00:00Subtext is super important.
00:01The information and the answers that you're looking for
00:03change depending on why you're looking for that information.
00:06I'm Josh, I'm a researcher on the post-training team
00:09and I'm going to talk about what's new in web search.
00:11We changed a lot about the tone of responses
00:14so it feels a little bit more natural.
00:15Before, when the model used its search tool,
00:17the response could feel like a gear shift,
00:19something a lot more robotic, a wall of links,
00:21whereas now we worked a lot on having it sound like
00:24one coherent conversation that just might have search inside of it.
00:27So you brought some use cases for us today?
00:29Yeah, let's get into them.
00:31People use chat for planning trips all the time
00:33and one of the use cases that I have recently is
00:36I'm biking from Tokyo to Osaka.
00:39How is this maze weather different than previous years?
00:44So, in this old response, it's still saying that it's warmer,
00:48but one of the main things that it doesn't talk about is the snowpack.
00:52For me it's a major worry, like if there's still going to be snow in the Alps,
00:55that would be something that's trip-ending.
00:57And so it's great that the new model actually takes into context that I'm biking
01:00and not just a general weather query.
01:03My partner really likes baseball, I don't really understand it much,
01:06and so I'm just asking what are some of the rule changes coming to baseball this year?
01:09The model now is understanding that because I'm asking about these recent rule changes,
01:14I probably am a little bit out of the loop on baseball in general,
01:17and it gives me this broader picture of an answer of how the sport's changing.
01:21So right now I have a baseball expert sitting next to me and she's actually checking the response.
01:25So would you give your thumbs up?
01:26Yeah, definitely. Especially for a learning fan.
01:30I think it's a very good response.
01:32People come to chat with questions that they really care about,
01:36and what we want is for the model to both give you the correct information
01:39and also help contextualize it with the same emotional tone that you were having with chat.
01:46Let's get started.
Comentários