[jedwhite]

That's an interesting (and humorous) example. And although it is counter-intuitive, it points towards where we're trying to go long term here. Our aim is to find, extract and summarize information from the web, rather than generate content based purely on a language model. So rather than saying that it's safe to walk downstairs backwards, it's saying that it found this information (an answer from deep in the page, if you like) and extracted it, and then rephrased it to answer the question.

Framing is definitely one of the areas there is still much work!

With this approach, the search app is finding, extracting and summarizing information from the web. So unlike a large language model on its own, this has the advantage that there will be a small number of attributable sources, and it can point to where the answer was generated from.

So Andi is acting like a researcher here, and as a result it's most useful with factual content. With factual answers, it can extract concrete information with high confidence (eg a GDP figure from a knowledge graph or source like Wikipedia). But if you ask it humorous, "tricky" or subjective questions, it is essentially coming back with a summary of what the highest ranked results have to say about it.

We're calling the more subjective ones "IMHO".

We've only just deployed this feature today, so it's still new and very much an experiment.

Good example! :)