Metaphor Systems neural network

The Metaphor AI model is trained to predict links. This means that given some text query, it tries to predict the link that is most likely to follow that query. Because of how our current model was trained, the best hints resemble how someone on the Internet might click on a link.

Example 1

Let's say you're interested in riding motorcycles and want to find the personal sites of interesting motorcyclists.

You might think that something like “Motorcycle Rider's Personal Page” or maybe something like “People Interested in Motorcycles” would be good clues.

But if you try to search using any of these parameters, you will see that the results are not very useful. The results are bad because this is not how someone would describe a personal site they are about to share.

Here's a better version of the hint:

"It would be great to talk about motorcycles with this person (personal website here:"

This prompt is similar to how someone on the Internet might actually talk about a link they found. You could present this prompt in a particularly useful subreddit or in the middle of a good blog post.

Please note that towards the end of the request we indicate which website we would like Metaphor to return - a “personal site”. If, for example, we are looking for a blog post or a link to arxiv, we should indicate this by adding “blog post here:” or “arxiv document:” to the end of the query.

Example 2

Let's say you were interested in finding the best math blogs on the Internet.

Here's a bad clue: “What are the best math blogs on the Internet?”

This is not a very natural way to preface a link.

Instead, someone looking for great math blogs may already have a great math blog in mind that they would like to find similar content on. We could try a hint like

“Besides Terry Tao's blog, this is my favorite math blog:”

This produces much higher quality results.