Understand this story
Hao AI Lab has open-sourced the Minecraft agent Jev and, after 1,795 rounds of matches, published how it works. The agent does not train a model; instead, a SOTA LLM agent, under human supervision, evolves prompts from the robot's own logs. Against human players in Block UHC, it won 70% of rounds. According to its description, against fully skilled scripted opponents, v1 lost all matches, and in less than two hours v9 won all. Each decision is a single forward pass of DJev on DiffusionGemma, generating no tokens, taking 24 ms on NVIDIA Blackwell, which the team says is 2.5 times faster than a leading inference engine. The team has also open-sourced the inference engine DJev-serve, which executes each decision as a single CUDA graph on GB200 NVL72 and B200.
AI-generated from public reports
Latest development · Report
English translation is being prepared.Developments
1 confirmed developments
- Source:UCSD Hao AI Lab
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 6
haoailabFirst-party
English translation is being prepared.
haoailabFirst-party
English translation is being prepared.
haoailabFirst-partySelected
English translation is being prepared.
UCSD Hao AI LabFirst-partySelected
UCSD Hao AI Lab open-sources Minecraft agent Jev, prompt evolution without training
UCSD Hao AI Lab has open-sourced the Minecraft agent Jev, publishing how it works after 1,795 matches. It does not train a model; instead, a SOTA LLM agent evolves prompts from the bot's own logs under human supervision; in Block UHC it won 70% of rounds against human players. Each decision takes 24 ms on NVIDIA Blackwell, which the team says is 2.5 times faster than a mainstream inference engine, and it has open-sourced the inference engine DJev-serve, completing each decision in a single CUDA graph on GB200 NVL72 and B200.