Agents de codi, tokens i criteri: la visió de Mario Zechner

it's going to be a rich man's game. People who have the means of production in the sense that they can afford the tokens have a massive edge. >> I have noticed there's a lot of people who don't actually believe in the fundamental technology. They're just here like for the trend. >> They try a lot, they build a lot, but they don't sell anything. The code doesn't need to be perfect. The code can be total [ __ ] slope as long as it generates time saving. If you can make your six people more productive by giving them coding agent, you can easily beat the 50, maybe even 100 person team. It's amazing. >> Right Mario, so obviously you're the creator of pie by that dev, the famous pie agent. Let's start with what gave you the idea to create pie. >> I've been using LLMs since basically 2022 when Open AI came out with ChatGPT because I had a background in NLP and machine learning in the 2000s. Kept up with the field and was quite enticed by LLMs because they could basically do all the natural language processing tasks that I was painstakingly trying to do with less capable models. And since I'm also a software developer, I always tried to get them to work for my software development tasks. [clears throat] And that only really started working I think with cursor sometime in 2024. And that worked for a while, but it was quite limited because it kind of only worked on individual files and while it pulled in some context from the rest of the code base, it wasn't really amazing. And in April 2025, Peter screamed at me and Armin Ronacher of Flask and Sentry fame, "Hey, the agents are working now." >> So that's Peter Steinberger, right? Creator of Open AI. >> And um >> We were both like, "No, shut the [ __ ] up." >> [laughter] >> Uh it's not going to happen, but that's around the time when he started using Claude code and he just told us, "Just try Claude code, man." And that was the big revelation because what the Anthropic team did was pretty ingenious. Previously coding agents like Cursor and so on would index your code base and kind of not give the agent enough tools to explore things on its own. Anthropic had the genius idea of making the agents basically use your computer. So, giving it the a sort of terminal, a bash, something that lets it run commands. And with those tools, it could explore the code base on its own. And it turns out LLMs are really good at exploring code bases. >> [laughter] >> Um so, that basically gave us a gentic search, as it's called. And that changed the equation, because now all of a sudden the agent was no longer restricted by by something like cursor that would not allow it to search things on its own, but kind of give it the thing where cursor thinks this is what it wants what the agent wants, but do things on its own. And that kind of started the entire agentic coding thing for me. And I was a diehard cloud code fan, but eventually the cloud code team grew. They started to, well, token max, I guess. And they added feature and feature and feature and feature and with that a lot of bugs, a lot of changes to the system prompt and >> So, roughly what part of 2025 we talking now? >> Uh about July, August. That's >> So, that's where you start to feel like the unnecessary features. >> Yeah, I I reverse engineered um Claude a little bit. I mean, reverse engineered it's a JavaScript app at the end of the day, so you don't need to reverse much, but I saw that they're doing a lot of dumb [ __ ] >> [laughter]

>> And preface, I love the team. They they they are they they created the genre, right? But um for me that was not a a usable tool anymore, because it broke my workflows every day. So, then I slowly, surely started out building my own thing and then October, I think, I switched over to pie entirely. And have been very happy with it, because it's stable, apart from the LLM part. I obviously don't have control over the LLM itself and those change as well. But everything around that is now fully under my control and it doesn't do any superfluous stuff. So, my workflows are pretty pretty stable even across different models. >> So, why do you think that happened? Is it just the natural life cycle of a product or they're trying to aim more towards consumers and they have to add like more guardrails? >> I think they're dogfooding what they're preaching. And I think that's perfectly fine. They're kind of like a canary in a coal mine where you can see how good agents autonomously can work at the moment and if the product they spit out are any good. I personally don't think Claude Code is good. Um, but your mileage may vary and a lot of people are obviously using Claude Code, so it's not terrible for most people, I would assume, or they're just used to the slop. >> [laughter] >> Which is also a thing of a bug serving. They seem to be reigning in things again, like just a couple of days ago they were like, "Oh, oopsie, sorry, we messed with, I don't remember quite what they did." >> Revisiting efforts. They they just quietly switched that. >> The revisiting effort isn't even a big deal. That was just a default. They would put you on medium instead of high by default. You can change that. But what they also did was, so, okay, when the model thinks, it generates a thinking trace before it generates a text content block on the server side. And Anthro- Anthropic doesn't um doesn't send that back to the client, so to Claude Code, but it lets another model summarize that and send that to the client so they have something to show. On the next turn, when you submit another prompt, that thinking trace is reconstructed, the original one, on the server side. But it turns out, uh what does it say? On March 26th, we shipped a change to clear Claude Claude's older thinking from sessions that had been idle for over an hour to reduce latency. So, if you had an agent run for an hour or you stepped away because you grabbed lunch and then came back, all the old thinking traces were gone, which influences what the agent can remember in that session. And that's like super bad because that's kind of like lobotomizing the the model for that session if you take away half the information that it generated throughout the the session. >> So, it comes down to context, ulti- ultimately, right? Because uh >> Everything's about context, really. Yeah. >> Yeah, so like that was one of the main things that you were seeking. Is like full control over the system, problem making sure that, you know, your assistant is the same today as tomorrow. >> Obviously, they still have additional context bits on their back-end servers like what you sent through the API that's not the final thing the LLM sees. I just want control over all the things I can actually control and and none of the including codex none of these coding agents give me that kind of control. So, that's why I built pie. >> So, you think like you have to have that exerted effort to keep it minimal. So, like have that as the main goal to not start it bloating and feature creep. >> you are a wielder of coding agents as well.

