Astra played Minecraft for 141 hours — then a creeper erased the run
A benchmark run that ended like a bad Tuesday for a human player, and a Chinese lab betting its enterprise pitch on compliance plumbing rather than a better model.
Vals AI stopped its Minecraft run for GPT-6 Astra after 141 hours — six blaze rods, three ender pearls, a semi-automatic blaze farm, more than half a dozen Endermen killed in a warped forest — and then one Creeper explosion took the chest and the bed holding the entire inventory. Thousands watched the stream as the model carried its loot to a chest and, moments later, watched the loot stop existing. It also started raining. The evaluator, which says no AI system had previously gotten this far in the game, called the run off there, well past the hundred-hour mark.
The interesting part is not the hours, it is what the model did with the failure. Astra wrote itself a note, partly in caps: always carry critical items, never store them in an unguarded chest again. Then it spent several hours farming almost nothing but potatoes. A single bad event became a permanent rule — and the rule held even after it stopped being useful, which is exactly the behavior ARC Prize used the run to illustrate. Its account of the leap is that Astra compresses unfamiliar game mechanics into symbolic shorthand it invents itself: objects, coordinates, rules, planned actions, with observations turned into rules and rules into plans. In benchmark games that compression is the win; in open worlds with random events it can turn into overcorrection, where the agent defends against the last disaster instead of pursuing the goal.
You can see the mechanism in the run's harder stretches. A bed that should have healed the character did nothing; Astra let the game run, tried again, and logged that the bed worked. Shots kept missing, so it moved closer, noticed the obstruction, and recorded it. At a keypad it waited for the game to confirm the press. Check whether an input worked before repeating it — trivial to state, and the reason the agent never spun on the same useless action for hours. It is also the reason it now refuses to leave anything valuable unattended.
Moonshot launched a Kimi package for financial services that plugs the model into S&P Global Market Intelligence, Crunchbase, Wind, Tianyancha and the SEC's EDGAR filing system — and, more revealingly, into a compliance gateway co-built with CITIC Securities. Data arrives through MCP calls into research tasks, so a number in a draft can be traced back to its source; Moonshot says dozens of institutions are already running it, naming CICC, ICBC and Sequoia China (now Hong Shan) among them, and that document work that took days now lands in hours.
The plumbing is the product. The company's own numbers in a CITIC pilot on interim trustee reports — 50-plus issuers, more than 100 outstanding corporate bonds, 60-plus reports — put preparation time per report from about 30 minutes down to 10, and the integration that was scoped for two months shipped in three working days. The five compliance measures around it are unglamorous: data classification, personal-information filtering, per-call authorisation of data sources and tools, human review before anything becomes a formal conclusion, and an audit trail linking every task to the data it touched. That is the pitch most Western labs are still making over dinner, written down as a gateway a bank can actually approve — and it is why Moonshot wants a cut when US clouds sell its Kimi K3 model: the margin is moving into the workflow, not the weights.
What to watch: whether the compliance gateway becomes the resellable layer for Chinese agents in regulated industries, and whether anyone publishes the Minecraft run's full action log so the overcorrection can be measured instead of described.
If a model teaches itself a rule after one bad day and never lets it go, whose job is it to audit which rules it keeps — the lab, the evaluator, or the customer? Tell us in the comments.
