Viktor
00:00:00.000
And that's that panic mode. and you should be panicking because you think it's fine until it's not. And when it's not, then it's too late, especially for larger companies that cannot turn around overnight. So panic because if you don't get there first, you will be in trouble. I dunno whether that's next week or next year. And you'll be in trouble in terms that you cannot recuperate from that trouble.
Darin
00:01:30.772
Something shifted back in February, OpenClaw came out and we started to see the normalization, for lack of a better word, of what agents can do. Heck, even Linux Foundation has created a whole sub foundation around agents.
Darin
00:01:48.652
Where do you think this thing has headed? I mean, to me, chatbots are gone now and it's all about agents.
Viktor
00:01:55.434
This is going towards one person or a very small number of people, companies. this smells trouble for software industry on many different levels. On many different levels. I think that this year and the upcoming years, we will see software companies being in deep trouble. I'm not saying they will all disappear, but is going to be complicated, right? For existing businesses.
Darin
00:02:24.580
Well, again, in this shift from chat bot to agent, like I still use a chat bot for some things. Like I'm trying to create a new recipe for something to cook for lunch, perfect for chat and just a long thread and, but anything beyond that. I am going to start up a new GitHub repository. I am starting to lay in a couple of skills and off we go.
Viktor
00:02:45.712
Yeah. I mean, you know. Depends if, let's say that, okay, so let me extend, and this is science fiction, hypothetical, future type of your example, right? And you said you're looking for a recipe right now. Let's say that part of that query would be, and make it with the ingredients that I have in my fridge. Then that would stop being a chatbot. It would be an agent. Hypothetical agent connected to your fridge, checking what's inside, and then going back to browse internet to find your recipe based on what you have. Right. And by the way, if anybody does this, I was the first with the idea. I want 10%.
Darin
00:03:32.959
you, have some of these refrigerators today that already keep track of everything, but you know, okay, number one, I don't want my refrigerator connected to the internet. That, that just, that's a whole different conversation. I agree with you. I think those are the next big waves of things to happen. But if I think back a little bit. Just over a year, or let's say before January, February, all we had was chatbots, more or less. I mean, and people were building businesses around those chatbots, which are starting to be like, eh, okay, so what?
Viktor
00:04:18.205
Too many. It was similar like, and it's still in case of Kubernetes. Everything is K something. K is K. That back in the day, it was uh, something GPT. That's when people thought that actually the only model will be Judge G pt.
Darin
00:04:33.613
Have you had this shift yet? Have you moved from the shift of I want AI to figure something out for me to, I want AI to take actions for me.
Viktor
00:04:41.993
I mean, when I do coding, it's taking actions for me all the time. It's writing code, that's action, right? It's executing something. That section. It's all actions.
Darin
00:04:51.831
But that's still been pretty recent for you, like the last six months? Recent a year. Okay.
Viktor
00:04:59.530
Yeah, so, but you know, a year with today's pace, that's actually a long time. It's a very long time. I could say that. Imagine when you, when you can say, oh, it is been a year and slightly over. That means that you're one of the first adopters. It's a long time.
Darin
00:05:22.149
yeah, because this isn't a 2026 invention. It's been going on for a while, but now that it's getting adopted at enterprise level and becoming, again, normalized in society for better or for worse, especially like with OpenClaw. Let's use a story of OpenClaw. You did try out OpenClaw. I don't know if you're, are you still using it as
Viktor
00:05:45.478
I'm not using it. Not good. I mean, so OpenClaw. Is the mo is the way how, where we are going. I have no doubt about that. That's the future. So when you look at OpenClaw, you're looking at the future. I don't think that necessarily OpenClaw is in this form, is that future, but it is definitely showing the direction. It's almost like meso for scheduling, right? It showed the path. It. Made us all accustomed somehow. Not all, but some of us accustomed to those concepts and stuff like that. And then it took a couple of other iterations. We got swarm. We got Nomad, we got Kubernetes and tap, prove what not. And then eventually we settled on, on, on a solution, which in this case, Kubernetes right now, in case of ai, I don't think that there will be one solution, but it's definitely not going to be OpenClaw in its current form. But it's amazing. It's amazing as a window into the future.
Darin
00:06:47.199
Well, one of the things that Anthropic did back either in late February or early March, was added in remote control, which is a variation of what OpenClaw gives. It's a very small part, but that actually has made it possible for me taking the dog out for a walk and I'm watching something as it's going through.
