So the CEO dropped this "AI-native" term again. It's starting to smell like "cloud-native" did back in 2013—a buzzword that means whatever the vendor needs it to mean at the time.
In our world, "cloud-native" eventually got a rough shape: containers, twelve-factor, declarative config, orchestration. You could point to a `Dockerfile` and a Kubernetes manifest. But "AI-native"? Is it just slapping a chatbot on your admin panel? Is it requiring every service to have an inference endpoint? I saw a pitch the other day where "AI-native infrastructure" meant they'd replaced alert thresholds with a model that guesses when things are broken. We tried that. It called every Tuesday morning at 3 AM a "critical anomaly" because of a batch job. So much for that.
I'm genuinely asking because I'm knee-deep in platform stuff. If we're supposed to build for this, what are the actual primitives? Is the "AI-native" app just a FastAPI service with a different kind of autoscaling config? Like this?
```yaml
apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
name: maybe-ai-native
spec:
predictor:
minReplicas: 0
scaleTarget: 50 # qps? or predicted-latency-percentile?
behavior:
scaleDown:
stabilizationWindowSeconds: 60
```
Or is it about the data plane being fundamentally different? Do I now need a service mesh that routes based on LLM-generated canary analysis? Because I can already feel the incident report: "Outage caused by the canary controller interpreting a sarcastic commit message as a rollout approval."
Genuinely curious if anyone has a concrete definition that isn't just "we use AI."