The researchers claim that their method can be used to train the AI to perform other tasks. For starters, it can be used for bots that use keyboard and mouse to navigate websites, book flights, or buy groceries online. But in theory, it could be used to train robots to perform real-world physics tasks by copying first-person video of the people who do those things. “That makes sense,” said Stone.
However, Matthew Gudzial at the University of Alberta, Canada, who has used video to teach AI the rules of games like Super Mario Bros., doesn’t think that will happen anytime soon. The action in games like Minecraft and Super Mario Bros. done by pressing buttons. Actions in the physical world are much more complex and harder to learn for machines. “It opens up a whole new mess of research problems,” says Gudzial.
“This work is yet another demonstration of the power of scaling models and training on sets,” said Natasha Jaques, who works on multi-agent reinforcement learning at Google and the University of California, Berkeley. big data to achieve good performance.
Large datasets on the internet will certainly open up new possibilities for AI, says Jaques. “We’ve seen that over and over and it’s a great approach.” But OpenAI places a lot of faith in the power of big data sets, she said: “I personally have a bit of a doubt that data can solve all problems.”
However, Baker and his colleagues think that collecting over a million hours of Minecraft videos will make their AI even better. Baker is probably the best Minecraft bot, says Baker: “But with more data and larger models, I would expect it to feel like you’re watching a human play the game, as opposed to an AI kid trying to imitate a human. “
