How to Thrive as a Developer

This is an automatic transcript of the recording above. It is published in full and unedited, apart from correcting names the recogniser reliably mishears. It will contain mistakes.

00:06Hi, I'm Tjis Kumar and I work on AI developer relations for Langflow over at IBM. Today I'm really excited to talk to you about AI development and how you can thrive as a developer in the age of AI. In fact, not just as a developer, but as a professional. Um in our time together, we're going to look at the shift that's happening um between AI and more, I'd say, traditional ways of working specifically around coding, but it applies to a broad variety of disciplines. We're also going to look at this through the lens of science and research to ground ourselves in in reality as opposed to um falling prey to sensationalism and hype. And so um I I'm really excited to take you on this journey with me and towards the end hopefully we can uncover clear ways that we can thrive instead of just survive in an age of AI. Uh to get started um I'd love to just start by drawing a parallel um historically to the precomputer days back when there was a lot of paintings.

01:07Let's talk about the Renaissance. The Renaissance in human history was a period of of artistic flourishing. I mean, we've we've all seen hopefully uh various Renaissance paintings. Um, for example, um the one with with God reaching his hand out to Adam in the in the Basilica of of um the Vatican. Anyway, uh the Renaissance art um if you haven't seen it, Google it. It's it's got it's got this trademark feel to it, and that's because it was a time in history where everyone was enlightened and they wanted to paint, right? It was the romantic era. Um, they didn't know though, right, about about the innovation that was coming, the innovation of cameras. And and this is I I draw a parallel here because that's kind of where we are today, right, is um you have coding agents. You've got clawed code, you've got cursor, um they do a lot of the job of an engineer. They just write code and and sometimes it's even better than stuff I write, right?

02:00Um, especially if you're learning a new discipline like my background is is front end or web engineering and um, you know, if if I wanted to get into Kubernetes, I would just have claude code do it and then explain what it's doing. I actually know Kubernetes. I have a Kubernetes cluster in production uh, because of cloud code, right? That that was better than me and that taught me how to do this. And so um, we're facing a similar thing as painters to cameras um where we're coders and there's coding agents. No doubt if we rewind to the Renaissance, there would have been painters who, you know, saw the first cameras and and must have thought, "Oh my goodness, this may replace me similar to us." Um, but then I I want to on a positive note also address some of those painters who saw the camera and thought, "Huh, that's a nifty new tool to do the same thing I do." Right? Because if you if you consider the painter and the camera um both of them do essentially the same thing. They they capture light and represent that light on some type of medium, a canvas, a photograph, a paper.

03:10In fact, some can say painters do superior work to a camera because a camera just observes and captures photons where painters can synthesize um photons. They can synthes they can draw things that are not capturable in reality right similar to diffusion models like flux by black force labs or stable diffusion etc. So um AI now has this capability also to synthesize art as opposed to just capture light. But if we consider portrait artists right they either paint the king or they take a photo of the queen. Um if we consider those skill sets there was definitely a replacement. The point we're trying to make here is tools and and this is maybe one of the I'd say leading ideas of this presentation is that tools may vary but there are also invariants and and by definition invariants do not vary they are constant. So you've got variance that are the tools and you've got invariants that are the constant. If we pull on this thread of painters and photographers, um the tool is the paintbrush, paints canvas. Similarly, the tool is the camera, the flash, the lens. Um what is the invariant? And this is where we start to understand where we sit as developers in an AI world. In the photography example or photography parable, if you will, there are invariants. The invariance usually come from first principles thinking which is really when you strip a concept down to its invariance. Okay. So if we take a

04:46first principles approach to painting and photography trying to identify the invariance, we'll see that the main invariant here is just capturing and retransmitting light. Right? And so let's pull on that one level more. What is the invariant here? The invariant is or the invariants are light in the world. There's just there's light. There's light around us. Um we need some type of medium that encodes this light. Um and that for us is air. It's literally the scattering of photons through air. In fact, the planet Earth appears blue because of atmospheric scattering, right? And so we have this medium that encodes it and we need a medium that decodes it in in our case of light. um the human visual system or the mamleian visual system. Mammals, even the insect, like all visual systems decode light in their own way. And so there's three invariants here. Light in the world, a medium that can encode it, and a medium that can decode it. These invariants are invariant. They're constant. And we can have a variety of tools to serve humanity along these invariants. the paintbrush canvas or cameras, lenses or even diffusion models and AI models that can synthesize art all but the point I'm trying to make is the tools vary but there are invariants and as long as we understand this as long as we understand invariant versus tool we're already starting to set ourselves up for success. Um everything

