πάντα ῥεῖ : In the Age of AI

I found myself thinking about the strange relationship between Greek philosophy, feather-shaped clouds, and — of course — AI. After a really long creative drought, December has finally arrived hopefully bringing fresh inspiration and new opportunities!

Welcome to Code Beyond the Earth! I'm Ece, the voice behind the Code Beyond the Earth. Today we’re going to talk about NP-hard problems, artificial general intelligence, clouds and how everything flows through time.

This winter, I decided to improve my photography skills. I still can’t bring myself to carry around my semi-professional Canon, unfortunately I don’t have quite enough courage yet :) but I’m learning the capabilities of my Samsung phone’s camera. Yesterday, today, and tomorrow, I remain obsessed with the sky: clouds, the sun, the moon, the stars… I’m amazed by the rise of Orion, the energy of the full moon, and the hidden shapes behind the clouds.

As a proud honorary member of the Cloud Appreciation Society, I’ve learned how to distinguish different types of clouds — and even how to spot shapes within them, like dragons, feathers, or cats! I’m not a fan of working from the office, but leaving the office before sunset lets me watch the beautiful autumn clouds and think about them.

Today, I wondered whether ChatGPT could help me distinguish cloud types, so I sent it a photo. Not surprisingly, ChatGPT did its best! And while I was enjoying the performance of one of the closest things we currently imagine as AGI, something made me pause and think… There was a time when detecting clouds felt like a real problem — something the Ece of 2018 was actively thinking and reading about. Has it been solved now? Has ChatGPT solved it? So I decided to ask ChatGPT how it distinguishes cloud types. I was on the service bus, the road was shaky, and after eight hours of work, I simply didn’t have the patience to read and figure it out on my own. ChatGPT explained why older cloud-detection methods were considered NP-hard, so let’s go over that.

Back in the pre–deep learning era, cloud detection was often treated as a messy optimization problem. Researchers would set up models like Markov Random Fields or graph-based energy minimization and try to assign each pixel a label: cloud, not-cloud, or cloud type, while keeping everything globally consistent.

The catch? As soon as you added realistic constraints such as non-linear interactions, spatial smoothness, or multiple cloud classes the whole problem slipped into NP-hard territory. Exact solutions were hopelessly expensive, so people relied on slow approximations: simulated annealing, belief propagation, heuristic graph cuts. It was clever work, but ultimately limited.

Then deep learning arrived, and the landscape shifted completely. Instead of solving a giant optimization puzzle, models began learning the mapping directly from data. Architectures like U-Net, DeepLab, SegFormer, Vision Transformers, Cloud-Net, and even operational tools like FMask or Sen2Cor simply look at an image and predict cloud masks in one forward pass. No NP-hard bottlenecks. No global optimization gymnastics. And they do it in real time.

So is cloud detection “solved”? For most practical purposes, yes. The old NP-hard formulations still exist on paper, but modern systems bypass them entirely. Rather than searching for a globally optimal labeling, deep learning learns patterns end-to-end, making inference fast, scalable, and surprisingly close to how humans perceive clouds.

This approach can be summarised as follows: many classical formulations of vision problems: segmentation, pose estimation, stereo, tracking are NP-hard. Deep learning replaces these formulations with learnable approximations, making NP-hardness largely irrelevant in practice. And this summary inevitably leads my brain to create an analogy: imagine a chess player who doesn’t actually know the rules of chess but has seen every possible move, and can instantly recall which move is suitable when shown a board state at time t₀. It’s an appealing analogy, even if it’s fundamentally wrong. That’s not how the story goes but before we correct it, I want to take you back to ancient Greece.

Behind today’s story stands the philosopher known as the weeping philosopher: Heraclitus. In his only surviving work, fragments, he famously wrote:

“No man ever steps into the same river twice, for it’s not the same river, and he’s not the same man.”

According to my seemingly convincing yet completely flawed hypothesis, a computer scientist might say to AGI: Everything that could ever be done on Earth has already been done — same old, same old. Let’s imagine, for a moment, that we could freeze the world, stare at the data, and relive the same day endlessly — the same inputs, the same outputs, the same river.

But what about the river?

What about πάντα ῥεῖ — everything flows?

That is a nightmare for someone who adores creativity, cherishes the small anomalies in the data, lives for possibilities, and is utterly addicted to the power of choice. My mind has begun to trick me, rolling me again and again into the same fear: nothing will change after the dawn of AI. But that fear triggers another thought: evolution has already warned us — organs that are not used eventually atrophy.

If the founders of AGI believe that rivers are frozen and every move on the chessboard has already been played, then what becomes of Master Kasparov?

Who will think beyond the borders of physics and chase the mysteries of 3I/ATLAS?

What happens to art if we endlessly redraw the Mona Lisa simply because it is “perfect”?

Do we really live in a world that repeats itself that much?

And this is where philosophy meets machine learning: the moment we assume the world is fixed, it changes. No model neither human nor artificial survives by treating life as a closed dataset. The river does not freeze. The world does not hold still for our convenience. This isn’t just poetic, it’s a real technical problem.

Deep models quietly assume that the world behaves like their training data: an assumption reality breaks almost instantly. This is why modern research focuses on robustness, domain adaptation, adversarial resilience, and architectures that can update themselves without retraining from scratch. Intelligence isn’t just accurate prediction; it is the capacity to remain aligned with a world that refuses to stay still.

Data drifts, contexts shift, and tomorrow’s clouds will never be identical to yesterday’s. Even if an AI had seen every possible configuration of pixels, patterns, and objects, the universe would quietly rearrange itself in ways that no dataset can fully anticipate. In machine learning, we have a name for this: distribution shift.

The moment the world changes, even slightly; the elegant illusion of “same old, same old” collapses. Models trained on the past stumble on the future because the river they learned from is already gone, replaced by another.

This is why AGI cannot be a perfect memorizer of all moves like my flawed chess analogy suggested. It cannot rely on an encyclopedia of everything-that-has-ever-happened.

Instead, it must learn how to adapt, generalise, and flow with the world — precisely what Heraclitus tried to tell us 2,500 years ago. The river flows, and intelligence - human or artificial - must learn to flow with it.

So as I keep photographing clouds, learning new things, and wandering between code and philosophy, I’m realising something comforting: the world doesn’t let anything stay the same; not me, not you, not even AI. My creativity isn’t disappearing; it’s evolving with every shift in the sky and every idea I chase. The river won’t freeze, and neither will the spark that keeps us curious. Maybe that’s the real gift: we’re meant to move forward, to change shape, to reinvent, just like the clouds I keep staring at.

If these ideas resonate with you, or if they stirred your own questions about change, creativity, and intelligence, I’d love to hear your voice in the conversation. What do you see in the clouds? What do you think waits downstream for us: humans and machines alike?

Join me in the comments; let’s follow this river of thought wherever it leads.

From my world to yours, Ece 🌍✨