existing projects that you then sort of brought agents into, like ours, right. You definitely have a deep understanding of how the project works. Right?

43:32

You're wrong. It's unnecessary because you know what you do? You point the agent at the project and you, like, get to know this bloody thing for me and map it out.

44:00

I guess you're right. Yeah.

44:07

And then build a series of. Build a series of skills that explain how to add a feature, how to test a feature, how to deploy a feature, and then use those skills as appropriate when working on this project. Like, the thing is, it can do that. It can discover it. You don't need to know the project at all. You can get a project, like, completely raw and work on it. And it's so funny because there was a while ago there was all these articles about all these old dudes who know Fortran are making like $400,000 a year because no one knows how to update the old systems. I'm like, they're doomed because this system is going to be able to do it just fine and it's going to be able to test it and everything.

44:08

But here's so, like, I'm trying to do a positive spin here. I feel like people interpret this as, like, tumorism. I think, interestingly, it's still never been a better time to be in this market. And a lot of people would say, like, why if my skills are redundant. But I would counter it to say that there's so much stuff to build, like everything just became up for grabs, like literally everything. And here's a case in point. So I've been talking to various large businesses in more of an advisory capacity around their IT spend, right? And one of the interesting tidbits is even those people have woken up to these entrenched sort of like weird tools they're using for various things in the business saying, you know, we're spending like $12 million a year on this contract for this very specific software that doesn't do much, but we're too scared to touch it. But interestingly, over the last couple years, a lot of these people have put the data in a data warehouse or in like, let's be honest, just a big database.

44:43

So they've already done, it can be

45:51

queried, they've done the hard bit. Now the easy bit is like, can you copy this software plus and run the testing and like run all the use cases through? So if you write a good enough spec, you can pretty much one shot these legacy systems, if the thing is

45:53

with, with browser use, you don't even need to write the spec. You're like, here's the URL, I'm logged in. Go figure it out.

46:08

Yeah, go figure out all the test plots you got to use. Plus I think that's like a secret thing. It's a technique.

46:13

Yeah, we need a skill for that.

46:19

So what I would suggest is that people that are fearful is that you can just go in and absolutely change the world, like change everything with these tools. And I kind of worry for entrenched companies. And at first my interpretation of the SAS sell off on the stock market was really that, you know, this is really overblown. Like these people are panicking for no reason. But I got to say, like it's going to compress margins, it's going to slow down growth and you've got to, you've got to have a reality check in five years. I think it's, it's having a huge impact when normies are waking up. And when I say normies, look at open claw today, right? The complexity of setting that up is reminiscent of setting up like a BBS system or something back when the Internet first launched. It's incredibly complex, way too complicated, and also quite frankly dangerous. But increasingly these things will become like eventually in your phone you'll have an agent that's capable of Doing all the stuff we're saying right now and on your behalf. Well, yeah, I know that, but come on, let me build the hype. And. And when you do, all of a sudden, like once everyone has access to that and that is widely distributed, it's over. It's all over. Because they're just going to be like, oh, I need to edit some like TikTok videos or whatever. Hey, agent, I've shot all this footage in my photos, like folder thing. Go get it and make some highly crack addict video from like, I don't know what happens at this point, but we're there. It's just now distribution is probably the issue and making it like accessible and easy enough to onboard into this stuff,

46:22

at least at a company level. I predict it's going to be the people who have always known what they wanted from their systems and from their company. And there's the gatekeepers, right? There's the IT dudes, there's the programming team, there's the product managers and all these people you got to get through to push your case, right? And often those people get ignored and they don't get their stuff done. Those people are going to be so unbelievably empowered. It's really, really going to change the nature of companies and who's driving things like product. Because it goes from having to deal with like timelines and sprints and all this sort of stuff. Oh, we've got to get this out in the thing to being like, I just added this feature anyway. Like, you know, I've just added this to the system. Works great. I've tested it, I've tested it with customers. It's brilliant. And that kind of thing like that is so realistic, it's crazy. Like, it's absolutely going to change the nature of things when it's really about the people who know what will work and what they want. And because experimenting is so cheap, unless you use Pro, you know, you can try stuff and go, you know what? That was crap. I don't like it. Get it out of there and do something else. And you know, if you think about things like safety and security, well, get some skills in there that make sure that when you're adding a feature, the system goes through known good processes for these things that test the security, it does a penetration test on it. It actually checks for real on the production server, how safe it is. You know, add it end to end, sorry, integration testing where it's like, I've actually tried this with a bunch of different kinds of data to Break it and it failed on this. And I've gone and fixed that. I've done that. You know, you can actually have recurring tests, server health checks. Like, you can have skills for all of these things. Set up recurring tasks that mean that even an amateur who doesn't know what they're doing, if they install the right skills and the right mix of skills, they're able to operate like a true professional without even knowing what they're doing. Like, I think that one of the things you and I have discussed is it isn't just enough to have like skills and MCPs with an agent. What you need is like a curated mix of those things. Like, like a sort of okay for like, if I want to work on like an online SaaS system, well, I need the following set of 12 skills. These MCPs. And I need to use this model as my master model and this model as my sub agent model. Bang. Now just ask it what you want. Pointed at the project. Ask what you want. And this thing's going to do a professional job. And I think then you'd have the same for white, different white collar works. Like, here's my leg. And I know I'm stealing your ideas here, by the way, but, you know, here's my legal guy, here's my construction guy, here's my maths guy, and they have all the different mixes of skills and mcps. And these are true expert agents who can do everything required of that role of them. And then your job is to ask.

