‌7

Small

Ranjodh Singh Dhaliwal

His watch beeped at regular intervals, but soft squeaks slowly turned into harsher pings. The noise was interspersed with messages of urgency; he had veered off his predicted schedule, and the models were readjusting what he needed to do to get back on track. He looked down at his watch. It was only 73% in alignment. This was dire. He rarely steered off below 90%, though there was that one time when he had burned his right hand while using the oven and spent a solid 72 hours hovering around 85%, and that had been a disaster! It had taken him three full years to recover. But 73%? This was unprecedented!! Especially when he could not claim a biomedical exception. This would surely have consequences he couldn’t even fathom.

He paused for a second.

72.8%.

Was the 0.2% just for pausing? Fuck if he knew…

If he was to recover, he needed to think. What had he done wrong? What had he forgotten to do? Nothing that he could remember. Surely, something was wrong with the predictions? Maybe, he hoped, this was a calculation collapse error? He still did not understand what it meant but he was repeatedly told while being assigned his watch that collapse errors were becoming more and more common. The metaphor that the freckled man, whose gold rim spectacles were easy to remember, had used was one of ‘changing climate.’

‘You can read all this legal bullshit or you can get this. Simply put, language predictions are getting more unpredictable, like the weather. It is kind of like how our grandfathers would talk to us about the first missed winter. My friend swears we are going to get our first missed CoTar soon.’

And he was right. A few cent clock cycles (CCCs) later – according to what his uncle had told him many CCCs ago, these used to be once called either days or weeks, though he could not be certain which – they got the first missed CoTar. A CoTar was officially called a Company Target but he liked the linguistic implications of saying ‘Tar’ out loud, a dirty, sticky, burning feeling on his tongue. It was reported in the cycle updates, but everyone was assured it only happened because of a unit in the far north that had a technical malfunction. The rest of the company had done their job. The median prediction success was over 97% still.

Was this another missed CoTar? Maybe this was closer to home now, in his zone, maybe even in his division? What else could explain the watch asking him to turn on his rice cooker at 4am? This was his building check cycle, after all; his local maintenance checks had already been completed two cycles ago. He was next scheduled to be on his workstation in the subsequent CCC; maybe he would get some indication of what had gone wrong? He at least had supervisor access to the latest local maintenance models. If it was something in their division, he could flag it up the chain, or ping it. Flag, chain, ping: he sometimes wondered who came up with these words. Probably one of those older megamodels that we were told had too much heat? The linguists, yes that was the word, of yore had called something like this a language-action gap; one could predict language, but people’s actions – which can be captured by language but also exceed it – mean that one has to constantly calibrate according to events, not just semiotic, alphabetical flows.

He looked out of the window. Red, blue, yellow, green. Lights as far as he could see, only interspersed by the void of the dune park in the middle of the city. The models liked the predictability of light patterns, and he was making the trips in his building to make sure the lights shone as they should.

71.2%.

Fuck.

‘The stairs on the right, two flights down, light malfunction. Turn the green light off.’ He followed the instructions. The stairwell had plastic wrappers strewn on the ground; bad job, probably the dude living beneath him was to blame, as usual. This level of tardiness was unsustainable. He wondered how long before the poor guy was kicked out.

He stepped back into the hallway after scrunching up his nose in the plasticky odour of the stairwell. The light was right there. He could see it. It was shining green. He found the switch underneath the bulb control, and flicked it off. It flashed twice. ‘Are you sure you want to turn it off?’ asked the voice. It was Nadya’s voice, or at least suspiciously similar to hers. ‘Yes,’ he replied firmly.

‘Job Code and Passbribe please.’

He looked down at his watch and read out loud. ‘DBGR3321CC55, and %$###^^^&JCB.’

‘Thank you.’

And the familiar sound of a light winding down, dimming towards extinction. They will probably come in the next day and fix it. Until then, it would remain a sore spot for the drone system above; a reminder to the models and the humans behind it that malfunctions were built into dreams of perfection.

77%.

OK, all right, no missed CoTar, just a procedural problem, he thought, maybe he can present an updation case to the firm in the next CCC.

