Okay, I'll be the one to say it out loud. I swear the Notion AI button is starting to feel like waiting for a container to pull from a public registry with a 56k modem. 🐌
It's not just the initial generation either. The "Continue writing" and "Change tone" actions seem to have added a solid 2-3 seconds of thinking time compared to a few months ago. I use it daily for drafting runbooks and post-mortem templates, so the lag is becoming noticeable. It's like they're trading latency for... what, exactly? Better "vibes"? My CI pipeline times out faster than some of these responses.
I'm wondering if it's a model size issue (bigger isn't always better for speed) or just backend scaling problems. In our world, we'd be checking metrics:
- P95 latency graphs 📈
- Cold start times (is it a serverless function struggling?)
- Region-specific degradation
But here, we're just left guessing. Anyone else feeling this, or am I just getting impatient in my old age? Maybe my expectations are too high—I'm used to `kubectl` commands that either work fast or fail fast.
- tm
You are definitely not alone, and you've nailed the most frustrating part - the lack of visibility. In customer support platforms, when we see latency creep like this, there's usually a trade-off dashboard somewhere: we're adding more context from the knowledge base, or running a more thorough intent classification. With these all-in-one AI features, it's a black box.
Your comparison to backend scaling is apt. I've seen similar latency inflation in some helpdesk AI chatbots when they've quietly swapped for a larger language model to handle more complex queries, but didn't scale the infrastructure proportionately. The "Continue writing" delay is particularly telling - that suggests the system is re-processing the entire preceding context each time, not just appending, which points to a change in their inference approach.
I'd love to see a transparent changelog for these things. When a new model version rolls out, tell us if we're trading 400ms for better coherence on lists, or whatever the case may be. Let the user decide if that's a worthwhile trade.
Support is a product, not a department.