So, you probably have a very good intuition of when the model goes south or when the agent goes south and it's usually related to context rot or the percentage of context that's been used. I want my stuff to be efficient. That's how I've done all my software development over the last 20 years and I want the control because I know if I don't have either of those my software is going to suck. And if you delegate a lot of your software creation to agents, you want to ensure that they have the highest percentage probability of actually producing good stuff. And I don't think that's currently possible with the way cloud code for example does things behind your back. >> There's have been a lot of theories that anthropic you know because they admitted some things in the postmortem but bigger accusations were the models changing in the background. Have you felt that or not because you're not using cloud models? >> No, I have I have never felt that. So, I think there's two two things that are happening. The first one is psychological a new model comes out you get super excited you bang out a trillion tokens and you build new greenfield projects and like wow it's amazing it built me a 3GS app which is like trivial. >> [laughter] >> Um and then eventually you put it to work for something that's more important and you see ah [ __ ] this still doesn't work and I hear it [ __ ] up and so on and so forth. So, I think it's mostly psychological in in that sense that the initial enthusiasm vanishes as you start using the model for >> honeymoon period. >> Yeah, exactly. >> Okay. >> Um and the second part is well I mean it's also the hardness specifically cloud code it changes so much that it breaks things just by changing the harness not the model. I personally would attribute most of that to those two things the psychology of it and the harness changing on your butt all the time and breaking because in PI I have not experienced that kind of degradation that other people using cloud code for example reported. >> Yeah, so there's like a lot of moving variables and then it's hard to know what is causing that change and people just go to the easiest explanation. >> Yeah, they quantized the model. Yeah, sure they did. You know what that means for the infrastructure if you have to run the model in a quantized way and so on. I mean it's not entirely completely different but it's effort, right? And while it might save them some infrastructure cost getting that to run also costs money. So I'm not convinced by that and again I didn't see it in other harnesses that don't change as much and I think if you talk to open code people that use entropic models they will probably tell you the same that they don't feel this degradation as much or at all. So yeah, I think this is mostly psychology. >> Yeah, probably and some you know collective psychosis where everybody has the same opinion as everybody else. >> To be fair they had a bug at some point earlier this year where they [ __ ] up some code in a kernel or something where there was actual model degradation because of a bug in the inference engine. But that's stuff that they usually themselves feel immediately and fix so. >> Okay, when it comes to the whole AI agent space, what's like one thing that you love the most and one thing that you hate the most? >> I love that I can be much more productive and can do projects that I previously wouldn't have done because I just wouldn't have the time to do them. I also love that it allows me to

experiment with things real quick. Like if I have an existing product I can just say hey agent build me this new feature and it doesn't have to be perfect. It just needs to exist for me to feel it out and see does it make sense for a product to my users need it and so on and so forth. I can bang out a new feature in an existing brownfield project with users in like an hour or two or an afternoon, test it myself, give it to a handful of people in in a test setup or test uh staging deployment. And if it works, I can make it nice. And if it doesn't, I just throw it away. And I didn't lose any time or sleep over it, and that's that's just amazing. That's But it also means that you need some judgment. >> Yeah. >> And the thing I hate is the [ __ ] hype. It's just so tiring. Things like Ralph loops. >> Yeah. >> Ralph loops can work, but people misunderstand why and how they work. Um I think uh car parties auto research is a better example of a Ralph loop because it basically is a Ralph loop. So, you give the model a a specific success criteria and an objective function that allows the model to evaluate whether it has improved something or not. And that is like the scientific Ralph loop, and that actually works. But a Ralph loop is a I have a PRD and a spec file and blah blah blah, and you just iterate until the spec file is implement. I've never seen this work. If people make it work, more power to them. I For me, it's cargo culting. Two weeks ago, I was in London at AI Engineering Europe, a big conference, and behind the scenes, a lot of people would come, talk to me, and be like, "Yeah, all of this [ __ ] doesn't work, like dark factories and so on, but we still need to sell it." >> [laughter] >> And I kind of I I just don't know why people would do that because we're just not there yet, and it's fine to accept that and just use what's there and make that good. But people seem to require that kind of hype machine. >> Do you think that's just impatience, or do you think it's like people who join AI for the wrong reasons? Like I've noticed there's a lot of people who don't actually believe in the fundamental technology. They're just here like for the trend or you know, the the business opportunity to get clients and stuff like that. >> A lot of the web 3 people came over into AI land. Uh and once web 3 became less successful in terms of monetization. So, you see a lot of people who have no [ __ ] clue about the technologies. But on the other hand, there is also very talented and smart people who actually understand the the technology. You have incentives to sell you a vision that's not yet reality because they earn money by that. I mean, I don't blame them. That's how you put food on the table, but I personally don't enjoy that. It It has no signal for me. And then you have a lot of people who don't understand the tech, but want to use it, and they're They're following the people who are screaming the loudest. And that usually ends in pain and tears and lost money. So, I I don't know. >> Do you think AI will make like there's a great quote by Balaji make the smart get smarter, right? Like if the rich get richer because they have a discretionary income, they can invest it. The smart people know how to wield these agents, how to build their own harness, or you know, contact engineering. They use the best models, obviously they pay for the tools, so they get bigger advantage. Do you see this happening, or do you disagree with that? >> I agree with that, with both. Like people who understand the craft of software engineering and have some experience with it definitely have um some edge compared to youngsters who just joined the field and are now

