Why do AI models feel like they're regressing?
Simple: mass adoption of wrapper tools (OpenClaw etc.) = flood of low-value compute.
Estimate: 80-90% of frontier model tokens are wasted on:
• Re-steering
• Debugging
• Arguing with the model
• Tech support
• Guardrail gymnastics
Models improved massively YoY, but the infrastructure is choking. Providers throttle → perceived performance drops.
This isn't the models getting dumber. It's digital shrinkflation at scale.
If you're serious about AI alpha, stop using bloated wrappers. Go direct to APIs or self-host where possible.