The Freshness Half-Life
Here's a lens AILmanac is built around, and the single most useful idea we can hand you about working with AI:
Not all facts about AI rot at the same speed. So you shouldn't treat them the same way.
The price of a model, the name of the "best" model, how transformers work — these are all "facts," but one of them changes weekly, one changes monthly, and one hasn't really changed in years. If you store all three in your head with equal confidence, you'll be confidently wrong about two of them within a month.
A model we find useful: give every fact a freshness half-life — a rough sense of how long until it's 50% likely to be stale — and then decide how to handle it based on that half-life, not based on how important it feels.
The core move: match the handling to the half-life
There are only three things you can do with any fact:
- Cache it — keep it in your head and act on it without re-checking.
- Re-check it — assume it may have drifted; verify against a source before you rely on it.
- Link, don't memorize — never store the value at all; store a pointer to where the live value lives.
The mistake almost everyone makes is caching things that should be linked. They memorize a price, repeat it in a meeting six months later, and are wrong. The freshness half-life tells you which move to make.
The tiers
A model we find useful is to sort AI facts into four decay tiers. The boundaries are fuzzy on purpose — the point is the gradient, not exact cutoffs.
| Tier | Half-life | Examples | What to do |
|---|---|---|---|
| Volatile | Hours–days | Exact prices, rate limits, queue/capacity status, live UI labels and button names, "is the API up right now" | Link, don't memorize. Never quote from memory. Always pull the live value from the official source at the moment you need it. |
| Fast | Weeks | Which model is "best" or "default" for a task, recommended settings, which features are in preview | Re-check before relying. Fine to hold a working assumption; verify before you bet anything on it. |
| Slow | Months | Which features exist at all, tool and product names, broad capability tiers ("can it see images?") | Cache, but stamp it. Keep it in your head, but attach a mental "as of roughly when" and re-verify when the stakes are high. |
| Durable | Years | How transformers predict tokens, why context windows matter, prompting principles, the Trust Ladder, this very framework | Cache freely. These are the foundations worth actually learning. Investing here compounds. |
How to read the tiers
The trap is tier confusion — treating a volatile fact as if it were durable. "Claude costs $X per million tokens" feels like a solid fact, so people cache it. But its half-life is days, not years. The fix isn't to memorize harder; it's to move it down a tier in how you handle it: link to the source instead of quoting a number.
The opposite mistake is rarer but real: re-Googling durable things ("how does an LLM work?") every time instead of just learning them once. If a fact's half-life is measured in years, stop re-checking it and internalize it.
A quick test when you catch yourself stating an AI fact: "What's the half-life of what I just said?" If the honest answer is "days," you should be pointing at a source, not your memory.
Why this matters more for AI than almost anything else
Most fields have a comfortable mix of durable and volatile facts. AI is unusual: a large share of the "facts" people repeat are in the Volatile and Fast tiers — model names, prices, limits, which thing is best this week. The volatile-to-durable ratio is high. That's exactly why so much AI advice ages badly, and why a freshness-first habit pays off more here than in slower-moving fields.
It also explains a counterintuitive rule: the more confident a tip sounds about a specific number, the more suspicious you should be of it. Specificity and freshness are often in tension. "Use the cheapest capable model" is durable advice. "Use model X at $Y" is a screenshot of a moment.
How AILmanac itself runs on this idea
This isn't just theory we're handing you — it's the operating system of the whole site.
- Volatile facts are never hardcoded into prose. Prices, limits, and current model names live in one place —
data/models.json, our single source of truth — and on the Models & Pricing page, which we keep verified. Everywhere else, we link to those rather than repeat a number that would rot. - Every volatile page carries a
<VerifyNote>stamp recording when a human last checked it against the official source, like this:
- The Freshness Dashboard scores every page against its last-verified date and your current clock, so the site can't quietly go stale without showing it. That dashboard is the Freshness Half-Life made literal: it surfaces which pages have likely decayed past their half-life and need a human to re-verify.
- When in doubt, we point at the official docs instead of copying a value. A link to a maintained source has a much longer half-life than a number pasted into a paragraph.
If you build your own notes, prompts, or internal docs on top of Claude, steal this directly: keep your volatile facts in one stamped, linkable place, and write everything else to outlast them.