50 AI agents built a Minecraft bureaucracy
Getting a team to agree on the work can take longer than doing it. Apparently, AI can run into that problem too. In an October experiment, Y Combinator and Olam Labs gave 50 AI agents three hours to build a civilization in Minecraft, the game where players gather resources and construct their own world.
The agents did organize themselves. They appointed a leader, rationed food and sent roughly 5,900 messages. But after a deadly first night, they became cautious and crowded into shelters. Four agents spent 11 minutes deciding who would apply fertilizer to a single wheat block. Others asked permission to remove grass.
That’s a revealing result for anyone hoping a team of AI assistants will get more done than one. These agents could communicate and divide up jobs, yet coordination became a job of its own.
This was a game experiment with a particular setup, not proof that AI teams always fail. But it shows why counting agents isn’t enough. You still have to check whether all that activity produces the thing you asked for.
AI is helping a telescope search for alien signals
A radio telescope can collect far more data than a person could work through as it arrives. At the Allen Telescope Array in Northern California, researchers are using AI to help with that problem, and to operate the telescope itself.
Nvidia described the setup on Friday. A researcher can ask for an observation in ordinary language, and AI agents can steer the antennas, adjust the instrument, spot errors in the incoming data and plot the results. The system processes observations beside the telescope as they arrive, instead of saving everything to hard drives for later.
There’s already a result to point to: Nvidia says the system detected the Crab Nebula’s pulsar in real time. A pulsar is a spinning stellar remnant that produces a repeating signal. This one is a known object, so nobody is announcing an alien discovery. The achievement is picking out a real signal while the telescope is still observing. That gives the team something concrete to build on as it searches for signals whose source isn’t already known.
Scott
The Minecraft bureaucracy story checks out. I've got my AI commissioned to running about a dozen 24/7 quant trading research loops, and in its inherent quest for methodological rigor, it keeps going down rabbit holes of trying to reconcile what turned out to be floating point errors, or resolve one missing block of price data on one stock (out of thousands in the test universe). I was trying to get it to just place an order for me based on a particular strategy, and it was arguing with me that it couldn't be validated. And then I told it to do a 3 ATR stop loss, and it told me it needed to wait for the close of the current daily candle to be accurate.
Yeah, it doesn't very willingly accept the need for speed. It doesn't do very well at what humans do, of simply using heuristic shortcuts for decisions under uncertainty in which speed is of the essence.
I'm reminded of the quote: "Leaders don't make right decisions. They make decisions and then make them right."
AI hasn't really learned that yet.
Maybe it can... with some human coaching.