AI is changing the productivity landscape.

There are great opportunities here, but 99% of “AI speeds up my work by 30x” claims are noise.

Extrapolating from “AI speeds up [X] by 30x” to “AI speeds up any work by 30x” is magical thinking. To see things clearly, we need to understand the psychological effects of FOMO (Fear Of Missing Out), and we need to see the shape of the problems clearly. For starters, speed isn’t to be confused with impact.

One month ago, I posted about ARC-AGI 3 benchmark, about fluid intelligence, and by extension AGI, still being a dream within a dream. Frontier AI models score no less than 1% out of 100. AGI-shaped problems are still way out of reach of latest advancements in AI.

Open research vs Commodity

There are 2 types of problem: Commodity (solved, AI can do) vs Open research (novel, requires fluid intelligence, AI is useless)

On another dimension, there are low-impact problems (low value), and high-impact (high value).

• Bottom: low-impact or low-value problems. Either everyone can do it (table stakes), or there’s no value in solving it.
• Top: high-impact or high-value problems. These are the solutions people want to have and are willing to pay for.
• Left: open research problem. Solution space is unknown. There is no building block. Everyday is exploration with fluid intelligence - that is why AI can’t do it, empirically can’t.
• Right: problems with known solutions or components. Either it is fully solved, which usually belong to the bottom right (table stakes). Or there are known components that can be composed to solve it. The composition is usually unknown, which is why it’s still high value despite the components being all table stakes. AI can help with table stakes component.

Valuable composition requires a cross of product and engineering skills, and the organisational authority to combine them, which makes it rare, but still very tractable compared to open research.

What LLM can do, and how to leverage it

1. LLM is still bad at open research. Any non-technical management that fails to distinguish open research from feature work will find their plans stall indefinitely, and no amount of “wishing” to LLM can make it happen.
2. Since I got to scope and architect my own vision, my discovered rule-of-thumb for productivity is “KISS” (Keep It Simple, Stupid). As long as I scope vision and features that can at least be broken down into components that everyone else is already doing, I can build across web, finance, infra, AI, and Data. I can cross domains freely if I don’t try to be ‘clever’.

Summary

My take on May 2026 AI-augmented development scene is - composition into value is the game, and where you can leverage AI the most. Everyone can vibe simple, commodity components now, and they’re now table stakes. The opportunities lie in composing these components into something of value.

As for open research problems, well… let the researchers do them, and please pay them fairly for advancing the frontiers of human capabilities. AI is a useful tool for them, but far from magic that you can just will things into existence by telling it what you want.