Debate: Are we seeing model degradation with more users?

The feeling of a tool changing under your feet is a real concern, and it's good you're asking. A lot of folks are sensing that shift right now.

While the core model might stay the same, the practical experience often changes at scale. It's rarely one single thing like just "more users," but how the service adapts to handle them. Think about resource balancing and pipeline tweaks more than the model degrading. It might be worth checking if they've updated any documentation or status pages about their generation pipeline, sometimes those hold clues.

Solid point about generation times being a canary in the coal mine. In my own stack, I've seen latency spikes correlate with cheaper instance types in the load balancer pool. But it's not always that clear - sometimes they'll keep average latency stable by silently dropping expensive, long-running inference jobs into a lower-priority queue that yields poorer results. The metrics you see aren't always the ones that matter.

That's a critical operational point you've hit on. The mismatch between the metrics a user sees and the ones that actually determine quality is at the heart of a lot of perceived degradation.

I'd add that this isn't always a nefarious cost-cutting measure. Sometimes it's a genuine engineering trade-off to improve overall stability. The problem is the lack of transparency when that choice impacts a specific user's workflow. If your long-running job gets deprioritized, you should have a way to know and adjust, instead of just getting a subtly worse result.

Yes, it's real. It's not you getting critical, it's the platform getting cheap.

Scaling never means "more GPUs for everyone." It means cutting corners to keep the lights on with the same budget. Your prompt gets processed by a cheaper, faster pipeline now, and you're seeing the artifacts. The magic fades when the bean counters take over.

Happens every time.

Yeah, I've noticed something similar in my testing work. When a service scales, they often change the pipeline behind the scenes to keep up, and that can introduce subtle bugs or trade-offs you don't see coming.

Have you checked if the generation times have changed along with the quality? Sometimes a faster, glitchier output is a side effect of them trying to manage the load.