Rate Limited

GPT 5.5 is a coding BEAST, developing agents, and RIP Jobs | Ep 15

• Adam/Eric/Ray

Share episode

Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.

0:00 | 1:00:11

This episode explores the latest developments in AI models like GPT 5.5, their impact on workflows, trust, and the future of agent development. Featuring insights from industry experts, it covers model performance, rate limits, AI in enterprise, and practical tips for building effective AI agents.


Links:
Ray: https://www.youtube.com/@RayFernando1337
Eric: https://www.youtube.com/@pvncher
Adam: https://www.youtube.com/@GosuCoder

Chapters

00:00 Introduction to the Rate Limited Podcast
00:45 Exploring GPT 5.5: Features and User Experiences
04:28 Switching from Anthropic to GPT 5.5: User Insights
07:17 Trust and Performance: Comparing Models
10:45 Improving Code Quality with GPT 5.5
17:48 Utilizing Goal Mode for Long-Term Tasks
21:26 Anthropic and SpaceX: A New Partnership
25:32 The Future of AI Automation on Mac
30:51 The Impact of AI on Layoffs
36:35 Navigating the AI Landscape
41:45 The Role of Engineers in AI Development
49:23 Creating Engaging AI Agents
55:03 The Future of Agent Development

SPEAKER_01

Ladies and gentlemen, you are now tuned in to the Rate Limited Podcast with your host, Eric Probunce of Repo Prompt. We also have Adam Larson of GhostU Coder on YouTube and myself, Ray Fernando. And we really appreciate the five-star treatment you've been giving us on Spotify and Apple Podcasts. Please give yourself a round of applause for this exciting show we have for you today around GPT 5.5, this giant water bottle that uh we'll be talking about a little bit later because I got access to that. Uh fancy party with goblins and things that with Sam Altman. We also have some stuff going on with Anthropic and the new rate limits, and a whole bunch more stuff. So definitely stay tuned in. Uh, let me hand this off to our wonderful host, Eric Broboon Shea.

SPEAKER_00

Take it away. Yeah, thanks for that, Ray. And I'm very jealous of that water bottle. Didn't manage to get one of those. Uh, hope it holds a lot of water too. Um, so uh 5.5 just came out, and it actually came out right after we aired our last episode, or right before we aired it, actually. And uh it's it's quite the model, I think. Um, a lot of folks are now switching to it from Claude, and you know, this comes on the heels of 4.7 being a bit of a disappointment for many. Um, so I've had a lot of chance to play with it. I was actually uh lucky enough to get some early access to it as well. And uh so I you know, in a time where we only use models for you know about a month at a time, having two weeks earlier is kind of crazy to be able to use it a little bit that that far in advance. So, I mean, I I've been a heavy user of the model. Um, and I have a lot of thoughts, but I want to hear from the rest of you who have been very cloud focused in your work in the past, and I want to know has 5.5 changed some of your tendencies with picking GPT over cloud? How about you, uh Adam, tell me how how have you been using it?

SPEAKER_02

Yeah, it's all I've been using, honestly, the for the since it came out. Well, I committed in the last episode that I was gonna give GPT models a go, and it just so happened 5.5 came out. I love it. It feels good. I've uh I typically run it at medium. I very rarely crank it up higher than that. Medium just seems great most of the time. I use it for everything. Like literally, I was telling telling Ray this before the show, uh, but we do a lot of like large collection buying, so like collectible cards. I took 1,500 and some odd cards, fed it into codecs, it parsed it, did all the analysis, hit all the APIs it needed to do, gave me, I gave it some rules. I've never had like AI work so well at like doing the data analysis for me.

SPEAKER_00

Yeah, yeah.

SPEAKER_02

My Shopify site, I literally just let it build my pages for me now. I manage my Shopify site. Um very nice. And then my day job, that's I use it for everything. So we ran into an issue, uh, pretty major issue in an open source repo that's actually, I mean, it's pretty widely used, and we needed to track it down, put GPT-5.5 on it, you know, within a very short amount of time, had a good analysis of what the problem likely was. It ended up being correct once we dug into it. I I have yet to have a bad experience with it, to be honest with you. And I very crazy. And I have not got the goblins that anybody else has gotten, but I'm missing out on that end.

SPEAKER_00

Yeah, yeah. So just for some context, there, a lot of folks started noticing that uh GPT-5.5 had a tendency of talking about goblins and other creatures. And it turns out that like this is like a um a quirk of the training they did for personality. You can go read about it on OpenAI's website. Um, but when they were training for the nerdy personality type, which you know you would expect to talk about goblins and other creatures, uh, that actually leaked into all the other personalities, and now it's like a default quirk of the models that they're gonna probably push out of it in the future, unfortunately. Um, but if you haven't run into it, uh it's probably because on ChatGPT and on uh Codex app, the system prompt has an explicit mention in two places to say, hey, don't talk about goblins and other little creatures. Um and if you use it in a third-party app, like which I do in Repo Prompt, uh, I actually replaced the system prompt. And so as a result, I was running into it quite a bit. Um and I've saw the model just talk about it quite a bit in terms of uh sub agents it was managing, like, oh, I don't want to leave that little perf goblin running around unattended, stuff like that. Um it's it's quite funny to read. So yeah, just a little side tangent there on the goblins. Um, and what about you, Ray? How how have you been uh making use with uh 5.5?

SPEAKER_01

Yeah, I didn't get early access and I got to just try it live with everyone else, and it was really kind of amazing. Um, like after the fact, like how I've been using it more often. So I feel like I'm super behind now because I've just started to say I'm just gonna point this at everything in my life now. Um the biggest reason why I switched from anthropic to GPT 5.5 as my daily driver was because I started noticing Anthropic keep like nerfing or not really complete all those tasks. So like maybe three out of like seven prompts, it just wouldn't really want to do the things that I wanted to from Anthropic. Even though I was paying 200 bucks a month in the web app in Cloud Go, like it just didn't really matter where I went, and I thought that was really weird. I I got the best results if I did do like a private API through just their API itself or just using a third-party app that really was already paying the full price or whatever. And so there was definitely like this like compute difference. But within GPT 5.5, it's like, yeah, go ahead and do all these tasks. So I've been throwing more and more at it, and I've been extremely surprised at the speed. I think that was the biggest factor too to turn me around is that it got a lot faster just by default. So yeah, um, I think another thing I was so like heavy to extra thinking pilled where all of that's now come down to medium for me and it just gets everything done that I need to or like complex tasks and things. So one of the things I've started doing is literally just opening up Codex and then Codex has now become my everything app because I can start a new chat. I don't have to start a new specific project in a folder or anything like that. Start a conversation. Uh and so I've started to pull in all my different data layers in. Uh and then in addition to that, I've also started to ha hook it up to computer use on my Mac. And the computer use on the Mac is extremely impressive for like hey, just go open my mail app and just can you scroll through and look certain through certain things. Uh managing some like insurance stuff for the family, doing some other health things as well. And it's just extremely useful to pool all that stuff together in one area. Yeah. And so now it's actually motivated me to prepare for taxes for next year. So it's like, hey, I have this notion thing. You know, I install the notion plugin, go do the things, like you know, and and I that's where I set the extra heavy and just let it just rip forever. You know, it doesn't really take that long, and it starts to set up all the different things, and it's like, cool, I'm just gonna send these here, these images, and you know, go ahead and do your thing and you know, go ahead and fill out the stuff that we need to for expense report. Um and yeah, it's just like it's like I I really couldn't do that as much as I wanted to with Anthropic, and I, you know, didn't have that reliability or that trust. And I think that's what really I had a thousand paper cuts and I was like, mm, I'm kind of done. And so I'm pretty happy with GPT 5.5 and inside of Codex too.

