Journals write the research

In Too Many Requests, you pay for research with Papers. It took me three tries to decide where Papers come from.

In the first try, a Research Lab turned spare GPU time into Papers. That worked, but it paid you for GPUs that were doing nothing. It said nothing about what you’d built.

In the second try, the Lab sent tests out to your GPUs and counted the answers that came back. The third try is the one in the game, and you can watch all of it on the wires.

A data source, like a Web Scraper or a Book Scanner, makes data.
A Research Lab turns that data into tests.
A GPU pool answers the tests, just like it answers customers. Test beads are paler, so you can tell them apart.
A Journal writes Papers for every right answer. Wrong answers write nothing.

So research depends on your build. Every GPU can answer customers, answer tests, or help train a model, and you choose with a wire. Time spent on tests is time taken from paying customers. That’s the tradeoff.

Papers come in colors, one for each request type. There’s also a rare sixth kind called Breakthroughs. Only a Frontier Lab writes those, and only a very well trained model can earn them.

Research works like a queue. You click things in the research tree, and the Journals work on them in order. This came out of a playtest. Players couldn’t tell what the labs were working on. Now the top bar shows it, and how far along it is.

I named a lot of the tree after real AI techniques, and the rest after the business around them. The joke is in the name, and the effect is simple.

Flash Attention. Every GPU pool serves a tenth more on the same power.
Scaling Laws. Just add more. Pools serve a tenth more.
Recursive Self-Improvement. Your models do the research.

Research that’s far ahead shows up as a dark shape with a short teaser, like what if it thought before it spoke or your own reactor. what could go wrong. The IPO and AGI keep their names the whole way, because they’re the promise.

Next time, the crash, and why there’s only one.