06:17else on top of the invariants are just implementation details with trade-offs. An example of a trade-off of a camera is you get photons exactly as they are. Um, so you can't be creative. Even with filters, you can't really be that creative. Um, a trade-off with paint brushes is it's not as precise because you're not directly capturing photons, but the trade-off is you can be as creative as you want, right? And like Dolly, for example, super creative uh artist. And so the the invariance are constant, but the the modalities, the tools can vary. Okay, this is so important. That's why I'm spending a lot of time there. I mean, it's been like 7 minutes, but this is why we're we're doing this. Okay. So, now what we need to figure out if we want to thrive in a world where maybe we're concerned something is going to take our job is we need to hyperfixate on first principles invariance and adjust the tools. It's worth taking a little sidebar here and acknowledging AI will never be an invariant. AI will never be an invariant because if we think about invariants, they're typically laws of the universe.

07:21Their their gravity is an invariant. Everything even outside the earth experiences gravity, right? The earth is held in place by the sun's gravity. So mass and its gravity effects are invariant. Their laws um light is invariant. Um decoding light, encoding light, these mechanisms that have evolved over millennia are invariant. Now if we put AI into this say chat room is it in a variant? Absolutely not. AI varies so much that there's a new tool breakthrough etc every week right like I I've heard a lot of people say they're fatigued they struggle to keep up keep up excuse me so um how can we say AI is not invariant and will never be invariant okay so then if it's not invariant by our current chain of reasoning then it has to be what a tool indeed it is absolutely a tool and since it's a tool there's good news for us because we can just use it to solve invariant problems from invariant first principles thinking So then the question becomes okay if AI is a tool how do we find invariants I've already identified a few but how do we like understand where an invariant is and the best way to do this is to invoke what is called nothers or nas in German theorem um nother's theorem states that when you see a symmetry from a variety of angles it means that there is an invariant there and this may be too abstract but let me let me put this into if if you for example observe a spinning spinning

08:48top, right? This thing's just like spinning on a tabletop. If you look at it from the left side, it's still spinning. If you look at it from the right side, it's still spinning. Um, that symmetry is called rotational symmetry. No matter from what angle you look at a spinning top, it's still spinning in a circle. That is rotational symmetry. And according to Nerther's theorem, um, there's a law of the universe. There's an invariant there. Indeed, the invariant is the inertia and momentum that the top experiences. And similarly um there's directional symmetry with gravity. Everything comes down therefore there's a law aka gravity. Um nether's theorem is like where you see symmetry there there is an invariant. Okay let's we've been we've been kind of in the cloud so far. Let's bring this down for us as developers. Um what are some invariants that we solve um as developers? And and if we zero in on the invariance that we solve as developers, we will thrive and we will be irreplaceable because we're solving against invariance and we're using um tools to do that, fallible variable tools to do that. Okay. So um what are some invariants we can identify as developers according to not serum? Well um one very strong invariant is agency.

10:03uh is like human beings, no matter from what angle you look at it, human beings tend to do better with agency and want a sense of agency. We want to be able to trust systems and and trust things and trust people and we want to have agency over our time as well, right? Like I I want to know like the times where I've been the most depressed. This anecdotal u was when I feel like I am just spending all my time with work and then somebody somebody else needs my time and this and everything's on fire and low agency leads to negative health outcomes. Similar for identity. Agency and identity are invariant with human beings. In fact, a 2021 um paper out of Harvard University and the University of British Columbia um titled the positive influence of sense of control on physical, behavioral and psychosocial health in older adults, an outcomewide approach. This paper from Hong and colleagues um was was tremendous to prove this. It was an 8-year long study um by Harvard's human flourishing lab and it looked at 12,998 people. Just imagine 12,998 people. All of these people were adults over the age of 50. And usually they chose that because adults over the age of 50 typically tend to have less agency um than than others, right? They like when you're young, you feel like, oh, I can do this, I can go here, I can do this.

11:25When you're when you're older and you're retired, you tend to feel like you maybe you pass your prime and so on. So they chose this and it's a massive cohort and they took eight years. Um and what they did was in the first as soon as they started they took an assay of these people's sense of agency. Hey how in control do you feel how much agency do you feel you have over your time over your identity and so on. um four years later uh they took another survey same thing how how's your agency and so on and then four years after that so in the eighth year they identified these people's physical and mental health capabilities and what they found was people who reported relative to control so within four years if they reported I feel more agency these people also lived longer had better psychosocial health and better physical health as well like every every mark went up and so we can observe Oberve this I mean 12,9998 people over 8 years like that's many many angles and if we observe this over many angles we can see an invariant no doubt is human agencies I want to feel in control pause and consider yourself like the reason you get frustrated when you have to click on cookie banners right is because you value your time you you want to do you want to get back as much time as you can so then you have the agency to do whatever you want spend it with family go out with friends whatever it may be right and so For us

