Dragan Stepanović

If you ever need to explain path dependence (or why you need to understand where a system has been to understand what is doing now or how to shift it) here's an illustration.

Brazil! ❤️‍🔥 Serbia! ❤️‍🔥

The Student List had no party machinery. It operated through university assemblies, protests and hiking between cities & towns. They had 2% of national TV political advertising airtime vs 93% for Vučić. This is one of the most remarkable grassroots democratic victories in modern European history.

The achievement of Serbian students overthrowing Aleksandar Vučić can't be overstated. In 2024, they began protesting. In 2025, they transformed a protest movement into civil dissidence. In 2026, they became a political party — and two months later, crushed a corrupt state-entrenched authoritarian.

As skills in writing code and safe refactoring deteriorate with increased use of GenAI, something I observed myself, GenAI becomes increasingly necessary, to the point of becoming a must.

This explains so many things... youtube.com/shorts/jzH18...

How the world changed for me with GenAI code generation: bad code slowly piling up, or bad to, at best, mediocre code piling up 10x faster while the team's mental model of how the system they build and own works dissolves.

I don't know about you, but Someone in the IT Services and IT Consulting industry from Berlin Metropolitan Area viewed my profile on LinkedIn.

Some good things to work on this week as collapses and power grabs surround and the operating instructions of the industrial growth domination and supremacy worldview are further revealed and clarified (remember apocalypse means "uncovering")

He had grown up in a country run by politicians who sent the pilots to man the bombers to kill the babies to make the world safe for children to grow up in.

There's a big difference between emergence (in complex systems) and stochasticity, and there's often confusion of the latter with the former when reasoning about LLMs and their properties.

Love Reality

Believe it or not, code review stops being a "bottleneck" as soon as you stop overproducing code.

"algorithm made me do it"

As a result, this is how I now judge LLMs: Are they a good pair programmer or assistant. Can they follow instructions, can they be taught good design principles, can they be taught not to do stupid workarounds? I don't care how genius the are, good maintainable code requires mastering the basics.

This is also why big up-front planning is not sufficient. You can't predict all the guardrails and architecture rules the agent will hit during implementation (although you should try to anticipate as much as possible IMO). Everything I see points to software engineers being augmented by AI.

And this is why you need to stay close to the code as an engineer. So you can actually understand and contribute to these discussions. You still own the design.

amazing article

One more day, one more "reading code when LLMs are spitting out so much code is impractical, so you should give it up" LinkedIn post.

The world is, no matter how we think it ought to be. You have to be with it. You have to let it be.

Too many of the same speakers at the same conferences.

interdependence > independence www.youtube.com/shorts/Uov2w...

Software delivery measurement vendors would like you to improve the metrics, but not to the point where you change the way of working (XP) to the point that you no longer need the vendor or the measurements.

The Toyota Way isn’t just about small gains. At its heart? Kaizen.

The purpose of all automation and cost reduction we got as a result of technological improvement since the start of the Industrial Revolution was to spend more time on social networks and consuming polarizing and addictive content. The purpose of the system is what it does. (POSIWID).

The right time to intervene in systems where an exponential, reinforcing feedback loop is dominating the system's behavior is when it seems it's too soon. "Let's wait and see" is a guarantee that it'll be too late.

“A numerical goal leads to distortion and faking, especially when the system is not capable of meeting the goal. Anybody will meet the quota (goal) allotted to him. He is not responsible for the losses so generated.” W E Deming, in “The New Economics”

The problem is that people deciding if developers should be writing code are very often the ones who have never had enough experience with writing code to have a good idea what the value of writing code is and the purpose it serves, but they are convinced they know.

I think we just passed oops.

The counterintuitive idea about managing flow is that the leverage is most often much less about speeding up slow parts and much more about slowing down parts that are running too fast when taking into account the whole system.

I've seen comments about accepting LLM-generated code in that we don't/can't understand it and excusing that because: "There's always code in the system that we don't understand"

Except, LLM-gen'd and human-written code are different. One was written by someone with a mental model. Modified by humans with another mental model (correct or not), etc. The other is generated by statistical, stochastic methods with constraints. These are not the same.

Does it ever hit home that the ascendancy of AI is taken to be assured, but the shifting of human society to fit into earth system limits is seen as everything from a head-scratcher to a longshot to an impossibility?