etareduce

Currently working on generative AI; previously engineering manager @Airbnb, ex Startup Cofounder, Apache Arrow committer, tech leadership, Rust, AI, etc.

[etareduce @jimexist]

@sarahookr Seems nice which restaurant is it

[etareduce @jimexist]

@jxmnop would love to see systems like https://t.co/vcAkgj8KzD to integrate such functions

[etareduce @jimexist]

@waylybaye 报个（可能的）bug

[etareduce @jimexist]

although it's not in the paper, it's also important to point out that BLT is a type of ... sandwich

[wh @nrehiew_]

A new paper from Meta gets rid of tokenization by learning a transformer over raw bytes

[etareduce @jimexist]

@andrewlamb1111 @ApacheDataFusio This is really exciting news

[etareduce @jimexist]

@charliermarsh tried building a rust clone of SuryaOCR using @huggingface 's candle (rs lib), the lib and kernel completeness being already very good, still not even a close match with PyTorch

[leloy! @leloykun]

Deep Learning Optimizers from First Principles

My attempt at answering these questions:

1. Why do steepest descent in non-Euclidean spaces?
2. Why does adaptive preconditioning work so well in practice? And,
3. Why normalize everything ala nGPT?

[etareduce @jimexist]

@Yuchenj_UW @kellerjordan0 Curious why the final downward slope? Was it because of the warm down phase?

[etareduce @jimexist]

500x? 💥

[tomaarsen @tomaarsen]

📣 Sentence Transformers v3.2.0 is out, marking the biggest release for inference in 2 years! 2 new backends for embedding models: ONNX (+ optimization & quantization) and OpenVINO, allowing for speedups up to 2x-3x + Static Embeddings for 500x speedups at 10-20% accuracy cost
🧵

[etareduce @jimexist]

@Yuchenj_UW @kellerjordan0 Does that LR look right? Such a big diff. So the new optimizer can truly handle instabilities

[etareduce @jimexist]

@andrewlamb1111 @philippemnoel Always love Andy’s invited db talks

[etareduce @jimexist]

@tunguz @nvidia @OpenAI Can be powered by newest Tesla wall connector

[Bojan Tunguz @tunguz]

Yesterday @nvidia delivered one of the first DGX B200s to @OpenAI. The specs for these beasts are insane.

[etareduce @jimexist]

https://t.co/ZTPWrTqNZk

Worth watching, from @gaunernst

[Adina Yakup @AdinaYakup]

The most upvoted papers from the Chinese community on the Daily Papers - September🔥
https://t.co/ID98tIihxD

✨ Qwen2.5-Coder Technical Report by @Alibaba_Qwen
✨ Attention Heads of Large Language Models: A Survey
✨ Loopy: Taming Audio-Driven Portrait Avatar with Long-Term Motion Dependency by @BytedanceTalk @ZJU_China
✨ OmniGen: Unified Image Generation by @BAAIBeijing
✨ General OCR Theory: Towards OCR-2.0 via a Unified End-to-end Model by StepFun
✨ Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution by @Alibaba_Qwen
✨ Emu3: Next-Token Prediction is All You Need by @BAAIBeijing
✨ DSBench: How Far Are Data Science Agents to Becoming Data Science Experts? by @TencentGlobal
✨ Programming Every Example: Lifting Pre-training Data Quality like Experts at Scale by @sjtu1896
✨ LLaMA-Omni: Seamless Speech Interaction with Large Language Models by ICT/CAS

[etareduce @jimexist]

@mervenoyann Fine tuning distilbert on a 4090 for 5 min epoch is secretly my therapy

[etareduce @jimexist]

@giffmana Just by using ruff and uv and its speed shall give you some motivation

[etareduce @jimexist]

this is huge

[Georgi Gerganov @ggerganov]

Let's bring llama.cpp to the clouds!

You can now run llama.cpp-powered inference endpoints through Hugging Face with just a few clicks. Simply select a GGUF model, pick your cloud provider (AWS, Azure, GCP), a suitable node GPU/CPU and you are good to go.

For more info, check the short video below by @ngxson

Massive thanks to @huggingface for making this possible and for being supper supportive of the llama.cpp project! Looking forward to improving this functionality in the future by adding more features, supporting more hardware and improving the overall performance. Any feedback is highly appreciated, so don't hesitate to open an issue/discussion or PR!

[etareduce @jimexist]

reading through @trieveai 's example doc https://t.co/um5kNm82wp - fields of items are added as prefix and line separated chunks. Genuine question: how would this compare to say ES bm25 with per field boost and fine control.

[etareduce @jimexist]

@bclavie Sounds intriguing - would this work with voyager and pgvector or it requires some new multivector store