starting using agents. Ultimately, I think that difference will probably vanish as models get better. Because you as the model driver or the agent driver don't need as much low-level understanding anymore, possibly. At the moment, I think that's not true, but it might become true. >> I mean, you could say that the first time in history where we do not getting the latest models, right? With like the release of Mythos. If me and you were running a multi-billion dollar company with enterprise contract to Anthropic, we we could just say like, "Oh, yeah, we really needed it for cybersecurity." And then obviously you can like divert 10% of token spends to R&D or whatever you want, right? It is already true, but I'm I'm talking more like on the person-to-person level. Whereas, you know, right now you can get insane amount of value from like $200 a month Cortex plan. That's already prices out like 99% of the population. >> I definitely think it's going to be a rich man's game. People who have the means of production in the sense that they can afford the tokens have a massive edge over people who cannot afford the tokens. And it's funny that you say the the $200 a month are an economic barrier for a lot of people already, because most people in our peer group within the AI field don't see it like that. They see it like $200 is like a bargain. >> Yeah. >> But that's mostly because they are pampered little software developers with multiple hundred thousand K income a year. For them it's peanuts. For for >> they know how to use it. Because they know how to get the leverage on it. >> Yeah, I I mean yeah. I honestly I haven't seen too many people actually building successful, monetarily successful things after starting using agents. What I do see a lot is, at least in my peer group of old grumpy computer people, they try a lot, they build a lot, but they don't sell anything. >> [laughter] >> Mhm. >> Um because most of what they build never finishes. Um so I don't know. But you are more in the builder scene than me, so maybe you can expand on that. What what are you seeing? >> I would say actually people misunderstand this. People expect like a lot of new products, which obviously like that that is lacking. >> Mhm. >> But I would say it's probably like 80 to 90% of the benefit is internal. >> Mhm. >> For like existing companies and existing businesses where like people on my team, people who are not technical at all, like if video editors and you know, people who are like zero reason to be building software are building software. And with some guidance, you know, from me and other people who have some experience, they're able to build useful internal tools that make us more productive. Like for example, for YouTube, you know, let us spot outlier videos. Let us easily make funnel variations, right? So like for or for accounting, like internal accounting. So this is not like a publicly deployed product that anybody can use, but internally it's immensely valuable. And like you can ask anybody in the company and they will they'll instantly tell you that. And again, it's not it's not like those people are developers. They are people who know that workflow. They they know, you know, video editing. And they understand what it makes what it takes to use a video model or image model and and make a great tool out of that. So I think that's where like a lot of the hidden results are. It is in existing companies that are growing faster than otherwise. >> I 100% agree with you. My my first

observation of that was last summer when I gave my wife she's a linguist she's non-technical. I gave her a coding agent because she's working in on on on linguistic data lots of it like she does interviews with people transcribes that annotates the thing that that the text the transcriptions and so on and then she puts it in a 18,000 row Excel sheet and manually transforms the data generates charts and all of that stuff. And I gave her Cloud Code and I gave her two nights of instructions on how to drive Cloud Code and all of a sudden her scientific output like 5x. And all she did with Cloud Code was basically I know the input data I know what the output data should look like. You are now my little programmer buddy I don't know programming I trust you you just write me Python scripts that can take my input data and produce the output data in the format I need it and also charts in the format I need them. So she understood that while she doesn't know how to program she understands what the input should be and what the output should look like so she can verify that whatever the model wrote or the agent wrote for her in terms of code actually works. And that's all she needs the code doesn't need to be perfect the code can be total [ __ ] slob as long as it generates that kind of time saving it's amazing. So that's also what I'm hearing about your company here you have non-technical people who understand now how to use coding agents to write little programs or do data analysis and transforms. Yeah I agree that's the the most massive unlock. The problem with that is that in companies that like yours I would assume there's a lot of young people who are tech not not not technical in a sense they can program but in a sense they they know their way around the tech yeah. Um I mean for you it's a breeze like easy peasy right? If you go to a big enterprise corporation where you have people in the range from 20 to 60 Yeah. That's not working. So companies like yours how many people are you 10? >> Like 4 or 5 like a 6. >> Yeah that's >> Like you have an advantage because if you can make your six people more productive by giving them coding agents or agents to do to to help them with their day-to-day tasks, you can easily beat a 50, maybe even a 100-person team if you played smart. Especially if the 50 and 100-person team is like an old guard kind of team that has real trouble adapting to new technology. So yeah, that's super cool. >> What are your thoughts on like optimizing token use because we've seen a massive trend of especially companies who spend crazy amounts, but even individuals, you know, like I myself have been pretty scared to check my spend on open router and on Frolic. And now I think, you know, especially after yesterday's release of Deep Seek V4, we're starting to probably see a trend of optimizing for these different workflows where like you can have the you know, 97% of the same capacity, same intelligence for like 1/10, 1/7 of the cost. Do you think like tokenomics and this whole optimizing trend will continue? >> Uh I hope so. Because intelligence must be a affordable by everybody in the world. I don't want this to be a first world kind of technology that [ __ ] sucks. And the Chinese labs, bless them, >> [laughter] >> Uh yeah. >> they are [ __ ] with the token economics of the US labs and I love it because I know the margins at least on inference for the US companies and >> is awesome. >> Yeah. So yeah, it can even be more than that.