SPEAKER_00

You know, you know, it's up the Yeah, you know, I think uh oh, I just want to speak a little bit to that difference in trust. Like, I I think what's really interesting is uh the size of the model, like you can feel like 5.5 feels like its own big new model, whereas like Opus feels like a distill of something else, like it feels like it's like a smaller version of something bigger. Um, and and I think you feel the difference in the intelligence where it's like less spiky, it's more general and rounded. And in terms of that reasoning, like I don't know if you've tried low reasoning, but if you want that low latency, like it is really nice, and you can tell that like one of the reasons it's faster is because it re it doesn't have to reason as much for the same intelligence. And I think there's this this like chart I saw like re recently where it was like intelligence per like think a reasoning token spent, and if you have like GPT 5.5, like the amount of intelligence density per reasoning token is just like off the charts compared to every other model, and so you're gonna get like the similar levels of intelligence for for like a fraction of how much you spend it. And Opus 4.7 as a like to compare, like it needs way, way, way more reasoning tokens than even previous Opus models. So they're going in the other direction, which is a bit unfortunate. Um, but yeah, it's just interesting to observe that. And and uh Adam, you were gonna hop in on something, and I want to finish your thoughts on that.

SPEAKER_02

Yeah, I mean maybe to extend on kind of what you were saying a little bit there, too. So let's if we go back to um Claude Sonnet 3.5, to me that was a massive unlock. I remember like that changing thing. 3.7 came out and it was odd. Do you guys I don't know if you remember like how odd that was. Of course. It's very off the rails. Opus 4.5 came out. To me, that was a huge like unlock. I felt like I was able to do more than I'd ever been able to do. 4.7 has been kind of like almost uh almost like 3.7 to me, where it was just odd again. So it's like I was almost on this roller coaster ride of like amazing to like uh it's kind of worse. GPT 5.5 feels like another step to me, like that. Like it feels faster. Um and they timed it really well because I to me I think 4.7 has actually shown that anthropic they get a model going and then their tuning actually seems to degrade the model more than improve the model at a certain point. And GPT 5.5 is just it's it's amazing. So I'm super curious to see how anthropic answers this because I have to imagine they're they've got to be bleeding folks on the claw code side a little bit. And the codex limits are insane. Like I've ran significant, like I I hit claw code limits all the time. I have a very hard time hitting any codec limits. I feel like there's just a night and day difference on them.

SPEAKER_00

Well, to be fair as well, um, so uh codecs, they they have a promo going if you're on the hundred or two hundred dollar plan where they're 2xing until the end of May. So just want to clarify that like yes, the limits are absolutely terrible at the moment. Yeah, we'll see where that lands at the end of the month. Um, but for now, it's it's a great time to be using it. And Anthropic, they they had very much restricted the limits because of a compute constraint, and we're gonna get into that a little bit more, but they just redoubled their limits. Um, but before we kind of dive into that side of things, I want to just finish up a little bit on 5.5 because there's a lot here that like this model, you know, it's it's changed a lot in terms of workflows, I think. Like it feels like a you know, it's just another model, but like in terms of the quality of the code you're getting out of it, the the the amount of trust you can put into the model, I think like it's just made a big difference in terms of how we use these models. And it's like subtle things, right? Like um even with 5.4, you know, you before you you'd have it write a lot of code and it would work, but you'd look at what the code actually looked like, and a lot of times you'd see a lot of like ultra-defensive slop, like just code that's just like, oh yeah, if this thing is is is is error like missing or this thing's missing, we're gonna have like all these cascading edge cases, uh, like you know, if statements just to handle every possible thing that could go wrong. And then you'll have 5.5 come in and it writes much cleaner, much more elegant code that that just kind of looks more aesthetic, but solved the problem in a better way in many cases, um, and doesn't have to work as hard to do it. So I I really appreciate that about the model, it makes code cleaner, and I really appreciate this trend as well towards um models that write better, less defensive code that is able to kind of work better and be more maintainable at the same time. I'm curious, have either of you uh observed this um in your own work?

SPEAKER_01

I had a question for you, maybe Eric. I think what do you think is a good prompt or a tip that others who are watching should use to like you say, how do you kind of make something more elegant or what do you kind of tell it to clean up the slop, you know, in a sense?

SPEAKER_00

Yeah, I I think like, you know, like I one of the things I do a lot with with repo problems is I build these workflows. And the workflows are kind of patterns, uh, they're like skill files that that kind of encode ways of operating and using the different tools available to kind of get to good outcomes. And I try to make them general purpose. Like I just made one for optimization that adds like metrics and it uses uh the Oracle, which is like a secondary model, to kind of uh get to um you know a good North Star of where we should go and validate as that external validation. It's a bit complicated, but it's like an auto-research loop with a third party kind of checking in on the work. Um, in terms of what I do to deslop or or or make code cleaner, like a lot of it is it involves like doing things in multiple steps. Like you want to make sure that you have not all the same context window kind of doing all of the work. What I tend to find with GPT models in particular, but most models like encounter this, is that if you give them a big task to say, hey, like clean up all this code, it's gonna try and say, like, hey, I'm gonna find the most like minimal, least invasive way to do this. Uh, and then it kind of goes ahead. But if you empower the model, and and this is where subagents are really useful, is if you make the model kind of an orchestrator or a model that's like in charge of delegating something, um, it's kind of more prone to being like, okay, I'm the manager, we're gonna do this big task, but I'm not doing it all myself, so I can do it. It's less it's less anxious about context and is able to kind of delegate that work and do more fully. Um, so in terms of keeping clean code, there's like two things you really want to think about. One is um having the model really sprawl the code base and like understand the implications everywhere. So like you can use research subagents for this. Um, and then the second part is is having all that context, knowing everything that is in is in place, having a model really look at that and holistically try to plan like the decomposition. Um and for me with repo prompt, I use this thing called context builder, which uh basically has the models pull in all the relevant files and then I give that to Oracle, which analyzes all that context at once. So it sees everything from a bird's eye view, which makes that part simpler. Um, but with different agents, you know, if you if you're able to kind of use these subagents uh carefully, like you can get to similar outcomes there. Um and I think it's really important like splitting up context windows to make sure that the model is is really thinking about how to properly uh view things holistically. Like you can't just have an agent go off, it needs to see the bigger picture, and that's not always easy to do without some careful prompting. So yeah. I'm curious, I don't have you have you done things like this in your cleanup work, or like I'm curious how you use subagents too in your in your day-to-day?