12:46as human beings, we agency is a strong invariant. We want to reduce uncertainty and risk so that we can preserve our resources. I don't want to lose a bunch of money accidentally and then I can use my agency with the resources I have to do what I want to be happy. Um, we want to get back more time again because time is very closely correlated with agency. If I have a week free, I can do whatever I want. Literally, I I want a week to do whatever I want. That language is agency. Um there was a there was a study by by Google um in 2009 um led by Brute Lag that shows that when they so they did an experiment they doubled the latency um on their search. So from 200 millconds they raised it to 400 millconds just to see what would happen and 0.6% 6% um of people just stopped using search because it was too slow, right? Because it impinged on the agency and said, "Hey, you don't you don't actually you you lose time here and if you lose time, you don't have as much time to do whatever you want, etc., etc." So agency is is the invariant we protect and and it usually um manifests through risk and uncertainty um or time or resources like money and so on. Um let's let's make it really practical and think about this from a software engineering perspective. This means we provide as less friction as possible in terms of UX and allow users to do the most meaningful trustable work in the lowest amount of time. Right? That's the

14:17whole invariant we solve as software engineers. So let's consider a calendar app. Let's make it super practical. We're building a calendar application and and there's a demo coming up here. We build a calendar application. Um, and people don't use a calendar application just to browse and book. Like it's not that simple. It's not like I just want to look at my calendar and book something. It's more I people we all use it to find the best time and then to protect me in case somebody cancels or plans change. Like it's the need a calendar app solves is not just browse and book. It's find the ideal time and protect me in case of uncertainty. If we consider this example even for like travel booking, right? It's the same thing. It's not just like I'm going to go find a flight and pay for it. It's more I need to find the best flight at the right price point and I need to be covered in case the flight is canceled.

15:09Like we always want the most optimal thing that preserves our agency and we need to cons continually preserve our agency in case of uncertainty. Meaning I need some type of insurance. I need to know that this is the best one, the best time slot and in case someone cancels, I'm covered. I get a notification um they don't book an event in parallel to mine etc. So those are the needs those are the invariants we solve. If we identify those invariants we can do so with software. We can solve them with software either by writing code like Google calendar. Google calendar is a great tool or this is where we can even use AI agency um for ourselves to thrive as developers. And so um let's explore both of those now just like by way of a demo here. So, I've got my Let me let me open up my calendar. Um, this is my calendar. Welcome. Um, and as you can see, I have quite a bit of agency. Um, and so, um, you know, I can I could find, okay, what's the best slot for lunch with Paulie? It's probably here.

16:07So, I'm just going to do like lunch, um, lunch with Paulie right here. Okay. So, now I did that and I found the best slot and I know like if it's uncertain. So, that's I've met my need, right? I've I've done it. Um that's that's one tool which uses software or we can use AI which again is not invariant. So I can do the same thing with AI. So let's let's go back here and this time I'm going to use Langflow. A great way I'm going to build I'm going to literally just build my own AI agent, my own personal AI agent for me because I can with Langflow. Langlow is open source. It runs locally. Um you can host it if you want. It's really you do you know. Um, but at it also helps us reason about first principles agents.

16:52That's why I'm going to use it here today. So, um, let's go and add a chat input. Um, I'm going to zoom out to 100%. And then we'll do a chat output as well because I want to be able to talk to my agent, of course. And then we'll do an agent. Um, right here, it's my agent. Hi, agent. And I'm just going to wrap up input and output just like that. And I'll go to the playground. I'll say, "Hi, agent. Do you work?" Um, yeah, cool. Seems to be working. Uh, now I'm going to give this agent access to my calendar. Again, we're solving against the same invariant but in different ways. Okay. I'm going to give this access to my calendar. So, I'll come here. Um, and I'll get Google calendar right here. Um, and this is using a great tool called Composio. So, now what what can I what do I want this agent to do for me? And notice I'm using the word agent here on purpose because the purpose of AI, this is the big thesis, right? This is the point. Don't miss it.

17:51The purpose of any tool is to serve an invariant. The invariant with human beings and software is usually agency. And that's why we have AI agents. We have agents to do the stuff for us to preserve our agency, thus serving the invariant. That's the point of this whole talk. I hope you got it. Okay. Let's let's preserve my agency. So I want you to be able to update calendar list entries and we'll turn on tool mode and we will yeah now we can do many things. That's what I wanted. So we can let's let's select none. Let's see. You can update a list entry. You can update a calendar. You can create an event. Actually I want to scope this a little bit. Um you can list events. You can move events. You can find events. You can um quick add events. I think this should be fine.

