GPT 5.4, NVIDIA GTC, AI Impact on the Job Market | Ep 12

This episode covers the latest in AI model releases, hardware advancements from NVIDIA at GTC, and the evolving landscape of AI's impact on jobs and software development. Experts share insights on GPT-5.4, inference hardware, and the future of AI-driven workflows.

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
02:57 Exploring GPT 5.4: Features and Improvements
06:08 The Role of Planning in AI Coding
08:59 Context Management in AI Models
12:10 NVIDIA GTC Conference Insights
15:01 The Future of AI Inference and Hardware
17:52 DLSS 5: AI in Gaming Graphics
24:52 The Future of AI in Gaming and Film
25:58 Understanding Open-Claw Strategy
27:45 The Rise of Personal Agents
28:32 The Changing Landscape of Software Development
31:18 Craftsmanship vs. Automation in Software Engineering
36:07 Job Displacement and the Future of Work
41:10 Optimism in the Age of AI
50:29 Skills and Context Management in AI
54:00 The Future of AI Interaction

Transcript

SPEAKER_00 0:00

Ladies and gentlemen, we have a fire episode today with your hosts Ray Fernando, Adam Larson, and Eric Provence. Together we are the Rate Limited podcast, and we cover agentic engineering, AI coding, and so many various topics. So if you're coding with us, if you're hanging out using these types of systems, this is gonna be the podcast for you. And we have decades of experience each of us in our own separate industries, and we come together to sit down and have these types of conversations. We're gonna be covering lots of really good spicy stuff, stuff that's been going on right now with NVIDIA GTC, some model releases, and kind of some other things that you should probably be tuned in for. So without further ado, um, Eric and Adam, welcome.

SPEAKER_01 0:42

Hey, and thanks for the intro, Ray. So it's been a little bit of a gap here um in since our last episode. You know, I had a little bit of an emergency with uh nice cold weather, uh rainstorms pouring ice all over the place. So uh unfortunately had to push this back a week, but here we are. Um, and we've got a lot of releases that came out in the last couple of weeks. Uh so the first one I want to touch on that we didn't really get a chance to talk about is GPT 5.4. So this is uh a big release from the OpenAI folks, and you know, as of the recording, they just released the mini model as well, which was a nice welcome upgrade. Um, but I want to hear a little bit from you all. Like, how have you appreciated this new model? Um, you know, what feels different from it for you? Like I know like the big thing that seems to be going around is that the model is like smarter and brings back uh a little bit less of the robotic talking to you and a little bit more prosaic in how it talks. And I think that's a welcome addition. But you know, Adam, uh you you you know, you're deep in the weeds with all these models. Uh how are you finding 5.4?

SPEAKER_02 1:40

I honestly think it's so much better than 5.3. And this you don't have data to back it up, but what you just said about the robotic nature of 5.3 is just such a turnoff, in my opinion. I think 5.4 make it's just much more enjoyable to work with. I did actually play around with the mini model as well, and it is a very competent, you know, super fast, like really, really good model. Um overall, I would say this is like the first time that they've kind of started to unify models a little bit more, too, right? So they used to do like codecs and then the regular model, so now they've kind of brought that together into one one kind of model that we could actually use. Yeah, overall, very positive. I think still Opus 4.6 is my like major coding buddy. Um but I'm gonna open it. I've started to give them a run for their money in my mind.

SPEAKER_01 2:31

That's interesting. And how about you, Ray? What are you what are you finding with the model? Have you switched over to it as your main coder?

SPEAKER_00 2:36

You know what's really funny? More often than not, I actually prefer 5.4 for planning now than I used to uh with Opus. Really? And I was huge into planning, and I think um there's a couple different experiments I've been doing, and I've been trying them in these different harnesses. And so I actually found, to be honest, Cursor to be the best harness for GPT 5.4 um for especially the tool calling tasks and grabbing context from the repo repos and stuff. Uh I did a massive refactor on my repo and added CLI capabilities to my apps, and because I want to have them available for open claw. And so it's just been really fun to just literally just do a whole re-architecture plan. And uh I find that 5.4, I have to switch between extra high or like medium. And so if I want to just choose high, curious. Um I I find that high and extra high are kind of the same, but I get a little bit more details out of extra high for like thoroughness for certain things, especially from like having it. Um most of the tasks that I have it sent out, like especially for the refactoring stuff, has just been like, hey, go take a look at these types of things. This is what I'm trying to do with the CLI. I want to understand like the data modeling. So it seems to do extra loops just to make sure that it grabbed additional files um that it needed to for those tasks. And a lot of times with cursors like launching these different subagents to do them. And um it's it's it's really, really fast. I I don't like even with the fast enabled to to get priority access, it just uh seems to be crazy fast. Obviously, it's a really high cost, but um, you know, with these research credits that I'm burning, um it's been kind of fun to just play around and see what it's like to be uh you know a tier one uh engineer with like infinite access to credits and stuff. So it's it's been really it's like okay, these guys really just can fly because you don't think about cost. And it's a weird dynamic when you're thinking about that.

SPEAKER_02 4:27

Yeah, I was gonna ask you about that, really. Like um because GPT 5.4 requires max mode and cursor. Like so you're so you're running it pure at token cost for all of them.

SPEAKER_00 4:40

Yes, yeah. And then I'll do extra fast like high with the okay. Yeah. And it it's it's it it can spawn off like four parallel tasks, and I I find it super useful, especially for research tasks. It is it reminds me of like the O3 model. Y'all remember O3 back when I was like I loved O3, it it it it's so similar to it now. It's like, oh finally, the model feels as intelligent as that, but it's faster. And it's like, oh, okay, this is really cool.

SPEAKER_01 5:10

So it's it's interesting you bring up O3 though, because O3, one of the big downsides of it was that it was much more prone to hallucination versus O1 and from versus like GPT-5, uh, which came out later. And I'm finding a lot, like, I mean, it's not like it used to be with O3, but like if you use 5.4 on extra high, especially, I do find it tends to spin out a bit more and hallucinate a bit more and like kind of go off the rails a little bit more. Um I do know that they they mentioned with the release of 5.4 that X high is like a notch more of reasoning than X High was on the previous models. So it reasons even longer. And I think like in some cases it's just too much reasoning, and and the model gets in its own way, especially if it doesn't have the right like way to retrieve information or like or it's retrieving the wrong information, like it'll it'll it'll kind of spin off like overthinking small things and kind of spend too much time on insignificant things. So you you end up like over resolving things that shouldn't be, you know, shouldn't matter that much.

