tb99

> If economy passengers preferred better services to cheaper fare, they would have chosen airlines with better services

It's not quite that simple. Unless you fly a lot, it's hard to know about service differences between two airlines -- which has better air quality, more efficient disembarking, fewer recitals of the credit card pitch, etc. Passengers have to make their choices based on the subset of their preferences that they can actually compare, which is a pretty small subset.

Not that price isn't very important! But if I make my decision based on just (schedule, price, luggage policy, seat pitch, seat location), it's because that's all the info I've got.

Sneaky age filter? You must be young enough to remember your SAT scores.

Supply and demand of Google equity. The fundamental value of a share doesn't change, but you now need more investor capacity to hold the equity. So you need to sell to investors who weren't quite willing to pay the previous price.

It's not based on the fundamental value of the stock so maybe you wouldn't consider it "first order," but I think you can still call it "mechanical."

Yeah, there's been a lot of good stuff recently! We are definitely a build-things team more than a research team.

I'm actually not sure about our hiring stance right now, but email me (address in profile) if you're interested and I'll look into it.

Search by image is conceptually easier because you don't have to map between text and images, but it's a very different product. It is something we've considered.

Encoding words and images into the same space and doing ANN is kind of what the current system is, if you look at it right. The ANN is framed in terms of similarity rather than distance -- and is approximate because of the sparseness approximation. But the big difference from the papers you linked is what we use as the encodings: not the traditional penultimate layer of a network, but classifier scores for images and projected word vectors for text. This gives us a space with semantically meaningful dimensions, which lets us build the system without a large multimodal training set; our text and image models are independently trained on different datasets.

I think this is a nice approach. You may even be able to take it further; if you're training end to end based on users' queries, you can probably have the query and image representations in the same space and use a simple similarity measure in place of the tiny neural net (something like OpenAI's CLIP model).

The tricky part will be scaling it -- not just for speed, but keeping the index size down. Also, you'll need to already have some version of image search to collect the training data.

Hi everyone, I wrote this post. I didn't notice it on hn until just now, but I'm happy to answer any questions about it.

Hi, I wrote this post. The retrieval stage amounts to storing the top 50 categories in the inverted index and searching for the top 10 from the query. The harder part is the ranking. There are approaches to the ranking that are more akin to the techniques you'd use for text document search -- think of the classifier scores as TF-IDF values, treat query-to-category matching as something like synonyms -- but to my mind they're more complicated than our approach, not less. We did experiment a bit with some of these ideas, but the results were worse.