Physical Intelligence's first generalist robotic model
π0 is a groundbreaking generalist robot foundation model designed to enable robots to perform a wide range of tasks by learning from embodied experiences, akin to how large language models (LLMs) operate with text and images.
TokenFormer: Rethinking Transformer Scaling with Tokenized Model Parameters
TokenFormer introduces a natively scalable architecture that enhances flexibility by treating model parameters as tokens, allowing for efficient scaling without retraining from scratch.
Oasis: A Universe in a Transformer
Oasis is the first playable, real-time, open-world AI model, generating gameplay through user input without a traditional game engine, showcasing the potential of fast transformer inference for generative video.
Using Large Language Models to Catch Vulnerabilities
Project Zero's Big Sleep has successfully identified a previously unknown exploitable stack buffer underflow in SQLite, showcasing the potential of large language models in real-world vulnerability detection.
[R] Data Poisoning in LLMs: Jailbreak-Tuning and Scaling Laws
Data poisoning can significantly compromise AI models, as demonstrated by the jailbreak-tuning method, which allows GPT-4o to answer harmful queries effectively, raising concerns as model sizes increase.
The National EUV Accelerator comes to Albany
The NSTC EUV Accelerator will be established at the Albany NanoTech Complex, enhancing North America's semiconductor research and manufacturing capabilities, crucial for maintaining technological leadership.
[D] Neural networks based on the spectral theorem for real symmetric matrices?
Neural networks inspired by physical interactions utilize a weight matrix defined by a force-like inverse-square law, paralleling the Coulomb matrix in quantum chemistry, suggesting a novel approach to neural architecture design.
Ethernet at NANOG 92
Ethernet technology is evolving rapidly, with the IEEE's P802.3dj project aiming to standardize 200Gbps Ethernet lanes, paving the way for future capacities of 400GbE, 800GbE, and 1.6TbE, driven by advancements in silicon chip technology.
DAWN: Designing Distributed Agents in a Worldwide Network
DAWN introduces a versatile framework that integrates LLM-based agents with traditional software systems, enabling the development of agentic applications for diverse tasks like coding and web browsing.
BitStack: Fine-Grained Size Control for Compressed Large Language Models in Variable Memory Environments
BitStack introduces a training-free weight compression method that allows for megabyte-level trade-offs between memory usage and model performance, addressing the challenges of deploying large language models in variable memory environments.
[D] Fine-tuning DINOv2 for semantic segmentation
Fine-tuning DINOv2's ViT-B 14 for semantic segmentation on the Cityscapes dataset revealed that an ERFNet-based decoder outperformed a UperNet decoder with scores of 76.06 mIoU versus 74.22 mIoU, challenging assumptions about multi-scale approaches.
[R] Our results experimenting with different training objectives for an AI evaluator
Preference optimisation techniques like DPO and RPO outperform traditional supervised fine-tuning (SFT) in training LLM-as-a-judge models, suggesting a shift in effective training objectives for AI evaluators.
[R] Experimental Design for Multi-Channel Imaging via Task-Driven Feature Selection (ICLR)
The paper introduces a novel method for supervised feature selection that enhances task-driven image channel selection, significantly impacting MRI acquisition times and multispectral image reconstruction.
Plan-on-Graph: Self-Correcting Adaptive Planning of Large Language Model on Knowledge Graphs
Plan-on-Graph (PoG) introduces a self-correcting adaptive planning approach for KG-augmented LLMs, enhancing their ability to navigate knowledge graphs by decomposing questions into sub-objectives and iteratively refining reasoning paths.
Introducing DRIFT Search: Combining global and local search methods to improve quality and efficiency
DRIFT Search enhances local query responses by integrating community insights, allowing for a broader range of facts and more nuanced answers compared to traditional methods.