SPEAKER_00 6:13

So it is sensitives there because I did notice that when I do like individual chats. So if I just want to accomplish one quick thing in the repo, I'd I'd noticed exactly that in extra high. It's like I don't really need this, you know. Like I just need to like I know exactly where this function is, like go go do that and like implement this, or you know, or take a look at these things. So I that's where I would switch back to medium, and maybe that's could kind of clarify that. So uh a lot of the times I've that found the best success for the extra thinking when I'm in the plan mode. So in that plan mode, what I liked about it as well is like my plans weren't actually my plans were much more succinct and much more um uh to the point in GPT 5.4 extra high, as opposed to like the Opus 4.6. I felt like Opus 4.6 was trying to sell me a little bit more on it, and that's kind of where I start to see like, you know, maybe I've been gaslit the whole time. Yeah. Uh so I don't know, but I always preferred now I prefer the plans in 5.4. It's it's it's really freaking good for me.

SPEAKER_01 7:13

I feel it it's interesting that you talk about the plans a lot. Um, because I actually saw this post on X the other day and it was from Dex Horthy. Um, and he's uh, you know, him and I are cut a bit from the same cloth in terms of how we think about context. And he was talking about how like, you know, when you have like really, really detailed plans that you're working with with models, a lot of the time you're actually not saving on what you're writing versus actual code. So your plan is like as detailed as the code would be, or sometimes more so. And so then you're like, Well, what are you reviewing? You're reviewing a like lossier version of the code. So then he's like, you might as well just review the code at that point. Um, and I think like in some cases that can be true, and especially if you're like overindulging on like super detailed specs, I think it's this risk there where you're kind of basically reviewing the wrong thing and you're you're kind of like letting the code run after. And I think there's an issue there. So I'm I'm curious what you all think. And and Adam, I'm curious if you've observed this at all when you're working with these detailed plan modes and spec-oriented development.

SPEAKER_02 8:15

You know, honestly, I hadn't thought about it from that perspective. For me, one of the biggest challenges with AI coding right now is the actual code review process. So what I was going through in my head is like, I actually prefer reviewing the plan, you know, one or two-page plan. I don't go into the detail some people do. What I care about are the high-level decisions that are being made. Where's the code gonna be? What are the interfaces for it? What are the things that need to be done? So I could see his point if people are going into like extreme detail and like really going into the weeds. But I think again, I could be I could change next week, but right now, like I enjoy iterating on the plan, on the high-level things, and then I less enjoy having to go through all the code it generates and trying to piece together how all that comes together in my head. That's where that's where I'm at on it, I think.

SPEAKER_01 9:12

Yeah, I think it's uh it's a careful equilibrium though. Like, you know, if you over-indulge on the plan, you know, you over-specify what you're doing. And then some people don't read what they're specifying, and so then you get like kind of like spec slop, and that's just like uh I've seen some plans not to call anybody hopefully they don't watch this uh podcast, but I guess I was given one I was given a spec sheet for something that was being built, and I opened it up 97 pages.

SPEAKER_02 9:41

And you could tell it was AI generated. I was like, there is literally no way I'm reading this thing. Like I'm not doing it.

SPEAKER_01 9:47

Well, I don't think they read it either, to be honest.

SPEAKER_02 9:49

No, it is it's like it's it's wasteful. Like honestly, there's just no need for that.

SPEAKER_01 9:54

But that's another topic. Oh, jeez. Well, alright. Well, on that note, let's move on a little bit to the other big release that that came out last week, which was Opus 1 million token context is GA now. Um and I'm curious, like, well, you all use Cursor a lot, which already had the 1 million tokens. Um I'm curious if you've all appreciated it in Cloud Code at all, or if you've tried it out and had any thoughts on that. Uh Ray, have you have you been playing around with it?

SPEAKER_00 10:22

Uh not not in Cloud Code specifically. Um but overall when I did use it in cursor for the 1 million token context window, I I didn't I didn't necessarily get any more benefit to having I actually appreciated more of the compaction that they do automatically now and in chats and stuff like that. Um and and there are very, very few cases, even for like this big refactor task. It's funny because Anthropic actually wrote a blog post about task decomposition, and then I've been studying that a little bit more after um being made more aware of this through the uh factory AI team. Um so it's it's really interesting that when you actually break the tasks down to much smaller bits and pieces, how um just like with a smaller context window you can just be super super efficient. And I feel like the intelligence of Opus really shines in that uh like that sweet spot of like 80k tokens in that context window where it's just it just has enough for whatever it's doing and and kind of gets in and gets out. But yeah, every time I've gone over 200,000 tokens or more, I haven't seen much of a like improvement. I haven't had much of a need, but I'm I'm wanting to m test this out maybe more inside a repo prompt because there is a different behavior when you throw all of the code inside of it and activate all the weights as opposed to just shoving in all these little bits and pieces of a prompt. And would love to hear your take as well on that. Yeah.

SPEAKER_01 11:45

Yeah, well, I mean, so the big thing with with large context is that like there's kind of two ways to kind of leverage it. So one is like you just keep going for longer, and if you're in an agent loop and the model's working and working and working, and you're just letting it bleed over, the thing that you risk is that the model starts to re um reread the same context multiple times and it starts to add noise. And if it's editing, say, like a really large file and it has bits and pieces of that file in, and then it's editing it, and then there's like different versions of the file in the context window, and then it's doing edits, you might see it start to fail edits a bit more because it's like recalling a replacement block for something that is no longer in the file, uh, and it has to reread context and then it does it again. Um, and so you you start to add a lot of noise in the context window and it distracts the model, and that's why you get a lot of intelligence degradation. You add ambiguity in terms of what the model's doing. But if you take a lot of context and you kind of organize it up front and give it to the model and spoon feed it, then it's able to kind of reason and attend to this large context in a much more efficient way. Um but even still, like if you want the best analysis, like you still want to really try and be efficient with that context. And and I don't think any of the models that are out right now are you know as good as they would be with lower context. You know, they they tend to perform worse, uh, which is a bit of an unfortunate thing. But it's interesting to try. And sometimes you do need um to give the model enough context to reason about something from a higher level, like uh just like have more resolution on the what the problem space is, because if you don't have enough context, the model is not gonna be able to know about certain parts of your code or certain parts of a spec or something that it it wouldn't it wouldn't be able to attend to if it wasn't there in the context window. So, you know, there's uh careful balance. Sometimes you you do want to push more, but like often if you can compress, like you're getting benefits from that. Um so yeah, I mean I I always try to go for less and um but uh speaking of context windows though, there is one thing that I wanted to to bring up is that uh just back to the 5.4 release. So, you know, um Ray, you know, when we when we first met, like a lot of the things I was doing with repo prompt, you know, is a lot around uh GPT Pro and GPT 5.4 Pro now, like they they fixed an issue after a long while where so you used to only be able to put in 60,000 tokens on the web interface for Chat GPT, and now it's unlocked, and I'm able to feed it like 160,000 tokens. And I've been able to do like deep plans and code reviews with this model, and it's been incredible to be able to just push it over there and and just really leverage the model uh with much more context. So it's it's actually a crazy thing.