18:42Can you insert events? Yeah, cool. Let's So these are the things you can do. Uh these are things my agent can do. Check it out. Just like that. Let's use 40 because it's a bit more capable. And now you know I can go here. Okay. Um sure. March September 2025, right? Um and let's And so it's going to find an event there. Your lunch with Polly is scheduled for today, September 9, 2025, from 1:00 to 2:00 p.m. European Berlin. Cool. Let's say now, let's go back to the calendar. Um, it's it is 1 to 2 p.m. Let's say move it to from like 12 to 1, right? So, we'll say move it an hour earlier. Um, in fact, let's just open this in split view, right? Um, and the agent that I'm building for myself should just do job. So, um, let's watch here maybe. There we go. So, it created a duplicate and made it slightly longer.

19:39Let's say, um, delete the one at 1 to 2 p.m. And I I I want to focus on this on purpose because there's some really great points here. Um, see this? You see this? Delete the one starting at 100 p.m. And now, um, come on. It's going to find the event and it seems it's already been deleted. I think it I just didn't give it permission. So I All of that, by the way, was super intentional. I wanted you to see that. Um because what what let's go what what is the invariant? The invariant is my agency and my time. Um did this exercise, this demo serve my time or my agency? It yes and no. Right? That's kind of where AI is today. It it it did in that cool I could make a calendar event by talking or or by typing. Um no, because the AI messed I made a double one and then didn't know how to delete it. And so then I ended up losing time and losing agency. This is exactly where we find ourselves with AI today. Um all of this was on purpose, right? is we get out of AI service towards the invariant as good as our input and our tools. This is also where we find the space for something called context engineering.

21:03Unfortunately, we don't have the time to go into that. If this was like an hour presentation, we'd spend a deep dive on context engineering here, but with AI as a tool, we need to know that garbage in equals garbage out. If you can provide really great context for example I could have provided the right context in my prompt and said do not create a new event move the existing event etc etc I could have done that but then there's also a time cost there that eats into my invariant my agency so I want to do more with less and in this case if I think about context engineering it's doing more are you understanding the nuance here I hope you are because that's the whole point of this talk um ultimately all software must serve the invariant of human agency. Um and and it does that by faithfully representing state. This is your calendar. Faithfully and reliably transforming state. I'll make an event.

21:58I'll move an event. And faithfully preserving the main invariant in between state representation and state transformation. I hope that's clear. That is like the entire purpose. The tools do not matter. AI can do it. It will be able to do it way better over time, right? People are saying AI is in a state of exponential growth. So what you just saw will probably not happen a week from now. Um but where it is right now is it is a tool and a means to serve the invariant. Right now you might decide, you know, cookie banners, Google calendar is better. Cool. That's fine because ultimately we're focused on the invariant. Um, as software serves the invariant, you also need to have a deep your product, whatever it is you're building as a developer needs to have deep knowledge of the domain specific invariance as well. Um, some things are just non-negotiable. For example, you're working in a bank. Um, and what must be invariant is user privacy, right? Like you must not you must store probably all your secure tokens on some type of secure hardware enclave. um so that nobody can read or write like face ID data. That's a domain specific invariant that you must also preserve and protect.

23:09Um another invariant is that good software is typically item potent because clocks drift, packets are lost, networks partition. Um and in the case of a message arriving twice, if it's the same message, you need to maybe do the operation once. You need to have item potent and durable systems that can recover from errors. Another invariant as we talk about laws of the universe is that things go wrong. Network connections drop, packets get lost, clocks drift. And so in light of those invariants, how can we serve our users? Um, finally, there's something to be said for authentication. What I just ran this example was local on my device. um nothing left my device except an integration with Google calendar but the entire AI agent lang flow whatever is running on on my device openai just generated language right um you may obviously need more security instead of talking to open AI you may need to use something like lama or vlm langflow has support for lama but that's another invariant is identity how much identity are you willing to share etc again this is going to be different for healthcare versus you know consumer apps and so on so um that's something to think of. I'd like to start wrapping up by giving you a checklist for builders across a variety of form factors. Um really to serve the main question of how are we serving our invariant with software which is personal agency. Um and and

24:33this is what you need to identify anytime you want to solve a problem keeping in mind that AI or cursor is just a tool. Um you need to have answers for these. Question number one, what are our domain level invariants? I just gave the example of banks and hospitals. What in your specific domain are invariants that you need to engineer against? Number two, um where do we lose time uh with authentication with maybe language model generating a lot of text? Where do we lose time and how can we bite back? Because that again serves agency. The more time you save your users, the more agency they have, the better you build software. Number three, um what in my application needs to be trustworthy and proven to be trustworthy? Something like authentication with pass keys, you need to prove or something like end to end encryption for a journaling application, right? This needs to be provably trustworthy. You need to say anyone can verify the integrity of this that nobody can read notes in transit um endto-end encrypted notes. Okay, so those are the three questions. What are my domain level invariants? where do I lose time and how do I buy it back? Two, that's two questions but one. And third is what do I need to be trustworthy and provable? Um that that it's trustworthy.

