hawkes.aiKey ideas
- The one idea — The pile froze at a date and the weights froze after training — past the cliff the model has nothing, and the machinery guesses anyway.
- How it is shown — The frozen calendar; a post-cutoff question meeting the cliff and the guesser filling the hole with plausible confabulation.
- The trap to avoid — Not knowing the cutoff of the model you use — recency questions need search/retrieval routing, not vibes.
- What it sets up — Tools that fetch the present.
Ask a model about last week and there's nothing there to find — so it guesses, fluently, from a world that ended months ago.
The one idea
The pile froze at a date and the weights froze after training — past the cliff the model has nothing, and the machinery guesses anyway.
The model doesn't know today. It's frozen at its last day of school — and everything you now know about training makes the reason obvious in one breath. The pile ended. The weights froze. The calendar stopped. Two freezes stack. The pile stops at collection day — the funnel's cliff from episode sixty-six. Then the weights stop at training's end — serving mode is forward-only, forever. And since training and finishing take months, the cutoff lags the release date: a model shipped today was schooled on a world from seasons ago.
How it works — the demo
The frozen calendar; a post-cutoff question meeting the cliff and the guesser filling the hole with plausible confabulation.
Here's where it bites. Ask about something after the cliff, and there's no drawer to find — but the machine is a guesser, and episode two's oldest law holds: some bar is always tallest. So it assembles the plausible: who probably won, what a version probably added, said fluently, confidently, from patterns that ended before the event existed. The cutoff isn't just a limit. It's a confabulation zone. The modern patches you've seen in products: search tools and retrieval, fetching the present onto the desk at ask-time — the model reading today rather than remembering it. They work, and they're coming in the agent acts. But keep the seam visible: under every patch, the calendar stays frozen. Recency is always delivered, never known.
The trap to avoid
Not knowing the cutoff of the model you use — recency questions need search/retrieval routing, not vibes.
Why it matters — and what’s next
Tools that fetch the present.
The hygiene, then — boring and effective. Know your model's cutoff; it's published, and it's the vintage on the bottle. Route recency questions to search explicitly instead of hoping. And when a time-sensitive answer arrives without a fetched source, file it as what it mechanically is: a guess from a frozen world. The trap is simply not knowing the date — trusting a frozen world about a moving one. And with the calendar understood, one episode remains in the training act: the strange fight over what scale itself does — whether abilities truly switch on, or whether we've been fooled by our own rulers. The emergence debate. Next.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:15 of narration
The model doesn't know today. It's frozen at its last day of school — and everything you now know about training makes the reason obvious in one breath. The pile ended. The weights froze. The calendar stopped.
Two freezes stack. The pile stops at collection day — the funnel's cliff from episode sixty-six. Then the weights stop at training's end — serving mode is forward-only, forever. And since training and finishing take months, the cutoff lags the release date: a model shipped today was schooled on a world from seasons ago.
Here's where it bites. Ask about something after the cliff, and there's no drawer to find — but the machine is a guesser, and episode two's oldest law holds: some bar is always tallest. So it assembles the plausible: who probably won, what a version probably added, said fluently, confidently, from patterns that ended before the event existed. The cutoff isn't just a limit. It's a confabulation zone.
The modern patches you've seen in products: search tools and retrieval, fetching the present onto the desk at ask-time — the model reading today rather than remembering it. They work, and they're coming in the agent acts. But keep the seam visible: under every patch, the calendar stays frozen. Recency is always delivered, never known.
The hygiene, then — boring and effective. Know your model's cutoff; it's published, and it's the vintage on the bottle. Route recency questions to search explicitly instead of hoping. And when a time-sensitive answer arrives without a fetched source, file it as what it mechanically is: a guess from a frozen world.
The trap is simply not knowing the date — trusting a frozen world about a moving one. And with the calendar understood, one episode remains in the training act: the strange fight over what scale itself does — whether abilities truly switch on, or whether we've been fooled by our own rulers. The emergence debate. Next.