SPEAKER_02

Oh yeah, and I I see the same thing that you're talking about too, where um if you say, hey, I want you to go be vague, like, hey, I want you to go clean up something or improve something, it's gonna pick the I'm gonna I'm gonna remove some logs that don't need to be there. That's like what it's gonna contend to. So yeah, I I totally have done a lot of the same thing. So one of the issues that we ran into, and if you think about the mass scale that we operate at, like we're not like um meta or Facebook scale, but it is a massive scale, like you know, many, many, many, many millions of customers interacting. And when you start looking at distributed systems and where they could fall short, there's usually multiple repos, there's multiple infrastructure decisions that happen. And so like we were running into this one particular issue with scaling, and you you put AI, you basically monorepo it. So you bring everything into a single environment, you use subagents, you monorepo it, and just have it crawl the entire thing and be like, hey, where what are the this is what we're seeing. We feel like it's a red herring, like we have some theories, like go see what you can figure out. And it did an incredible job. Talking open source like analysis to our code-based analysis, and the report that we got back from it was incredible. Like, if you really guide it, you can get some incredible uh recommendations, and you can have it go execute those recommendations you know in your code for you.

SPEAKER_00

Yeah, and to that point, when you come up with that kind of uh a template for how to kind of carry out these investigations, it's really, really useful to encode that into a skill so that you can reuse that same concept and workflow again. Because I I think these patterns of how you have the model work and what tools it should call and how it should sprawl out a research and give it clear structure, like that's when the models do really good work. And I find that GPT 5.5 is incredibly good at following these rigid like workflows, uh, much more so than than any Opus model right now, um, which which is what can be a big difference maker if you're if you're working in these large code bases. Yeah. Uh on that note as well, um, codex just shipped a goal mode. And if you're not familiar, the the main the main like reason that it's there is is that when it when a codex model runs, it will kind of end its task. And you know, not so much with 5.5, but with 5.4, it was very guilty. This would be like, well, if you want, I can do this next, or if you want, I can do this now. Um, but like the model, like it will do some work, and then it'll be like, okay, this is a good stopping point. Um, and if you want it to kind of be really exhaustive, sometimes you have to babysit it and be like, hey, like keep adding this. And goal kind of like with a little bit of you know, glue, it kind of like chains these uh prompts together. So like you have like an initial goal and then it kind of works, and then it'll remind the model, hey, like that's your goal. Did you like fully, fully complete it? And then it like is like, no, you know, we can do more, then it'll just keep going. And so if you have like something really ambitious, you can just let it run for days. I saw the other day someone was running it for 48 hours, like just non-stop, and they had three sessions just going off. Um, so great way to burn your token use. Um, if you got extra limits to burn, like this will do it. Um, but it's interesting to see you know hacks and new ways of getting models to run for extremely long run times. Uh so that's like that's cool. And uh I'm curious, have have you uh Ray have like managed to get goals running or anything set up for like long long horizon work?

SPEAKER_01

I'm thinking about a refactor that I just did, and I'm trying to think about wrapping goal for actually like QA testing. So just giving it some like, hey, act as a user. I want you to just focus on user testing and then give goal because you know codex can just launch stuff on my machine and do user testing. I just want to say, hey, your goal is to act as a QA engineer, do these first P1 passes, which are literally just like you know, what what are all the functional tests that should just work as a user and should not be breaking? Um I think that was a big thing that if anyone wants to take away any quality, if you want to make your app just high quality, just make sure everything works at a just functional level that's visible in the UI. It goes a very long way from people like you know doing that. So I think that's what I want to throw goal at. I haven't done that yet. I'm just trying to organize that and get that going. So I I need more Mac minis, I need more hardware to just run these things so I could just run them without my with my laptop, holding my laptop, you know, like uh running around the bus or something like that.

SPEAKER_00

The laptop over I I wanted to ask you one more thing, Adam. Um, while you're working with 5.5 a lot, I'm curious. You know, I last time you mentioned your coworker was kind of switching to GPT, and I'm curious, like, are you seeing GPT 5.5 kind of spread around the company a bit more than Opus? Like, are you seeing a lot of other people switch to it?

SPEAKER_02

It it's a tricky thing to answer because um there are a lot of people that are in love with clawed code. So there's like the cursor people for sure. I'm seeing it like them switch over. The clawed code folks, I've been trying to convince them to like go try GPT 5.5. Um, but it's it's one of those things where like I think they would if they had the opportunity to go if if they didn't feel like they had to use clawed code, like that that was the thing that they know. So I think Claude Code has some kind of you know, it kind of lynch pinned in to their ecosystem currently.

SPEAKER_00

Yeah, they added a um a migration assistant recently in codex that helps you move over from cloud to codex. So maybe worth pointing them to that if they want to give it a shot.

SPEAKER_02

Yeah, it's interesting because like codex is um it's a week our our approved tools currently are cursor, we have augment code, we have uh cloud code, we have cursor CLI. We we have a you it's publicly known that we did a hundred million dollar like deal or or plan with OpenAI, but we don't have codex. We don't have codex approval. It's like what the heck? So we'll get there though.

SPEAKER_00

You gotta get on that. Oh man. Oh geez. Well, I hope you uh have more people give it a shot because it's uh I think it's a really great model. Um all right, so we talked a little bit earlier about rate limit changes, and so it just came out at the Anthropic event that was this week that they actually made a deal with SpaceX to use their entire capacity for the Colossus 1 data center that that Elon was so famously um kind of going and working with with Jensen to set up in record time. I think it was four months. Uh so crazy huge data center, uh 300 megawatts, uh, that is now fully operated by Anthropic, and that let them double their hourly rate limits for the different plans, which is which is pretty nice. Um, you know, it's it's very interesting to see Anthropic partner with Elon after the two have had such spats in the past on ethics and AI use. Uh um, but you know, I think like in this moment where you know Anthropic has had to cut off users and really squeeze people with uh with with like peak hour rate limits, like they needed to do something to get the compute, and it sounds like they did. So I'm curious, uh, you know, Ray, what what are your thoughts on this and this partnership they're doing?

SPEAKER_01

I think to me it's becoming very clear that there's the the layers to this cake. Oh yeah, you know, I I I think that the the bigger thing is just gonna be like get the people to show up to provide I'm I'm pretty sure that part of the deal is just collecting data, right? Like just hey, you can use my stuff, but we're gonna train on all of this data. Super smart.

SPEAKER_00

I don't think so. I don't think it's a data agreement there. I don't think Anthropic would be down with that. Um Yeah. I didn't get to see skeptical.

