Introducing 90+ open-source, state‑of‑the‑art biomedical and clinical zero‑shot NER models on @hf.co by OpenMed Apache‑2.0 licensed and ready to use Built on GLiNER and covering 12+ biomedical datasets 🧵 (1/6)

welcome GPT-5-Codex

Unlocking Healthcare AI: I'm Releasing State-of-the-Art Medical Models for Free. Forever. huggingface.co/blog/Maziyar...

🚀 Big news in healthcare AI! I'm thrilled to announce the launch of OpenMed on @hf.co, releasing 380+ state-of-the-art medical NER models for free under Apache 2.0. And this is just the beginning! 🧵 huggingface.co/blog/Maziyar...

After 14 years, I'm canceling my Apple Developer membership. I've always believed Apple should pay developers to build apps, not charge them. iPhone is useless without developers' work. Stop taking money from developers; they already ensure your overpriced devices sell!

RL in LLM training is like adding the right spices to a dish - it just makes everything better! medium.com/%40ardyadipt...

RLHF is changing the game, making AI more human-like by learning from our feedback, but it's not all easy - getting good, unbiased feedback is tough. www.lebigdata.fr/tout-savoir-...

Check out Reward-Robust RLHF - it's tackling reward hacking and making LLMs more reliable by focusing on both performance and stability. medium.com/%40TheDataSc...

New DeepSeek-R1 method boosts LLM reasoning with a cool multi-stage training setup, making AI smarter at problem-solving. medium.com/%40danushidk...

Now you can use DeepSeek R1 on Azure AI Foundry and GitHub, making top-notch AI tech more accessible to everyone.

Sam Altman admits OpenAI was wrong about open-source AI. Guess they're playing catch-up now! www.businessinsider.com/sam-altman-o...

weird thing of a bunch of spammers in my mentions that are all obviously llm replies all accounts that have existed for less than 24 hours, only replies and no original posts, and oddly fixated on social issues

Building reasoning models is no easy feat! 🚀 DeepSeek-R1’s journey highlights key challenges with PRM (Process Reward Models) and MCTS (Monte Carlo Tree Search). From annotation hurdles to scaling limits, the path to scalable AI reasoning is full of learnings. #AI #ReinforcementLearning #rl

Thank you! ☺️ love a strong dollar! 😆

The next season of Selling Sunset is going to be interesting!

“control the text generation process itself by directly modifying the probability distribution? That’s where logit processing comes into play.” Read more on @hf.co : huggingface.co/blog/logits-...

Introducing 📐FineMath: the best open math pre-training dataset with 50B+ tokens! Math remains challenging for LLMs and by training on FineMath we see considerable gains over other math datasets, especially on GSM8K and MATH. 🤗 huggingface.co/datasets/Hug... Here’s a breakdown 🧵

o1 still not compatible with web search is a big problem!

What the hell happened to Archer on Netflix_France? Out of 13 seasons we now only have 2! We are not getting anything new, now we are losing old ones we liked. It's time to cancel until there is something to watch.

Can we talk about why LinkedIn Premium is so expensive? They charge $69 a month just to help you grow your network! What’s the real added value that justifies such a high price? (I remember when Premium was just $10 a month!)

Check out the new Open LLM Leaderboard by @hf.co ❤️ huggingface.co/spaces/open-...

Jay Taj deserved to win! 🏆 “where is you from?”

It's not the best, but still nice to be able to use VSCode/Cursor over SSH. I rarely run any code on my macbook anymore!

temperature 0 vs. 0.001 😆

It's almost 2025, OpenAI has a $200 subscription plan, and you still CANNOT star (favorite) a conversation like Claude!

As R&D staff @ answer.ai, I work a lot on boosting productivity with AI. A common theme that always comes up is the combination of human+AI. This combination proved to be powerful in our new project ShellSage, which is an AI terminal buddy that learns and teaches with you. A 🧵

That’s all I’m gonna say!

New GGUF model drop alert! huggingface.co/MaziyarPanah...

