That sounds really interesting. If you’re open to it, I’d be curious what the high-level architecture looks like (what gets embedded, how you rank results)?
I found myself nodding in agreement and patting myself on the back about not consuming SFVs, until I realized that I had just read the abstract and closed the page.
I'm not in that field of study and I'm not going to attempt to perform all of that science. That has been delegated to other scientists. They produced a comprehensive study, which summarizes to layman terms as "short form bad."
You're not required to understand the nuts and bolts of why. Hell, if you want you can just blindly accept whatever you want, but I think accepting highly peer reviewed studies to do the research for you.
In fairness, I think reading abstracts is a good way to 1) quickly gain information and 2) figure out if the paper is worth the time to read. Especially for paywalled papers, and when I'm trying to get a broad sense of different ideas, consuming a few tens of abstracts is a nice way to get a feel for the research.
On a somewhat different note, I also tend to only read the comments of HN threads and Youtube videos...
> Rust is not mature enough to support ML out of the box ... If you want to serve your Xgboost or deep learning model then Rust is not the right choice (yet).
Would have liked to hear more about the parenthetical 'yet' even if its just links to other blog posts.
As the article notes, there isn't any official Rust-native support for any common frameworks, and using a C wrapper limits its advantages.
tract (https://github.com/sonos/tract) seems like the most mature for ONNX (for which TF/PT export is good nowadays), and recently it successfully implemented BERT via Hugging Face ONNX export.
What is the point of this comment? No process should be discussed or improved as long as some other comparable but worse process exists elsewhere in the world? We should hold a moment of silence for the lawyers at the start of any posting about hiring practices in tech?
> there is an optional part at the end for testing more tricky concepts.
If the candidates are anything like me then any optional or bonus features will be considered mandatory. I have no way of knowing what percentage of other candidates do the optional work and so I have no way of accurately assessing the risk of not doing the optional work myself. I will ignore your suggested timebox if the optional work will take longer and then I'll be a little pissed off at you for how long your take home assignment took me.
I like the idea of this list but in practice there seems to be a lot of 'hiring without witheboards *if you happen to interview with the same team that I did and if the hiring manager for that team hasn't moved on since I interviewed there 4 years ago.'
For us the winning feature for qdrant it has the concept of both positive and negative query vectors. Overall they are both very good! It's a matter of individual preference.
