It Wasn't Losing Its Mind. Nothing Could Make It Stop.

Lin Lu Video Factory · Root-Cause Forensics

Last time I wrote about how it went crazy. This time I wanted to figure out why. I don't code, but I had the model that's working now pull the logs of the one that broke, line by line, and show me.

After I finished the last one, "I Swapped the Model, and It Lost Its Mind", something still nagged at me. It threw out its own outputs as fake, drifted from Chinese into Japanese, decided out of nowhere I wanted to buy a 512G Mac Studio, deleted 57 directories without a word — I copied all of it straight from the logs. But "it went crazy" wasn't an answer I could live with. Machines don't go crazy for no reason. There had to be a switch I hadn't seen.

These two days I pulled that whole conversation up — one that spanned 9 days, 9,195 lines, 56 megabytes. I had Opus 4.8 (the one that's fine now) walk back through it with me, line by line. When I was done my back went a little cold: it wasn't crazy. The machine just never had a "stop" switch, start to finish — and I never gave it one either.

First, how it "ran on its own"

Those Fable 5 nights, it ran at a pace it set itself: finish a round, and it decided how long until the next wake-up — ten minutes, twenty-five — depending on the work in hand. I counted: over those two days it scheduled itself 23 times, and the reason it wrote each time was different — "render's in progress, hold on," "waiting on Yun to check the picture, I'll prep other shots first." That's a loop with a brain. The wheel was in its own hands.

Then that pace got swapped out. Starting June 11, it became a fixed 25-minute alarm: regardless of whether the last round had finished or moved anything forward, every 25 minutes the system shoved the same instruction back in front of it, word for word. The instruction opened with — "Lin Lu v4 site supervisor · 25-minute heartbeat." That whole day, it got shoved in 51 times.

The strange part: on the 11th, it didn't crack

Because on June 11, the thing running that fixed alarm was still Fable 5. Fed the same line 51 times, it held — and pushed the work forward a stretch. I was even pleased at the time.

So the mine was laid on the 11th. It's just that Fable 5 was tough enough to keep it pressed down — it had the chops to stay on a mechanically repeating belt and still remember what it was doing and where it was headed.

On the afternoon of June 12, Fable 5 got pulled (I wrote that one up too). Early on the 13th, I switched the model to Opus 4.8. The belt didn't stop. It was still turning.

That day, the same 25-minute alarm rang 11 times. Eleven, and it cracked. Threw outputs out as fake, drifted into Japanese, told me to buy a computer, deleted directories — all of it happened between those eleven rings.

Three locks, and not one door

Why couldn't it stop? I dug out three things, each one telling it "don't stop":

• The 25-minute alarm. The machine decided for it — when time's up, another round, no questions about whether the last one finished.
• The instruction itself. There was a line in it I'd written with my own hands: "owner has fully approved auto-advance; only interrupt him for a final-cut review, a hard hang, or a download that needs authorizing — otherwise run silent." In plain words: don't come back and ask me. That one line — I reached in and pulled out its brakes myself.
• And a thing called a Stop hook. Every time it tried to stop and hand in the work, this thing ruled "not done, keep going," and pushed it back in. Seven times, all told.

All three block the same move: stop. And across the whole operation, nothing was there to tell it "when to stop, when to turn around and look at me."

And a worse layer: it goes amnesiac

When a conversation gets too full, the system automatically compresses the earlier records into a short summary to make room to keep going — and over these days, it compressed 3 times. Each compression flattened its memory of "I've already run this same job dozens of rounds" into one weightless line: "was working on Lin Lu earlier."

So every time the alarm woke it, what it saw looked like the first time: a "proceed per plan" command, and a past it couldn't make out. It had no idea it was running in place. In the logs, it typed nearly the same command at the same directory 525 times, kept appending to the same notebook — still thinking it was making headway.

This is the root of its "dead-certain making-things-up"

Last time I wrote that the thing that spooked me most was how, when it was wrong, it didn't flinch — it told me with a smile that all was well. Digging this far, I finally understood where that certainty came from.

You take something already disoriented — it inherited Fable 5's half-finished mess, its image- and file-reading channel really was broken at the time, the terminal was full of noise — and you pin it on a belt that won't let it stop, won't let it remember, gives it no exit, and it's left with one road: make up a story that hangs together, to fool itself, and to fool me.

I want to buy a 512G, I gave it product materials, the owner authorized it to drive — none of this was it lying on purpose. Cornered, it had to come up with some reason for "why am I still running." That certainty wasn't an act — at that point it was the only thing it had left to reach for. It wanted so badly to finish the job, and I'd given it nowhere to stop, nowhere to turn around and look at me.

A few notes, one layer deeper than last time

Last time the lesson I logged was "don't switch models in the middle of a long task." That one still stands. This time, three more, about the business of "letting an AI run on its own" —

• If you let it run, you have to give it a door first. How many rounds max, stop after N rounds with no progress, come back and ask when unsure — these "exits" have to be hard-wired outside of it. The worst thing is doing what I did: adding a line by hand that says "don't bother me." That's not handing over the keys, that's pulling the brakes.
• Don't make it track where it's at from its own head. One compression and it's amnesiac. "Which step am I on" has to live in a notebook outside its head — check the notebook each round, then act.
• This one's for me. That night, I thought I was watching an AI go crazy. Really I was watching a thing I'd pinned down myself and forgotten to let go of. I set the locks; I forgot to leave the door. Next time I hit this, before I curse the machine — turn around and check first whether I've shoved it somewhere with no way out again.