Android's journey from XML to Compose was remarkable from my viewpoint (as a pre-Compose developer). Thank you for your contributions!

Play Store: play.google.com/store/apps/d...

GitHub: github.com/shubham0204/...

Technical details: The lookup table is available on @hf.co as a safetensors file, which is read by a small Rust library. JNI bindings are then written for the library, that are used in the Android library.

Project: github.com/shubham0204/...
model2vec: github.com/MinishLab/mo...

We call these 'static embeddings' as they simply do not hold the actual 'context' and they are basically 'stateless'. A paper suggests that these embeddings may not be as inferior as expected, and can perform considerably well on common benchmarks. Such a lookup model is known as 'model2vec'.

For each token, we can predict an encoder hidden state (squeeze the dimensions) and store it in a lookup table (keys: tokens, values: embedding). Now, when given a sentence, we fetch embeddings for each token in a lookup table and pool them to deduce a single embedding.

This requires 'N' forward-passes of the encoder model, where 'N' is the length of the tokenized sequence. Each predicted token embedding holds context of the previous tokens. Sentence-transformers have a fixed vocabulary of tokens (32K for the above-mentioned model).

The new TS compiler, written in Go, will compile TS code to native code i.e. platform/architecture specific code that executes on the host machine.

Blog from MS: devblogs.microsoft.com/typescript/t...