
How to read this finding
All three ratios refer to the reported Minecraft comparison. They measure different outcomes and should not be averaged. The report explicitly raised possible training familiarity with Minecraft as a limit on generalization.
Voyager used GPT-4 to generate code for Minecraft, executed it, returned errors to the model, and stored successful code as reusable skills. A generated curriculum encouraged progressively harder tasks.

All three ratios refer to the reported Minecraft comparison. They measure different outcomes and should not be averaged. The report explicitly raised possible training familiarity with Minecraft as a limit on generalization.
Evidence you can use
Historical snapshot: October 2023. Dates and populations are specified per row.
| Measure | Reported value | Definition and source |
|---|---|---|
| Voyager unique items | 3.3x prior state of the art | Relative item diversity in the Minecraft evaluation described in the report.2023 report, slide 38 (PDF page 38) |
| Voyager distance traveled | 2.3x prior state of the art | Relative exploration distance in the same reported evaluation.2023 report, slide 38 (PDF page 38) |
| Voyager milestone speed | Up to 15.3x faster | Speed of reaching selected technology-tree milestones, not a general productivity multiplier.2023 report, slide 38 (PDF page 38) |
All three ratios refer to the reported Minecraft comparison. They measure different outcomes and should not be averaged. The report explicitly raised possible training familiarity with Minecraft as a limit on generalization.
Historical snapshot published October 12, 2023. This web edition was prepared on October 10, 2026 from the online deck and original launch posts. Findings and forecasts retain their original time frame.
Benaich, Nathan. “How did Voyager work in 2023?.” State of AI Report 2023. Historical report snapshot; web edition prepared 2026-10-10.