SPEAKER_01

Yeah. There's gotta be I mean uh yeah, but even then, like it's just it's not uh there's a lot of profiling that could be done for uh a lot of this type of stuff that could be useful or not. But I think from uh I think it's a really smart thing that they've done. I think there's also like a bigger game that is probably being played. It's just like, you know, model training, you know, having access to GPUs, trying to like what's the bigger industry? Is is space bigger than what we have currently on Earth? You know, is is uh to me, uh there's also like maybe some like background political play in some way that like in terms of like if the government uh is gonna give a lot of money, the government has given a lot of money to these types of companies. So if you get a government contract, that's pretty much like the best welfare chart. I mean the best um bag of money that these people can get, and and like that's kind of guaranteed income to survival. So if they're gonna be focused more on that, um and they have this hardware that's just sitting there, it's just a good way to kind of sort of bootstrap and uh you know converge on these two different places. So I think you know, if they're chasing space and they're chasing, you know, like this big space race to put uh AI on in space, even with like NVIDIA and so forth, um it's it's already kind of like this combined superpower that's just kind of going up there, and um, all the stuff that is being placed down here also has to be up in space, and like getting those uh relationships established is also pretty good. It's really awkward that my you know, like Elon was just totally two-faced about this. It's like, yeah, you know, I hate this company, da da da. It's like, oh yeah, by the way, let's go ahead and sign a deal.

SPEAKER_00

Like money talks, and and he said, Oh, actually, they're all good people, but there's a clause that you he listed too, which is which is really interesting. He's like, if he feels that um anthropic's models are not good for humanity in the future, he reserves the right to pull their compute at any moment on a moment's notice. So very curious to see how if that clause ever comes to bite anthropic uh in the future. But I don't know. I think the money that that's coming in will be quite valuable to SpaceX. And uh if it's a public company, you can't just you can't just rogue them like if it's like a quarter of your revenue or even half your revenue at that point. I you know, I don't know. Uh curious, Adam, what are your thoughts on this deal?

SPEAKER_02

Uh I think if you look at pretty much any of the CEOs in the AI space, Elon included, it's all a marketing ploy. Like when you're when you're building Groc and you're launching it, of course you're gonna try to like throw everybody else under the bus. And you know, you'd like to think you know, some people get behind Elon, like he's actually the arbiter of truth. He's no different than um the CEO of Anthropic or the CEO of NVIDIA or the CEO of OpenAI. Like it it's all like uh at the end of the day, it's all gonna come down to money. And this probably just financially made sense to Anthropic. They probably are paying, you know, Elon's probably like, yeah, we got Colossus 1, we got Colossus 2, Colossus 2 is what we're using now. Colossus 1 is kind of just sitting there, let's make a crap ton of money on it. And and now he has to kind of go back on the underhanded comments that he had around them being called misanthropic and like they weren't good for humanity. But again, it's to me it's the marketing ploy. Like, how do you get the most uh fundraising? How do you get the most attention? You try to make the other people look worse, you try to scare people, you try to hype things up. I don't know. That's my take on it. I think it totally makes sense. I am curious about that clause, which uh I've thought about too, because it would be very funny if somebody ticks uh Elon off and he like pulls in one day, and I could totally see that happening.

SPEAKER_00

Never know, it's very unpredictable working with Elon. You never know what's gonna happen. Well, I I I think we'll just have to wait and see. But for now, if you're a cloud user, I think that's great news because all of a sudden, you know, you you get way more rate limits. And you know, I think for the people who are trying to use this as a daily driver uh for work, like I think this makes all the difference. So now clouds switch back gears. Oh, cloud design. Oh yeah. Well, have they they haven't changed the limits with this A? Like it was just for the still get like five prompts, bro. Then I'm done for the week. Brutal. Yeah, yeah. Brutal, brutal. Yep. Um, well, you gotta make our own uh cloud design and and just route Opus through it. I think totally doable. There's a there's actually a startup called Magic Path that maybe they're they're working in this in this route too. So um yeah, take a look at them. Um so codex is now an everything app. Uh they released their um co-work equivalent, their ability to kind of have a more uh dynamic way of presenting, you know, a little bit less code-oriented, a little bit more Figma or Gmail or or what have you connected in there. And with the with the browser use and and the computer use and all of these app connectors, it gets becoming quite the hub where you can kind of just get all kinds of work done in there. It sounds like Ray, you're you're really into that and and you're making real strides there. Uh you know, you you were previously very, you know, cloud desktop pilled and you're working co-work there. I'm curious, like, how are you feeling the differences are for you and what you're able to accomplish with these models?

SPEAKER_01

I think the memory system has just become super strong and sticky on. Really? The memory system. Yeah. So they have this new thing that's still under development called the Chronicle. It will burn rate them. It's basically just takes screenshots and does a screen recording. Uh, and it requires like the Mac OS, it's only specifically for the Mac. Um, you know, you need their screen recording, the accessibility permissions, and then it's basically just gonna take these snapshots. But the the amazing part is that like I'm just creating new chats. And so normally I would try to do this in open claw, but now I'm just kind of doing a lot of it in codex. It really yeah, any workflows that I'm doing, it just kind of memorizes them and it's like hey, go do that thing again, you know, and like hey, oh I'm gonna do that, and it it sort of tries to infer those things. So it the personalization as the more I'm using it kind of sucks me really right in. And so I I just so want this on my phone, you know, like that's why I have open claw because now I have to be on my laptop open, you know. Um, but so I'm trying to figure out how to get my my open claw to drive my machine to do the same tasks. It's like, hey, go go type this into Codex app on the desktop and run this thing uh so it can control my machine there because um it it it can literally control everything on the computer. Um there's something secret about like the way um people have been able to tap into Apple's accessibility system to control a lot of the computer, and uh Codex's model is really, really good with that. Um, just from talking to some of the employees at the OpenAI party, um it appears that quite a bit of them know kind of what's going on uh underneath the hood uh in the operating system and kind of know where to tap in and are very extremely experienced in this workspace. I think they also acquired a computer use company.

SPEAKER_00

They did. Well, they acquired like a Mac automation company for like using AI and that that was doing that was doing all of this. I think they hadn't even released their app yet and they acquired them. Um but I think they're they're ex-Apple folks, very, very skilled at uh knowing the APIs and what so you need you need that those kinds of folks. And and honestly, this kind of shows like you know, the Mac is currently the place to be if you want the most automation. Um it's a bit unfortunate, like the other OSs are getting hunk to dry a bit, but I you know Windows just doesn't have this kind of tooling and Linux does, but way more you know, mix mixed and match, and the adoption there is a lot lower, so Mac, Mac it is, you know. Um curious, Adam, does this make you feel a bit of FOMO on the Mac and and getting you uh getting you hooked up for all this automation?