SPEAKER_00 14:22

So for those who aren't familiar, yeah, you can pay your 200 bucks a month for your GPT Pro plan. And the pro plan allows you to run that model, which can run for 45 minutes sometimes or even an hour. It just keeps churning and doing stuff. And you you feel like the output from the pro model um you know outweighs a lot of the other stuff. And and so you you I would love to hear your operations that you like to like pick apart specifically for the pro since some people will be trying this, right?

SPEAKER_01 14:50

So like uh I'm curious what do you find that the best inside of repo prompt for well yeah, I mean yeah, so like often if I want like a really detailed like optimization or if I'm looking for like a deep code review or something like that, like and I and I'm not like as latency sensitive, I'll like kick off a prompt where I just generate like export the context and then feed it out. Um, you know, and and you know, I can't like uh officially like there there's there's some terms of service issues around automation around this, but like you know, there are people using, you know, like all the agents can use web browsers now and can feed context and you can do all kinds of stuff now. So feel free to experiment with what works for you, but um you can create these loops where you push context off and then come back to it later. And uh, you know, if you're if you're pulling it manually, that's all fire again. Like if you're just starting tasks with automation, that's that's like totally fine, I think. Um but yeah, interesting, interesting to be able to like push things off. And you do feel the difference with the pro model that can go with with such depth and can browse the web and research things as well. And it's it's like a full agent in there now. It's it's kind of nuts. Um so anyway, just a little side note on the pro models, which are worth coming back to. Like they they can they can do some stuff that you don't usually get out of the normal models. Um, now another thing I want to talk on this week uh is uh GTC. So that's Nvidia's conference. And um, you know, Ray, we've we've had we're lucky to have him here, who's been on site on the floor, um, and he's been experiencing this firsthand. So I want you uh to to dive into it a little bit, Ray. So what's it what's it been like being there? What what are your main takeaways from the event? Like where's the AI industry at, like in your in your experience there? It's kind of the epicenter of the whole space right now.

SPEAKER_00 16:31

So yeah, I'd say this conference is really focused a lot on the enterprises right now. There's like so many of these different things, but um also like on infrastructure and pretty much up and down the entire stack. And to me, I I didn't realize how embedded NVIDIA is with a lot of these companies. You know, literally I call it from like the electron up to the app layer. Uh you know, just like the actual power grids and like infrastructure and like all the way up, and if it means that they have to create software or combine whatever they need to do to get whatever the strategy is, and I think uh the biggest takeaway is just that, you know, they want to be like the inference king, you know, and basically a lot of the work that they feel is gonna be happening in the future is gonna be going through their GPUs and all of their architecture. Uh and if they don't have an architecture that's that they already have something on the roadmap that they're building and putting together, and that was kind of like uh I the level of detail also was also not only expanding in computers and AI stuff that we're doing here, but like robotics and like other fields that are like completely surrounding you know, all like basically any economically viable task, which I thought was interesting. And then um, you know, like the bigger narrative was just about you know, like you know, tokens and then like how are you gonna get them, and then companies that are currently paying them, you know, they're like showing off you know cursor and these other big providers who you know um are either making models or providing them, but then if they upgrade to the latest versions of the hardware, then it's gonna be much more efficient, and then they can still charge more margins. And I think the one example is like you know, a company can still make $150 billion more in terms of margins if they literally just get the latest version of the hardware, you know, like that's crazy.

SPEAKER_02 18:13

Yeah, I did want to talk to you about the hardware rate. Yeah, maybe you you the the Vera Rubin platform, I'm probably pronouncing that poorly. But um it almost seemed like NVIDIA made a shift. I only watched some snippets of things, so you have more context. But it seemed like they made a shift from training to like inference this time. They talked pretty heavily in the ones that I've seen where they were talking about higher like inference throughput per watt, so there's like a power focus. There's also more where before it was a lot on like training. There was a lot of focus on training and passwords and a watt. Did you see that same thing? And then did you hear much about the Vera Rubin platform or have takeaways from that?

SPEAKER_00 18:51

Yeah, uh they seem to be going towards that. I don't think they've released a lot of details in terms of like what's gonna happen for the rollouts, like especially they they brought in the Grok inference folks and that how they're integrating them, and then they're gonna make some basically this the software stack of things, because I feel like there's this long tail of okay, now that you have you know these big old like you know, pre-fill tasks, you know, so the pre-fill parts of um you know what needs to be calculated and everything like that, and then access to memory, which is like a second stage thing, and um you know, like where where grok can fit into the picture of this. So for those who aren't familiar, grok with the Q, um, it was an extremely fast inference provider. I think you still go to grok.com and and try some of their models. They're really, really fast. Um, of similar to likes like Cerebrus and some of these other providers where um the bigger focus is like the memory being really close to you know where all the inference is being done. Um but it requires much more of these units to be kind of stacked together because like it's it's you know it's pretty much almost like large chip.

SPEAKER_01 19:49

And the memory is more constrained, so you can't run as big of models, which is the interesting thing with the Nvidia chip, you know, putting it together because they're like specializing certain tasks in the pipeline, as you're mentioning, uh, to be using this this stuff. And I think it's around like processing the the KV cache and like um you know, I I'm not familiar in the deep the depths of it, so if you want to talk more on that, you can. Um, but yeah, like finding the right workflows where you can still use big models and still benefit from these speed ups. That's the big thing there.