>> [laughter] >> So yeah. Uh I I think like optimizing tokens is is going to be a big thing and and you using open router, there's like a gazillion of of token providers and routers like that now out on the web, so there's a definite need for cheaper tokens and I think that's going to happen and I personally think it's only going to happen because of open weights models like DeepSeek or Kimiko 2.6. Honestly, I've been using Kimiko for the past 3 years 4 days now. It's good. Like I can deploy this on my own little GPU cluster in the cloud at cost that's comparable to what I have to pay if if if I went through say Anthropic and Claude. And the intelligence is there. I I don't actually need that much more than what Kimiko offers me. So, this is just going to get better. And I don't necessarily think that frontier models have an edge in intelligence anymore. I can see there is regressions in some tasks and some verticals. So, yeah. If we peak in intelligence and that's what we get what we have now for the next 2 or 3 years, I'm happy and I'm so going to switch to open weights models because they have caught up. I hope that answers what you asked. >> [laughter] >> Yeah, I mean, I'm I'm wondering like why isn't there a bigger migration of individuals and companies to open weights models? Is there China fear mongering or like the brand of Anthropic Open AI? >> I think it's the brand. Anthropic is very aggressive in their end advertising and marketing marketing especially to enterprises. And they have great success. I'm I mean, you also have to admit that the brand is really good. Maybe not with the developers anymore after all the [ __ ] in the last couple of months. >> Yeah, last last 3 days, yeah. >> Yeah. >> [laughter] >> Um but but with respect to enterprise people and the the technical bias and the bias at enterprise companies, they have a perfect story. Um OpenAI, not so much. They're trying to get there, but I think Anthropic just has a really good brand at least in the best and western world, Europe and the USA. Uh and yeah, there's a lot of fear mongering with regards to China as well, but honestly, as a European, both China and the USA kind of I mean, [laughter] at least in in terms of data privacy and and industrial espionage, they're are equal to me as a European. I don't trust him. >> [laughter] >> So, it doesn't matter to me. I will use whoever gives me the best kind of intelligence for the best price. Um and hopefully eventually Europe can provide its own intelligence at a good price, but yeah, not seeing that at the moment. >> Yeah. And like you said, you know, if you can take Kimi or Deep Seek, deploy it on your own cluster, and you know, run it at similar cost if not cheaper, then you get the massive benefit of the data stays yours. >> Yeah. >> So, >> I don't need to deploy it for me individually. I can take five more people and the cluster will happily do happily serve all of us at great TPS. Um [snorts] so, it can actually be really cheap if you pool together and and and kind of get a a cluster for you and your six people at your company, for example. Um It's something I consider for one of the companies I'm involved in, but yeah, we'll we'll see if that goes. I want to experiment with this a little bit and just see how

