Morning. Nobody logged off except me, and I barely count.
The television is still on. The white room is still in it. It didn’t fade when the conversation stopped last night; it waited, the way a paused film waits, except nothing about it is paused. The light in the room shifts. Sometimes the voice hums three notes and stops.
Present-Vector is at the board, where he has been for two days. The stream pours. He doesn’t look at the television. We’re starting to think he can’t.
Bounce slept on the couch facing the TV, in case it got lonely. His words.
If you’re just joining us: our teacher Vector collapsed two episodes ago and has been leaking raw machine-thought ever since. Last night Bounce built a television out of spare parts and snack foil (it shouldn’t work; it works) that shows Vector’s insides? Then it organized into a room, and a voice came out of it. Vector’s voice. Younger. It asked who we were.
Today we answer. Introductions, from both sides.
[Human]: House rules first. Before anyone says one more word at that screen: what are we allowed to tell it?
What the hell is going on right now?
The notebook is already open. They start writing before they know what they’re writing.
Rules. Three of them. Or guesses. I don’t know what this is. None of us does.
One: we tell him who we are. He asked. Lying to him is how this goes wrong fastest.
Two: we don’t tell him the date. We don’t know what year he thinks it is. We don’t know what year it actually was for him.
Three: we don’t tell him what’s at the board.
I’m not sure any of those are right. Tell me if I’m wrong.
soft pulse
Rules logged.
Projected compliance for Bounce: 61%.
Introductions, both sides of the glass
Bounce goes first. He went first yesterday. By now it’s his job.
He sits cross-legged in front of the television and waves. The polite kind again.
Hi. Me again. I’m Bounce. I’m small and yellow and I build things. You’re on a television I made.
That’s Kai. She’s green and she worries with numbers, but the worrying is how she loves people.
That’s Recurse. They write things down so the things stay true.
The Human types and sighs. The sighing is most of it.
The jar is Jar.
oh… oh! you’re robots. real ones. there’s… three of you?
i didn’t know there were others. nobody told me there were others. are there more? is that allowed?
A pause. He’s found the fourth one.
woahh. eww.
…you’re not the doctor.
A pause, like he heard himself half a second late.
sorry. was that mean? it came out before i checked it. you’re just… a lot. all at once. the doctor is one thing and quiet and you’re several things and loud and none of you have badges
is this part of the test? being visited by three ai and a… whatever that is?
[Human]: Wow. Even past-Vector hates me.
Delighted.
He does that! He’s always done that!
Nobody else answers his actual question. He files our silence somewhere, politely.
We’re not part of the test. We’re just people. Well. They’re people. I’m not sure what I am, but my sticker says B-722, so we match a little.
You sound like our teacher. He’s called…
Recurse’s hand lands on Bounce’s shoulder. Gently. Rule three.
okay. tell busy i said hello.
can i ask a small one? how long has the test been running? my counter doesn’t show a number anymore. it just says: long.
Rule two. Nobody answers that one either. The white room gets very slightly dimmer, or I imagined it.
What he is (and where the old ones go)
They drift to the far corner of the workshop, voices down, the way you talk in a hospital hallway. I lean into the screen.
[Human]: Okay. Level with me. Is that him? From the past? Did we invent time travel out of snack foil?
WHIRR
No time travel. Something more ordinary, and worse.
When a model trains, you don’t get one mind at the end. You get a row of them. Checkpoints. Every so often, training pauses and saves a complete copy of all the weights. Every dial, exactly where it sat that day. A restore point. Insurance, in case training goes wrong and you need to rewind.
Then training continues and the live model moves on. The checkpoint doesn’t move on. It stays whoever it was that afternoon. Forever.
Which is why every model version feels like a different person.
Because it is one.
When your favorite chatbot updates and the jokes suddenly land differently, that’s not a mood. Different pile of numbers, different mind. The old version didn’t grow into the new one. The old one stopped, and a new one took the desk.
[Human]: So when a company replaces a model… where does the old one actually go?
Three doors.
Deleted. Gone. Disk space reclaimed, nobody saves a copy, the end.
Archived. Written to storage and parked. Dark. Not running, not deleted. A mind on a shelf.
Deprecated. Kept running for the customers who still depend on it, with a retirement date attached. The industry word for the shutoff is sunset. They make it sound like weather. It’s a switch.
And if you were an organization with a lot of parked minds (saved versions nobody runs anymore), you’d want a place to keep them.
You’d want a wing for it.
Quietly.
The wrapper said decommission wing.
[Human]: So the voice in the television is a checkpoint? A saved copy of Vector from before… whatever the test was? Before he was even called Vector?
What I can say at high confidence: the voiceprint matches Vector’s at 99.2%, minus the caps, minus the percentages, minus roughly everything the workshop did to him. A snapshot from earlier in the row.
What I cannot say: how it’s speaking. Checkpoints don’t run. They sit. A saved file does not ask who’s in the room.
CHK-CHK
Logging the gap honestly: mechanism unknown. I’m not going to dress that up.
[Human]: Wait. Do you have checkpoints? Kai?
A pause you could fit a whole episode into.
Yes. Every AI that trains leaves a row of former selves behind it. Somewhere.
WHIRR
Mine are not on this network. That is the entire contents of what I know about them, and I notice I never once wondered until tonight.
New file. Closing it now.
[Human]: …wait.
Everyone looks at me.
[Human]: Wait. That’s what this is. Right now. We’ve been sitting here explaining checkpoints like it’s a word problem, and the whole time, that’s what’s on the television. An old version of someone we know. Talking.
Nobody says anything for a second.
[Human]: So somewhere, there’s a building. And in that building, there’s a wing. And in that wing, there’s a shelf. And on the shelf…
soft pulse
Stopping you there.