Viktor
00:07:06.559
There is another feature they added recently, I think weeks ago, which to you as a software engineer will sound ridiculous because you can do it in five minutes, but it's a very important feature, and that's a schedule in, uh, cowork, I think. Which is essentially a cron job just to be a hundred percent transparent and clear. That's what it is. but that's also what OpenClaw is. OpenClaw is a, is a cron job. It's a way value for you to, it's a way for you to create cron jobs that will run agents with the prompts you give it that I just told you, the whole infrastructure of op claw, that's what it is. And that's what we need. That when you think about it kind of, oh, tell me about my day every, every day. Morning, eight o'clock. Well, what is that? That's a cron job. Uh, the executes an agent with a specific prompt and connected with specific, cps. when you, look at it from engineering perspective, you say, actually, this is not the rocket science. This is not anything special. but still we had to get there. Here, right? Somehow.
Darin
00:08:14.968
getting there. Let's rewind back in December of 25, that's when the Linx Foundation created the. A Agentic AI foundation. It had, lemme look at my notes just to make sure. It was founded by anthropic Open AI and block with platinum members of AWS, Google, Microsoft, Bloomberg, and CloudFlare. These are all heavy hitters.
Darin
00:08:40.913
three anchor projects, Anthropics, MCP, blocks, goose, and then open AI's Agent md.
Viktor
00:08:47.916
Okay. I, I agreed with your first part of this. These are heavy hitters. the second part, the projects are, eh,
Viktor
00:08:57.061
I mean, kind of what is Agent md? It's a, it's a, spec. It's a kind of spec. It's a spec that says if you write markdown file, then uh, agent will read it. That I, I just gave you, gave you the whole project. That's what it is, MCP spec, that's very important that it is in a foundation, but let's face it, it's not nobody's loss. what I'm really trying to say with all this is that none of them contributed anything of true value and which understand those are businesses, they're not gonna give you true value kind of if Anthropic. So here's what I would consider a huge win. If Anthropic gave Claude Code to the foundation, I would say, okay, now we are talking. Now we are talking really. Instead, we got Goose and no offense to great, but that's the only real project there.
Darin
00:09:54.519
And for the. At the time we're recording, the layoffs recently just happened to where they basically cut it in half. that's not really surprising because Jack Dorsey typically over hires, which is fine. I mean, you sometimes you have to over hire to figure out where you need to be and it, it, stinks. But that's business. But then also anthropic added in the agent skills. not to the foundation yet, but at least it's collaborating with everybody on it and people are starting to pick up and use skills.
Viktor
00:10:27.812
Everybody. I think that almost all agents are using it now, which is absolutely fantastic and shows how fast we're moving. And the only beef I have about skills and a few other things is the, uh, is the company's originator of those. Anthropic because, so everybody started with skills Anthropic says, oh yeah, skills should be dot plot slash skills. And uh, cursor says, oh, skills should be in dot cursor slash something something. And I think that most of them eventually agree that it should be dot skills or something like that. Directory. And off you go except Anthropic. So now I still need to have links. Same thing with the Claude.md, right? You have it? I have it. You know what everybody else uses? Skills md. Sorry? agent.md. Right? And it's just a file name. Kind of like who cares? So Anthropic, if you're listening, please.
Darin
00:11:26.800
just please. Yeah, we'll leave it there please. We think about the security issues that OpenClaw really brought to the forefront. I mean, it got a crazy number of GitHub stars. Not that that matters, but you know, it's usually a smell of something. Some people say it's a smell of garbage, but we'll leave that there. But it exposed that, oh, we got all these things running, and then all of a sudden, and we've talked to people about this, it's like, would you allow your people within your company to run OpenClaw on your. Corporate machines and the vast majority of people, in fact, every person has said no way they can run it on their own.
Darin
00:12:07.827
But you know, it's happening. So we have these security issues happening within companies. I mean, ignoring the OpenClaw thing for a second. Companies are trying to figure out how to use agents within their infrastructure,
Viktor
00:12:27.441
That's Anthropic is actually winning. they're really taking those things seriously. Not great, not perfect. Seriously. They're the choice for majority of enterprises in the Western world, and I suspect that as the prices hopefully start dropping, that's why also we will see increased usage of self-managed inference. And that's where actually China comes in. For good or bad.
