[lmeyerov]

I've been calling this Software Collapse, similar to AI Model Collapse.

An AI vibe-coded project can port tool X to a more efficient Y language implementation and pull in algorithm ideas A, B, C from competing implementations. And another competing vibe coding team can do the same, except Z language implementation with algorithms A, B, skip C, and add D. However, fundamentally new ideas aren't being added: This is recombination, translation, and reapplication of existing ideas and tools. As the cost to clone good ideas goes to zero, software converges towards the existing best ideas & tools across the field and stops differentiating.

It's exciting as a senior engineer or subject matter expert, as we can act on the good ideas we already knew but never had the time or budget for. But projects are also getting less differentiated and competitive. Likewise, we're losing the collaborative filtering era of people voting with their feet on which to concentrate resources into making a success. Things are getting higher quality but bland.

The frontier companies are pitching they can solve AI Creativity, which would let us pay them even more and escape the ceiling that is Software Collapse. However, as an R&D engineer who uses these things every day, I'm not seeing it.

[zozbot234]

> Things are getting higher quality but bland.

"Bland" is not a bad thing. The FLOSS ecosystem we have today is quite "bland" already compared to the commercial and shareware/free-to-use software ecosystem of the 1980s and 1990s. It's also higher quality by literally orders of magnitude, and saves a comparable amount of pointless duplicative effort.

Hopefully AI will be a similar story, especially if human reviewing/surveying effort (the main bottleneck if AI coding proves effective) can be mitigated via the widespread adoption of rigorous formal metods, where only the underlying specification has to be reviewed whereas its implementation is programmatically checkable.

[titzer]

The dark side of this is that everyone has graduated to prompt engineering and there's no one with expertise left who can debug it. We'll be entirely dependent on AIs to do the debugging too. When whoever controls the AIs decides to enshittify that service, we'll be truly screwed. That is, if we can't run competitive models locally at reasonable efficiency and price.

I don't know how this will play out, except that I've been so cowed by the past 15 years of enshittification that I don't feel hopeful.