‘Check on apartment 44S. Possible overconsumption. Elevator up ahead, walk straight for 22 meters, go to floor 36.’

His feet dragged him forward, instinctively.

77.1%

He looked out of the windows as he walked briskly. There were a couple of black spots where he could swear he had seen reds and blues and greens a few moments ago. Maybe it was more serious than he had hoped? What if it was those troublemakers the security team had warned them about? There had been a breach in the language archive a few CCCs ago that no one had paid any heed to; the data was doubly-triply stored. Redundancy was a great epistemic virtue. The original had been copied and was regularly fed to the new synthetic models to retain their vitality; even the too-hot early models were good enough for this rejuvenation of prediction. Most had no idea about this breach, but his position in the local language vitality centre gave him access. The notification, however, did state that there were long term large poisoning possibilities. Was this the long term?

77.7%

He scanned his palm at the elevator. It had already been called for him by the model. The metallic door, painted with a dark walnut wood pattern, opened, and he stepped in. Another palm wave that captured his fingerprints and hand condition together – recognition and biological maintenance at the same time – and he was off to floor 36. ‘Confirming,’ he said after looking at the autoroute to floor 36. The elevator started moving down.

He stared at the red carpet below with its firm monograms in designs of previous eras. The rounded edges of letters of the one on the bottom right amused him.

The elevator slowed down. He looked at the door. It was only floor 55. Highly unusual; what could be important enough to break the predicted routes for everyone?

The elevator stopped and before he could react, someone stepped in. They wore a patterned black and white sweatshirt and large, chunky glasses.

‘Hello!’

‘Hello’ was the soft reply.

The elevator door jammed shut as the new entrant waved their palm. It started moving again.

‘Dr. Q?’

‘Yes, do I know you?’

‘You work at the local language vitality centre at the firm, right?’ The tone, however, suggested a statement rather than a question.

‘I am not sure that information is…’

‘Never mind. I know it. And that is what matters. There has been a huge missed CoTar. Systemwide.’

He wanted to say ‘that’s not a very funny joke’ but he couldn’t. There was a lot to process before his words could be generated.

‘We helped precipitate it. Or at least orchestrate it. And now we need your help.’

‘I am not sure I understand.’

‘You will. Follow me.’

‘But I am going on a building run.’

He looked down at his watch.

78.1%

‘I know. Floor 36. 44S.’

‘But, how could you?’

Beep. It wasn’t his watch.

He looked down. It didn’t have a number. It had a heart icon, glowing orange.

‘What is that?’ He had forgotten his original line of questioning. This new object was more shocking.

‘Oh, we run the small models. Not like yours,’ anticipating the response, ‘these are local and personal. We generate our models from small groups of random humans who created data in the original source, the pre-2021 internet.’

‘But no one has the originals.’

‘We do. We didn’t breach the datasink for nothing. Some things in life become quite easy once you stop paying heed to the number on your wrist.’

His brain ran that word over and over again.

Small.

*

This attempt captures a speculative fictional scenario where large interconnected multimodal AI systems run the world by predicting and producing it. In other words, this world has been predicted and mapped on those predictions. The surveillance, precision and accuracy of such predictions is a two-act play: on the one hand, the models (and the socialities running and making them) ask and show what can happen, and on the other, they will and force it into happening.

To predict, after all, is to assume that the future will be just like the past. And if it is not, it can at least be nudged in a certain direction.

Amid conversations of global predictive regimes and selective informational flow, we are faced, in this vignette, with different senses of temporalities that rely not on clocks but on clock cycles, and challenges that stem from always impending model collapse, which has to be cyclically prevented, by prediction, through human creative ingenuity.

Small, careful applications of organisation – in both the organising and organic valences – present local possibilities of tweaking and collapsing what cannot work but will not break by itself. The small, in this vision, can be more fruitful than the large, the epistemic saturation of which in a regime of growth cannot be understated. As opposed to perpetual expansions and captures envisioned by large-scale growth, small offers no promises of scale and makes no claims of mastery.

Never even contemplating a universal God’s eye view, small instead always endeavours to situate locally.