Chat no longer provides hyperlinks to the files that it is recommending updates to.
When instructed to provide hyperlinks, they don’t work.
When given very clear prompts it ignores the instructions and does the exact opposite that it was instructed to.
Starting new chats is useless, as new chats maintain the same errant behavior as prior chats.
.cursorrules are being entirely ignored.
What’s going on?
It’s definitely broken. I thought maybe it was Claude, but something else is going on here. Cursor + Claude 3.5 Sonnet used to be an OP combination, now it just breaks half the time, deletes code unnecessarily, lines of code are replaced with placeholder comments like // existing functionality goes here
On Claude.ai I was seeing notices about it using concise mode due to demand. Cursor has been less responsive for me lately too, so maybe I’ll try the new 4o.
Plausible sounding speculation I heard during the first “Claude is getting dumber” debate had to do with differences when running on NVIDIA vs. AWS Inferentia (and/or quantization during high demand.)
Amazon is their largest investor and one of the stipulations for new funding was to train future models on their chips. 
According to The Information, the deal with the Claude chatbot maker hinges on a key condition: Anthropic must adopt Amazon-developed chips for training its AI models, a shift from its current reliance on Nvidia-designed hardware.
https://www.tomsguide.com/ai/amazon-eyeing-multibillion-dollar-investment-in-claude-maker-anthropic-heres-why-its-a-big-deal
Also:
Quantization Matters
Open source models like Qwen 2.5 32B are performing very well on aider’s code editing benchmark, rivaling closed source frontier models. But pay attention to how your model is being quantized, as it can strongly impact code editing skill. Heavily quantized models are often used by cloud API providers…
Quantization matters | aider
I’ve been struggling since at least a couple of days ago. I’m running into issues where cursor suggests code that is no different than the existing code. I also run into situations where the apply button does nothing. Not sure if I’m on drugs, but it also seems to be be losing focus, but I may be allowing that with bad prompting?
Yes. Apply is crapping out.
Which is no fun, as then a new AI has to be trained.
AI Chat is a total mess this morning. Must be hungover.
I am wondering if AI’s continue to train and pick up the bad habits of humans…
Nope it is mostly a cursor specific issue. I recently switched to windsurf ide. Not trying to brag or anything but it gives way way better results than cursor. I am yet to come across a situation were that ide just goes off the rail in an insane way and starts deleting important files like cursor. Both tools use same llm. But the results are vastly different especially in complex large projects. Cursor has just so many app breaking bugs. The team needs to slow down fix it.