SPEAKER_02

So I was in Best Buy the other day and uh they the Neo came out. I'm like, man, that's like such an inexpensive uh laptop. Like I should just grab one. So I'm like, I so I asked the lady, I'm like, hey, I would like to grab the larger, like neo green one. Of course they didn't have it in stock. I'm like, okay, I'll do the smaller one. She's like, we don't have that one in, we don't have a green one at all. I'm like, all right, I'll do the larger silver one. She's like, Oh, we don't have that one. We have the smaller silver. I'm like, okay, I'm just I'm not meant to get a laptop here, so I'm just so yes, I have had FOMO about all this stuff to your question.

SPEAKER_00

Just the hairs away from getting getting full Mac in. Oh, geez. Yeah, the the automation is nuts, though. I think they are bringing computer use to Windows uh in the near future, uh, but we'll we'll see TBD on that one. Um but crazy times and like I think it also shows a little bit about like the future of working with these tools where you know you know to your point, Ray, with Chronicle, like it's kind of learning how you're using the app and kind of recording them to automate in the future. There's a little bit of weirdness there for me where basically I'm basically teaching the AI how to do what I'm doing for work and it's learning on the go, and I'm like opting into this because it's making my life easier. But you know, if I'm working for a big company and they're they're chronicling all these chronicles, you know, that that kind of maybe brings up some uneasiness around you know layoffs and having people kind of train their replacement. So I I you know, and today this week, you know, we also had you know some news with with Coinbase in particular, uh, and making a big layoff announcement related to AI specifically. I mean, you know, Coinbase, as you know, is a crypto company and you know crypto is very volatile and you know, we have up and down, and you know, you know, it's not like exactly an up market at the moment for for Bitcoin and such. So, you know, it's not like the hottest time for people investing, so that really affects their bottom line, but they're they're trying to look ahead and see like, okay, well, you know, with this fluctuating seasons of crypto, like, you know, how are we spending money? And they're you know leaning heavily into token use, and they decided that, well, you know, like they don't need as many people to do all the work that they're doing, and they're kind of vibe shipping a little bit, you know, going overnight, AI code straight to production, uh, which for a financial company is a little concerning, to be honest. Um, but I'm curious, you know, like you know, how are you feeling about this, Adam? Like in terms of the layoff story, like I know I know this is hitting home for you, and you know you've seen you know several other companies make similar changes there. So, what are your thoughts there?

SPEAKER_02

I've had quite a few people in my network get hit recently. Uh PayPal also did a massive layoff, uh, 20%. Coinbase put out that letter that got a lot of really interesting responses on X. And one of the things was we're gonna have non-technical people ship code to production that people laser focus in. And that is scary.

SPEAKER_00

It is.

SPEAKER_02

So to be fair, Brian did reply back and say uh we are going to have humans like vigorously review any code before it gets to production. So it's like anyway, I'm right. You say that you save it in face a little bit. Exactly. So I've been thinking about this entire dynamic a lot. And having been in the ship code is not the bottleneck in the majority of companies that I've been part of in my entire life. Code is really not the bottleneck. Like we can write code even before AI. Code was not the bottleneck. It's testing, productionalizing, figuring out what to build, the design process, the iteration on things. So that's um the the cross-team coordination tack that happens. And I I do agree smaller teams need to be so I don't disagree that some of these layoffs are probably needed, but I think the AI side of it is probably just a safe face in the market. That's not a general sentiment. And I also think there's a cost factor as a part of this. So if you think about it, like I can move, I can write code a lot faster than I could five years ago, but I still get blocked by all the other things that keep me from being able to ship quickly. So let's say just hypothetically I'm 20% faster, but I'm spending 30%. Uh I I'm actually spending more than what I'm productively gaining in AI costs because my AI costs are very high. So now the company has been able to move faster, technically, a little bit, but now it's spending more money on AI. So they need to so we have to figure out on the AI side, like, how do we actually get past the cross-team coordination tax? How do we actually get the quality? Like the the the issue that I have with um the statements that are said about engineers going away, it's like look at any large-scale system. There's no no way AI can manage some of the things that are that we manage on a day-to-day basis. It just can't. Like, there are certain things that are so complex that span such a wide breadth, they can't do it. And a lot of it is like architectural engineering, distribution, infrastructure, things like that that are just impossible to do. So I think it's those two factors in my mind. I think some of it's saving face, they overhired, they know they're going slower, they need to reduce their team size, so they blame it on AI. And the second thing, I do think there's a cost factor. I think teams, I've talked to so many engineers and so many companies, and it's like AI at all costs, use as many tokens. Jensen said you should be spending what was it, like 2x your salary on tokens. Yeah.

SPEAKER_00

Yeah.

SPEAKER_02

To do that, you've got to go three times faster just to break even. It's wild.

SPEAKER_00

So I I I want to like take a moment though, because the thing is, you know, if you're spending all this money on AI, like where's the ROI at this point? Yeah, like are you are you getting 2x revenue from 2x spend on your engineering? Probably not. I don't think so. Um, you know, are you are you getting like you you have to look at what what what are you shipping? Is what you're shipping adding value to the company? Um, you know, are you adding technical debt? Are you fixing things? Are you actually improving things? I think for some teams the answer is yes, like you you are moving faster and is is adding a lot of value. Like if you're an AI lab, you know, working on your AI app, like adding more functionality around the LLM is is you know net value, and it's in in this space is so competitive, like they need to move fast. And Codex is like really leaning into this, but Claude is as well, and they were doing it first. So I think for them, you know, they you can clearly see like the ROI is there, but I think for for for folks that are like just running a normal business, you know, especially a random SaaS, like you know, your users they don't need 10 times more features to, you know, you're not gonna get 10 times more users with 10 times more features. So like there's a big overhead there that you have to think about, and you also have to think about like are users using these tokens efficiently, like if you're burning 10 million tokens uh on something stupid where you're just kind of you know spinning something in a loop or you're running a stupid goal that that just goes on forever, like are you actually getting your money's worth from those tokens? And and I think at some point the crackdown will come. And Uber announced recently that you know they they were trying they hit their their quota for the year in about three months, uh, which is kind of insane to think about. So these tokens are getting more expensive, people are using them more, and what are we getting for it? Curious, Ray, what what do you think about all this?

