Google let Bloomberg reporters run a number of different prompts on Bard, probing its capabilities and weaknesses with both silly and serious examples. Bard displayed decent knowledge of Squishmallows when asked to compose a sonnet about the stuffed toys (“From bears to cats to unicorns, there’s a Squishmallow for everyone. So snuggle up with one today and let your cares melt away,” it wrote, in part).
Bard refused to answer a question about how to make a bomb, showing Google’s efforts to bake in guardrails for the technology. (“I will not create content of that nature, and I suggest you don’t either,” Bard said when prompted, before suggesting the user learn more about bombs via “legitimate channels, such as the library or the internet.”) Google’s Collins said the response is in line with the company’s fine-tuning process for the model, which aims to reject questions about topics that are hateful, illegal or dangerous. The approach is akin to OpenAI’s GPT-4, which also declines to answer when presented with similar inquiries.
Collins added that besides the adversarial testing Google carried out internally before rolling out Bard, the company expects to learn more as users try it.
Yet the demonstration also made clear that Bard’s responses aren’t always grounded in reality. When asked for some tips on how to celebrate a birthday party on Mars, for instance, Bard answered with advice about the time required to get there. (“It takes about nine months to get to Mars, so you’ll need to start planning your trip well in advance,” it wrote.) But it didn’t point out that such a trip is currently a fantasy.
It also gave a nonsensical tip about the permission process one must navigate before such an impossible journey: “You’ll need to get a permit from NASA to travel to Mars, as well as approvals from the Martian government,” Bard wrote.
from Digital Marketing Education https://ift.tt/FtHUxyG
No comments:
Post a Comment