SPEAKER_00 20:18

Yeah. Yeah. Yeah, I think there's I think there's just massive green field opportunities for these companies to um you know figure out the workflows and and the the things that need to happen for for inference in general. And so yeah, once you've trained the models, now you've got to serve them to the world and scale them up. And I think that's kind of where Nvidia's like strength is because they've been doing this for such a long time and trying to figure out how they can get that into their ecosystems. And to me, I you know, obviously, like the the top-tier customers are the ones that are gonna get these like latest infrastructure stuff, and then everyone else is gonna be rolling out. Um but I'm I'm I'm really curious to see um you know, like that was kind of like a lot of buzz that was kind of going on right now. Yeah.

SPEAKER_01 20:59

And and how does it feel like on the ground there, like talking to the folks there? Like, how are they feeling about this? Because I think like there's the whole story of like Nvidia coming in and and like obviously gonna keep pushing the the ceiling on what you can do with these with these with this hardware and what kind of models can be built. But like you're in you're in a space that we're like you know, there's all these industry folks, you know, gathering and um you know the world's changing so fast. I'm curious, like, if you've had like really particularly interesting conversations around this.

SPEAKER_00 21:25

I think a lot of folks are still trying to like it just it feels like this, like uh, I don't know how you say this, like um, it feels like a Taylor Swift concert. You know, like you you go to a concert and like everyone's excited to be there because they're there for the headline star, and it's like that's all you kind of see right now. So uh some of some of the other conversations I've had is just more of a like, you know, what what are people's strategies for how they're gonna scale it up and you know, hardware wise, and what are they gonna put in and and and what does this look like for their employees? And and I think people are still like There's various stages of discovery that I think people are just kind of having and I think that's kind of what's cool that they have like an in-person event where everyone can actually talk to each other, um, which is hard to do over Zooms and stuff like that. Yeah.

SPEAKER_02 22:12

Did you also see the uh DLSS five announcement? Eric, and you and I both spent some time at game dev.

SPEAKER_01 22:17

So Yeah, yeah.

SPEAKER_02 22:18

Yeah, that that was an interesting announcement with very polarizing opinions both ways. Some people hating it, some people loving it. Uh was there did you get to run into any of that, Ray?

SPEAKER_00 22:30

Not really. I mean, I I I I talked to some people about it, but I I feel like it's kind of the same thing. Everyone's kind of they I think they just need to see more of it, and then it's probably gonna just normalize by itself. But yeah, I would love to hear what you guys think because I think the audience would love to hear you guys' background and opinions on this.

SPEAKER_01 22:46

Yeah, well, one thing that I've been seeing more of today is like watching game devs like scrutinize the footage, and and like the thing that you notice when you do that is that there's actually like like Nvidia did a good job of keeping consistency in there, but like there's still like a lot of subtle details that are changed by this this like AI literature. So if you're not familiar with what DLSS5 is, is basically like they're saying it's like an AI-powered, like lighting re-lighting engine uh that changes lighting in your in your game so that like the frame is like much more photorealistic. And the thing the thing about it that is a little bit uncanny is that it actually ends up just changing the image in subtle ways. Like you're not actually just changing the lighting, you're actually just redrawing a frame in a different way based on some source data. And I do think like this mixture of like 3D models and and and like AI is like the way to do AI in-games and you know, real time uh going forward. But yeah, like when you're starting to change the artistic base layer, like even though you have controls, it starts to be a little strange and and it's it's it's uncharted territory. I'm curious, Adam. What what are you thinking about this?

SPEAKER_02 23:56

Same page. Some of the things I was like, dang, that looks awesome. Others were like, they changed the way the character looked drastically. That doesn't even look like the same like person that they probably designed and whiteboarded out and like had figured, you know, planned through the entire process. So I don't know. And I also I don't know if you did this, but I went frame by frame through some of the videos, and you could still see some of the classic issues with DLSS with like image breakups or the soccer ball, for example, on the bottom left. They really were focusing on the person, but the soccer ball was all distorted with like different patterns and stuff on it. So it still has some of the same problems, yeah, but I don't think people are paying attention to that. They're paying attention to like the people that are changing. Yeah, I I I agree with you. I think this is the way it's going. It's inevitable. Like we can render so much more if we use AI to process graphics and we're rendering it a smaller size and able to blow it up and do a cool thing. But I I don't think we're gonna be on it. I think it's probably gonna be another 10 years or so before we like fully crack. Um I guess some people would say we've we've cracked it now, but for me, I notice it because I've been in game dev and I I can I can tell immediately when I'm on DLSS or AMD's version of it. I'm very curious if DLS 5.5 can fool me. If it if it fools me, I'll be it'll be interesting.

SPEAKER_01 25:18

Yeah. Well I don't know, you know, it doesn't need to fool everyone. It it doesn't need to be perfect, it just needs to look good enough. Yeah. I we'll see. I mean, like we're we're starting to get into weird territory now where like you're just regenerating frames and when it starts applying to movies and and all kinds of things, like I think that's where it starts to get disruptive. Like, you know, you know, with this kind of technology, if it progresses, you know, enough, like you get to a point where you can just record your entire movie in CG, you don't need real actors anymore, and then all if it looks photoreal, like you know, why why would you bother doing real sets anymore? You know, so you guys you start to get to some crazy, crazy offshoots from this technology.

SPEAKER_00 25:58

Question for you guys. Yeah, what is your open claw strategy? This is uh from GTC. I'm just trying to figure out what your open claw strategy is.

SPEAKER_01 26:09

Yeah, it's a big, big thing. It's it's important to think about actually. It's a funny line, it's a funny thing. Um, and the lobsters are crazy, but personal agents are here. Um, they're coming. Uh, they're medium useful right now. Like you can do some things that are fun, but you can tell like all the labs are really pushing towards this. And and what is what does having an open clause strategy mean? I think it's really just like how can you make your software usable by agents? And and like that's why people are thinking about CLIs, that's why people are thinking about you know MCP. Um, how can you expose things more easily to an agent that just runs on someone's computer? Um, so I think it's an important thing to think about. I'm curious, Adam, like does your company have uh open claw strategy?

