hawkes.aiKey ideas
- The one idea — Video models "forget" because they re-derive each frame from limited temporal memory instead of truly tracking objects — so anything not strongly pinned drifts, producing flicker, morphing, and identity loss.
- How it is shown — A shirt that changes color, a face that morphs, an object that pops in and out — each traced to the model regenerating content rather than remembering it.
- The trap to avoid — Expecting long, perfectly consistent clips from today's models — consistency degrades as the clip gets longer, so you design around it.
- What it sets up — So which models fight this best — and what do you actually pick?
AI video models don't track objects. Each frame is re-derived from limited memory of the last few — so anything not pinned down quietly mutates.
The one idea
Video models "forget" because they re-derive each frame from limited temporal memory instead of truly tracking objects — so anything not strongly pinned drifts, producing flicker, morphing, and identity loss.
Watch an AI video and you'll catch it: a red shirt drifting to maroon and back, a face subtly becoming a different person, an object vanishing and reappearing. It looks like the model forgot what it just drew — and that's almost exactly right. Once you know the mechanism, you'll see it everywhere. Here's what forgetting really is. The model doesn't store a fact like the shirt is red. It re-derives every frame from noise plus a limited window of nearby context. So anything not strongly pinned gets re-guessed slightly differently each frame — small differences accumulating. The failures come as a family, and naming them helps. Flicker: rapid frame-to-frame jitter. Morphing: a face gradually becoming something else. Identity drift: a character clearly one person at the start, subtly different by the end.
How it works — the demo
A shirt that changes color, a face that morphs, an object that pops in and out — each traced to the model regenerating content rather than remembering it.
Plus objects popping in and out. Four symptoms, one disease — regeneration without memory. Two roots. First, a limited temporal window: the model attends to only a few nearby frames, so a detail from seconds ago is out of view. Second, the no-referee weakness from the hands episode, now across time — nothing insists this is the same shirt. It carves what's locally plausible; it never tracks objects. What helps? Bigger temporal context. Better architectures that carry identity across frames. Keyframe anchoring — pinning frames it can't wander off, the next episode. And shorter clips, since drift compounds with length.
The trap to avoid
Expecting long, perfectly consistent clips from today's models — consistency degrades as the clip gets longer, so you design around it.
Why it matters — and what’s next
So which models fight this best — and what do you actually pick?
The newest models hold a character far longer than the ones that made the memes. Improving — not solved. The trap: expecting a flawless sixty-second shot from one generation, then being baffled when the character melts halfway. Consistency degrades with length — a property of the method, not a bug. So design around it: short shots, cuts, anchors, and re-rolling takes that don't hold — how filmmakers have always worked. So forgetting isn't a mystery — it's regeneration without memory. And every video model is fighting this exact problem differently: some hold identity better, some sustain longer clips, some give more control. Which raises the question: who forgets least, and what should you pick? Next: Veo, Seedance, Wan, Kling — the landscape.
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 2:30 of narration
Watch an AI video and you'll catch it: a red shirt drifting to maroon and back, a face subtly becoming a different person, an object vanishing and reappearing. It looks like the model forgot what it just drew — and that's almost exactly right. Once you know the mechanism, you'll see it everywhere.
Here's what forgetting really is. The model doesn't store a fact like the shirt is red. It re-derives every frame from noise plus a limited window of nearby context. So anything not strongly pinned gets re-guessed slightly differently each frame — small differences accumulating.
The failures come as a family, and naming them helps. Flicker: rapid frame-to-frame jitter. Morphing: a face gradually becoming something else. Identity drift: a character clearly one person at the start, subtly different by the end. Plus objects popping in and out. Four symptoms, one disease — regeneration without memory.
Two roots. First, a limited temporal window: the model attends to only a few nearby frames, so a detail from seconds ago is out of view. Second, the no-referee weakness from the hands episode, now across time — nothing insists this is the same shirt. It carves what's locally plausible; it never tracks objects.
What helps? Bigger temporal context. Better architectures that carry identity across frames. Keyframe anchoring — pinning frames it can't wander off, the next episode. And shorter clips, since drift compounds with length. The newest models hold a character far longer than the ones that made the memes. Improving — not solved.
The trap: expecting a flawless sixty-second shot from one generation, then being baffled when the character melts halfway. Consistency degrades with length — a property of the method, not a bug. So design around it: short shots, cuts, anchors, and re-rolling takes that don't hold — how filmmakers have always worked.
So forgetting isn't a mystery — it's regeneration without memory. And every video model is fighting this exact problem differently: some hold identity better, some sustain longer clips, some give more control. Which raises the question: who forgets least, and what should you pick? Next: Veo, Seedance, Wan, Kling — the landscape.