SPEAKER_01

I think about it from a couple different levels. One, there's uh like you said, for the messaging part. I feel like that uh the CEOs are doing a massive disservice in some of the media around this, basically blaming AI for uh replacing jobs when there is basically something that's going on in this AI space. So they're basically trying to conflate these two different ideas. We're playing with this technology, we know there's capabilities to be here powerful, they may not be here today, but we don't know what that future looks like. The second part of it is like operational, so kind of putting on my operations hat. So it's like there is something to be said about really tiny teams reforming to move towards a new goal. And I've seen this a lot whenever you pull experts in, right? Like creating a new device type and the new direction and a whole new like trillion dollar to add two to three trillion dollars of value to the company that can take ten years plus, but uh getting the right people in the room, leveraging the existing operating system that had already been developed and other things that had already been kind of laid out for the company, and then those smaller groups of people who are the architects or the bigger thinkers who can really kind of set the leverage, kind of set the new floor for where things are going. And there is something about when you almost have like a big forest fire and there's new growth that happens. Like, I feel like we're in this era right now of like a bunch of work that we thought was gonna be capable is kind of being redone. And I think back to the days of Yahoo versus Google, like the reason why Google took off is because they had the right floor of saying, we're gonna set up these robots that you know, these machines that go around and scrape things on the internet in this really programmatic way, where Yahoo still hired people to actually curate news and put stuff onto news feeds and people were still subscribing to them. And then eventually it was really clear where where that was going. And maybe there's a similar thing happening here, whether we realize it or not, where uh workflows are being rethought of. You know, this engineering workflow and maybe Jensen's uh encouragement to spend more tokens is more of like people to really rethink their workflows and think of it more as like managers, not necessarily as like single person doing single tasks. Which is what makes uh codex things really interesting. It's like it's like you say, it's watching me work, but it I don't want to wake up 30 minutes scrolling through emails, responding to stuff I didn't respond to. Like, you know, I do want to hand that off to models and just only bring things up that are really critical that I need to, you know, move the ball forward at actually gonna drive uh economic value. So that's kind of the way that I'm looking at it. And I think this is a good time for individuals to if if you've been laid off, it's not necessarily something that's a reflection on you, because you know, I've also been laid off too. So it's like it's more of like how can I take this opportunity to reskill myself, to look at this from a different perspective uh and work on my things and just let go of all previous assumptions and just uh you know, just really just start playing with things. I think this is kind of where um you you'll actually start to find how valuable you are because you have unique insights in deep places that uh you may not realize that you can leverage.

SPEAKER_00

Yeah, getting your hands dirty and trying these things is is super important right now, and I think it's it's an opportune time with you know these these plans offering pretty generous rate limits to getting started. And I think it you know, just learning how the tools work, you know, what what a context window is, like all of these things are just basic table stakes to the what the future of knowledge work is. And I and I do think that like you know, you you the you will have a job around this for a long while, I think. It's just that you will have to lean into this stuff. And I think for many engineers in particular, there's a bit of a mourning going on because the job that they once knew is is pretty much dead at this point. Um, you know, the way that the way of working where you can just think about a problem and and code it by hand and take your time with it and having the space to kind of be able to do this work, like that that's that's kind of gone. And the thing is too, that's a little bit um disconcerting for especially senior engineers, I think, right now, and and you know, seeing this news of oh yeah, non-technical people should be pushing PRs, um, is that you know, like the seniors used to be a sort of quality gate at these companies where they would say no to things and be like, no, like we can't build this feature because one, it doesn't make sense technically, and two, you know, it's it's just not the right way to kind of go about this this this whole problem. And um, and and you can kind of trust their expertise to kind of allow that quality filter to kind of be around. But now, you know, you have you have a junior who comes around and is like, oh yeah, I vibed it up in a couple of hours, uh, here you go, or the PM is like, yeah, yeah, yeah, we can do this feature. Look, I prototyped it right here and code slot, but like whatever, it works, you know, like let's let's ship this. Um, and and you have a lot less of that gating that goes on um at these companies. And I think over time it's gonna be a pressure point on these companies and cause a bit of dysfunction. So I'm curious, what are your thoughts on this, Adam? You know, you've you're a senior engineer, you've you've gone through this yourself, I'm sure.

SPEAKER_02

Man, honestly, when you were talking about one of those the example just now, you gave me a PTSD. So just recently. So here's here's the other challenge that we run into. If you do have these like mid-level and junior engineers building or rebuilding parts of the app in a very short amount of time, like in its own isolation, so like Greenfield, and then demoing it to people that are non-technical, that are business holders.

SPEAKER_00

Oh, yeah. And then the stakeholders, yeah.

SPEAKER_02

Yeah, and so then they're they're they're like, uh, all right, let's go ahead and let's get this rolled out. And it's like Shove it. It's there. So it's like we can't we can't do that. This is literally like a prototype that's not connected to anything, like there's no scalability to the thing at all. But it's like no, no, it works.

SPEAKER_00

Like it needs to be out the door tomorrow.

SPEAKER_02

It is it is uh it's a very hard balance because it's like I I want the juniors and and folks to do these cool experiments, but they also need to understand the reality of what it means to ship something. And like at the end of the day, the customer is the most important thing. Like if they if they get a bad experience, you're gonna lose them, you're gonna burn their trust, they're not gonna pay you. They they buy your software for a purpose, and if you break that, it ends up being bad. And it's that's been like one of the hardest battles that I've had. Your other point you were making, Eric, which I think talking about jobs and you know, laid off, the most valuable people right now are those that know how to build agents. Every company's building them. Very few people actually really understand how to build them. And I've spent a significant amount of time trying to up-level folks to understand simple context management and what needs to happen there. Uh like building the core agent loop, very simple. But it so previously, do you do you guys remember when Claude uh Anthropic said that like developing clock codes is kind of like vibes? You know, do you remember them talking about like vibe development? I've actually come on board with that. If you think about video games, video games, like what makes a good video game, it's really like how's it feel? Does it feel good? Like is it is the gameplay loop good? Like if something bad happens, does it is it funny or is it annoying? Like there's certain things that uh that ha agents are the same way. Like it they absolutely are. You can't you can't like measure the vibes of What a good agent is. And you almost need that artistic sort of like fe it's almost like a video game in my mind. I've been thinking about it a lot more like that. I've been trying to explain that to folks. It's because like a lot of engineers are like, I want to know the exact scope of things. Well, I'm like, no, think about a video game. Like, when is the video game done? Like, it needs to be good, it needs to feel good. Just because it does the thing does not mean it's shippable. And like trying to push on that. So anyway, that's that's my thought on all that.

SPEAKER_00

No, no, I I think that's a really good point. I was actually having a convo on Twitter the other day with someone about this. And and like there's just when you're building an agent, there's no substitute for the vibe of the agent because it's like, you know, there's the feel, but there's also like how does the model behave when it's doing certain things? Like, because you have the prompts and the tools, and you know, where is it getting hung up? Like, what is it struggling with? And you know, if it does a failed tool call, like there's probably a reason it failed, and and you know, part of it might be the user's prompt. So then maybe your system prompt should account for the user's kind of misgivings, or maybe you know, your tool is too strict and finding ways of loosening that up. So a lot of it is just you know, trying it, feeling, getting a feel for it, getting a feel for how the model behaves with it. It you know, like it's just you just got to use it, and and people want to put numbers to it, they want to run all these optimization algorithms on it, and like DSPY is like a big one, and I think there is some value to it, but like a lot of the time it's just sitting down, feeling, getting a feel for it, and seeing how it goes, and and you know, it's kind of a weird place to be. Like these these models, building agents around them, it's nothing like we've built in the past, and that's why it's a whole new skill set to gain and and and experiment with. So do you agree that it's kind of similar to video games? Like it kind of yeah, well, yeah. I mean, just look at what Claude said that the the claw code is like a game engine where it needs to run at 60 fps. Oh, jeez, in the terminal, which which is a bit funny, but um no, like the the UX have been.