48:08

You know, I think too, the one thing we've probably in this conversation leaned into like building software and code quite aggressively. But I think the interesting part about this is like, anything you can do on a computer, the computer can now do, right? So I, for the first time. And again, these are all like, I don't want to keep moving the line here because I feel like, you know, we have a tendency to do that. Like, oh, wow, it can do that now, but it can't do this. Here were the things, like early on in the show we were like, can we rebuild Trello like that? That would be major. Like the whole app end to end with authentication. We just achieved that. Now let's look at like number two, right? So a lot of people talk about filling in forms. It's the bane of my existence. Like filling in forms, filling in PDFs I get sent from like big vendors that you still need to like annotate on the, you know, the Mac annotate thing. And I don't even know how to do it on Windows. So I don't. But now I just get. Click on the notification light, see the email, click plus into context. It brings the email in and the attachment. I'm like, can you fill this in? And it already knows, like, for example, my finance assistant already knows, like, my bank deed, like, all the details it needs to know. And it just like, fills it in and then it can just send it. Send it back. So, like, it can, like. I know. I can't believe I'm saying this, like that excited that it can fill in forms. But it saves so much time. Like, I just don't. When I get a form that I feel in my. Can you do this? Like, I don't want to do it. And that's. That's how.

51:04

Well, the thing is with tasks like this also, it can afford to be so much more thorough than you can. Like, for you, it's an. It's a distraction. It's like a task. You're just gonna smash it out as quick as you can for the model. It might as well do everything properly.

52:57

Documents as well. So, like, I had a huge, like, vendor form to fill in with a proposal into certain boxes, right? And with the agent, I'm just like, can you go do that? Like, put in what. What we want into the boxes, edit the document. It's still a docx file. It has track changes if you want, and you can send it over to them. And it's just handling that interaction it can do. Present, like all that knowledge work. I still think it's a collaboration because humans are in the loop, right? Like, you're trying to communicate with other humans. So it's not like it's just handling it all magically, but it's just taking the painful stuff away now.

53:10

Well, like, look at the example I gave you of that mcp, the one that the system made for me when it was done. I'm like, hey, can you document this? Like, how someone would actually install this in various systems? Then can you email it over to them and let them know that it's ready? And it's like, so that is like a big part of a project. Like, you finish a project, but then you've got to sort of communicate what's been done, how to use it, all that sort of stuff. But you can professionally and comprehensively now do that just by asking. It's just like, hey, I might as well do this properly because the agent's going to do all the hard work for me. Please do it. And because it's in its memory because it knows the nuances of it and your communication style and all that sort of stuff. It can just go ahead and do that part of the process for you. It's remarkable.

53:52

Yeah. I don't know. I think a lot has changed and it's also really hard to digest because I think that we have this ability right now to like, we're really bad at predicting the future. I think everyone is historically right, maybe not you, but everyone else is. And so the way you sort of think it'll play out is probably not how it will play out. But I do think increasingly, as you start to use agents and look, they're not there, by no means perfect. And there's a long way to go still. But it's just very clear to me now that we're on a path where all knowledge work changes completely. The economics around it will completely change. And the only question I have now is, like, how rapidly is the change? So is it five years? Is it 10? Like, how long does the penetration take? And that ultimately now is the question. But if you're betting on the horse of, like, this stuff's going to go away, it's a bubble, it's not actually getting better. Like, this won't be able to replace me.

54:39

Put it this way, if it goes away for everyone but me, I'm going to be sitting there running every business in the country, because if only I had it, I would be.

55:44

I question why they don't just not release any of this now. I'd just be like, all right, we're done. Thanks.

55:54

Yeah, actually, AI doesn't work, guys. Don't worry.