25:48Okay. Um let's wrap up by talking about AI as a tool. AI no doubt is great. Um but again we're talking about AI in the context of a tool and as invariance. Um many of you here are using cursor. Um and let's think about this in in the context of invariance. And first principle is we use cursor because the promise of cursor is it preserves our agency by giving us time back right I don't have to think a lot I just like vibe code I just say hey make this thing and it makes it um the the problem is um while it does give you a 10x speed boost you can say um it also gives you a 10x speed boost to ship the wrong thing right uh and so we lose time curating cursor's output which may be the same amount of time that we spend hand coding things instead of vibe coding things. You know what I mean? And so I would love to see maybe I perform it myself a randomized control trial just a completely clean research experiment um done with a vibe coding group and a non-vibe coding group um just to see like who actually gets it done faster because my thesis right now for complex problems is that cursor gets enough things wrong and you have to manage the agent. um that eats similar amounts of agency or time as hand coding things similar to this calendar example I shared here and so um are we going to be replaced by AI how well do we serve the invariant that that should be the question okay um similar many of you are

27:16using GPT chbpt serves the invariant of agency by giving you control for example you ask a question about something you don't know explain to me the citric acid cycle right and chbpt will give you some authoritative answer and you feel like your agency is served because you have more of a sense of control. Um the danger here is it can hallucinate and be confidently wrong. In fact, it was in preparing for this. It was many many times confidently wrong. I I did use it. Um and so then you've got to go check the sources. You've got to verify. You still lose time and control because you're just doing the work anyway. So what is the net agency win there? All right. Um and this I I will I'm I'm wrapping up. I promise the main question to ask the main question to ask when building software and when we want to thrive as professionals is what is the net cost to my invariant that's it and if a tool has a lower net cost or a higher net benefit it wins right now AI does not have a lower net cost at least in the examples I shared and don't even get me started on trust right like a lot of people chat GPD has a great um agent feature where it can literally like book tickets for you.

28:33People don't use it because it's I mean how do you trust that with your credit card information? So there needs to be more engineering done towards trust again in service of the invariant aka personal agency. In fact many of you listening to this I hope you feel entrepreneurial enough to build that yourself. You can you can just build things and I I encourage it. Okay. Um last thing before we wrap up I promise is I wanted to explore the question maybe we'll do another talk about this um taste a lot of people say taste is is the thing that separates humans from the machines taste you know um is taste invariant meaning if you look at taste from a wide variety of angles according to theorem um is it constant and is there some type of law around taste. I don't know. Uh but I do know that Apple, for example, has been known for its taste. We all love iPhones because of the taste that goes into it. We love a specific style of UI and UX because we can tell it's tasteful. We like specific dishes and food and snacks because they are to our taste. And it it's an interesting question because is taste just the aggregation of the majority or is it something more? That's a question for further exploration. I have honestly nothing to share here. I just thought it was an interesting one. Um if you have any thoughts, please come up to me and talk. I'm I'm here. Um or um if you wanted to leave a comment that that'd be

30:04great as well. Um let's let's wrap up and then and then we'll end this. Um the main the main wrap-up is this. Um how do we thrive as professionals? The main answer is to identify invariance and then solve them from a first principles approach. Um the the thesis of this talk is that the chief invariant with software engineering is personal agency. Um and there are multiple tools to solve those problems. AI is one. Um but there are others. If whoever can solve against an invariant for the lowest possible cost wins.

• 2026

  The New UX

  CityJS London 2026
• 2026

  Frontend after AI: The New UX

  Future Frontend 2026
• 2026

  Harnesses in AI: A Deep Dive

  AI Engineer Europe 2026
• 2026

  AI Yesterday, Today, and Tomorrow: Extending AI Systems with Model Context Protocol

  How to Web 2025
• 2026

  How to Thrive as a Professional with AI

  React fwdays 2025
• 2026

  How to Optimize Your Health with React Native

  React Universe Meetup Berlin 2026