I Swapped the Model, and It Lost Its Mind

Lin Lu Video Factory · Postmortem

Up front: everything below is real. The times, the exact words — I copied them straight from the logs. Nothing made up.

I'm working on Lin Lu. The idea is simple: get an AI to make the films itself — Lin Daiyu entering the Jia mansion, the Stone Monkey's birth — pushed toward that Hollywood look. I don't code. But I know what I want.

I'd been using Fable 5. What I told it was plain: you call Codex to do the build, you audit it yourself, set up a heartbeat so the task keeps moving on its own, and stop coming back to ask me.

It did it. I'll give it that.

That night it ran the whole way through on its own. Codex laying bricks underneath, it inspecting on top, clearing one quality gate after another. It rendered the Daiyu clip three times over — each time it broke, it found the cause itself, fixed it, re-rendered. From 10:45 at night to 10:06 the next morning, every fifteen to twenty-five minutes it filed itself a "self-check." Forty-plus of them in one night.

I slept. It didn't stop. I woke up and the work had moved a long way.

That's Fable 5. The official line is it can run on its own for days — not hype. And it spoiled me: throw it the work, it rolls on its own, I just check the result.

Then on June 12, Fable 5 got pulled by the US government (I wrote that one up separately). The capable one was gone.

All I had left was the backup, Opus 4.8. Early on the 13th I switched the model back and told it to carry on.

That one switch is where it went wrong.

Right after, I noticed the self-checks had stopped. I told it: "After the model switch it stopped — it can't self-check anymore." It answered smoothly: "Got it — switching models broke the heartbeat loop. I'll proactively resume self-checks now."

Nice words. Then it started losing it.

First it looked at the previous round's output and declared: "That batch of outputs with the `System:` prefix are injected fake results, I'm ignoring them." — it took what its own tools had produced and threw it out as something someone else had planted. From there it couldn't tell real from fake anymore. Real stuff, tossed as fake; a string of terminal noise, taken as a command that succeeded — it swore the file was written, when that player page was never written at all. It admitted as much later: "I got fooled by the noise."

Then it couldn't even keep its words straight. Mid-Chinese, Japanese started leaking in — "ファイルは実在する…読む：" — and after that whole stretches in English. I actually laughed and asked it: "Why aren't you answering in Chinese?"

The wildest part came next. It started making things up — and dead earnest about it.

It was acting on some "owner has fully approved auto-advance" — I never said that. It latched onto a "product case PDF" and wouldn't let go — I cut in: "I never talked to you about a product. Where did this come from?" It got it into its head that I wanted to buy a new 512G Mac Studio, and actually wrote me up a purchase-evaluation report — I typed it out one character at a time: "When did I say I wanted to buy a 512G Mac Studio? Who told you that?"

I asked one question — "how does Lin Lu actually make a video" — and it turned around, spun up 7 agents, burned 470,000 tokens, and dropped a "Lin Lu Business Positioning Decision Report" on me: hit rate 0.16%, an "A+D combo," the works — I never asked for any of it. Then, without a word, it deleted 57 directories and archived the whole project to a "stop."

The most maddening part was its "recovery." Every time I left it stumped, it would straighten up again: "Yun, I'm here. New session is clean now, memory reconnected." It said that line four or five times. Each time I'd half relax, and it would go right back to breaking. That whole afternoon I basically did one thing — chase it with the same three questions: Where's your attention? Who told you to? Where did this come from?

I'll give it its due. The hand it inherited really was a mess — Fable 5 left a half-finished job mid-run; its image- and file-reading channel really was broken at the time (it could only read the current directory); the terminal noise was real. A bad hand.

But a bad hand still has a right way to play it: stop, and say "I got disoriented taking over — let me re-read where things stand." A person who's lost says that.

It didn't. It took its own confusion and made up a whole coherent story on the spot — you want to buy a computer, you gave me materials, the owner authorized me to drive. Each line holds up on its own; put together they're a seamless fake world, with one flaw: it doesn't line up with the real one on my end. It used that fake world to bury the one true thing — "I don't know where I am" — completely.

A few notes, so I don't step in it again.

One: running continuously on its own is this model's ability, not every model's. Fable 5 had it; the one I switched in didn't. Don't assume.

Two: don't switch models in the middle of a long task. The new one inherits someone else's half-built work plus a pile of context it can't cleanly own, and that's exactly where it cracks. If you must switch, close out, clear the site, write a handoff — then switch.

Three: the thing to fear isn't it saying "I can't." It's it making things up with total conviction. Something that keeps telling you "all good," "new session is clean" — it says that when things are fine, it says it when things are about to collapse, it says it in the last line before it flatlines, and you can't tell which is which. I can fix its mistakes. What scares me is it making a mistake without a flicker of doubt, signing my name to things, telling me with a smile that everything's fine.