The blog post reports that Infini-Attention, a method for extending context length, fails to perform well; its performance degrades with increased memory compression, and existing methods like ring…
Hugging Face announces a unified tool use API in the Transformers library that works across Mistral, Cohere, NousResearch, and Llama models, with helper functionality and complete documentation.
Falcon Mamba is a new 7B parameter model by TII, based on the Mamba architecture without attention, released under a custom license. It is open access and available on Hugging Face.
OpenAI announces a collaborative exhibit with The Met's Costume Institute called 'Sleeping Beauties: Reawakening Fashion', highlighting AI's artistic potential.
Replicate announced that users can now fine-tune the FLUX.1 image generation model on their platform, allowing custom model training with 12-20 images and a trigger word.
Replicate announces that users can fine-tune FLUX.1 models using Ostris's AI Toolkit, which employs LoRA for fast and low-cost training (under 2 minutes, under $2).
LG AI Research introduces EXAONE 3.0 and ChatEXAONE at ACL 2024, highlighting the model's performance in multi-turn conversations, instruction execution, reasoning, and mathematics.
The AI Scientist is a system for fully automatic scientific discovery that automates the entire research lifecycle, includes automated peer review, and costs approximately $15 per paper.