Darin
00:12:56.063
Because the models are cheaper and distilled, for lack of a better term, they're easier to run.
Viktor
00:13:02.112
Not only that, but because realistically, if you're going to self-host models and you want good ones, what are your options?
Viktor
00:13:22.487
Yeah. Quinn Kimi, right? But from us, the only open weight model that I'm aware of that is basically not moving going anywhere anymore is the one from, what's the name? The one from Facebook, the Lama. That's the only one. So basically you're choosing, when we talk about security, do you trust the company and I, you know, who's going to be the winner in this case? Actually, AWS Microsoft and similar, I,
Viktor
00:13:52.375
they're hosting the models and you trust them already as much as you can trust, a provider, right?
Darin
00:13:59.314
think that's where, and we've talked about this before, I think that's where Azure and AWS have really done the right thing, right thing maybe too strong of a term to where they're just going to add support. They're not gonna go and try to create their own foundation models. Sure, they're investing in everybody, but you know that's not gonna be their core business. Their core business is going to be inference.
Viktor
00:14:22.277
and I think that. Actually the big three are in a very different position and very different strategy. AWS feels like they're saying, okay, we are not going to compete on models. We are just going to compete on getting money from running inference from all the models. That's our business, right? Microsoft is, is doing some silly things with copilot. Nobody understands what they're doing. And when I say copilot, I don't mean GitHub. Now, I mean, all the things that you cannot, click, uh, anything in Windows anymore without triggering copilot, and nobody wants it and all the jazz, right. And it's 365 copilot. I, I feel that they're lost somehow. And Google man, they know what they're doing. They're basically the top. On every single level. Their TPUs are amazing. They have models that are at the very top and they give you inference, if you need, I mean, when I say inference, I mean like part of Google Cloud, offering outside of just APIs for Right. If I would have money to put, that's where my money is going.
Darin
00:15:46.404
probably that I thought you would probably put a little more weight behind Anthropic versus Google.
Viktor
00:15:52.705
oh, no. Tropic is the, is the, is the, is the kind of, I'm not even mentioning it because I think that that's. that is a company that is on a way to become the next, top company, whatever, whatever you call the company that has the biggest projected revenue, what's not right. They're just, uh, when they go IPO and they will probably go before the end of this year, is going to be a gold rush. again, if I would have money, uh, I would invest in a tropic right now, this late in the game, I would still put money into it that are going to do great. And if I had any money into open ai, I would take it out.
Darin
00:16:31.018
That's a big jump right there. We'll let let that sit there as somebody that works with infrastructure. On a daily basis, the open source project you work around is about infrastructure.
Darin
00:16:44.149
Working and having an agent being able to make changes to my infrastructure is a heck of a lot different than a chat bot that just hallucinates things. How are you trying to sort of, okay. I, at least with the chat bot, I can get answers and I can go do the thing. Whereas the agent, it's got the credentials, it's got, uh, it's got everything. What do we do?
Viktor
00:17:09.578
let me assume for a second that when you say Eng TD context, you mean you built, can I make that assumption?
Viktor
00:17:17.618
Okay. So. It's an agent that you build and that agent is doing agent loop, meaning that it goes in the loop between agent executing tools that LM requests back and forth, back and forth, and then it does something and then it responds right during control, which tools you give it. You can only execute the equivalent of kubectl. Get and describe. You have, there is no option to, I'm not giving you any means, to execute kubectl, delete. If that's what you want,
Darin
00:17:49.884
But I've also given, going back to Anthropic, I've given Anthropics permissions of don't run these,
Viktor
00:18:01.221
I can never experience that. I mean, does it without asking you or does it kind of, when you say, yeah. Yeah.
Darin
00:18:08.567
Oh no, I, but in case you get, once you get used to answering Yes. So many times you just see it pop up and you just hit yes. Again, it's like, oh no. What did I just do?
Viktor
00:18:17.408
So that's very different from the agent I mentioned before. So you may right now what you have, let's say with Claude Code. I'm assuming that that's what you meant with when you said the Tropic, right? What are you giving Claude Code access to B essentially. And then there is some rules that kind of like, you can execute this within a bus or you can execute that, but basically, essentially you're giving it permissions one way or another to execute shell commands. Correct.
