It’s late May 2026 and from my vantage point it’s quite unclear whether AI is a speculative bubble, or a Manhattan Project/Singularity arms race. Here I want to take the prosaic, pragmatic angle of understanding (and discussing) what generative AI is good and not good for, right now. It’d be silly for me to predict what hundreds of billions of dollars of capex riding on a quickly changing scientific field could make possible in a few months.
To me, what generative AI is definitely good for right now:
- Copying: this is their fundamental ability. They can take an existing corpus (be it text or code) and add to it or modify it. They can mimic quite effectively the style that’s already in place.
AI is also great at copying from scratch. That is: ask it to do something that has been done many times before, but from scratch. The AI will rely on its training to create an artifact that, at first sight, looks very much like the real thing. This is particularly powerful for evaluating prototypes of ideas, particularly in code.
An instance of copying is expert-level tool use. For example: AI is better than most people at searching the web, or running commands on a terminal.
- Spotting issues
AI is great at noticing possible issues, whether in text or code. They spot possible abnormalities. An extreme case is how effective they have become in detecting security vulnerabilities. But the same point still stands for normal code review, or for exploring issues in chains of idea (or natural language grammar).
I think this is the negative impression of copying: they’re so good at copying that they also notice where something doesn’t fit the mold, since copying is about applying a mold to create something new. But what’s interesting is that they can detect abnormalities or remarkable things in novel contexts, by extrapolating what they already know. AI models can definitely reason with analogies.
- Giving feedback about ideas
LLMs are great partners for discussing new ideas. They seem to grasp the intent quickly, and they can provide context of other ideas that are related to what’s being proposed. I’ve found LLMs to be a good sparring partner to challenge new ideas and chains of reasoning.
To me, what generative AI is definitely NOT good for:
- Simplicity
In my experience, even with careful prompting, a minimalist stack and crystal clear prompts, AI is unable to create elegant code. It might be able to write a single elegant function. But it completely lacks the ability to grow systems in a way that reuses and adapts the whole. Essentially, LLMs see the codebase through a keyhole (even if the entire codebase fits in their context window) and they can only accrete.
I might be a victim of the availability bias here: recently I mostly vibe coded a project and when I tried to simplify the code, rather than turtles all the way down, I found knots all the way down. This has also been the case in other projects where I kept LLMs on a far tighter leash.
The consequences are stark: with little human intervention on the code level, LLMs tend to create massively complex codebases. Massively complex codebases are bad on many levels. For small, personal projects, the extra complexity is usually not a problem. But for software that has to grow in either features or users, complexity is generally the death of a thousand cuts. The next couple of years will be interesting, because we’ll see a lot of AI built software that’s massively complex, and we’ll see in reality if that type of software becomes impossible to sustain economically.
- Originality
As good as AI is at discussing new ideas, it’s quite weak in proposing new ones. And especially about deciding on whether the ideas it proposes are really good.
It’s also bad at sounding original. It’s quite easy for humans to spot text that has been generated by AI, which is quite remarkable.
Synthesis/compression requires originality. And originality requires you to do something beyond following a statistical average of what’s done before. So it makes sense that these are the two weak points, and that they are closely related. Humans, however, create truly novel things out of what already exists. Today, it doesn’t feel like AI has tapped into that.
- Skin in the game
AI has no guts and no adrenals. This can be a great hindrance for delegation of responsibilities (as opposed to delegation of tasks).
If you’ve made it here, I’m really keen to hear your thoughts on this matter, particularly if you disagree. Scary as it would be, it would be also revolutionary to have AI that can create simple software and original ideas. If you’re already experiencing that, I’d really like to know about it, and how you achieve it.