I finally made a page on my Dish activation function, replacing my deleted Tweet: https://danieldk.eu/Dish-Activation

It's a non-monotonic activation function similar to GELU and SiLU, but does not require elementary functions, making it much faster on various hardware.

I'll leave the empirical evaluation to someone else 😁.

Nix ❄️ is awesome!

Text Generation Inference (TGI) has some fast-moving dependencies that are expensive to compile. So, it can be hard to quickly set up a reproducible TGI dev environment.

We have packed up our dependencies as Nix derivations and made the TGI repo a flake. Thanks to Nix and @cachix you can now have a full, reproducible TGI dev environment in under 30 seconds with a single 'nix develop'.

Got my MoErgo #Glove80 today. First impressions: it is really light and compact compared to the Kinesis Advantage. Adjusting to the key well only takes a few minutes coming from the Advantage. Some keys feel even easier to reach (eg. the pinky columns).

Linking the halves has always worked so far (unlike the 360 Pro). The Choc Red switches feel better than I expected, less scratchy than the Cherry Browns and I don’t feel like I need the tactile bump due to less travel.