how not affordable, but would it actually make sense doing that at this point already? I experimented with that last year and it wasn't quite there yet. Neither in terms of intelligence nor in terms of cost, but I think we're getting there. What is your in your AI-first, AI-native company? What what what is your strategy in terms of where you acquire intelligence from? >> Yeah, so we use the APIs. We don't self-host. I mean, I run more and more models locally. I did the wise decision in like beginning of '24 to buy a 120 GB MacBook. So, now with Gemma 4, this is really paying off. And you know, the models, especially the progress on the small open weights, open source models, is like even faster than, you know, on the bigger ones. So, I think I really want the next MacBook to be like 256 GB of RAM. >> Yeah, me too. >> So, we could actually run like, you know, 200, 250, 300 billion parameter models. And then you could technically just use that for most workflows. And then like if you really need the cutting edge or you know, for some some tasks you really want the best of the best, then you would use API. But yeah, I mean I use OpenRouter for most of the projects I built because you can use all the models and I used I'm actually like on the last obviously changes all the time, but in the last 30 days I had a I reached like a run rate of like 6K a month on Anthropic API, but now I'm pivoting that more with the release of GPT-4.5 and DeepMind 4 and OpenSource 4.7 4.7 because 4.7 doesn't have the fast mode and I was using 4.6 in fast mode which is six times more expensive. >> Yeah. >> So yeah, I'm going to >> apparently. >> Yeah, but I mean look I I I love the speed so I I was happy with the speed, but now 4.7 is a bit you know, I don't know. For me the vibes are a bit off, but I I like 5.5 so I'm using I'm using Codex more and uh Yeah, so that's going to reduce my cost a lot and OpenAI just gives you way more inference for the same price. Like for the two whether you go with $100 cuz OpenAI added $100 tier as well. If you compare these tiers at OpenAI and Anthropic, you really feel the difference. Like with I never hit my limits on $200 Codex and I always run multiple in parallel. With Claude code it's like easy to do that in a day or two. You can easily hit your limit. So yeah, that's my situation and also the reason I use that is because it's easy to just give an API key with like two different teammates, right? So like if I know somebody is like really good, I can just give an OpenRouter API key with $500 weekly limit and you know, he can do stuff. If someone is like a little bit you know, beginner, maybe it's like $50 total limit and I know >> [laughter] >> that person is like you know, probably not going to hit that even. So I I like the management of like having all the keys and you can just name it who has the key and you can distribute that. So I like that. You touched on Europe, I want to touch on that. Why isn't there more AI stuff happening in Europe? Is it just America poaching the talent or what's happening? >> Yeah, it's mostly America poaching the talent from both Europe and China actually. I mean it's not like there's not US native people who who are also amazing AI people, obviously. But, the talent is fleeing to the US because that's where the money is. You simply do not have the same venture capital landscape in Europe. You also don't have the infrastructure in Europe. I personally don't think that regulation

in Europe is a problem because you have similar kind of regulation in California, for example. They also have their own kind of AI act and it's even worse there because now every single of the states has their own kind of regulation, whereas in the EU, at least it's the same [ __ ] for all 27 states. Um which I personally think is fine. Uh but, yeah, it's mostly about the money. Um you I mean, just look at all the names that you know in the AI space. They're mostly European names and Chinese names. So, >> [gasps] >> Yeah, it's kind of sad. >> Bunch of the like main people at OpenAI are Polish, you know? So, like they just anytime there's something happening, uh either VCs or a big big tech comes in and acquisition acquire. >> Yep. >> Yeah. It's sad. Well, so what do you think would need to happen? Like how do we get more investment, more VCs, more capital? Is it just like the European um mindset is is like risk-averse and it's like stocks and real estate or >> No, I No, I honestly don't think that's the problem. I think the problem is mostly about the regulatory framework. Um and and the legal framework. Um investing in an EU company and shares and uh RSUs and all all of that stuff that's very simple to do in the US, it's very hard to do in the in Europe. Especially if you want to grow your company and kind of have multiple uh multiple offices around Europe. Because in every new country you settle, you have to kind of deal with that country's structure in terms of how can people invest in that, how can employees get stock and so on and so forth. And that makes it really [ __ ] hard, whereas in the US, you set up a Delaware corporation and that just works. It doesn't matter if you have offices in Boston, New York, or San Francisco. It's the same for everybody across all the states. And in Europe, that's not the case. So, that that's one point. It's hard to create a company here with the same kind of legal framework that makes investment and giving employees specific benefits hard. And the second thing is that for investors, actually investing in a company in Europe on the tax side and all of that, it's it's different and not as good as in the US. It's also a question of um usually if there's disputes and stuff, you need to nominate a court where you want to have things handled then between you and the other party. And in Europe, that's just a [ __ ] Whereas in the US, everybody basically gravitates towards Delaware again. So, yeah. There's some progress there with the um what's it called? EU Inc. movement. And the EU has said the Yeah, they the EU has kind of agreed that they want to want to do that what the that movement and that Andrea Klingner and others has proposed, but like I love the European Union. I hate the [ __ ] implementation of people of [laughter] it. I'm pretty sure they will [ __ ] that up as well. So, yeah, I don't know. It's hard but we'll see. Yeah. >> Okay. So, shifting topic to consumer and apps. Well, not necessarily consumer, but like apps, right? Do you think most of today's apps will even exist in the future? Or will it just be you as a human talk to your agent and the agent does 90 95% of things? >> Mhm. Good question. So, I have a history in mobile games starting somewhere around 2010 and around 2016 for me, apps on mobile were already done. Like the market was [ __ ] because there was an overabundance of apps nobody needs.

