Ok, I thought I was going insane. The last two larger coding tasks I gave Claude Code it left about 35% of my request completely undone or done sloppily.
I because of this, the next task I gave it on the larger side, I ran its work through Codex which identified 7 glaring unfinished parts of the task.
The trend was starting the part of the task but then leaving a "skeleton" of what I has requested without any of the actual working parts.
The way I would describe it is a kid cramming his 3 month project into a Sunday evening for Monday's due date.
I created an account today to ask "Why?" -- Why are you using this tool? It's consistently producing subpar work, to the point that you're using _another_ (probably equally inferior tool) to compare the previous output?
This is something I see all the time with AI consumers and I am continuously baffled. If anything else (autocomplete, intellisense, etc.) produced this much garbage it would be immediately abandoned. Why is there such a high tolerance for the chat bot equivalent?
i run claude code pretty heavily for overnight sessions and yeah the inconsistency b/w runs is noticeable. same prompt, same codebase, wildly different quality depending on the day. the frustrating part is when it half-finishes something and you come back to a mess you now have to untangle. still the most capable coding agent i've used but the variance is real.
Anyone expecting a higher tier subscription to be announced since this current reduction?
Cynicism aside - I do wonder what the future will hold given that current token burn rates aren't sustainable without VC cash. Anthropic even pushed us to use haiku for claude code for "many" tasks in our enterprise training, so I'm wondering if it's not a company need of sorts to reduce the burn?
yes I have been facing this issue and all aspects of task execution has become lower in quality, I have moved to opencode.ai for all my routine repeated tasks for now -- a bit slow but works well on the free model, looking for alternative for difficult questions to run in terminal
11 comments
[ 2.7 ms ] story [ 28.8 ms ] threadIn reality as they scale up, the models lose nuance and become noisier. The boosters do not want to admit this.
We need highly-specialised models/interfaces. Not one thing and trying to force-fit it.
https://news.ycombinator.com/item?id=47664442
I because of this, the next task I gave it on the larger side, I ran its work through Codex which identified 7 glaring unfinished parts of the task.
The trend was starting the part of the task but then leaving a "skeleton" of what I has requested without any of the actual working parts.
The way I would describe it is a kid cramming his 3 month project into a Sunday evening for Monday's due date.
This is something I see all the time with AI consumers and I am continuously baffled. If anything else (autocomplete, intellisense, etc.) produced this much garbage it would be immediately abandoned. Why is there such a high tolerance for the chat bot equivalent?
Cynicism aside - I do wonder what the future will hold given that current token burn rates aren't sustainable without VC cash. Anthropic even pushed us to use haiku for claude code for "many" tasks in our enterprise training, so I'm wondering if it's not a company need of sorts to reduce the burn?