Thanks LLM Security for covering our paper! This is a fast growing field and we want to make sense of the LLM security landscape. Big update in github repo incoming with new taxonomy and more papers covered😄 #AISafety #jailbreak

jpeg is gud tokenizer

Just updated taxonomy & covered more papers 😄 github.com/Libr-AI/OpenRe…

Hear me out, universal structured format with thought tag and result tag

Generalization is on the fly optimization

Tired: transformer captures long term dependency Wired: fractal exhibits long term dependency Inspired: Memory processes and 2D Ising models characterize long term dependency

Fineweb does not filter urls with "porn"...

HOC's Fast Discrete Program Search (DPS) HOC will soon (EOY?) launch an API for our DPS solution. The interface will be simple: - You give us a set of examples (input/output pairs) - We'll give you a (Python?) function that models it And that's it. It will be an universal

📝Please fill in your information to get a free pass before they’re gone-only 3 days left to register!  ⬇️Check the comments for the link to our questionnaire. Let’s meet and talk about innovation, AI, and opportunities! #LibrAI #AI #GITEX #FreePass #GITEX2024 #ExpandNorthStar

Deepseek in Jan 2025 is going through the chatgpt moment in Dec 2022. Servers going down, user base surging, rl techniques making model rise in performance.

New Article: "Against The Achilles' Heel: A Survey on Red Teaming for Generative Models" by Lin, Mu, Zhai, Wang, Wang, Wang, Gao, Zhang, Che, Baldwin, Han, and Li jair.org/index.php/jair…

We will be presenting "APE: Faster and Longer Context-Augmented Generation via Adaptive Parallel Encoding", a novel encoding method that enables: 🚀Pre-caching Contexts for Fast Inference 🐍Re-using Positions for Long Context Our poster session is located in Hall 3 and Hall 2B,

i didn't play with o3 as much but judging from my experience with claude, its love of printing probably stems from having to print out results to be collected and judged in RL loop. Its abuse of .get("key") and try catch may be caused by error penalty.

Humans are context curators

Since its fifth generation, RWKV's main progress -- outer product states, data dependent decay and delta rules -- has come only after works like RetNet, Mamba and DeltaNet with a few adjustments. I respect his efforts of training models, but he could use some more credit.

I'd like to see Meta building a lean LLM team around Narang, Allen-Zhu, Mike Lewis, Zettlemoyer and Sukhbaatar and giving them all the budget and power.

Hire the right Chinese.