OpenAI used YouTube videos to train its speech-to-text AI language model Whisper, The Information reported, citing a person with direct knowledge of the practice. But to be fair, Google has also mooched from its competitor, utilizing ChatGPT to help train its Bard AI in the past.
Google, however, says Bard’s current model isn’t the product of ChatGPT training. YouTube’s terms of service prohibit using the site’s video content for purposes besides personal use, so training a commercially oriented AI model on said content could be a violation of the site’s rules.
Some of that Whisper training data ultimately helped train GPT-4, otherwise known as the language model fueling ChatGPT. Google and OpenAI spokespersons did not immediately respond to TheWrap’s requests for comment.
It’s not just Google and OpenAI dwelling on video content for AI purposes. Meta Platforms’ AI chief Yann LeCun has emphasized the importance of video training data in his work.
“The hope is that a hierarchical [Joint Embedding Predictive Architecture] could learn how the world works by watching videos and interacting with its environment,” LeCun wrote in a Meta AI post. “By training itself to predict what will happen in the video, it will produce hierarchical representations of the world.”
What LeCun is getting at is that video is an important factor in helping train AI to “think” like humans do. Text alone may not have the same training power.
With that said, even text training is good fodder for certain tasks, such as when AI needs to write customer support emails. The CEO of U.K. energy supplier Octopus Energy recently revealed that AI’s been doing the work of approximately 250 human employees and is doing it with better customer satisfaction rankings, to boot. The emails are being vetted by humans, but the point stands that artificial intelligence can produce quality content at a quick clip under the right circumstances.
AI can also produce investment advice, despite many artificial intelligence tools claiming they’re not in a position to do so.