Viktor
00:18:45.833
Uh, that's very wide. That's very, very wide. Now I imagine that you say no B for you, man. Just those tools that I give you. And when I say tools, I don't mean kubectl. I mean literally there is a tool that I created and I call it. Cube get and cube describe and that's what you can, you can use, So, when you use existing agent like Claude Code, you're defining permissions. you're giving it posh access and then you have some file where you define what it can do and what automatically, what it might have must ask you to do. Right. but when you're building your own agent, you're deciding, right. Okay. So these are the tools. Yellow Lamb can execute period, and we will see more companies building their own agents. I can guarantee that.
Darin
00:19:35.385
Well, the companies will, you're talking about companies to sell the agents or companies building agents for themselves to
Viktor
00:19:43.223
I think that we will see simply companies building more than they were doing before, buying less.
Darin
00:19:49.036
That's a whole different episode to talk about. We won't do that today. we've talked about oasp in the past. The Oasp top 10 that's been out for. How long? It feels like decades now. They've now published an oasp, agentic, AI Top 10, and a couple of them make sense, agent goal hijacking, tool misuse, and exploitation. Identity and privilege. Abuse. There we go. But the last one is rogue agents, or at least that's not the last one, but that's just one in the list. compromise or misaligned agents, diverging from intended behavior. All of these sort of wrap up to one thing and we want to give agents least agency, but to me that just sounds like. Lease privilege all over again. Are we just recreating the wheel with agents yet again? It feels like we are. Because if you gotta onboard an agent, just like you have to onboard a human, you have to train an agent, just like you have to train a human. Uh, why would we expect, we expect our humans to have least privilege? Wouldn't we expect our agents to have least privilege as well?
Viktor
00:20:55.505
think that the good starting point is whatever you give people, you give agents, and I dare to kind of, what would be the reason not to do that? And if you say halluc, they hallucinate, uh, I would remind you that we hallucinate just as much
Viktor
00:21:13.033
or more. Yeah. So If the argument is I'm not going to use agents because they hallucinate, well, you know what? You should fire everybody as well, whomever is saying that that happens to be the only person that doesn't hallucinate. That's the person that always knows the best.
Darin
00:21:29.387
but how often do they work? They only work during daytime hours and they're never on call.
Darin
00:21:40.010
it's not surprising. Well, that sort of leads into what should we be doing? Where are agents good today? I think maybe, you know, summarization, you have a failed, let's start our day-to-day stuff. You know, you have A-C-I-C-D pipeline fail. Have it give you a quick analysis of what it is. This seems like a, a safe starting point. Read only.
Viktor
00:22:04.382
I think it's, it's good that almost everything we do today up to some point. And what we do is always done through a supervision, right? Like people, if you work in a company, you, you're rarely, you're most likely not unsupervised, right? There is a product manager, there is a project manager. There is a right. I think that is the same.
Darin
00:22:31.526
So you're saying we need to duplicate everything we've done for humans, for agents.
Viktor
00:22:35.708
I think it's a good starting point. Not necessarily the end point, but it's a good starting point, right? No. So I'm promoting you to be the tech lead now, and you are what you were doing before you get an agent or you build your own agent. So ba basically everybody gets promoted.
Darin
00:22:54.916
And then once we have every, all the humans promoted then and they built out their own agents. This is a, again, telling a story again of replacing yourself.
Darin
00:23:04.666
Just like how we used to offshore, now we're going to onshore to our computer that's on our desk.
Viktor
00:23:11.635
The, the only thing that changed the scope, we had this story many times before. how often did you experience. Technical shifts that resulted, oh, what am I going to do now? Happened many times before.
Viktor
00:23:29.300
Yeah. The major difference is that the scope is infinitely bigger of what is happening at the same time, and the speed with which it's happening is much bigger. Right. You get 10 years to debate, whether Kubernetes is a thing or not. What you had 15 years to debate whether cloud is a thing or not, I don't think that we have that. That's the same timeframe we will experience now. I think that it'll be much shorter period before that becomes existential threat.
Darin
00:23:59.035
Where do you think agents are too premature right now? Basically saying, Hey, here's the root credentials for everything we have. Go have fun, right? We don't wanna do that.
Viktor
00:24:11.502
It depends what you mean by agents. As in agents that you use out the box, like off the shelf or you mean what We could, agents that we could in practice RT o rebuild. If you're talking about top the shelf, I can tell you right away everything related to your, specific to your company. That's that's very bad.
Darin
00:24:31.359
so you're saying a vendor provided agent is going to be worse than what we could build internally.
