The year is coming to a close. It’s a great time to reflect on wins, losses, learnings and preparing where to adjust for the future.

I’m writing this as a guide—what I’d tell myself a year ago.

You are probably launching too late.

At Hugging Face, things move fast. Very fast. We ship and launch products all the time. Keep in mind, this is with millions of users. The most valuable metric is usage and user feedback. Everything else tends to be procrastination.

It’s best to ship something, even if not 100% ideal, to get an idea whether it resonates with users. If it works, double down. If it doesn’t, pivot, iterate, then launch again. Most importantly, don’t be afraid to cut bait and kill products that aren’t getting the level of usage it should.

The traditional one-day launch is dead. Products are being launched all the time. Most likely, no one will see or care to look at the first thing you put out. You need to keep showing up.

A great reminder that nobody cares. Just win baby.

It needs to be Simple.

We’re drowning in AI slop. There’s a v0 or Lovable for anything you can imagine. Yet some of the biggest winners of 2025 did one thing and did it exceptionally well.

The best products I’ve used this year are remarkably simple. No feature bloat, just one value prop they absolutely nail.

A few that come to mind:

• Claude Code: give Claude some tools and a terminal
• Conductor: run Codex / Claude Code in parallel on your Mac
• Greptile: AI that reviews your Pull Requests

They all center around one workflow—easy to market, easy to communicate, easy to get value from.

Jevons Paradox is legit.

Jevons Paradox states that when technology makes a resource cheaper or more efficient to use, consumption of that resource actually increases rather than decreases.

In the early LLM days, the consensus was that the price of intelligence would trend toward zero. This is playing out—but we keep finding more ways to throw tokens at AI. While cost per token is plummeting, total spend is skyrocketing.

Source: Tokens Are Getting More Expensive by Ethan Ding

Some people are burning through 250B+ tokens in just a few months. As models get cheaper, we don’t use less—we use far more.