SPEAKER_02 26:49

Yeah, we definitely have not called it an open claw strategy. Um at least I haven't heard that. But I will say, you know, there there's a shift for sure to build like agentic workflows and connect agents together, which I think at the end of the day is what we're trying to do. It's the kind of the premise of open claw. Um personal agents, uh honestly, I think it's gonna be a whole new sector. There's gonna be startups popping up if they haven't already, and and it's going to be a massive space. The complexity is enormous though. When you start the things that I'm dealing with on a daily day-to-day basis to make sure that because like safety, security, all that stuff is so important to make sure you get right. Like if you're a company and you're releasing a product and you're letting agents do work on behalf of someone, like you as a company are beholden. Like you are responsible for what that agent does. Like it there's a lot of risk to it. So I'm I'm excited to see where it goes. I would say take any workflow system that you can imagine for any role from being a banker that needs to approve a commercial loan to a real estate agent who has to kind of like dial out to someone if they put a request in to go view a house. I could see all of that having personal agents drive in in the future. I think that's gonna be really interesting to see that transition happen.

SPEAKER_01 28:14

Basically talking about like, you know, the the the feeling that a lot of software engineers are feeling right now, where you know, the work of being a developer, the coding work, is automated now. And so you have to ask yourself, like, where where is the joy in the work that you you have? Like, where do you find meaning in the work that you're doing? And and in that video, it was uh by someone named Mo. Um he took he compared the work of building software now from going from like an artisanal craft into uh the work of assembling a sausage, you know, like you don't really think about what goes into the sausage, you just kind of deliver the sausage and you know you can you can you can ship it, you can sell it, but like, you know, what are you building now and you know finding that meaning there? So I'm curious if like there's been uh a drop in the the the feeling of satisfaction you get from working on software since AI has come around. And and I know Adam, you've been like an artisanal crafter before and been heavy in agents for a while now, but I'm curious if you've stopped to think about like that feeling since then.

SPEAKER_02 29:16

I mean that video was awesome. I appreciate him putting it out and got a lot of people talking. Um I think it really depends on where you are personally. So I would say for the first like five to seven years of my career, I l I was right where Mo was. Like I loved writing the code. I loved like every line, you know, was just part of like my being to the point. And this is gonna show some immaturity as a young engineer, where if I found out something I built was gonna get thrown away for whatever reason, you know, wasn't it it didn't find product market fit, like it hurt me. Like it was like internally soul crushing because I had worked so hard. I like poured everything into it. I got maybe I got burnt out so much from that that really what I started to actually like is his sausage example. I could care less how the sausage is made, but I want the best sausage possible. I want people to come around the world to actually eat my sausage. Like I care about like how I design that sausage at a high level. You know what I mean? Like that that's the thing that I care about more than anything. And and that's where I ended my career. Like, I don't want to make UI, like I don't want to spend time hand coding another API that I've done a million times before, or add a button to a screen, or whatever it ends up being, all the different systems you work on, with like 80% of it being monotonous a lot of times. What I care about is the outcome. Like, how do we get to the point where we have software that people use? It does the job better than anyone else in the market, and that's what I find joy in. And I love the systems, I love the big system planning and thinking through it. So I don't resonate with Mo, but I could see myself ten years ago being in that exact same spot, just to be totally honest.

SPEAKER_01 31:05

Yeah, and Ray, have you felt anything similar like that? I know you went to being a manager early on, so like I know that the the the decision matrix of how you think about building and products has has been a little bit more detached from the lines of code for a while now, but I'm curious if you feel feel this shift yourself.

SPEAKER_00 31:23

I I could totally understand it, and I think I look at other people in the industry too who have like are starting to see this in waves. I feel like there's a wave like four months ago or early on that people were kind of saying this as well, and and then being burnt out, and then they kind of just stop. And then they just kind of embrace the thing and realize like this is a tool that they could use and so forth. And yeah, I I think being on the product side so heavily, uh, but also being kind of in the weeds to having like been in the trenches and and actually see teams like like like what Adam was saying, like being on the other side of it's like, okay, we actually have to surgically remove your code from this release because we're gonna announce this thing like next week and it's gotta be gone. And like it is very painful to like work with teams that way. Um and and then like you know, hide things or do whatever you need to do to get stuff out. Um because it just isn't ready for the release. And and that's just kind of what happens. And I I think about that like emotional feeling because it's it's there's like a craftsman that there's like this effort that you put in as a human that and I don't know if it's more like male oriented, because as a male, I feel like I'm a provider, and so I do this stuff, and this is how I get recognition, and this is also how I make money and provide for my family. So if that that underpinning is like going away from me, then like what is what is it that I'm providing for my family? Like, you know, there's like this deep emotional core that I don't think was stuffed, like that I feel like is coming out. It's like if that uh identity gets removed, then what do you do next? Right. And if this is your identity, you know, that you've spent so much time crafting, then you know, it it it it's um yeah, it's totally like an you know an existential crisis, and it's not it's like people would just say, Oh yeah, just go draw art or something or do whatever. It's just like I don't think you understand, you know. I I I I I kind of get that, you know, there there's that craftsmanship and uh it's to me like uh being on the other side of it too, it's like okay, now that I've zoomed out a lot more, you know, how how can I like keep seeing these little layers that keep growing and growing and growing and and um you know become maybe uh a different type of person or how can I how can I grow from this? And I think if anyone's probably going through this right now, it's probably a good idea to just ask yourself and just kind of step back a little bit and see, you know, just take a look at the realities and say, you know, what what is actually really happening in front of me? If this is actually true, then where can I provide more value and where can I bring my expertise into something uh to start kind of driving value and and start to play? Uh, you know, because if you think about this framing in terms of like what's being removed from you, then you're gonna just only focus on that. Um if you start to, you know, once you kind of get over the first stages of like this reality is actually shifting from underneath me, maybe it can kind of get you thinking in a very different direction about your life. Um and you know, I think when you spend more time in the positive thinking, I feel that can get you to be in in a much more productive space. But um, I think that's it's it's a very real emotion, I feel like. Um, you know, if you have those emotions, write it down, let it out, you know, make sure you talk to someone. Yeah.

