Building with AI in 2025

Everyone’s talking about AI. Most of the conversation is noise.

Here’s what I’ve learned from actually building things with these tools over the past year.

What works

The boring stuff works best. Document processing, content summarisation, code assistance. Not glamorous, but genuinely useful.

The key is specificity. Vague prompts get vague results. The more context you can provide, the better the output.

What doesn’t

Anything that requires genuine understanding or creativity is still hit-or-miss. The models are good at pattern matching, less good at novel thinking.

Also, don’t trust them for facts. They’ll confidently state things that are completely wrong.

Where we’re headed

I think we’re still in the “faster horse” phase. People are using AI to do existing things more efficiently. The more interesting applications — the ones that create entirely new possibilities — are still emerging.

The next few years will be about figuring out what those are.