AI and R&D: where does real human value lie?

Lessons learned 8/11 on Cursor after 15 years in software

By default, AI is good at what it knows. The training process conditions LLMs to compress information from the dataset, and therefore to easily retain everything that frequently appears in it: Hello Worlds, common design patterns, etc.

The corollary is that results won't necessarily be great for anything innovative.

And that's actually good news

R&D work can be roughly broken down into:

1. Gathering the state of the art
2. Implementing the existing solution
3. Implementing associated data and metrics
4. Improving or reinventing the solution (the innovative part)
5. Comparing metrics obtained on the same data

Almost all of these steps can be accelerated using AI. The innovative part of R&D must be done by hand, but the time saved on preparation and evaluation is considerable.

What it means in practice

AI won't replace innovation. It will free up time to innovate.

Today, a researcher or developer spends an enormous portion of their time on repetitive tasks: reading documentation, implementing known baselines, setting up data pipelines, writing measurement code. AI can drastically reduce that time.

What remains — the ability to ask the right questions, imagine new approaches, make unexpected connections — is precisely what makes an expert valuable. And AI, by nature, can't do what doesn't yet exist in its training data.

The real skill is no longer knowing how to implement, it's knowing what to implement.

---

Originally published on LinkedIn.