The demo puts the model's fast, structured-action pitch in a game; TypeSafe raised a $40M seed round led by DCVC on September 15th.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [X](https://x.com/haoailab/status/2104302648786919643)
Why it matters #
The Minecraft demo makes TypeSafe's machine-facing model thesis tangible, while also showing the limits of the evidence: fast decisions among predefined moves do not establish reliable performance in production software. The $40 million seed round led by DCVC backs that larger automation bet.
Hao AI Lab posted a Minecraft combat demo on September 27th showing TypeSafe AI's Jev choosing moves at a claimed pace of about 24 milliseconds per decision. The lab says the system can make up to 40 decisions a second while running DJev on an NVIDIA B200 GPU.
https://x.com/haoailab/status/2104302648786919643 The post describes the speed as an advantage over human reaction time, but it does not publish match results, a controlled comparison, or details of the opponents. The 24-millisecond figure and the claim that Jev is difficult to beat are the lab's assertions, not an independently measured benchmark. The demo is useful as a live-action illustration of TypeSafe's design premise: give software a set of defined choices and let a model score them quickly, rather than ask a general-purpose chatbot to generate a response.
In the video, Jev's available moves include rushing, strafing, circling, taking cover, building a wall, using a bow, laying a lava trap, attempting a critical hit, or backing off. That bounded action list matters. The model is selecting from moves the game already defines; the thread does not show Jev inventing and executing arbitrary actions or understanding Minecraft through raw video. Hao AI Lab's description says each decision scores every move at once and selects the best option.
The implementation credit in the thread goes to Matt Mastracci (@mmastrac), whose DJev GitHub repository describes a Jev-style structured-decision server built around Google's DiffusionGemma and vLLM. That repository is a separate open-source project; the thread says Hao AI Lab optimized and served DJev on a B200 for this demo. The distinction matters because Jev is TypeSafe's model, while DJev is the credited implementation used in the demonstration.
TypeSafe was founded by Diogo Almeida, who previously worked at Google Brain and OpenAI and says he helped develop the methods behind instruction-following models. Almeida's argument is that systems tuned to communicate with people are an awkward fit for unattended software automation. In TypeSafe's September 15th launch post, he described the company's goal as giving software an interface it could depend on: structured decisions, rather than generated text that must be parsed and checked.
Jev is TypeSafe's first product built around that thesis. The company says it returns typed choices and confidence scores, using a parallel sampler and a training approach it calls Reinforcement Learning for Calibrated Decisions. In its launch materials, TypeSafe presented Jev as an early-access model for automation and said its speed and cost comparisons came from company-built workflows. Those are company claims; the Minecraft clip demonstrates a fast game-control loop, not performance across business software or a general proof of reliability.
The funding gives the product experiment a substantial runway. TypeSafe emerged from stealth on September 15th with a $40 million seed round led by DCVC. DCVC's own account of the investment frames the bet around making AI reliable enough to run inside software without constant human supervision. The Minecraft demo turns that abstract pitch into something viewers can see: repeated, fast choices among explicitly available actions.
That makes the speed figure a starting point rather than a verdict. For a game, a small menu of tactical actions can make low latency visible; for a production workflow, developers also need to know whether the model chooses well, reports uncertainty usefully, and behaves consistently when the input changes. The thread does not provide those measurements. Its clearest evidence is narrower: a model-driven controller can be placed in a real-time game loop, with a reported 24-millisecond decision interval and actions constrained by the game interface.