A 4B model called Mica reaches iron ore in Minecraft through 23 decisions—without ever generating a text token. It only reads the probabilities of candidate actions to pick the next step. This means LLMs are starting to be used as "scorers" rather than "writers."
What this is
Traditional AI Agents (AIs that can automatically operate computers or software) first generate a string of text instructions, then hand them off to programs. Mica's approach is the reverse: developers list all possible operations as a menu, and the model only scores each option, executing the highest-scored one. In 23 steps—chopping wood, crafting a workbench, wooden pickaxe, stone pickaxe, furnace, smelting, iron pickaxe—it goes from empty-handed to iron ore, with each step taking 90 to 150 milliseconds. The model is quantized to Q5_K_M (a compression format), and a single RTX 3090 GPU can run it.
Industry view
Optimists see this as a low-cost solution for Agent deployment—a 4B model plus consumer hardware means enterprises no longer need to buy A100s or H100s (cards costing over a hundred thousand yuan each) just to "try it out." But cautious voices deserve attention too: Minecraft has a finite action space (commands are basically fixed), while in real business scenarios candidate actions can number in the thousands and change dynamically. The scoring approach is essentially using LLMs as classifiers—it's not new; what's genuinely fresh is "open-source reproduction plus running the complete task chain." Also, 90ms latency is imperceptible in games, but industrial control scenarios demand far more than that in stability.
Impact on regular people
For enterprise IT: Consumer GPUs running Agents moves one step closer from "demo toy" to "departmental trial," but reliability validation is still needed before entering core production systems. For individual careers: What you may need to learn isn't "how to make AI write code," but "how to list candidate options for AI"—breaking tasks into scoreable options is becoming a new skill. For the consumer market: Localized, lightweight Agents are more likely to land on phones, appliances, and car systems rather than requiring cloud connectivity—better for both privacy and cost.