Загрузка видео...
Не удалось загрузить видео
Astra began to saturate a computer-use eval we haven’t even released yet. Astra was not only able to kill several hostile mobs in Minecraft, but it made it all the way to building a Nether portal in less than three hours. We believe that if we ran it longer,... show more
30,212 просмотров • 6 дней назад •via X (Twitter)
Комментарии: 16

Astra killing zombies

Astra does the trap door trick to mine more efficiently

Astra boats back to land

Astra enabled “Keep Inventory,” allowing it to progress faster. However, its ability to operate in Minecraft represents a huge leap over previous models.

Can’t you make it where it pauses during its thinking? I feel like it would make it a lot better.

if an open weights model is ever able to do this anytime soon, that would be very impressive. but this is otherwise very cool for an ai to actually be able to play minecraft like this, very impressive showing from openai

Astra building a 3x3 door was my AGI moment

the next benchmark has to randomize the world on every run. a fixed three-hour Minecraft trajectory becomes a demo fast. change the spawn, inventory, goals, and hazards to see whether the agent learned the game or one path through it.

Impressive

MCSpeedrunBench could be interesting

It's over if Astra gets on Cerebras.

Wow. Next step might be something like the recent open world Zelda games. Wondering if it can do these too (or challenge runs that require deep planning ahead)

The long-horizon trace is the real benchmark: survival after deaths, inventory state, recovery from bad plans, and the number of human interventions. Publishing the harness and full replay would turn a three-hour demo into a comparable computer-use eval.

Release it! this way we can compare it with other models

nether portal in under 3 hrs.

more tokens. Beating the game would make history
