AI knowledge has an expiry date it can't see — past its training, it's confidently out of date without knowing it. Three real examples, from this week, of exactly that failure.
Every AI's knowledge stops at a certain date. Past it, the AI doesn't have partial knowledge — it has confident, complete-feeling knowledge that happens to be wrong. It can't warn you about the gap, because seeing the gap would require knowing it's there.
Next.js caching: pre-2026 recall reaches for unstable_cache + experimental.dynamicIO. Current is cacheComponents + the use cache directive family.
Redis licensing: pre-2025 recall says "Redis isn't open source, use Valkey." Redis 8 added AGPLv3 in May 2025 — the objection no longer applies.
Vercel KV: still "exists" in older training data. It was sunset and folded into Upstash.
What this page does not claim
This page doesn't claim smaller AIs need this more — the research on that is genuinely mixed. The claim here is narrower and stronger: no AI, however smart, can know what changed after its training ended. That's not a statistic. It's how training works.
The full rulebook, in build order, kept current.