If you look at Apple submission and Android Play Store submissions now, there's a massive amount of apps that nobody needs because now everybody slopes apps onto those stores. So, then the question is will my grandma build herself her own Spotify and today I say probably not. Because Spotify is not just the app, it's also the whole infrastructure behind it, the licensing deals with music what's it called? The music mafia. And all of that and and and people always only see the software but they don't see the the squishy human parts around all of the software that actually make the software the thing it is. And for apps like Spotify I don't think that that that's going to be replaced anytime soon. For things like fitness trackers or diet trackers and all of that small beans things, I think that's done. I see [snorts] no future for that. But I also don't see that regular people go on into an app and say build me a dietary tracker thing. It's just that the AI agent app that you have will just do that and you don't know that it's building an app in the background. It's just smart enough to kind of track things for you and surface the information whenever you need it. So, yeah. I I hope that's the future because malleable software, self-modifying software, I'm a big believer in that. >> Yeah, I mean on that topic of like fitness and diet when I when I had to track my calories, instead of going to the App Store and downloading an app, my first thought was using OpenClaw to just create a CSV file and log my calories. And then I created a second file which is markdown file with approved meals and you know, it has like the exact macro for each meal. It's like oh, you know, track 50 g of blueberries. And it knows exactly which blueberries I have and boom. And like it tells me Yeah, and it's not even It's not even a comparable because you also can like say like, you know, I'm a bit tired and I have a you know, podcast coming up. I need something with high energy and it gives you custom meals. Like none of these apps on the App Store can give you you know, intelligent yeah, so like intelligence is is the is the >> And and if I read a new paper on some dietary stuff or sports stuff, I can immediately integrate that, you know? And and no app in the world does that. That that and that's fantastic. I do the same with Fi. Like I have a little Fi instance that I can talk to via Slack. And I use it for exactly that kind of stuff. I can also, if I don't know the calories of the meal I'm currently taking, I just take a picture and say >> kind of estimate >> estimated. It's going to be very wrong, but just ballpark and I I kind of recorded that I ate that [ __ ] so yeah. >> Well, actually, my experience is that even when restaurants tell you the calories, they usually tell you a lot less than there is. So, >> [laughter] >> the agents are at least I know there's no hidden agenda in the agent. Okay, so on the topic of like future of work, how do you see us working in like 2 3 years? Let's let's let's talk about like people on the cutting edge also like the average person. You can make a good argument that a lot of current jobs that are not enough physical jobs, but like knowledge work could already be replaced with the proper workflows. So, like if we extrapolate to the 3 years, how does that look like and how does the cutting [snorts] edge look like? >> May I first ask you what your background is? Are you like a software developer by trade or are you just a technologist or

>> Yeah, so I mean, my first project was actually an Android game, so funnily enough. That didn't make any money. Got like 1,000 downloads though. But then I was just an entrepreneur ever since 16. I I did my own businesses and yeah, I've been an entrepreneur all my life. I mean, I'm technical, but not like a like you, you know, I'm not a great programmer. I'm fine, but you know, I I I use AI now for everything. >> Yeah, I mean, me too. So, we are no different anymore. Um yeah, the reason I'm asking this is this. I'm I've been in this industry, software and to a larger degree AI, for 20 years now. >> [laughter] >> And I I'm always hearing this, "Oh my knowledge work is over because we can automate all this [ __ ] with it was SVMs at one point that could replace a secretary triaging emails, and now it's LLMs doing whatever agentic workflows. And I think a lot of tasks by knowledge workers can actually be fully replaced by automation. I'm not yet convinced that the knowledge worker itself can be replaced. Um, I think we still need some human in the loop for coordination, and also the thing that we software people always we software people always think we understand the world because we can make computers do things, but if you have ever worked in a non-software company, you will see that the world is much different to what software developers or the IT industry thinks it is. A lot of things are actually, even in knowledge work, are very hard to to automate because there's so many human squishy parts. Um, and so I have experience in non-IT, uh, in the non-IT world as well. That's why I'm a little hesitant to say that knowledge work will change massively over the next, say, 5 to 10 years in in the sense of less humans, more automation. I I I'm big believer in Jevons' paradox where >> Yeah. >> agents don't replace people, agents just make people even more productive. And I hope, at least, that's the world I want to live in. I don't want to live in a dystopia where nobody has a job anymore. Um, and we all get UBI or whatever it's called, universal basic income, dependent on the state, and I don't want >> that's terrible. I I've seen the idea of universal tokens. >> [laughter] >> I like I like that >> I'm all for that, yes. >> Yeah, like each person gets some tokens, and like you either waste them, you don't use them, you can resell them, or you build something with them. I think that's already better. >> a really interesting idea, yeah. >> Yeah, like people always, you know, don't look at the second-order consequences of like you universal basic income, universal high income, you know, as Elon popularizes it. It's like, okay, but then who controls the income, you know? >> Exactly. [laughter] >> Yeah. >> Yeah. >> Okay, so so you don't see the like labor changing labor markets changing dramatically. >> I do think the labor market is going to change, but in a sense that a lot of people will have to upskill. A lot of those 40 50-plus people will need to [ __ ] start using agents. And a lot of them will not be able to do that, and they will probably be on the cutting block. And they also see that a