SPEAKER_01 34:36

Yeah, I I want to expand on this a little bit because you know, like there's an Andrew Yang article that I just came out today as of this recording, and it was talking about you know the potential job displacement coming in, and he's talking about like maybe 20% of like white-collar workers getting pushed out because a lot of work is getting automated. Because you know, it's you know what we're experiencing is you know, software engineers is like, you know, the code is getting automated, you know, we're still finding work to do to get this code built, and there's more code to build. So the you know, we feel like there's still there's still a role for us here, it's just a different role. But like there's a lot of work out there where you know you're sitting in an office and a lot of what you're doing is is quite manual, but you you know, you're able to provide for your family, and this work is is you know sustaining for you, and you know, there's a lot there. Um, but as agents come for different jobs and different you know tasks in this this area, like we have to ask ourselves, like, you know, okay, well, like at one point, like there probably will be some job displacement. Um you know, we're not quite there yet because like the AIs are still falluable and you know it's still LLMs, and LLMs are not perfect, and they have limited memory, and you know, but things are changing fast. You know, we got more hardware with bigger context windows and you know, more things to think about there. Um, but I I'm curious, you know, like we had a chat earlier before the recording here about this, and you know, there's different ways to look at it, and you know, and and I'm curious, like I I want you to just dive in a little bit, Ray, on like that other side there, because you know, you you had some good points uh earlier, kind of kind of pushing against this this view of the shift to the economy and to the jobs that you know we're potentially seeing here. Do you want to just expand your your thoughts there?

SPEAKER_00 36:12

Yeah, if you can kind of refresh me again, I kind of like that for some.

SPEAKER_01 36:16

Well, yeah, I I mean listen, like I I think you know the the data right now like is showing currently that like there's more work for software engineers. There's not enough data showing that like that that these jobs are getting fully displaced. Um, but you know, I think like it's still so early and the automation is only starting to kind of touch the economy, so it's really hard to say for sure, like to even have data to even point to to kind of counteract this claim.

SPEAKER_00 36:43

So yeah, oh that's right. Yeah, I think it's just about adoption for AI in these companies and whether it's actually displacing them or not. And there's just a lot of uh to be honest, uh, I think a lot more people are complaining that like what Adam was saying is that they have more code to review now. Uh now that there's more AI generating everything, um, there's too many features on the plate. There's too many bug requests, there's too many, you know, very detailed specs of you know, bugs and everything. So it's like it's almost like getting a magnifying glass and just kind of magnifying everything. And it's kind of funny because now it's like having more money and more problems, you know, like that saying, it's like, oh, you're gonna magnify everything in your life that you really never kind of got really clear on. Um and and I think uh I think we're probably overhired for the pandemic or before the AI race really just got defined and and really kind of got into fruition. And you know, the reaction now is like, okay, now that we have this extra people and we don't really need them, and we see that AI is coming in in the horizon, the people who are being laid off are not necessarily being laid off for AI reasons per se, but it's just or they're using that as AI reason, but it's not like physically actually happening. It's I think they see it's going to happen, and um it's interesting from the block CEO um Jack Dorsey. Is it Jack Dorsey, right? Yes.

SPEAKER_03 38:02

Okay, yeah.

SPEAKER_00 38:03

Basically just making the opinion saying, like, okay, either I let everyone off, you know, a at a smaller cadence, and eventually 40% of y'all will be gone, or just we'll just do it all at once, right? And so if you kind of start applying that same lens everywhere else, you know, across these bigger companies, um there's a lot that's coming in the pipeline, and um it's probably not a bad time to actually start getting into these systems because by the time they do get realized by these companies, it could be five years until they bring these types of you know open claw-like things that are more defined. If you're if you're defining those types of features and leading them up independently, you know, you can you can drive an organization inside of a company that's super valuable. Um or just your own independent company, right?

SPEAKER_01 38:47

Yeah, yeah. Well, hold on, you have something to say, Adam, and I want to hear what you're what's on your mind there.

SPEAKER_02 38:51

Yeah, you made a lot of really good points here, and I think there's a lot of things we could peel back. Um I I tend to agree with you that a lot of the things we're seeing are I would say marketing or uh kind of like positioning in the market, rather, about being AI and being able to say, hey, I can let go 40% of the group uh because AI. I tend to think what's probably true is that the company has overhired and needed to lose some weight. And I I still feel bad for everyone involved, but I tend to think there's a disconnect between the message that's actually being shown to the public and then the reality. If you were to go talk to the employees on the ground, they would tell you a very different story of how much AI is actually speeding up things. There's a lot of complexity to it. We all know, everybody that's listening to this knows the complexity that we have to deal with. But that doesn't mean that there are other roles outside of software engineering that aren't going to be highly impacted. And I I've already seen that. You look at any sort of writing or technical writing role, those route, those jobs are being highly impacted. And that is going to be true throughout every because it it's such a like a very easy thing to immediately get agencies. I used to own an agency that would build websites for people. And I think it was uh Andrew Yang's article that actually talked about this, where I would charge, you know, you charge thousands of dollars to build someone a website, that's non-existent now. You pay $20 subscription, you can get a website up and running. Granted, you don't get the support and all the kind of the stuff that goes around it. So you have to kind of like think about the up-leveling of what you want to do as a business and where you can go with AI and not be opposed to it. And part of like the entire journey that we're gonna go through is we're going to have all of the I think about back when the internet came around. Because I was around during that time. I can't I think about when crypto came around, I think about all these like big technological technological shifts. Ultimately, at the end of the day, nothing really changes. We just figure out new ways to actually utilize people to do more things. So I'd like to be a little bit more on the optimistic slant that for the people that want to up level themselves, like the sky's the limit. You're using AI, you're figuring out how to do things, I think you'll have a role for the foreseeable future. But there are engineers I know that are have a open to work on LinkedIn today that are very opposed to using AI for whatever reason, they are going to have a very hard time. And it's going to be hard for them to find a role or get back in the industry. Anyway, Eric, I hand it back to you.

