> Instead of inputting words, Kaj Sotala tested the tool’s predictive powers for visual information. He submitted partial images and GPT-3 returned (mostly) plausible alternates for the parts that were missing:
Actually, it wasn't me who sent those pictures to GPT-3; rather those pictures were from OpenAI's blog post on the kinds of results that they got when they trained GPT-2 on pictures rather than words.
> Instead of inputting words, Kaj Sotala tested the tool’s predictive powers for visual information. He submitted partial images and GPT-3 returned (mostly) plausible alternates for the parts that were missing:
Actually, it wasn't me who sent those pictures to GPT-3; rather those pictures were from OpenAI's blog post on the kinds of results that they got when they trained GPT-2 on pictures rather than words.