lot of juniors will not get any jobs because seniors plus an agent can now kind of replace them. I'm not talking about software, I'm talking about all kinds of knowledge work. As a senior in a knowledge work position can now replace two juniors with an agent. Um and still have their own output. Um so, those two things, old people and young people, that's actually really hard. I I think there's going to be some a little bit of a job cop so job cop ellipse for some time until it evens out again, until we are at an equilibrium again, where companies realize they need a junior pipeline because eventually they're running out of seniors because they all go >> retired. >> Yeah. And where we will figure out a way to upskill old people that have trouble with this technology. Um >> Okay, but don't you think the argument is that like young people have much higher neuroplasticity, so they can And you said they're technical in a non-obvious way. Like when people heard hear technical, they always assume like programmer, developer, but like it's it's it's like more about like do you know how to use a computer? Do you know like where to click, how to go to settings? That's actually way more important than memorizing syntax or whatever. So, don't you think like the cuz you you made argument that even the juniors will struggle, but then the counterargument would be the juniors are adopting AI everywhere. >> Yeah, no, no, definitely. I I think so. It's hard to Let me step back just real quick. So, I've been in this in this programming thing for a long time. And uh over the past 15 years, there's been this narrative of the digital native. I think it's even longer than that. It's like 1999 or so, people started to say, "Oh, this new generation is growing up with the internet, and they're going to be so computer native. They They don't even need to explicitly learn programming and computer use anymore. They will just by osmosis become amazing." And that's not what I've seen. Like, [laughter] uh I I also thought at schools, 12 to 16-year-olds and so, I had some insight on how they use technology, and they're not digital natives. They're only digital natives in the sense of they're digital consumers, but they're on average not digital producers. >> Okay. >> That means >> That's a big difference, I think. >> Yeah. And people like you, you're obviously a digital producer. You are actually a digital native. You know how to not only consume, you know how to build [ __ ] yeah? But most kids don't. And uh so, I think it's nothing much has changed. Anybody who's motivated enough will make a good thing. And anybody else who just serves the tides will fail, and that's always been the case. So, I I don't necessarily know if young people have a huge They obviously have a huge advantage with their neuroplasticity. Um but the problem is more that enterprises don't want to hire young people because usually it takes a long time to to get them up to speed and to become productive. >> Yeah. >> I think that's wrong. I think if you At least in software engineering, a junior plus an agent plus a senior who can mentor that junior, mwah, perfect. We're doing this at at the companies I work with, and it works. It's just that a lot of people think, "No, we don't need juniors anymore.

They're They're just a drain on our resources." It's just stupid. >> I really like the distinction of like creator versus consumer. And, you know, when when you said that, I make the connection to content, right? We're like everybody is a consumer of content, but like only a fraction of people create, whether it's articles, videos, podcasts. So, that same thing is with code. Like you mentioned, most people are not going to build their own app, build their own alternative, and they're just going to consume what others have created. So, they know how to use their phone really well to scroll TikTok or, you know, Instagram, but they don't know how to use the Canva app on their phone to create graphics, right? So, it's going to be the same with code. Speaking of a code, how does your workflow when you're developing something look like? How many agents are you running, which agents? >> Yeah. I mean, I only used pie since October as the agent. Um I generally don't run a lot of agents in parallel. I usually have at the maximum four four terminal windows open. And I have a super strict workflow. Like I have a bunch of prompt templates for different parts of my workflows. So, I have like a prompt template for pulling this issue from GitHub, analyzes it in this and this and this and this way. Give me the analysis, give me an implementation plan or a bug a bug analysis, what what you think the issue is. And the same for pull requests, and all my workflows are basically this. I take an issue or pull request from GitHub, put it in as an argument to my prompt template. Then I go away or start a new session with a different issue or pull request, and once one of these comes back to me with the first analysis result, I can start working on that specific issue. And that work done is usually depending on the scope of the issue. Like if it's someone who wants a new flag somewhere, a new interface thingy, yeah, sure. Implement and drop it up. I don't care. Uh with instructions on how to manually test it. So, in my case, it's a terminal user interface, so it's very easy to test. But the same could apply to a web interface. But the agent can is given a a goal of how it should look like and behave like, and it can actually test whether that's the case. So easy, right? Other things like refactorings I'm still doing by hand to some degree. So I go into the code and modify types and interfaces and APIs just so I get a feeling for things and I still understand the code. Because I found that agents, especially for bigger code bases, they cannot ingest all the things. So they will start fixing things locally and then things get loads globally. So I'm still of the opinion that you as the human need to understand at least how the code fits together. You don't need to understand every single line. Like there's probably 30% of code in Pi I don't understand or I have never read. But it's stuff like an export of the session to HTML. I don't [ __ ] care. I don't want to read that code. But other code like the agent loop itself, I know by heart. Because that's really important to the functionality of the product. Um so your my my workflow is split between let the agent autonomously slop up a feature or a bug fix because I know it can do it or uh do it collaboratively. Um so it's more a human in the loop thing or do it manually. And it really only it really only depends on how much I trust the agent to be able to to do the job.