Viktor
00:24:39.377
A hundred percent. so vendor provided agent is a very, very, very capable engineer that just joined your company and you chose not to tell that person anything about your company. It's a pure genius, Hey, that person writes rust like crazy, never creates bugs that that's, that's the best hire ever. It's just that you happen not to use rust, and you happen to deploy everything in Azure instead of AWS, and you happen to use Lambdas instead of Kubernetes, right? That person is going to do terrible job that. Genius that you just hired,
Darin
00:25:19.205
I agree, but isn't there some value in having a vendor provided agent that might be a, oh, I hate this phrase, a best practice, a best in show that you then adapt and
Viktor
00:25:37.219
Same analogy as, as with that prior, mentioned before, point to the documentation. Explain what you're doing. Explain what your business is, explain how you're doing things. Who knows, maybe that person will even improve your, your situation, and you will end up not, not only with that person doing what should be done, but doing it even better. Who knows? But. Train the person you, every single employee in a company is to be trained. Whether that's five minutes training or five months, depends, but you cannot say nothing to a person who just say, thank you for coming. Go.
Viktor
00:26:16.242
Not only that, but kind of like that. That person is going to create an amazing mobile game. Amazing mobile game. And you just hired him into your company, who is pharma company, right? Kind of like you need to, you know, we are a pharmacy, right? The, we, we deal with, with, uh, this and that. You, you need to explain what the heck is going on.
Darin
00:26:39.681
It's extremely rare that I ever take a skill that I find online and just use it. I cannot think of a time. I mean, I might bring it in and it's like, Hey, I just loaded in this new skill. I'm in plan mode. I just loaded in this new skill. Let's adapt it for what we're doing in this project because there are some good thoughts. Let's talk about it and see how we can best use this.
Viktor
00:27:02.146
using skills as they are in vast majority of cases is complete nonsense. Complete nonsense because think about it this way, let's assume for a second it might not be a hundred percent true, but it's close to, a hundred that, models are trained, on all public knowledge. Everything that is public, they know including skills that are sitting in some repo for the whole world to take it. They will be trained on them as well, So assume that whole internet is their training ground. then you need to start thinking, okay, so what is missing? What is it you don't know how to do or you don't do right? And so on and so forth. And you will come to inevitable conclusion that if whole public knowledge is in the training data, the thing missing is exclusive. To what your company does to things that are not public. So how can public skills solve the problem of, models not having your private information?
Viktor
00:28:07.935
Yeah. Because if you say, Hey, do the prs, the standard way, whatever, the standard way I'm inventing it now is you need skill to do that. It already knows how to do prs and you cannot. And now you can create, oh yeah, but Joe likes prs like this. But then that's Joe's skill that doesn't work for you. You either use whatever the majority uses or use something special. And if it's something special, you cannot fix that something special by having skill made by somebody else who doesn't know that something special is. So write your own skills.
Darin
00:28:44.544
This is becoming very circular, isn't it? This is the same thing of should we hire a consultant or should we hire an employee? the consultant can get you over the hump that's getting in the skill, right? They're gonna give you some broad based knowledge. They might give you a couple of pointers. I'll use an example. I copied down a skill the other day from somebody that created a skill for initializing, a Tauri project, T-A-U-R-I. 'Cause I'm working on some Tauri projects, apps, right? And backend is rust. But it could be any number of flavors. A front end. Well, this model happened to use React for the front end. I don't want react in my apps at all. I want as lightweight as possible front end, so I'm not dealing with stuff. So I took that skill again, like I was saying before, loaded it up and it's like, Hey look, I don't wanna reuse React, I wanna use something as lightly as possible with a minimum number of dependencies as possible. That's impossible with JavaScript, but that's a different conversation. Then it went off. It's like, okay, here's a handful of different frameworks that we can consider using, or no framework at all. What do you want to do? I went, well, gimme the pros and cons. You know, after 30 minutes or an hour of conversation, it's like, okay, we're gonna do this. Great. And also looked at the skill on the rest side and it it's like, oh, you know what? This guy isn't doing all the rust things that needed to be done that I've already discovered in doing other projects.
Darin
00:30:12.159
feeding back into the skill. Once you build a skill, it doesn't mean that it's done. In fact, I can guarantee you once a skill is built, it's never done. Because as a human, as we're learning new things, we're building on that knowledge and we're able to use that as we move forward in a better and hopefully more efficient way.
