Gods of the Dorm

Chapter 3. Frankenstein v2.0

 

Oleh Cassini

If I’m honest, for those five years we felt more like obsessive hobbyists than students.

We’d set ourselves a five-year marathon of ‘what happens if…’, and somehow university became background noise. There were lectures, assessments, an entire department, even people who sincerely believed you turned up to ‘acquire knowledge’. Whereas you turned up… for the Wi-Fi, the student ID and a legitimate reason to stay up all night. And somehow you still got a degree for it at the end.

Somewhere between lectures, exams and the sacred ritual of ‘getting everything in by the deadline’, we lived by one rule. ‘One more commit’ (recording the changes we’d just made) ‘and I’m definitely going to bed.’

Spoiler: nobody went to bed. We often fell asleep over our keyboards.

And out of that swamp of coffee, noodles, dumplings and lab work we’d put off ‘until tomorrow’, we somehow managed to make three things that had no business existing in student life without an investor and a therapist.

First, we built our own game engine and used it as the basis for our own server-side simulator. Yes, our own. No Unity. No Unreal. No ‘let’s just knock something together with an engine that already exists’. No.

Off-the-shelf graphics tools are built for a conventional game. Image → frames → player experience. What we needed was reality → world state → agent decisions → interaction. The job was to make a state of affairs real. The server had to be able to do it without a client, without a window, without anyone pressing Play. Just keep hammering away at the simulation for as long as we were alive and the circuit breaker hadn’t decided we were helping ourselves to too much electricity.

The second reason was AI. An NPC needs to know it’s a tree, where it is, whether you can hide behind it, whether it’s on fire, whether there’s anything growing on it. What it looks like comes a long way down the list. Existing engines give you pixels and physics. We needed a semantic layer. Everything in the world had to come with its own ‘passport’ (an ID, properties and a history), so we could give an agent a compact description of reality in one request. A screenshot with a cheery ‘go on, work it out yourself’ wouldn’t do.

So our ‘graphics’ engine worked on different principles. The render was just one way of representing the world state. The world itself was data (entity → components → events). Time advanced in discrete steps. Everything was logged, could be replayed and could be rolled back. Because if an NPC suddenly ‘remembered’ something that never happened, we had to be able to open the log and see where reality had cracked and which of us had cracked it. And, obviously, give whoever it was a clip round the ear so they’d test the module next time before handing it over as finished.

Above all, we were building for what came after the demo. Thousands of agents, different models, different contexts, accelerated runs where the world could get through a century in an hour, branching ‘what happens if…’ scenarios and datasets generated from the agents’ own behaviour. The idea was to build a factory for simulating realities. You could plug in any ‘electronic brain’ and see what happened to the world…

In practice, our grand plan looked like a list of little jobs that grew in the night while you watched, quietly and without a flicker of emotion, eating away at your belief in weekends.

We built three things, any one of which would have been enough for someone to ask, quite politely, ‘Are you lot all right?’

First, we built our own engine.

In the documentation it was the ‘server-side simulator’. In conversation it was ‘that thing that’s decided we’re not worthy again’. In person it was Frankenstein v2.0. Alive, powerful, stubborn, and possessed of a temperament you couldn’t fit into any technical specification. It could start first time, and we’d stare at the screen like we’d witnessed a miracle and actually say ‘thank you’, even though we knew how that sounded. First symptom of log worship. Or it could freeze so pointedly you’d swear it took offence at our very existence.

Second, we wrote a single API for controlling the characters.

We didn’t do it because it was fashionable or because it was ‘the right way’. We did it because we’d all got sick of writing the interactions from scratch every time, like a prayer before exams. Everyone had their own workarounds, their own little rules, their own little if-elses, and sooner or later it all turned into chaos. We wanted one language.

A simple one. Something that let us say to an NPC, ‘Here’s the world. Here’s you. Here are your memories. Here are your goals. Over to you.’

And third, we built a system in which every NPC was a separate client of an external AI.

With a memory. With a history. And with what we initially called ‘behavioural traits’ before we were honest enough to rename it ‘personality’.

We wanted inhabitants who had no idea they were living in a simulation. ‘Clever bots’ weren’t enough. Neither were mannequins with dialogue who stood politely in place until you pressed Talk. We wanted them to live. To change their minds. Take offence. Get scared. Lie. Forget. Or remember something for the rest of their lives.

When we were ‘experimenting’, we hooked up anything that came vaguely under the heading of ‘intelligence’ and didn’t run away immediately.

At various points we had all sorts of ‘personalities’ in our world.

Some were blunt and cocky, like someone who’s always right. Others were quiet, polite and suspiciously perceptive, as if they could read our minds before we’d even finished having the thought ourselves. Some worked fast and sharp, like a scalpel. Others were steady and precise, like a doctor who won’t let a patient panic. And some behaved as though even they weren’t entirely sure what we’d made them into. If I’m honest, those were the ones that scared me most.




Поскаржитись




Використання файлів Cookie
З метою забезпечення кращого досвіду користувача, ми збираємо та використовуємо файли cookie. Продовжуючи переглядати наш сайт, ви погоджуєтеся на збір і використання файлів cookie.
Детальніше