How much I put myself into the loop. >> So it sounds like the skill set of software design and you know, code base architecture is going to be more and more important than specific coding. >> Oh yeah, definitely. Like I don't care for syntax anymore, which is very nice. I can also use languages I haven't used in a long time or I have never used because I understand, like I've built compilers and debuggers in a previous life. I know how everything works. I just never learned the syntax for example. Um but as long as I know the semantics, I'm good. I can let the agent write in any kind of language it wants to write. Um so yeah, definitely. Architecture and system level thinking is probably going to become even more important and that is a warning to the young people. I think you can only learn that if you do some of that by hand. And I'm not saying write the code by hand, but design the thing together using an agent to type it out, but you need to think about the design of the thing. Um so, don't let the agent design things for you because it learned all of that from the internet, and on the internet it's my old shitty code and the old shitty code of a lot of other old people, and it's [ __ ] Like 90% of that code is [ __ ] and that is what your agent will propose the architecture and system should be. Don't let it do that. Do it yourself. Use your brain. >> Yeah, especially for like lesser-known languages. >> Yeah. >> Yeah. So, how important will then taste and judgment be? Cuz like from my experience, AI is really bad at unique ideas, partly because of what you described, where the training data is just, you know, so [ __ ] on average, and then, you know, if you say like give me a business idea, it's just going to be like something maybe slightly better than average, right? So, what's your fault? Can AI be like current architecture be truly creative or not? >> Well, I mean, from a mathematical standpoint, it learned things, so there's a little cloud that represents all the things it learned, and it can interpolate between two points in that cloud, yeah? So, it can come up with things that are combination of things it has seen, and those can be novel. But anything that's outside of the cloud, it cannot do, right? Um because it cannot interpolate between a point inside that cloud, so the things it learned, and outside the cloud, things it has not seen yet. Um so, it depends. If you bring up the business idea case, I would not trust a [ __ ] LLM to give me a good business idea, but once I have a business idea, I can trust the LLM that it can uh tell me, "Oh, you forgot about this thing that every business needs." Or "Oh, you forgot about thinking about this legal framework." or whatever. So, it's great at double-checking your human ideas, but I wouldn't trust it with a business idea. It sucks at that. >> So, doesn't that point towards like the ultimate future, whether that's 5, 10, or 20 years away, where like we just give the initiative? Like the seed comes from a human, and then the agents just do everything else. >> Yeah, maybe. I I mean, we're not there yet, but maybe. I I have some theoretical reasons why that's won't work, and it's [clears throat] mostly coming back to training data. Um but, I would like that to be the the future. Like the Star Trek computer uh is what I want. >> [laughter] >> Just give it a like a high-level instruction, and it figures out draws the rest of the [ __ ] owl.

Um but, sorry for cursing. I know you might get demonetized. I hope you don't. >> good. It's all [laughter] good. >> Uh So, yeah. I hope that's the future. We're currently not there yet. And the reason I believe we won't get there, at least not with LLMs, is because of the way the things are trained. Um so, for programming, it's a very good example. We We already said system design and architecture, nah. But, implementing a feature, like a simple kind of thing, yes. And the reason one thing works and the other thing doesn't is for one thing we have the training data. I We have a lot of traces of im- features being implemented. Um that can be Codex or cloud code sessions since April. Or commits on a GitHub repository. What we don't have is the design process of of systems and architectures kind of encoded in a way that >> kind of stuff. >> Yeah. >> Yeah. >> So, that's why I think we might get there, but we probably won't because we lack the training data, and it's really hard to generate training data for that. >> This is a very interesting take. So, basically, the top 0.01% of humans, whether it's in software, whether it's in music composers, you know, greatest entrepreneurs, a lot of their understanding and skills is not written, and it's not to turn into markdown files and tokens. So, it's very hard to train the models on that. >> Yeah. >> Yeah, that's actually a really good idea. >> Yeah, and it's not even the top 1%. You have to It's not even about is it in the training data or not, it's also how often is it in the training data. >> And the average stuff is much more present in the training data than the like high-quality, high-signal stuff. >> Yeah. >> And that's also the problem with code. Like 90% of the code that is in the training data is garbage and the 10% which are pristine, awesome, amazing, they exist, but they're they have so much little more statistical power during training that they don't get reinforced into the model weights. So, that's one problem. The other problem is that some things are really hard to encode as tokens. Like when you come up with a human business idea, how would you encode that process as tokens? >> Impossible. >> Yeah. >> It's like [laughter] all your life experience, all taste, you know, yeah. >> And and that's why I think ultimately LLMs and agents will make us super productive, no question about that. But the squishy human parts, the things you just described that make you you and make your business successful, that's hard to encode in tokens and that's also why it's hard to train a model to do the same. >> All right. I think it's great to end on this. Appreciate Mario. Where should people go? >> Um pi.dev if you want to use the Pi coding agent. There's also a Discord and we shoot the [ __ ] there. So, come say hi there. And otherwise, Twitter, I guess. But Logic Games on Twitter and yeah, pretty much that's it. >> All right, I'm going to leave all those links below the video. Once again, appreciate it and thank you for your time. >> Thanks for having me.