SPEAKER_01 41:37

Yeah, I mean, I I think a lot of our perspective is very colored on the software side because that's where we're experiencing it. Um, I just I I I I think there's just too many people in different roles, like all over the place, you know, doing doing small tasks or spending a lot of their day not doing too much. Um and I I you know whether like whether we want them to be up leveled or not, like there's a lot of folks who just don't have the ability to be up-leveled to this degree. And I like there will be job displacement, I'm I'm quite certain. If only because it takes less people with AI to do the same task as you did before without AI. Now there might be more work to do, but I don't think that's like universally true. Um, you know, I think that like if you're building software, there's just an infinite number of things to build, but like at a certain point, like you have to think like, is there enough? You know, you can't just add infinite features, you know, even to any software. Like there's gonna be a diminishing returns on what you're doing, and at one point your users don't even understand everything that your thing does anymore. So, like, you know, like the there's a ceiling to the returns on adding more stuff. Um, and I think we'll start to have tougher conversations in the next few years. Um, but as it stands right now, like AI still needs a lot of oversight, and you know, as smart as the smartest models are, like they're not fully autonomous, they're not able to do everything, they're not able to make decisions, they just don't have enough context. Um, but I don't know where that goes in five years, like it's hard to say. Um, and I think it is important to have these difficult conversations sooner rather than later. Um just to go back to GTC as well, like one thing that was interesting is um I was listening to um the interview with Jensen Wang and um Ben Thompson from Stratekery, and he was going on about you know having how labs are being kind of like overly dumeristic at the moment, and how they're like kind of touting capabilities and fear-mongering around what the models will do in the future. And you can tell like Jensen doesn't really feel that like he's not like as AGI pilled as you would say like Dario is, who's the CEO of Anthropic. Um, and I think it's interesting to see the the dynamic there. So, like the folks building these models are sitting here, like they they they feel like they've like unleashed a god onto the world. It's not quite there yet, but they can see the god is coming and they're they're very afraid. For what happens when the god hits the streets. And then you have Jensen who's like, no, these are like machines and and like you know they're they're do certain things, but they're not gods, you know, and you know, like he feels like there's just like a big gap in the expectation of what these things are. So I I don't think you can kind of point to anyone's opinion right now as to what will or won't happen and take that as definitive. Um but I you know, like all like I think at the end of the day, the the thing that's most scary is the uncertainty is that we don't know you know what's gonna happen in five years. Like the world used to be so much more predictable. You can kind of plan your life out, you know, 10 years out, 20 years out, and that's just no longer the case. Um, and I think you know that's scary for a lot of people, a lot of people, and yeah, for everyone. Like, we just like predictability in our lives, and that's gone.

SPEAKER_00 44:47

Well, so this is like March, middle of March 2026. And open claw was just like sort of mainstream in like February. Yeah. End of end of January, beginning of February, where they just kind of start to hit this explosion, surpassing all these stars. I mean, what w uh with what you see that people are kind of grappling with right now, I mean, what does the next month look like for you guys? I mean, just right around the corner.

SPEAKER_02 45:12

Can well, let me let me ask you something, Ray. How many jobs do you think OpenAI is hiring for right now? Like if you had to if you were to go to their careers page, how many open positions do you think they have?

SPEAKER_00 45:22

I would think they have quite a bit, right?

SPEAKER_02 45:25

Like, give me a number that you think roughly.

SPEAKER_00 45:27

Like hundreds, I think. Yeah. Yeah.

SPEAKER_02 45:29

Yeah. So the way I've been thinking about it, they have 630 open. So the way I think about it is I'm monitoring the companies that are should be the leading indicators of when things are gonna go poorly.

unknown 45:42

Yeah.

SPEAKER_02 45:43

You know how like when you're when I first started flying, like one of the things someone told me is like always keep an eye on the flight attendant. You know, you hit some turbulence or whatever. If they're calm and they're fine, like you're fine. Don't panic. Don't you know? So it's like open AI anthropic, when they stop having job opening, they start laying off, they start moving through things a lot faster than they are now. That's like a canary in to my in my mind, like in software. You know, it's you see the things starting to happen that are gonna trickle down everywhere else. We don't see it at the companies that should be the most primed to take advantage of it. We're definitely not seeing it in the other companies. We're seeing efficiency, by the way. We are getting faster, we're able to do more. But there's also a lot of new things we now need to build because of AI that didn't exist two years ago. So there's some really cool opportunities we have ahead of us. So anyway, my current state is I'm monitoring the canaries. I'm gonna keep going head down, building what I need to build, and just enjoying the ride while I while I'm you know blessed to be a part of it, honestly. As long as people want to keep me around to work on it, I'm gonna keep doing it.

SPEAKER_01 46:44

Yeah. I I think the the fear though with this canary thing is that like you have the labs saying that they are gonna hit this thing in the next year. Um, I think OpenAI is different. OpenAI is a lot less like, oh, we're gonna replace everyone. They've been a lot more upfront about working with businesses and trying to grow the pie. But Anthropic is very pilled on this AGI thing, and they're very pilled on just the idea that like they won't need the sweet job anymore, they won't need a lot of this this work that these people they're hiring for, you know, are doing. And and you know, if you listen to them, they're saying, like, well, look, we raised all this money. Of course, you know, while this is still true that these humans have value, we're gonna put them to work and accelerate to get there. Um, because that's gonna give them get them there closer and faster. But like, if in the next year that changes, you know, it's a big difference. Like, I don't personally see this happening in the next year, like so quickly. Like, I think there's just always like ways to evolve the work that a human's doing, and and LLMs are just not a human replacement. They like if they're incrementally smarter, incrementally more capable of processing context, like they can do specific tasks better, but those tasks still don't fully overlap with what a person does. Um, you know, human human learning and human adaptability is just hasn't been reproduced by machines. And and like, you know, I think it's kind of diminutive to think that you know, within a year that you know, human experience is completely obsoleted. Like it just feels like you know, yeah, there's just like a problem with your view of humanity at that point, I think.

