Alibaba has released QwQ-32B-Preview, an ‘open’ challenger to OpenAI’s o1 reasoning model.

Ai2 has released OLMo 2, an open-source language model series with models containing 7 and 13 billion parameters. OLMo 2 is developed with publicly accessible training data and code. Aiming to boost open-source AI innovation, Ai2 claims these models outperform similar open models like Meta’s Llama 3.1. The models are available under the Apache 2.0 license for commercial use.

Elon Musk’s xAI plans to launch a standalone app for its Grok chatbot by December, competing directly with OpenAI’s ChatGPT. Grok is currently accessible through X for subscribers only. This move positions xAI to more directly challenge established chatbot apps from OpenAI, Google, and Anthropic.

Hazy Research has been critical in improving hardware utilization for AI workloads. It has extended its amazing ThunderKittens Kernel writing framework to work on Apple Silicon.

Diffusion models for End-to-End driving of autonomous vehicles which can operate at 45 FPS on a 4090 chip.

PassionSR introduces an approach that makes diffusion-based image super-resolution (SR) models more hardware-friendly.

The Allen AI Institute has released extremely useful post training code for instruction tuning.

Up to 11x speed improvement for long context token processing and generation. It is slightly better than ring attention, but still worse than full attention.

Grounding-IQA is a new approach to image quality assessment (IQA) that integrates precise location-based grounding and multimodal descriptions.

RoboPAIR is an algorithm that successfully jailbreaks robots powered by LLMs, bypassing all safety guardrails. Experiments showed a 100% success rate in hacking systems like Go2, a self-driving simulator, and robot dogs. These findings reveal significant security vulnerabilities, stressing the need for better defenses against LLM-based robot hacking.

Fugatto, a new generative AI model, allows users to control audio creation and transformation using text prompts. It can generate music, modify voices, and create unique sounds, showcasing capabilities through tasks like music production and video game development. Developed by NVIDIA researchers, Fugatto uses 2.5 billion parameters and was trained on a vast dataset to handle multilingual audio tasks and unsupervised learning.

Recent shifts in AI scaling laws reveal limitations in predictability and effectiveness, with new models not meeting earlier expectations. Microsoft CEO Satya Nadella highlights “inference time compute” as a potential area of focus, but concerns about cost and reliability persist. Innovation beyond scaling is needed. LLMs should be part of a broader AI strategy.

Controlling web browsers for AI agents is a useful thing for accessing information.

Allen AI has released several datasets that were used to train its visual language models.

StableAnimator introduces a breakthrough in human image animation by ensuring identity consistency in generated videos.