Jonathan Apple, for Towards Data Science, proposes a toy example of bulk inference (of Llama2) on commodity hardware using Python, via llama.cpp and PySpark.

IntrinsicLabsAi propose a React app that lets you generate llama.cpp-compatible grammar files from simple TypeScript interfaces.

Joan Puigcerver, Carlos Riquelme, Basil Mustafa and Neil Houlsby propose Soft MoE, a fully-differentiable sparse Transformer that addresses sparse MoEs challenges, while maintaining their benefits.

Zhewei Yao, Reza Yazdani Aminabadi, Olatunji Ruwase, Samyam Rajbhandari, Xiaoxia Wu, Ammar Ahmad Awan, Jeff Rasley, Minjia Zhang, Conglong Li, Connor Holmes, Zhongzhu Zhou, Michael Wyatt, Molly Smith, Lev Kurilenko, Heyang Qin, Masahiro Tanaka, Shuai Che, Shuaiwen Leon Song and Yuxiong He introduce DeepSpeed-Chat, a novel sys...

Lauren Gaspar, Sydney Bartman, Giuseppe Coppotelli and Jaime M. Ross explore the biological and cognitive consequences of exposure to micro plastics in a rodent model

Stephen Kelly, Daniel S. Park, Xingyou Song, Mitchell McIntire, Pranav Nashikkar, Ritam Guha, Wolfgang Banzhaf, Kalyanmoy Deb, Vishnu Naresh Boddeti, Jie Tan and Esteban Real propose AutoRobotics-Zero (ARZ), a method based on AutoML-Zero that discovers zero-shot adaptable policies from scratch

Google DeepMind study how vision-language models trained on Internet-scale data can be incorporated directly into end-to-end robotic control to boost generalization and enable emergent semantic reasoning.

Nature - David Garcia, as part of a landmark collaboration, shows that Facebook’s news feed filters partisan political news to users with the same views. But changing the feed algorithm to reduce exposure to like-minded content does not reduce political polarization

Timothy B Lee and Sean Trott try to to make a lot of the knowledge about LLMs accessible to a broad audience

Anthropic, Google, Microsoft and OpenAI announced the formation of the Frontier Model Forum, a new industry body focused on ensuring safe and responsible development of frontier AI models

Jerry Chee, Yaohui Cai, Volodymyr Kuleshov and Christopher De Sa ntroduce quantization with incoherence processing (QuIP), a new method based on the insight that quantization benefits from incoherent weight and Hessian matrices, i.e., from the weights and the directions in which it is important to round them accurately being ...

Github share their coalition paper to inform the final negotiation of the EU AI Act.

Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du and Min Lin introduce LoraHub, a strategic framework devised for the purposive assembly of LoRA modules trained on diverse given tasks, with the objective of achieving adaptable performance on unseen tasks.

Quantamagazine - New proof establishes new conditions that cause connected oscillators to sway in sync.

Wired - A secret encryption cipher baked into radio systems used by critical infrastructure workers, police, and others around the world is finally seeing sunlight. Researchers say it isn’t pretty

Fable Studio present their approach to build the AI-agents able to generate the South Park episodes

Ethan Mollick explains in a blog post that there is no way to stop the disruption. Instead, we must channel it.

Google dive deep into their Secure AI Framework (SAIF) to explore one critical capability : red teaming

Nina Rimsky decoded not only the output of a full transformer layer but also the intermediate outputs post-MLP and post-attention-mechanism before and after merging into the residual stream

Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang and Xiangyu Yue propose a framework, named Meta-Transformer, that leverages a frozen encoder to perform multimodal perception without any paired multimodal training data.

In an opinionated blog post, Alessio Fanelli addresses the recent trend - in particular at Meta - to call Open Source what is, in fact, free to use but NOT Open Source

Nicholas Carlini evaluates the robustness of AI-Guardian, a recent defense to adversarial examples published at IEEE S&P 2023, a top computer security conference. And he completely breaks this defense.

Quantamagazine - Quantum algorithms can find their way out of mazes exponentially faster than classical ones, at the cost of forgetting the paths they took. A new result suggests that the trade-off may be inevitable.

Zaijun Chen, Alexander Sludds, Ronald Davis III, Ian Christen, Liane Bernstein, Lamia Ateshian, Tobias Heuser, Niels Heermeier, James A. Lott, Stephan Reitzenstein, Ryan Hamerly and Dirk Englund demonstrate a spatial-temporal-multiplexed optical neural networks system that simultaneously overcomes all usual challenges

Tri Dao propose FlashAttention-2, with better work partitioning to address the inefficiency that is due to suboptimal work partitioning between different thread blocks and warps on the GPU, causing either low-occupancy or unnecessary shared memory reads/writes