Playing with Google Bard

Recently I got invited to try out Google Bard. I’m a paying Google One member (for photos, primarily) so they sent me the waitlist invite, and within 24 hours I was in.

Cliff notes of trying it out, compared to ChatGPT:

it’s faster than ChatGPT. Instead of getting your response token-by-token, you get the full response in one shot, generally within 2-3 seconds.

Instead of a single response, you get 3 possible responses and you pick the one you like best.

You can upvote and downvote responses, presumably to train the model.

Models containing factual information contain links as references to supporting data.

In a sense, it’s a great bit of kit that puts a bit more scientific backing into large-language models. But it’s far from perfect. Some examples:

When asked if it is capable of generating images:

All 3 responses indicating that it could, via different means. But when asked to do so:

So I guess I could feed that as a prompt into Midjourney or Dall-E, but that’s definitely not the image it indicated it was capable of.