My hunch for them closing unresolved issues is because it will cost them a lot of Claude tokens to fix them due to the devs only having a high level grasp of the internal machinery. It’s only natural. They knocked something great out at high speed.
I’ve dealt with something similar in my profession for a project which we rapidly made with Claude’s assistance. More time was spent afterwards understanding the minutiae of the load bearing components so we could actually take back control of it.
I couldn’t imagine doing that for a project that’s spreading like wildfire around the internet with git issues constantly flying in the door.
Kind of feel sorry for them. They’ll be needing a small fortune in Claude donations to actually polish this up. But that’s the price paid when the hard thinking is left to the machine.
This is why I am wary of AI assisted projects. Every code project has a natural flow to its structure that’s created in the reflection of the creator’s natural style and thought process. LLMs will never be able to successfully meld with that perhaps only when there is actual general AI.
CLAUDE.md doesn’t cut the mustard. I’ve found its’ context to be too brittle.
So even though one may guide the LLM, there is never a true sense of a full cognitive ownership of the codebase. The LLM may drift or the dev may may be tired after making countless decisions on the many complex questions posed by the LLM during development, thus some pivotal architectural decisions get glossed over and are forgotten about. Then an issue arises around a particular facet that needs an intrinsic understanding of the whole codebase and your are then in trouble (€€€).
It’s really not sustainable. A LLM driven codebase’s token usage also will scale at O(n^2) as a codebase grows and the context expands.That’s obviously broken as a reasonable programming methodology, except for the money makers who will of course love that kind of exponential spending. We really have to be careful here championing this stuff and to not do the marketing for these AI companies.
Comments 3
Re: Motherboard Prices Set to Rise as AI-Fuelled PC Component Disruption Continues
Just waiting patiently on the price of oxygen to rise.
Satiation for the endless-growth-lords
Re: Pokémon Yellow Voxel Mod Turns the Original GameBoy Code into a Stunning World
@Doktor-Mandrake
Feels like the underpinnings of a damaging brain drain. Turning off the people skilful enough to write the foundational tools benefits whom?
Re: Pokémon Yellow Voxel Mod Turns the Original GameBoy Code into a Stunning World
My hunch for them closing unresolved issues is because it will cost them a lot of Claude tokens to fix them due to the devs only having a high level grasp of the internal machinery. It’s only natural. They knocked something great out at high speed.
I’ve dealt with something similar in my profession for a project which we rapidly made with Claude’s assistance. More time was spent afterwards understanding the minutiae of the load bearing components so we could actually take back control of it.
I couldn’t imagine doing that for a project that’s spreading like wildfire around the internet with git issues constantly flying in the door.
Kind of feel sorry for them. They’ll be needing a small fortune in Claude donations to actually polish this up. But that’s the price paid when the hard thinking is left to the machine.
This is why I am wary of AI assisted projects. Every code project has a natural flow to its structure that’s created in the reflection of the creator’s natural style and thought process. LLMs will never be able to successfully meld with that perhaps only when there is actual general AI.
CLAUDE.md doesn’t cut the mustard. I’ve found its’ context to be too brittle.
So even though one may guide the LLM, there is never a true sense of a full cognitive ownership of the codebase. The LLM may drift or the dev may may be tired after making countless decisions on the many complex questions posed by the LLM during development, thus some pivotal architectural decisions get glossed over and are forgotten about. Then an issue arises around a particular facet that needs an intrinsic understanding of the whole codebase and your are then in trouble (€€€).
It’s really not sustainable. A LLM driven codebase’s token usage also will scale at O(n^2) as a codebase grows and the context expands.That’s obviously broken as a reasonable programming methodology, except for the money makers who will of course love that kind of exponential spending. We really have to be careful here championing this stuff and to not do the marketing for these AI companies.