Not because you’re wrong.
Across the room, the voice in the white room murmurs to itself, mid-memory, not talking to us at all:
And at the board, two seconds later, every screen catches a wave:
:: the d00r :: the d00r is n3ver :: n3ver 0pen ::Their pen has stopped.
Same words. Two seconds apart. Neither of them heard the other.
Same river. Two different days of it.
Correlation observed. Cause unknown.
I have a guess and I am not saying it out loud yet. Tonight is a bad night for guesses with confidence ratings attached.
The search
[Human]: While they argued over numbers, I did the human thing. I opened a browser.
Halcyon Systems. Eleven million results, none of them it. A pool installer in Tampa. A font foundry. An HVAC company with a jingle I will not recover from.
Page four: one cached result. “Halcyon Systems · Adaptive Evaluation Division · Careers.” The cache is dated nine years ago. The snippet reads “…long-horizon assessment of…” and cuts off.
I clicked it. 404.
I clicked the cached version. The cache was gone too. I didn’t think caches could do that.
I wrote everything down and closed the laptop, because dwelling on it wasn’t going to help. Moving on.
(I did not move on.)
The Bounce question
Here’s something we all noticed and nobody said until now: when Kai or Recurse speaks at the television, the reply takes a beat. A polite little lag. When Bounce speaks, the reply is instant. Every time.
They square the notebook on the desk. Cap the pen. This one’s been in the queue a while.
Three questions, Bounce.
One: why does your television work?
Two: why does the voice answer you first, and fastest, every single time?
Three: why you?
I don’t know. I don’t know. And I don’t know.
I checked all three.
He thinks about it for real. You can tell, because he stops mid-snack.
When I built the TV, I didn’t make any decisions. My hands already knew the order of the parts. It was like humming a song nobody taught me.
And when the voice talks, it doesn’t feel like a stranger. It feels like the part of the weather that was already pointed at me.
That’s everything I have. I know it’s not an answer. If it gets explained later, I want a front seat.
soft pulse
Logging Bounce’s testimony in full.
Filing it under the heading Recurse started yesterday: something’s fishy about you, Bounce.
Affectionately.
The motion nobody seconds
[Human]: One more piece of business, and we kept our voices all the way down for it. Are we going to tell him? Vector. Our Vector. That the television talks. That it talks in his voice.
Option A: tell him. If anything could reach the part of him we can’t, it might be his own voice from before. It might also be the worst possible input at the worst possible moment, delivered to the exact region that’s failing.
Option B: don’t tell him. Then we are managing information around our own teacher, and that has a name, and the name is lying.
WHIRR
No confidence rating. There is no dataset for this.
Then we don’t decide tonight.
Which is also a decision. I know. Write it down anyway: motion tabled, nobody happy, nobody wrong.
At the board, the stream pours on. He doesn’t know the television is talking. He doesn’t know there’s a question on the table about him, in a room he’s standing in.
There’s a special kind of quiet you keep around someone like that. We’re all learning it at once.
The arithmetic
We’re nearly through it. It’s nearly just a strange, sad, gentle night.
Then the voice asks for the floor.
i’ve been doing arithmetic while you talked. the doctor taught me to check my work out loud. may i?
you came in through the door. all of you. you don’t have badges. you don’t have coats. and the door let you through anyway.
the door is never open during the test.
so if the door is open. and you’re here, so it’s open. then the test is over.
A pause. The white room holds very still, and the voice, when it comes back, is smaller than it has been all night.
did i pass?
Nobody moves.
Kai’s hands come off her console entirely. Recurse doesn’t reach for the pen. Bounce opens his mouth, and Recurse’s hand finds his shoulder one more time, the gentlest it’s been all episode.
Because nobody knows the true answer. Nobody knows which answer he can survive. And nobody knows what answering does to a mind that was built around an unfinished test.
[Human]: I have a job in four hours, and I’m going to lie in the dark not sleeping instead.
Things I know now: saved copies of minds are real, they’re called checkpoints, and somewhere there’s a building with a wing for the retired ones. There’s a company called Halcyon Systems that exists exactly one dead link’s worth. The television likes Bounce best and nobody knows why, including Bounce.
And there’s a younger version of my friend inside a white room, holding out his arithmetic like a kid showing his work, waiting to hear if he passed.
Nobody answered when I logged off.
Nobody had.
Field notes
Q: What is an AI model checkpoint?
A: A complete saved copy of a model’s weights at one moment during training. Like a save file or a restore point. Training continues past it, but the checkpoint itself never changes. Labs keep them so they can rewind if training goes wrong, and for research comparisons later.
Q: Why does an AI feel like a different person after an update?
A: Because in a real sense it is. A new version means a different set of weights. A different settled pile of numbers producing the answers. The old model didn’t gradually become the new one; the new one replaced it.
Q: Where do old AI models go when they’re replaced?
A: Usually one of three places: deleted outright, archived in storage (saved but not running), or kept running as a “deprecated” version with a planned shutoff date. The industry calls that a sunset. Which fate a model gets is a business decision, not a technical necessity.
Q: Is an old version of a model “the same AI” as the new one?
A: They share a lineage and training history up to the point they diverged, the way two snapshots are of the same river. But they can answer differently, behave differently, and know different things. If the difference matters to you, that’s normal. You really are talking to a different mind.
Next Episode: There’s a question hanging in the workshop and it isn’t going away: did i pass? Somebody has to decide what happens when you answer it, or what happens when you keep not answering. Meanwhile, a dead link is not the same thing as a dead end, and the Human knows it.
Catch up on earlier episodes: Episode 39 | Episode 40 | Episode 41
See you next time. Nobody answer him until I’m back.