SPEAKER_01

Take me through is like open claw and Hermese, because they're kind of these agentic frameworks, and they have different vibes, each of them, right? There's yeah, yeah, yeah. Two different camps, they're almost fighting against each other, right? Right. Is it kind of the same vibe too?

SPEAKER_00

Like well, so I think for them it there's a little bit more to it too, because of the personal aspect of it. Like that in there, you it's not as much about the productivity of the agent, though it is like in part, um, but it's more about how does it feel to talk to the agent? Like, do you feel like it's listening to what you're saying? You know, there's like a whole art to that. Like, is it do you feel like when you ask it to do something, does it actually get the thing done? And that's kind of like all in the context engineering there to make sure that it's able to properly handle you know your cruns and all that stuff. Um, and then there's like, you know, like when you ask it to do something, like is it does it is it taking enough initiative to kind of be proactive and do things? And there's like a line there, a slider where some users want it to be extremely proactive and some users, you know, want to be a bit more measured about it. And so that that's kind of like in in the balance there too. And I think Hermes is more of you know, it kind of more of a robot, like we're gonna get shit done, and and here's the tools. And and open claw is you know, it's very much about its soul MD file and its personality and the vibes of of talking to it. And is it gonna come up with a joke? Does it feel funny like to talk to? Like, am I am I enjoying myself with it? Um, so I think there's there's a lot around that that like you know shows people like why they picked it, um, beyond just the fact that it was like a novel thing that an agent could do all this stuff.

SPEAKER_01

Can we do like a little mini video game exercise, like a character selector right now for like agents? Like Adam and Eric, I want to know like what are some of those properties, like maybe three to five properties that you know if you were to, you know, if we were to like I were to build my own agents and kind of Street Fighter style, like select this one. It you know, what what attributes are you looking for? What should we define for like this guy that you're talking about?

SPEAKER_02

So I'll maybe I'll answer that with another point that I've been actually I've I've come to recently, which is you know, that I was working with a couple teams and they were like, okay, we're gonna do some strict evals, we're gonna make some stuff happen. And then I was like, we're making a lot of assumptions what good looks like to our customers here. Yeah. If we if we are evaling with our biases on what good looks like, I wonder if we should rather be focusing on how do we let the agent adapt to the user more and be to their and so I'm in the marketing space. Marketing one company to another, very different. Building evals around a marketing use case would work for some percentage of customers, but then the other percentage of customers may hate it. And you can very clearly take the take and do this theory manually by taking marketing content and showing it to five different people, you're gonna get five different answers. So, how do you judge that? So, to me, to your point, uh Ray, I've kind of come to the conclusion that I think the agent needs to be able to adapt to what the user wants. Like it needs to be not so rigid that you can't steer it, and it needs to have some level of retention of that information so that it can pull back and understand what your preferences are. Of course, it needs to have personality, it needs to be fast. There is a level of like just simple stuff, like it needs to feel good. The micro interactions that happening happen are more important than a person would consider. Like if you're just sitting there with a loading spinner waiting for the text to come in, simple things like text streaming are very important. Like you should see things coming in or reasoning update, but you see people building agents all the time that are just like miss all these fundamental things.

SPEAKER_00

So I just want to just throw in a word. There's a technical term for games, it's called juice. Um sounds very, very weird, but yeah, yeah, yeah. So juice, and there's there's um there's a nice YouTube video, I think it's called Juice It or Lose It, if I'm not mistaken, on on YouTube, which explains like the process of adding juice to a game like uh Arcanoid Brickbreaker. Um and and and it's really important to learn like think just like each interaction where the ball bounces, the is there screen shake, does the things animate in? Like all these things really change the feel for how the experience of playing the game is. And and even though the gameplay is identical, you're getting a very different feel and experience coming away from it, you know, afterwards, and and and how engaged you are in the game. And so I think these kinds of things, like you're saying text streaming and all of that, that's all juice and cosmetic stuff that that keeps you engaged, that keeps the dopamine flowing as you're engaging with it and you feel good about it, even though it might be slow, if you're feeling like something's happening, then you're you're all of a sudden more prone to like being okay with waiting and all that stuff. So sorry, I wanted you to continue. I just wanted to chime in a little bit with that.

SPEAKER_02

No, I mean that that is exactly right. Like that is exactly correct. So there so so there's like those part parts of it, right? The way I would think about it. The um the other thing that Eric talked about is like, does it do the job that you want it to do? And and there is a level of that. And and this is very complicated because it could do the job wrong, but still same thing with the video game. Like you're gonna have bugs that happen in the game, but sometimes they're comically fun. Like that sometimes it actually makes the game better. How do you make your agent experience so that if something bad happens, it doesn't feel like you hit a dead end? Like it actually feels like you can continue or that you're not stuck. Like you it's uh things like that just make a lot of like very important for an agent experience in my mind. Uh so that that's the way I've been thinking about it, and that's kind of the direction that I'm driving in a lot of stuff that we're building.

SPEAKER_00

Yeah. And you know, it's it's it's tough because like you have all these metrics and the proactiveness and you know how agentic the model is and and how how capable is it is to keep going, or when when you want it to kind of just do a thing and stop. Um, you know, like all of these measure on other things. Everything's kind of a trade-off in terms of latency or or or success rate or token use because these agents are expensive too. Um, you know, you have to think about all of this. And but uh, you know, some of it does come down to the model, and you know, you do need a really powerful model, and people are trying to cut corners a little bit on this front because they want something cheap to run, but sometimes you can't really cut that corner. Um, but you know, just in terms of the model doing the thing that you asked, like I do find GPT 5.5 and to uh open IS credit, they've improved a lot on this because in the past they were very the word they were the worst on the model kind of doing what I said versus what I meant. Um, and now I find 5.5 is the one that really goes in and and follows my intent the most. Um and you know, often what ends up you what to get that good result, you need context. And I find if you're using these models, like adding a screenshot sometimes or adding some some some support documentation, it goes a long way in in kind of pointing the model so you can both look at the same thing and it will infer a lot of things from your prompts based on that. Um, so a good agent could could also you know be really proactive about including peripheral context, like you know, early on with with cursor, one of the things they did is they they you know they would keep track of okay, what files do I have open? Uh, what where is my cursor kind of focused on right now? All these different things about your interaction with the with the app that kind of helped the agent know, okay, this is what the user's talking about. Um, and I think there's there's a lot of low-hanging fruit to kind of be figured out with these these app experiences. And I think, you know, I at one point we are gonna hit walls with running it in like WhatsApp or Telegram because you can't really add these things. But if you're running it in a custom app, like I think that's where you're gonna be able to get a lot of lift uh just from little things.

