# When ideas expire

Author: Florent Martineau
Language: en
Published: 2026-09-25T00:00:00.000Z
Updated: 2026-09-25T00:00:00.000Z
Canonical: https://florentmartineau.com/en/blog/when-ideas-expire/

> It’s the cache invalidation of mental models that I care the most about. That’s where I really see the lag.

[David Heinemeier Hansson — Rails World 2024](https://www.youtube.com/watch?v=-cEn_83zRFw&t=265s)

At the Google cemetery, an engineer offers to revive old projects. The dead volunteer, but he turns Google+ down.

GOOGLE CEMETERY  
With AI, some projects deserve a second chance!  
ME! ME! ME!  
Google+ : Me too?  
I said some.

At a hackathon in 2017, I presented CoursetomiZed, an adaptive learning project. The goal: move quickly through concepts the student already knows and spend time explaining the rest. The project won first prize at [EDHEC’s Pedagogical Innovation Challenge](https://www.edhec.edu/fr/news/ledhec-organise-son-1er-pedagogical-innovation-challenge).

Then came preparing the courses. An expert had to write several versions of each explanation for different student profiles, then map out the concepts and their prerequisites. Every variation had to be prepared in advance. The workload was too much, and the project failed.

With [Learn Your Way](https://research.google/blog/learn-your-way-reimagining-textbooks-with-generative-ai/), introduced in September 2025, Google starts from a textbook and generates explanations tailored to the student’s level and interests. No more writing every variation by hand.

Perhaps CoursetomiZed wasn’t a bad idea. It was just too early. I’m considering bringing it back to help with my return to university.

How many other ideas have we put in a drawer without ever thinking of taking them out again?

## Our assumptions have a shelf life

In his Rails World 2024 keynote, David Heinemeier Hansson talks about **“cache invalidation of mental models.”**

There’s a fable about monkeys being punished for climbing up to get bananas. They teach the others not to climb. The punishment stops, but the ban remains. Generations later, the monkeys still won’t climb up for those damn bananas!

DHH’s example reminds me of it. Developers adopt tools to work around browser limitations. Even when the browser can do without them, we keep using those tools. And there we are, losing hours to a webpack configuration and questioning every life choice.

[DHH — Rails World 2024, starting at 3:11](https://www.youtube-nocookie.com/embed/-cEn_83zRFw?start=191)

[DHH — Rails World 2024](https://www.youtube.com/watch?v=-cEn_83zRFw&t=191s) · Starts at 3:11; the cache remark is at 4:25.

A new AI model comes out every 28 seconds. Can our habits keep up?

## “We already tried that”

We rule out an AI feature because the model makes too many mistakes, responds too slowly, or costs too much. A few months later, a new model might do the job. But we have to try again.

“We already tried that. It doesn’t work.” Sure. But when? And with what?

The model fails at a task, so we split it into smaller tasks. The system becomes more complicated, but more reliable. A new model might handle the whole task just as reliably and at a lower cost. Simpler, cheaper. A win all round. But we have to remember to simplify. Boris Cherny takes this reasoning further: at Anthropic, he says they build for the models they expect in six months. **Why spend six months working around a limitation that might be gone by launch?**

[Boris Cherny — Y Combinator](https://www.youtube-nocookie.com/embed/PQU9o_5rHC4)

[Boris Cherny — Y Combinator](https://www.youtube.com/watch?v=PQU9o_5rHC4)

What didn’t work yesterday may work today. And what was necessary yesterday may no longer be.

An assumption that holds today may be wrong in six months. What about assumptions from fifty years ago?!

## Let’s bring some ideas back to life

We’ve put decades of ideas in the trash 🗑️. Some had good reasons to end up there. Before clicking “Empty Trash,” let’s see whether a few deserve to come out.

I’d like to revisit AI history from this angle. What did we abandon for lack of data, computing power, or human time? And what still stands in the way today?

Take expert systems: they apply rules to draw conclusions. CoursetomiZed worked this way to estimate a student’s knowledge. Those rules are expensive to write and maintain. Could an LLM automate that work? I’ll return to this in a separate article.

## Conclusion

In 2017, CoursetomiZed needed too much preparation. Can AI solve that problem and make CoursetomiZed viable?

“I already tried. It didn’t work.” Yes, in 2017. It’s 2026: I can give it another look.

**Which of your projects would you revisit today?**