56:00

Yeah, sorry. It was all. It was all a bubble and then just sit and just like. You know, there was that book we both read that sort of suggested that they had developed AGI and they just used it to make a lot of money by building stuff. Like, I think they built like Netflix, like a new Netflix clone and deployed it and made really addictive shows to people's interests and like, did all this stuff to basically, like, milk it before anyone else had it. So anyway, interesting times. You got to check out Microsoft Teams. It's a. It's a big deal. There were some other announcements in the week. I'll just quickly go through them. Gemini 3.1 flashlight. It has a million context window. Pretty amazing for such a small model. 65k output tokens. It can handle text, image, video, audio, PDF. You got to give it to Google, Google in that one and. But it can only do, sorry, text output only, which makes sense but text, image, video, audio, PDFs in, it's basically free as always. It's super fast and it's an eighth of the cost of Claude Haiku. I haven't had a chance to like extensively test it at all, but if the claims are correct, it's a really interesting model. Now as you said earlier, we use these fast models as workhorses throughout sim theory and they are the best. There's no better models for it. They're really affordable. You can just have functions that like, you know, I will name your chat, I will do this. All these magical functions in your app, they're reliable, they're fast, they're cheap.

56:03

And so what I've actually done is make a skill for this. So when I work on my like side projects and I'm doing stuff, it's so useful these days to have those universal functions where it's like oh, I want you to use AI for this. I want you to in the project I want you to use AI for this like when you're making decisions or whatever. And so what I do is the skill knows, oh, here's how to use like in my case I was using Gemini Flash, whatever the last one was, but I'll probably switch it to this flashlight now. But having these models that you know are cheap but you can use to make like your app seem so much better in terms of its intelligence and its decision making and, and it's just a sort of catch all for solving problems. These models are ideal for that and I think just baking them into your process of developing things is really valuable.

57:41

Yeah, I mean it's just crazy how much I rely on these to solve problems fast with intelligence that in the past you would have to have regex or crazy match settings for. And it's interesting because I guess a lot of the coding models were trained prior to AI actually existing. They often will revert to like really dumb like pattern matching or like manual things that you would have done in the past.

58:29

They don't.

58:57

And this is where skills are going to helpful. It's like you can train it and say like actually always use like intelligence where possible. Here's how to use it and you, you give it a guide and that

58:57

it's a great point because that happens to me all the time. Something isn't working as well as you, you think it should. I look at the code, I'm like what in the world? It's like sort of trying to manually match the exact example I gave. And then I ADM admonish My bot. I'm like, you're an idiot. How could you even think this was a good solution? And then I go, use the AI. You've got a. You've got a universal tool here who'll solve it for like, a fraction of $0.01. Like, use that, and that'll solve it perfectly. So.

59:09

Oh, wow, that's a lot to take in. But, like, I don't think it should be exhausting to people. I think it should.

59:37

I'm exhilarated. Like, I wake up every day now and I barely even sleep because I am literally issuing commands. I'm like a delegation machine. I'm like, I'm just excited to unwrap all my presence. Like, even after this podcast, I want to go check my five tabs and see how everything went. You know, it's literally, like, exciting because at any point, you can be making things, deploying things, updating things, doing things you've always wanted to do in your software. Like, it's a brilliant, brilliant time to be doing this stuff. And I think the very important point you made earlier is that it's not just code. It's like, okay, I've got to do administration work. I've got to do compliance work. I've got to do other things for my job and my life that need to be done, and I can do them just so comprehensively and quickly. And we spoke for so long, so many episodes about, oh, you've got to take the time to gather the context. But now I just pointed at a folder and I'm like, or an email or whatever, and go, here's the context. Do it. You know, it's remarkable.

59:44

It's. One goal I have, too, is to prepare, like, knowledge work scenarios for people and do some videos once the new versions out, because I want to show them. I find it very hard because a lot of that work is sensitive to talk about on the podcast without basically, like, you know, breaking NDA. Like, you know, I. I just can't talk about certain things. And I feel like that makes it sound like we're full of. But so what I do plan to do is once this version's out and. And Hummingbird, make a series of YouTube videos for the audience that want to learn, like, about our workflows because it's the most common request. It's like, how do you guys use this? And so, like, I can give a quick example up on the screen now for those that watch. So in the new version, I've got the split tab scenario here. And you can see on the left, this is where. And keep in mind my themes. Like an old school IBM terminal. I love it but it doesn't have to look like this.

1:00:43

You need to get one of those clicky clacky keyboards to match. Yeah.

1:01:39

PS2, I've got like in this tab on the left it's like a bunch of servers are running to host all my crazy ideas. Like this is it building out updates for me on Microsoft Teams. And on the right I've got a document that I was just testing like research and document in the new model for OpenAI. To the left of that I've got my Sam Altman Flappy Bird game which was to too unhinged to share on the show unfortunately, but it worked. And yeah, and so there's like a really nice executive summary document here where it's put in like charts, done research, collected numbers, put in sources, everything like that. And so I, yeah, so generally this is how my day will pan out, right. On the left I'm kind of working on like coding related things and on the right I've got like emails going, knowledge, work documents, spreadsheets, now present like whatever I need on, on that sort of part of my split brain. And I think that, you know, these are the use cases I really want to demonstrate. But I, as I said earlier, I just find it really hard.