SPEAKER_02 48:24

Um two quick add-ons to that, because I think you're right, Eric. It's like marketing, the marketing that's happening does not like fit reality to me. And this has been true pretty much for every major like cycle that I've been through. So it's like I think at some point those are gonna converge together, and we're gonna figure out what is reality, and we're gonna realize the marketing was overhyped. I could be wrong. I've been wrong before. Like it could be the reality, I just don't see it yet. And the second thing is I really am not a big fan of going down the rabbit hole of like YouTube Shorts, but every once in a while I get I go down the rabbit hole of YouTube Shorts specifically around like dumb things AI does. And there's a there's a YouTuber on there, uh, Father Phi, I think. I probably pronounced it wrong. Anyway, he'll post something where he asks like questions to Chat GPT or Grok or whoever, and he asks one, he's like, Hey, I need to get my car washed. It is right across the street, let's say maybe you know, a quarter of a mile away. Should I walk there or drive there? And the AI said, You should walk there. You know, it's a sunny day or whatever, you need to go out there. And then he's like, Hey, I got to the car wash, uh, but uh my car is back where I started. I didn't, you know, so it's like the logic of these things still are very flawed. Like, you know what I mean? Like it it they are they'll gaslight you a little bit, they're trying to figure out the answer that you want to hear. Now there is reasoning that's starting to get better, but like a human would never make that mistake. Uh human would be like, of course you're gonna drive your car, you want to wash your car. Like you're not walking to the car wash to wash yourself. So I think there's a there's a lot of work to be done, and there's a lot of things AI is really good at, but you still need human in the loop. So I think there's a lot of like human intelligence, you know, human capabilities that we're gonna use that that AI will augment us. But I just in my lifetime, I don't see it fully replacing us on the engineering side.

SPEAKER_00 50:16

To that note, I would love to ask you about skills because that's become super mainstream now, and I feel like that's a big part of the open claw adapting to the environments that the LLMs don't have. Um what's your been your experience there and and has that been useful for your businesses and people that you've been working with as well?

SPEAKER_02 50:34

100%. I mean, it is a great way to manage context. I was actually a little bit more skeptical of them when they first came out. Anything new I have to kind of like feel out. I think it's actually a great architecture to build around. I think it allows you to it allows you to not pollute context with all the things that your AI needs to be capable of, and it can go find the instructions to do it, along with the details around the tools and the workflow that needs to follow. I think it's absolutely fantastic, and I actually see some evolutions of that coming forward that will probably even make that better. And one of the things I've been thinking about is a lot of these um that all the tools are still a lot of times will be registered or will still be in the the context. I still think there's probably a future where we actually are doing like tool pruning or dynamic, like tool selecting based on having an intent. You figure out the intent of the job, you pull in the skill that needs to be done, you pull in a series of tools that need to be done, and then you can have a much broader agent with better context management. Still purely theoretical, but I believe that's the way things are gonna start evolving.

SPEAKER_01 51:40

Well, if you if you take that to what you're doing right now with with like sub agents and and skills, like you can just spin off a sub agent with the skill that that pulls in the tools it needs with tool search, and you know, Cloud Code can do pretty much that right now, you know. So you have like a purpose-designed agent to do that work.

SPEAKER_02 51:58

Yeah, you lose the conversational aspect of it, though. That'd be the that's the piece that I struggle with. I think it's like you still want the the conversational aspect with it. So you almost want the like the main chat to morph into the thing that I'm trying to solve, you know what I mean? So I can conversate back and forth with it. Sub agents lose that sort of like if I want to accomplish a particular job, I want to brainstorm with the AI, then I want to go actually generate some ideas, then I want to go make the thing for me, then I want it to sync it to whatever system I need to sync it to. And it usually isn't like what I would say is like an autonomous agent, usually more of like a workflow driving thing.

SPEAKER_01 52:37

You you should play around with the Pi agent. Um, you know, there's there's a lot of um there's a lot of interesting workflows. So the Pi agent, you know, I think we talked about it on a previous episode, but it's just like a super hackable agent harness. I should try that. Yeah, and and the thing is like there's there's really interesting plugins. So there's this guy named Nico, um, who's been you know such a prolific extension developer for Pi. And um one thing he was experimenting with was you have the agent kind of go off on it, like go down, go down a road, um, and then you can have it kind of send a message back to itself in the past and prune out all the conversation history um that that was kind of carried out since then. Because you know, Pi has this tree architecture for its context window, so you can kind of branch off in a conversation into a branch and and kind of explore some road, and then you can kind of bring back the context back to the root and and and kind of like that's so cool, throw away all the stuff that you you did, but keep a summary of it. Um you're kind of able to go down this rabbit hole. So there's a lot of experimenting to do with this kind of workflow, and you know, he's also got like um like basically a whole instant messaging system, so you can spin off sub agents and have a chat log between them, and you can watch them kind of talk to each other and you can participate in the chat too. Um, and then you just it's there's all kinds of ways to work with models now.

SPEAKER_02 53:54

You want to talk about my dream job is just experimenting with that stuff all day long.

SPEAKER_00 53:59

Yeah, yeah, yeah. Hey, yeah, you should become an introducer, Adam. I think that uh that's in your future.

SPEAKER_01 54:04

Just do all that. Oh man. There's a lot to build, a lot to play with, and I think we're just scratching the surface right now because you know, like the way that we design these harnesses, there's like there's like two vectors, right? There's like, you know, one is like maximizing the productivity and the intelligence of the model, and the other one is like maximizing the ergonomics for interacting with the model. Um, and so there's a lot to play around there, a lot to build, and you know, we haven't seen the final form yet. Um, so hopefully we know we kind of lend like push ourselves a little bit more into the ergonomic side of it and and really try to empower people to build a lot more and do a lot more and keep in the loop and understand more of what's happening. So there's a lot to build there. Um anyway, I think this uh this is an interesting you know little digression there. I think there's a lot to explore. The world's changing fast, and like I yeah, I think we just have to like stay involved, stay in the loop, and uh avoid avoid just kind of throwing our hands up and saying we're useless now. Um, because I think that's that's where you know that's where we're gonna get lost if we do. Um so it was a good time. Thanks uh for getting together, guys. And uh thanks for tuning in. If you appreciated this episode, don't forget to subscribe and to give it a like and share it with your friends and let us know what you think. Um want to hear more from all of you.

SPEAKER_00 55:13

All right, we are aiming to be the number one AI podcast. So make sure you give us five stars. Five.

SPEAKER_02 55:19

We are the number one AI podcast, Ray. We are the oh yeah.

SPEAKER_00 55:23

If we ask Claude Opus, yeah, we should definitely like give ourselves our own gold medal and stuff. Yeah, you're absolutely right.

SPEAKER_02 55:28

That's our new open claw plan is we're gonna have like it go spin off a bunch of stuff to make us the number one AI podcast. There you go. Go listen, yeah.

SPEAKER_00 55:36

That's a good idea. Yeah.

SPEAKER_03 55:39

All right, gentlemen. Until next time. See ya.

SPEAKER_00 55:42

Till next time.