04·Drift observatory · re-measured weekly
"It got dumber" — now with evidence
The model behind an endpoint changes: silent checkpoint swaps, quantization changes under load, serving-stack regressions. We re-run the same probe suite on the same configurations every week, so "降智" stops being a feeling and becomes a number with a date. Latest run: Jul 6, 2026.
Weekly Intelligence Index · dashed rules mark serving-change events
A·Event log
Attributed serving changes
Speculative-decoding aggressiveness increased for throughput; quality dip still in effect.
Quantization silently switched int8 → int4 during a capacity crunch; reverted after three weeks of community reports. The canonical "降智" case.
Serving-stack regression in medium-effort routing under load; acknowledged by OpenAI and fixed four weeks later.
Silent checkpoint update — math and coding scores improved without an announcement. Drift cuts both ways.
Chat-endpoint prompt-template bug degraded instruction following for two weeks.
B·Biggest movers · 4 weeks
Recent shifts
| Configuration | Δ index |
|---|---|
| Kimi K2 Thinking · Groq | -1.2 pts |
| GLM-5.2 · DeepInfra | +0.6 pts |
| GPT-5.5 · OpenAI | +0.2 pts |
| GLM-5.2 · Fireworks | +0.1 pts |
| GLM-5.2 · SiliconFlow | +0.1 pts |
| Claude Opus 4.8 · Anthropic | -0.1 pts |
| Claude Fable 5 · Anthropic | +0.1 pts |
| GLM-5.2 · Z.ai (official) | +0.1 pts |
Method: mean of the last 4 weekly runs minus the mean of the 4 before, per configuration at its default effort. Run-to-run noise is ±0.15 index points; we only attribute an event above ±0.8.