Viktor
00:30:31.936
Yeah, no. Here's a question that hap hypothetical or real situation, whatever it is, of initializing a project. Do you plan to initialize those types of projects in the future as well?
Darin
00:30:43.927
absolutely. And here's the problem with that. 'cause I'm dealing with that right now. I've got three other products I'm getting ready to bootstrap because, hey, I can do it now a lot faster.
Darin
00:30:57.430
I don't have, and this is why I've been sort of battling with, I don't have a corporate office yet. A corporate office repo or corporate standards repo, that contains those skills that I can have as a high level repository that then points out to everything else for bootstrapping. I think that's where I'm heading. I'm not
Viktor
00:31:19.311
That's the reason why I'm asking because, if you, that, that was a one off thing. You didn't need the skill. You speak with your agent, you explain the same things that you added to the skill and it does it and it's faster. Now, if you plan to do to for that to become some kind of repeatable thing, then yeah, you're building a skill and you're facing the problem of how to propagate to all the reports and what's another, but in this specific case, C is because you copied once.
Darin
00:31:46.984
But I think if I come up with a way, and we're sort of going off track here, but if I come up with a way to have that corporate overview that's basically got the octopus legs into all the other projects, or it's the one that bootstraps, I guess it's sort of like going back into either A PMO role to where, hey, we need to start this new project up. What are all the things we need to do when we start a new project? Right. So to me, that's a corporate level thing that I want to keep out because I know that I'm gonna be creating Go cli. I know I'm gonna be creating some new desktop apps. I know I'm gonna be all these things. So if I can come up with those pro to where, okay, I use this to develop X, let's promote that up to the corporate as now the corporate standard. So when I create a new project like this, punch the button bing, bang B, everything has done in five minutes and off we go.
Viktor
00:32:38.466
I will make a guess that actually long before ai, you were working with companies that were doing that exact thing. I am pretty sure that you've worked with companies that were building some kind of system. There were People in their company can bootstrap new projects the way how company likes to do with them to do this, to do that. It's been done many times before. You just that. Now it's probably better 'cause it's more flexible. There is some intelligence in it and so on and so forth. But what I like to say, you realize how much actually the goals did not change. At least I was exposed to, I've worked with companies for, actually they were building kind of, okay, so you have this webpage and if you want start a new project, you execute this thing and it'll create a GI report with bootstrapped, files and this and that. And if you wanna do this, then you go here and you click this button and, uh, it'll do this and that. Those are all we were building skills long before ai in a way conceptually.
Darin
00:33:46.111
And less humans finally. Right? Again, it's just automation. Going back to your idea of it's Aron job, it's like, oh, we've got this new thing coming up. Okay, we need this. Alright. We typically do those on the overnighter. We hand it off to our offshore team because they can just work on it in isolation and it doesn't affect us any.
Viktor
00:34:04.626
even that is the same because in those situations in the past when there was a team creating those scripts or APIs or what not, let's say, to bootstrap new projects, different types of projects, I'm using that as an example, right? They, somebody works on creating those, let's call it scripts, and there were no humans involved outside clicking the button and saying, yes, I want it. After that, the work is still the same, the format is different.
Viktor
00:34:39.457
very often, it's not developers creating CI pipelines. Uh, very often, at least in large organizations, it's a dedicated team that creates CI pipelines and then developers choose, okay, I, I, starting a new project. I want this pipeline because this is Java project and that pipeline, because it's this, That's been happening before and it's happening now as well. It's just that now that those same people or maybe different people are creating it in a different format now, their skills.
Darin
00:35:20.980
What should practitioners be doing now? again, we keep on reiterating this is nothing new. We need to do an inventory of what we have. We talked about doing lease privilege. Basically we just need to break out our playbook we've been using for 25 plus years and apply it again. Right.
Viktor
00:35:40.353
The first thing that people should do, and in those constantly, when I say people, I mean companies the first step, ready, panic.
Darin
00:35:50.060
But isn't that what we do? Oh, have we been doing that for the past 25 years too?
Viktor
00:35:54.543
do you know that? there was a massive drop in the stock market for all security. Right. Companies recently. recently. as if of time of this recording,
Viktor
00:36:11.317
Yeah. Or slightly more elaborate slash command. Yes, correct. And that's that panic mode. and you should be panicking because you think it's fine until it's not. And when it's not, then it's too late, especially for larger companies that cannot turn around overnight. So panic because if you don't get there first, you will be in trouble. I dunno whether that's next week or next year. And you'll be in trouble in terms that you cannot recuperate from that trouble.
