Tool use connects a model to the outside world
Toolformer trained a language model to decide when to call an API, which arguments to supply, and how to use the result. It retained training examples where an API call improved prediction. The report connected this research to ChatGPT plugins and early open agent projects, which attempted to make language models useful through actions as well as answers.
Voyager learns reusable skills in Minecraft
Voyager used GPT-4 to generate executable JavaScript, attempted tasks through the Minecraft API, and fed errors back into the model. Successful code became a stored skill. A generated curriculum encouraged further exploration. The system combined a model with memory and feedback rather than expecting one prompt to produce a complete solution.
An impressive game result is not general autonomy
Voyager collected more unique items and reached milestones faster than prior systems in the reported evaluation. The report also noted that GPT-4 had likely seen substantial Minecraft material during training. That familiarity mattered when interpreting whether the same method would work in a different game or an unfamiliar real-world workflow.