Thinking about doing the same but from other reason.

While the cascade feels smarter than cursor it often introduces frustrating changes. Cursor might not be that smart but seems to be much more predictable.

dumber but predictable > smarter but frustrating

For all those less fortunate: www.youtube.com/live/4toIHSs...

I was also thinking about microrepos or rather increased modularization. It might improve issues with “collateral damage” but I’m anticipating more problems with proper module api usage. It seems solvable though with proper module interface docs.

Well... just today I went full caps "DO NOT TOUCH X ..." and started few sentences with "How TF ...". It was all good in the end but the road was very bumpy.

I'm actually laughing out loud 😂

Same here. Lots of tests, frequent testing, frequent commits. But it’s frustrating it wasn’t like this before.

I think that in the last ~10 days something changed in the agent part that applies the patches. The diffs can look good but applying them often results in totally different changes as if they were “interpreted” and not just applied. I often observe damage to the logic that was not in the diff.

Is this Windsurf?

Paper📰: Evaluating Language Models as Synthetic Data Generators ( arxiv.org/abs/2412.03679 )
Code💻: github.com/neulab/data-...
Data: huggingface.co/Data-Agora

It’s excellent! For me your account is top-of-mind when it comes to AI/ML news.

www.bnnbloomberg.ca/business/tec...

One of the best pieces on the topic I've read this year.

Endless ruminations on this: @josiecox.bsky.social

I agree with most of this but I think that Steve is not extrapolating enough.

The future is not just about ‘better models’ it’s also about better workflow on the technical side (better agents) and on the process side (better code structure / work division).

We are still early in this.