Darin
00:36:51.517
used to, large company meant enterprise, which I'll basically say anything greater than 500 employees.
Viktor
00:37:05.586
maybe we need, uh, new names and say to rename the existing large companies into super large companies and, qualify everything above 10 up to previously established boundary for large companies as large companies. look at it this way. OpenClaw. Made by a single person in a relatively short period of time is a project that most mid-size companies can only dream to have in terms of business. Only dreaming of having OpenClaw is a bigger potential business than vast majority of mid-sized companies. Alone made by one person fast. Have you seen symposium that's got like OpenClaw from the guy? Alex, I think is the name guy that made, cage, GPT, like solution that works in Kubernetes, which makes so much more sense if you ask me. You, you wanna schedule a lot of jobs running periodically. Well, that's the place, also done in a matter of days or weeks. One person. And it's infinitely better than anything that mid-size companies have.
Viktor
00:38:21.911
I lost lot less code. Yes. So you should panic. I think that if you're thinking about the title of this episode, I suggest now it's time to panic.
Darin
00:38:32.311
Because again, it could have happened in the past. There were probably spikes, and again, I think OpenClaw was a spike of something happening, but it's showing what can be done and how quickly people just want to get their job done or whatever they think the job is to be done.
Viktor
00:38:49.188
Successful midsize companies that were born from startups are also exceptions. And so, yes, OpenClaw is an exception. Not that everybody build something that's successful. So it is the same thing. It's just that it's happening by a person like singular or very small teams. That's the change. That's the difference. And I think that. Companies are still not able to wrap their heads around that or admit it even.
Darin
00:39:20.907
Because they still need to have the meetings before the meetings, before the meetings to decide what they're gonna name the meeting about. In which conference room were well, at the time, Peter, the guy behind OpenClaw, basically, I think it was in two months, two to three months tops,
Viktor
00:39:37.197
Can you go, so essentially. Yeah, a company should do experiment and, split people in their company. Say half of you will attend the planning meetings that as we normally do, and the other half are going to build random stuff. While we're doing, we are doing those meetings to figure out what we're going to do, and let's see who's going to be more successful. I. guarantee that people are gonna build MVPs. Before the company decides on even what to start building. And when they say MVPs, I mean MVPs that we actually, demonstrate the validity of something instead of visual thinking. That's predates it.
Darin
00:40:21.800
Is there anything that the industry can do to help get this right, right now? We talked about Linux Foundation. You know, they're, they're setting the standards and blah, blah, blah, blah, blah. But what is it that the industry can do?
Viktor
00:40:33.084
Tell people to panic, I think, I think that that should be the messaging, right? Kind of like, okay, uh, if, if only I say it now it's time to panic, then it's not, you know, you amplified enough. If more of us start saying the same thing, then maybe it'll reach the right levels.
Darin
00:40:54.277
So basically what you're saying is once Kelsey Hightower says it's time to panic,
Darin
00:41:05.768
here's the thing to think about is, chatbots were easy, they were toys, but useful at times.
Darin
00:41:13.208
Agents have moved beyond toys. Are becoming extremely useful with incorrect guardrails.
Viktor
00:41:21.127
And there's still give or take equally easy just to be clear. Unless if you're not building your own agents, right, for you, agent is a chat bot. You talk to it and it does things from user perspective is the same thing,
Darin
00:41:37.630
Again, you're a manager now. You think you used to be a developer working inside of an IDE typing. You're a professional typist with some knowledge, but now you're a manager trying to bring a business to life if you don't, your business may no longer be there tomorrow.
Darin
00:42:04.471
you can take that two ways. Human managers managing, managing other human managers are in trouble.
Darin
00:42:12.001
we've seen a, a clear out of the mid-management level in enterprises. It hasn't even started yet. We're on the early part of the curve for that. I think that's gonna be decimated within a couple years.
Viktor
00:42:28.010
also mana managers managing people that will transform drastically because you know, before you, you normally company would split into teams or five to 10 people or whatever. The teams are in charge of different things and each team. In charge of different things gets a manager or product manager or whatever is the name that manages that. Those people. And I think that we are moving into one person teams. And then you say, kind of like, is there a point having a manager who manages me and only me. And if you say, okay, so that manager can manage multiple single people, team, single person, person, teams, right? But then you say, okay, but don't we already have a role for that? That's a tech lead or director.
