An AI beat Factorio in 13.5 hours, by hand-carrying what it couldn't automate

GPT-5.6 Sol launched a rocket in peaceful mode. When its late-game production lines didn't connect, it shuttled supplies between factory areas itself.

In brief

Ryan Madden gave GPT-5.6 Sol a Python library that enforces human limits like reach and fog of war. The run ended in a rocket launch after 13.5 hours of game time, at a low thinking level.

Commenters split on what it proves. One watched another group of models play with biters on and sees the same planning gap; another read the trace and noticed the model treating the player character as just another automation unit.

What the run actually showed

The rocket launch
The rocket launch

Ryan Madden's post describes GPT-5.6 Sol, running in Codex at low thinking, playing a peaceful Factorio game with no biters. A game manager let it act through a Python library that enforced limits human players face. Earlier runs had used Claude Code and a Bash library. Madden had planned to stop after blue science but kept going, and the run finished with a launched rocket in 13.5 hours of game time.

Total resource production with research milestones
Total resource production with research milestones

He is candid about the weak spots. Scaling problems appeared after blue science, and Sol never connected some late-game production systems, covering for it by repeatedly ferrying supplies between areas. The run was peaceful only, so combat was untested. Its predecessor, GPT-5.5, had largely failed to use transport belts.

Is hand-feeding cheating?

"Looking through the video, around 12m one can see the Sol trace report the cancellation of a conflicting 'courier' (hand-feeding) script. Rather than implement 'manual' effort, it seems like Sol considered the Engineer to be just another unit of automation!" — a commenter who watched the run video

He adds that it is outside the spirit of the challenge, though not far, and a hard loophole to close when an agent uses a scripting layer.

Peaceful mode and memorized guides

One commenter listed two limits of these demos. Peaceful mode removes response time as a factor, and the models have read every guide and video ever published. Given a similar game without that training data, he thinks it might fare terribly. Another added that full factory blueprints are available to copy from sites like factorioprints.com, so chaining them together might nearly beat the game.

The version with biters

"I've been watching streams of 9 different models playing Factorio together and it seems the answer is — not very well at all. They run into the same issue they have elsewhere in the game as well as in coding tasks: inability to create structured plans for the future." — a viewer of a nine-model Factorio livestream

He describes the loop in that stream: get attacked, build a few turrets, get attacked again, without ever strengthening defenses unprompted. That pattern is the gap Madden's peaceful setup was designed to avoid, and the next test on his list is other models and thinking levels.

Get breakdowns like this every day

The top Hacker News stories, unpacked — by email, Telegram, or RSS.

Next story · 6 of 11 Read more
🚀Tech

Vitalik left the X app but not X, and lost interest in the daily fight

He kept posting through a multi-platform client that merges X, Bluesky, Farcaster and Lens, and says the main benefit is the missing algorithm.

★ 308 · 💬 228 · vitalik.eth.limo · 2 min