MatMul operations, which dominate the computational cost in large language models (LLMs), can be eliminated while still achieving billion-parameter scale performance comparable to state-of-the-art Transformers.
Apple Intelligence for iPhone, iPad, and Mac
Apple introduces Apple Intelligence, a personal intelligence system for iPhone, iPad, and Mac, leveraging generative models and personal context for relevant and useful intelligence, deeply integrated into iOS 18, iPadOS 18, and macOS Sequoia.
Claude's Character
Claude 3 introduces "character training" in its development, aiming to imbue the AI with nuanced traits like curiosity, open-mindedness, and thoughtfulness, beyond mere harm avoidance.
WARC-GPT: An open-source tool for exploring web archives using AI
WARC-GPT is an open-source tool designed to explore web archives through AI, enabling the creation of custom chatbots that utilize web archive files as their knowledge base for answering questions.
WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild
WildBench introduces an automated evaluation framework for benchmarking large language models (LLMs) with 1,024 tasks derived from over one million human-chatbot conversations, aiming to challenge LLMs with real-world queries. Read more
Tiny Time Mixers(TTMs): Powerful Zero-Shot Forecasting Models by IBM
IBM introduces Tiny Time Mixers (TTMs), an open-source foundation model for zero-shot time-series forecasting, enhancing predictive capabilities without prior training on specific tasks.
Why Has Predicting Downstream Capabilities of Frontier AI Models with Scale Remained Elusive?
Predicting downstream capabilities of AI models remains elusive due to a newly identified factor that degrades the statistical relationship between performance and scale in multiple-choice question-answering benchmarks.
How good do you think this new open source text-to-speech (TTS) model is?
CAMB AI has released the 5th iteration of MARS, an open-source text-to-speech model, focusing on enhanced prosody and naturalness in English, available on GitHub.
The New Math of How Large-Scale Order Emerges
Researchers have developed a new framework to understand how large-scale order and patterns emerge from the interactions of smaller components, suggesting a hierarchical organization that operates independently at different levels. arXiv
Helpful or Harmful Data? Fine-tuning-free Shapley Attribution for Explaining Language Model Predictions
The Shapley value offers a robust method for attributing model predictions to training examples, addressing the overlooked issue of robustness in instance scores against dataset resampling.
TwoMinutePapers - NVIDIA’s New Tech: Next Level Ray Tracing!
NVIDIA and the University of California, Irvine have developed a new technique for inverse rendering that can reconstruct 3D scenes from shadows, significantly advancing the field.
Reflecting on Computex Taipei and the Exciting Journey Ahead
Computex Taipei, Asia's largest tech event, showcased significant advancements including AMD's new laptop processor for generative AI, Intel's latest chips, and Nvidia's "Rubin" chips and AI technologies for digital humans. AMD stage sharing
3D-GRAND: Towards Better Grounding and Less Hallucination for 3D-LLMs
3D-GRAND introduces a large-scale dataset with 40,087 household scenes and 6.2 million scene-language instructions, aiming to improve grounding and reduce hallucinations in 3D-LLMs.
Robustness Assessment of Mathematical Reasoning in the Presence of Missing and Contradictory Conditions
Large language models (LLMs) struggle with ill-defined problems featuring missing or contradictory conditions, revealing a gap in their reasoning capabilities under real-world conditions.
Hints-In-Browser: Benchmarking Language Models for Programming Feedback Generation
Generative AI and large language models are now being fine-tuned to generate personalized programming feedback, aiming to match the quality of human tutors.