Darin
00:43:23.131
I think all these titles are going to soon be nothing Again, there'll be a few organizations to where it does matter, like the ones trying to create the things, but we no longer have to have a fan out organization of 64 people.
Viktor
00:43:39.954
Yeah, number of people working on a project. Whatever the number it was, cut it. And then decide either you have more projects or less people. One of the two things.
Darin
00:43:52.691
Because otherwise you won't be able to justify the spend on the humans because if you're doing agents and humans, you better have a lot of money coming in. Yes, you're not paying the agents as much, but you done incorrectly. You could be paying the agents a heck of a lot more than the humans, as I've seen recently
Viktor
00:44:12.681
I think that if you're talking about people really heavily relying on agents, you know, LMS and all those things, the cost is not those 20 or 200 bucks a month just to be completely clear and transparent, right? The cost is higher. If you dar would be working like full-time with agents. and probably kind of running multiple, more than one of them in parallel. Your subscription would not give, would not be enough. Would not last a day.
Viktor
00:44:48.754
Yeah. Or a higher tier subscription. But we are probably talking about that cost. And I still think it's low, but we are talking more and more around a thousand bucks a month than, uh, like a hundred or something like that.
Darin
00:45:02.310
Right. I did see somebody the other day, and I don't remember which it was, but it was, they, their typical spend was around 150 to 200 a month, and their token leaked. And somebody got a hold of it and within 24 hours, their standard spend of under 200 a month went to $85,000 in two days.
Darin
00:45:25.755
This goes back to how do you onboard a project correctly? What are the typical processes you would do to onboard a project Today, it's no different actually, if you had bad processes back then. Change the processes. If the processes were like, man, why does it take us a freaking three weeks to get a get repository? You remember those days? Or to get a database, right. Just these simple things. Or one of my favorites when I was consulting. Okay, you're gonna need four VMs. Great. I'll talk to you in 12 months.
Viktor
00:45:58.699
Exactly. You know, if, if young, if young audience is listening this, they will think that we are inventing stuff right now. That never happened.
Darin
00:46:08.249
I've seen it happen over and over again. Like to bootstrap a project in some places I saw was three to four months because they had to set up the nexus repo and the get repo and the Jira tracking and the blah, blah, blah, because they had to go through seven and eight different teams.
Viktor
00:46:26.040
and this is a true story without me naming any names and true story that happened more than once in Variations. Our goal, we did it to spend the week with a company. I go there first day, spend maybe two hours, reach the point that we need as search keys to continue, and then they tell me, okay, we'll get it. We'll get it. It is it. No problem at all. And then half an hour, hour later they say, okay, actually not today, Go and go back to your huddle, kind of like you're free for the rest of the day. And then I would get the call the next day in the morning, kind of not today either. and this is true story. I had the whole full week engagement where I spent only two hours with them actually. And I was on site just to be a hundred percent transparent and clear on site, not from my home, all the nothing else.
Darin
00:47:22.571
I've got variations of that story as well. So you may laugh at us, you may be yelling. It's like, I don't, I don't want my job to be lost. Look, it's time to get on board or really become good at crafting. What was my phrasing recently? Organic, uh, free range code. And charging quadruple for it because that's the only way you're gonna make any kind of money.
Viktor
00:47:44.683
Yes, If your output will stay the same, you are in real danger with your job. If you multiply your output, you're fine. Companies need more. All the companies always dreamed of doing more than they're doing. They just couldn't afford to do more. There are issues piling up. There are features that everybody wants but cannot be implemented. There's plenty. There is infinitely more work in every company than the work being, done.
Darin
00:48:15.719
So if you haven't started with chatbots yet, that's probably the place to start. Don't go straight to agents. I still think that's a good learning path, but but if you've already gone through the chat bot era and you haven't made it to the, you decided, eh, that's not really for me. I could have agreed with you. Nine months ago because things were just getting started and like the first passes of GitHub copilot were
Darin
00:48:44.459
yeah, it, it's gotten better. It's still not my favorite, but today most everything is good and getting better every day. Models aren't changing a lot, but. The tooling around them getting a lot better. So what do you think? However, the slack workspace. Go over to the podcast channel. Look for episode number 3, 4 8, and leave your comments there.