My take on AI

The past

About four years ago, a dear friend of mine showed me my first programming-related AI. I gave it a simple “Hello World” and then tasked it to slightly modify the code.

The AI broke the code. Not in a subtle way, but so obvious that anyone knowing how to read failed compilation errors would’ve been capable of remedying things.

So I set it aside. About a year and a half ago, another good friend of mine showed me opencode. It took me a day to wrap my head around its mode-based interface, but then the productivity boost hit. In a single day, I finished three side projects I’d never found time to sit down for.

The present

Fast forward to October 2026, and the state of open-weight models is wild.

Tiny models that can do basic programming tasks (LFM2.5), models that are good enough for most home use (Qwen3.8-27B), models that are fast, and affordable with regard to $/1M tokens (GLM-5.3-Flash, DeepSeek-V4.1-Flash). Even a model that recently proved the quasi-Riemann hypothesis (OpenAI).

So the question is: What to make of it?

To me, it’s a three-edged sword:

• As a utility, it amplifies what the wielder can do. There’s Bjarne Stroustrup’s old quip about C and C++ — “C makes it easy to shoot yourself in the foot; C++ makes it harder, but when you do it blows your whole leg off”. In a sense, the same applies to the current LLMs. If you don’t know about the topic at hand, it can come up with something that looks decent enough but is actually riddled with flaws; if you do, it’s an incredible boost.

• It is great at speeding things up. To be blunt, there are few people who can sustain 200 words per minute of typing for hours on end. Yet there are plenty of inference providers where models can attain those speeds.

• The major cause of concern to me is the negative environmental effects of the infrastructure required to run them. Be it noise, water use, or the accelerated burning of fossil fuels, the hidden cost is high and long-term.

As such, my present stance is to use it to automate things that would otherwise take the longest. Anything that can be done quickly by myself will be done by myself instead.

The future

There’s also the notion that AI models are still in their infancy. A normal LLM’s adherence to any programming language syntax isn’t a guarantee, just a likely result due to good training. The next frontier, in my opinion, will be models that can extend their tokenizers and modify their reasoning based upon newly gathered information. That would permit them to come up with far smaller contexts and improved code generation/modification.

My guess would also be that multi-token prediction (MTP), DFlash, DSpark and Uno are stopgaps and in the long run, we’re more likely to see those baked into the output stage of a normal mixture-of-experts (MoE) model.

The takeaway

Four years ago I wrote the whole thing off over a broken “Hello World”. These days it’s a regular part of how I work. What changed wasn’t my mind — it was the tools quietly becoming worth using.

My take is simple: treat AI like any other tool. Judge it by results, not promises, use it where it saves you the most time, and keep enough of your own understanding to notice when it’s wrong. The models will keep changing; the need to know which end of the tool you’re holding won’t.