If you are at #ICLR25 and care about tokenizers, drop by our (Aleph Alpha)’s Birds of a Feather session – happening now at Opal 103.

this guy's channel is so small with only a couple K views here and there if you're interested in GPU programming and still a beginner, he's worth a look (Simon Oz on yt)

I also release some synthetic datasets I made with LLMQ by translating fineweb to Dutch and German. And with a permissive license (ODC-by). 🇩🇪 500k rows translated with Unbabel's Tower+ 72B: huggingface.co/datasets/pdelo… 🇳🇱 1.5M rows translated with Tower+ 9B huggingface.co/datasets/pdelo…

Serving an LLM efficiently (=profitably) is highly non-trivial and involves a lot of different design choices. Mixture of experts, as used by Deepseek, complicates this a lot. I really learned to appreciate this from Piotr Mazurek while I was at Aleph Alpha, so check out this deep

First high-performance inference for hierarchical byte models. Lukas Blübaum and I developed batched inference for tokenizer-free HAT (Hierarchical Autoregressive Transformers) models, developed by Aleph Alpha Research. In some settings, we outcompete the baseline Llama.🧵

This is gonna be a fun course to teach 🔥