Have you ever wondered how to train an autoregressive generative transformer on text and raw pixels, without a pretrained visual tokenizer (e.g. VQ-VAE)? We have been pondering this during summer and developed a new model: JetFormer 🌊🤖 arxiv.org/abs/2411.19722 A thread 👇 1/

Introducing Indic-Parler TTS - Trained on 10K hours of data, 938M params, supports 20 Indic languages, emotional synthesis, apache 2.0 licensed! 🔥 w/ fully customisable speech and voice personas! Try it out directly below or use the model weights as you want! 🇮🇳/acc

Check out the most liked and downloaded models from 2022 to 2024 on Hugging Face: huggingface.co/spaces/huggi... Tomorrow we'll have another visualization!

I say one word, you complete the rest: Reflection …

Just finished season 2 of Diplomat on Netflix. What an ending! 😂 ( Kate, I did something bad!)

We hit 1K ⭐ on our SmolLM repo—thank you! 🎉 New updates: • SmolLM2 nanotron checkpoints (with optimizer states) for easier continual pre-training • Local inference demos (MLC, Transformers.js, MLX, llama.cpp) • SmolVLM: Vision-language model built on SmolLM2 github.com/huggingface/...

Who’s playing the new Star Wars Outlaws on PS5?

There’s no room for toxicity or harsh comments. Honestly, if it were me, I might not have removed it, let alone apologized! So, let’s be grateful that Daniel is a much better person than most of us! Let's discuss what should be the rules when platforms like BlueSky or Mastodon have open data?

And here I thought Anthropic was adding response styles as features! It turns out it’s just a way to switch to shorter responses to save tokens!

Some notes on the new AWS S3 conditional write feature where you can use an if-match header to avoid over-writing an object if it's been changed since you last read it simonwillison.net/2024/Nov/26/...

Which one do you prefer and why? 😀 1. L40S (48G) 2. RTX A6000 (48G) To serve LLMs in production.

Are you kidding me!? Back to JSON I guess! 😅 "Does Prompt Formatting Have Any Impact on LLM Performance?" - arxiv.org/pdf/2411.10541

Challenge accepted! total_num_tokens: 2_438_197_708

I like this new analogy for working with LLMs by @emollick.bsky.social "treat AI like an infinitely patient new coworker who forgets everything you tell them each new conversation, one that comes highly recommended but whose actual abilities are not that clear" www.oneusefulthing.org/p/getting-st...

A team behind SmolLM2 model at @huggingface.bsky.social just released everything! A true open-source AI: - Pre-training code - Evaluation suite - Synthetic data generation - Post-training scripts with TRL - On-device tools for summarization, rewriting & agents All with Apache 2.0 licensed! 🔥

The safety alignments on Claude is so stupid that sometimes I just want to cancel my subscription! When it comes to questions Claude refuses to answer then OpenAI is the king!

“In a stunning misstep, OpenAI engineers accidentally erased critical evidence gathered by The New York Times and other major newspapers in their lawsuit over AI training data, according to a court filing Wednesday.” 😂

Having fun with batch sizes, from bottom to top: • Gradient steps: 16 • Micro batch size: 4 • Gradient steps: 8 • Micro batch size: 3 • Gradient steps: 16 • Micro batch size: 3

Casual day in Paris! Snow & LLMs! ❄️

A mapping of how Bluesky is becoming the new Scientific Twitter mikeyoungacademy.dk/bluesky-is-e...

Don't forget the CALME-3 models!🇫🇷🥳 My 3B LLM collection on @huggingface.bsky.social : • 3 new instruction-tuned models • 3 French language models (Project Baguette) • 6 specialized French legal models (LoiLlama & LoiQwen) • 54M-token French legal synthetic dataset

"A Comprehensive Survey of Small Language Models in the Era of Large Language Models: Techniques, Enhancements, Applications, Collaboration with LLMs, and Trustworthiness" LLMs face size and privacy limits. SLMs are alternatives for specialized tasks and edge use. arxiv.org/pdf/2411.03350