1:01:43

The point is that all of this stuff, it's like having, I think you mentioned it's like having all the ingredients for the cake, but the extra ingredients is that proprietary data. It's the company's financials, it's the company's internal policy documents, it's the like, it's the company's employee data, it's all of those things like you know, course information or something. Once you add in that proprietary data in the form of an MCP that, that can be mixed in with all of the tools you've just demonstrated. That's when people's eyes just go, oh my God, this is going to solve literally all of my problems. Especially when it can take actions in those systems too. You suddenly in a world where you just command it and it can run your business, it can run your organization, it can plan, it can and inform you about what's going on. You can take the time to do these amazing things and then you can build these agent apps on top of it to actually make these processes repeatable.

1:02:49

It is interesting though because at the start of the year we were talking about cognitive overload, right? And this is still a huge problem. Like you have a break Which I must admit, because of it right now, I rarely do because I'm too excited. But when you do, you come back and sometimes your own context is your own. Like you've reached max context in your brain. So I do think one solution to that problem next is to have a command center overview where it's like we use some of these fast models to summarize what's going on in each tab and where your attention's needed. Notifications kind of work for that, but you don't want to go through all of them. So it's like a nice summary of like, hey, you should probably focus on this. So I think those are the relationships with agents that are probably going to come next. Like, there's a lot of problems to be solved around workflow here that are not yet that obvious.

1:03:45

Yeah, I definitely agree with you. I find myself a lot of times I've set off, say five. Five is about my limit. But you know, four or five different things to do, I get distracted or it's the afternoon and I, you know, I'm not around and then I come back and then I've got to remember what were all the things I did, you know, recently that I need to review and, and, and make permanent. So I think that, yeah, that there needs to be more work done on that area to, to make that a bit more straightforward, a bit less stressful, I think.

1:04:34

So on Microsoft Teams, I'm going to set up a channel called Trello Feedback. Yes. Get used to all my great names and so if you do have feedback and you want to use it'd be

1:05:02

funny list like the podcast and everything ended up with us just being sued.

1:05:13

It probably will at some point. So yeah, you can, you can check it out. Please join us on.

1:05:19

Did you add a sidebar with a calendar and calls and activities?

1:05:25

Yeah, yeah, I'm iterating. Chris.

1:05:28

It literally, it literally just showed our calendar coming soon. It just showed up on my screen. I hadn't seen it before. Wow.

1:05:30

Yeah. Microsoft Teams, I mean, everyone's talking about. It's one of the most hyped apps of all time. All right, any final thoughts? That was a huge cop. Wait, I bleeped the wrong bit. But anyway, you get the idea. Sorry, sorry. I think you're just brain dead now. The AI does all your thinking.

1:05:35

It's true. I just want to delegate more tasks. I want to get the release out so everybody else can do this.

1:05:58

Yeah. So our goal just to put it publicly out there is not next Thursday, but the Thursday after Australia time. So you'll get a podcast, hopefully, and you'll get the, the new release.

1:06:04

Yeah, and I, I wanted to come up with a creative punishment for us if we don't hit the deadline. You suggested shaving heads, but I don't do that.

1:06:17

I don't want to do that. I'm struggling to keep the hair on my head as it is.

1:06:24

So I think if people like in the comments maybe come up with some sort of punishments, hopefully we won't have to actually do it.

1:06:27

Yeah, drop your, your punishment suggestions below if we don't hit our self imposed.

1:06:34

But I think realistically the reality is we've been using this stuff every day for months. It's. It's there. We just, it's just that the, the polish that we, we want to deliver quality to people and not have to backtrack. I think that's the real point.

1:06:39

I think also the challenge of deploying this to so many people now, whereas in the early days of SIM theory, when there was barely any users, we just yoloed everything. But now people have serious workflows on this stuff and rely on it in organizations, so. So we just can't yolo this one. And I think that's where the time creeps coming in. But at the end of the day, I think what that means is we'll leave people with something that is truly special and stable and functional and, you know, hopefully you'll see what we see in as well. We'll see you on Microsoft Teams. I'm going to be there. I'm going to be there for the next like 24 hours. Maybe we'll make Discord next. I mean, I mean, it could be a thing.

1:06:52

We didn't do it. We didn't do a 5.4 diss track. Maybe we need that for next week.

1:07:36

Yeah, I just. I don't know. I thought I'd give it a rest. I might attempt it next week. Or you know what? We could make an iPhone app called Diss Track Maker where anyone can make a diss track. Maybe we should do that. I mean, that's easy. I'll start a tab now. Let's go. All right, we'll see you next week. Thank you so much for listening. We really appreciate you. Goodbye.

1:07:40