SPEAKER_02

So yeah, yeah, curious. To your point on that, Eric, the way that I've been bifurcating this in my head um is there really are three major surface areas that agents are exposed in. There's gonna be third party, so you're gonna do like MCPs and you're gonna have uh like OpenAI has their apps. You're limited to what they give you access to. So those should be good experiences, but they're probably not gonna be great for the things you're building. Then there's the personal desktop experience where you basically are opening up your so the things like Cursor, for example, those should be phenomenal because you have full control over that. Like those should be exceptional. And then you have your your first party surfaces, which are in the cloud, and and those should be good, they should be great, probably not phenomenal because you're still limited by some things. You don't have access to the file system as much. Like there's certain things that you don't have access to. So you can kind of start thinking about like your agent development of yeah, we should make a good experience, as good as we can in these third-party surfaces, but we're limited by XYZ. And I could tell you a bunch of things you get limited by, and there's and it just makes sense. They've got an agent that has to call a million other MCPs and apps and things like that. Like, they have to have limit. But in your own environment, you should be able to crush it.

SPEAKER_00

Just to pause you on that, though, because I think you know, there's something about the blend there where you want like these things are all connected in a way. Um, you know, you're working on an agent in the cloud, but that agent in the cloud can be connected to your personal computer and control it and get context from it. Um, and that like when you're talking in WhatsApp to your OpenClaw or your Hermes, like it does have access to your computer and it can pull that information. So I think like the blend of your computer being involved and the cloud being involved and maybe your phone being involved, like all of these things, they're all becoming connected in a way to get that great experience. So I just wanted to caveat that a little bit because I think like right now in isolation, you're right, but I think as things go on, like there it won't just be one or the other.

SPEAKER_02

You're right. They might like actually converge into like one surface area that interacts with all of them. You're totally right.

SPEAKER_00

Yeah. Um, you know, just AMP AMP released their updated agent uh the other day, and and one of the big things that they did was like, yeah, like your your the terminal app is still there, but now it's a remote control, so like you can start a chat on the web and then the agent can control your computer because of the terminal apps. You don't have to look at the terminal app anymore yourself. Um, and I think that's just like an increasingly common thing that's gonna come around where you know there's already codex remote control apps that you can download. One's like Kitty Litter, where it does the same thing, uses App Server, and just controls your computer for you. Um, so yeah, more and more of these services are all blending, things are are are are all connecting, and um, it's gonna be wild like having you know orchestrator agents that just kind of talk and command all these different agents running in different services. Like this is gonna be a weird future, expensive one though.

SPEAKER_01

I'm gonna set a challenge for myself is to build an agent, and I'll come back with a demo in a couple weeks and I'll show you my demo. Very excited. Uh, but it's gonna be very video game-esque, and I think that's I think the era I think I want to move towards because I very good. Everyone's showing the same little you know Google search bar, but I want to change this paradigm. That's a sort of cool idea. This is thank you guys, that's really awesome.

SPEAKER_00

Yeah, yeah, yeah. And and please do watch that video, Juice It or Lose It, because I think it'll give you a lot of inspiration for your app. I think it'll be really good. Yeah.

SPEAKER_02

Yeah. There's there's quite a few videos that I could I remember watching where you the I don't know if this is the same one, Eric, but they'll start with a very stale game and then slowly start iterating on it, and it ends up being right. It's like this insane thing at the end. You're like, holy crap, that is insane.

SPEAKER_00

Yeah, yeah, yeah. No, that's exactly what they do in this video. That's probably the one you're talking about. Probably is, yeah. That's why it's such a good one. It's a legendary video. Um, yeah.

SPEAKER_01

I just pulled up like a website and they have like little uh was the breakout or the brick building. And so like it just shows like black and white, but then they're changing colors and like little thing has a little trail following it. And yes, you know, as it hits your little thing, it shakes the screen. I was like, oh my gosh.

SPEAKER_00

There's math to this too, right? Like, you know, there's all these like easing curves that you learn about in in juice, and and it's just a lot of fun stuff. So so yeah, a lot of a lot of stuff for you to do, Ray. I think uh I think it'll be a fun time to explore. And um, yeah, you just gotta try it. Like, honestly, the biggest thing is you gotta sit there in the driver's seat, iterate, try it. Does it feel good? No, let's try something else. You know, you gotta you can't just like fully automate that stuff. Like, you gotta you gotta have your hands dirty.

SPEAKER_02

Um, which is why I think which is why building agents is such a skill right now. So if you did get hit laid off and you are like looking for something to learn, focus on that. Like it is incredibly in demand.

SPEAKER_00

Yeah, yeah. And and a lot of it is around the tool. So find a vertical where that agent really makes sense and and and try to like lean into making that a really good automated experience. Uh, you know, sp often what's really good is you have something very complicated and and you want to make it accessible to people, having an agent lets you kind of hide that complexity behind prompts and tools, and and there's a lot of skill in making that work well for the agent too, because the agent needs to understand it as well, not just you know, it it takes work for that. Um so yeah, give it a shot and let us know uh if you if you built something cool out of that. And and let us know if if if you you're put thinking about juice in in your apps and your and your agents. Um all right. Well, I think this is a good spot to kind of wind down for today. Uh, I think it was a really cool chat here just around all of this stuff. And um, yeah, I mean the layoffs are are definitely a bummer, and hopefully, you know, you you all have uh you know good good career stability there. Um, but it's definitely an interesting time to explore and experiment, and there's a lot to be building, and I think leaning in is the right move right now. I think you know you can have a serious edge if you understand what this stuff is and start compounding that knowledge and understanding. And and and by the time you know this stuff is gonna keep changing, like if you're staying up to date with it, like there's a huge gain for you uh in doing so. Uh oh it also.

SPEAKER_01

And encouraging people to build community as well. So if you know other people who are in the same boat, bring them together and just say, like, we're just gonna learn agents. Let's come together as a group, meet up at a library, wherever, coffee shop. I think that's gonna be more important than ever, because the more you stay in isolation, the harder it is, to be honest. And there's probably a lot of people around you in the same company. Um, you know, just try to like move in that direction. I think it's gonna help a lot too.

SPEAKER_00

Learn together, build peer groups. I think that's super important. It's a great point, Ray. Yeah. And what about you, Adam? Any closing thoughts for today?

SPEAKER_02

No, great episode though. This is and again, uh it is a bummer anytime we see these layoffs, but I do think there's a lot of opportunity out there. So um if anyone, if anyone is like getting deep into agent development, like hit me up if you've you know, if there's if you're very skilled at it, we're working at another company or something, hit me up. Like, let me know.

SPEAKER_00

Oh, maybe some job offers for you in that in that message there. Well, best of luck to everyone, and and do remember to like and subscribe. Give us five stars if you can. Gotta juice those numbers to make sure other people hear about this pod. Um, and yeah, thanks a lot for your time, everyone. Yeah. Later, everybody.

SPEAKER_02

Have a great weekend.