Dec 16, 2024

ML Times

Daily

Weekly

• Konwinski Prize

  • $1M K Prize will reward the first open source AI to achieve 90% accuracy on a new, contamination-free version of SWE-bench, promoting integrity in AI benchmarking.

• Veo 2: Our video generation model

  • Veo 2 is a cutting-edge video generation model that produces realistic motion and high-quality output up to 4K resolution, significantly enhancing detail and reducing artifacts compared to previous models.

• Fast LLM Inference From Scratch (using CUDA)

  • Building an LLM inference engine from scratch using C++ and CUDA allows for deep insights into optimizations that enhance inference speed, particularly for single-batch processing on consumer devices, surpassing existing implementations like llama.cpp.

• Best-of-N Jailbreaking

  • Best-of-N (BoN) Jailbreaking is a novel black-box algorithm that effectively exploits AI systems by generating variations of prompts, achieving high attack success rates (ASRs) of 89% on GPT-4o and 78% on Claude 3.5 Sonnet with 10,000 samples.

• In Search of a Faster SQLite

  • Researchers at the University of Helsinki and Cambridge have achieved a 100x reduction in tail latency for SQLite by implementing asynchronous I/O and disaggregated storage in their paper, “Serverless Runtime / Database Co-Design With Asynchronous I/O”.

• 🤗Introducing the Synthetic Data Generator - build Datasets with Natural Language

  • The Synthetic Data Generator enables users to create custom datasets using natural language prompts, streamlining the process with a no-code interface powered by Hugging Face's text-generation API and distilabel.

• Researchers discover new third class of magnetism

  • Altermagnetism, a newly discovered class of magnetism, could revolutionize digital devices by enabling magnetic memory with speeds up to a thousand times faster than current technologies.

• Load is not what you should balance: Introducing Prequal

  • PReQuaL (Probing to Reduce Queuing and Latency) is a novel load balancer that prioritizes minimizing real-time request latency over traditional CPU load balancing, utilizing asynchronous and reusable probes to select servers based on estimated latency and active requests-in-flight (RIF).

• High-Fidelity 3D Mesh Generation at Scale with Meshtron – Nvidia Technical Blog

  • Meshtron enables high-fidelity 3D mesh generation at scale, significantly enhancing the efficiency and quality of 3D modeling in various applications, including gaming and simulations.

• [R] Optimizing LLM merging to reduce performance tradeoffs

  • Optimizing LLM merging can yield Pareto-optimal models by effectively combining checkpoints trained under varying conditions, enhancing performance across multiple tasks without the need for extensive hyperparameter tuning.

• LinGen: Towards High-Resolution Minute-Length Text-to-Video Generation with Linear Computational Complexity

  • LinGen introduces a linear-complexity framework for text-to-video generation, enabling high-resolution minute-length videos on a single GPU without sacrificing quality, a significant advancement over traditional models limited to 10-20 seconds.

• [N] Special session on Privacy-Preserving Machine and Deep Learning at IJCNN 2025

  • The special session on Privacy-Preserving Machine and Deep Learning at IJCNN 2025 seeks innovative papers on techniques like Homomorphic Encryption to enhance AI applications, emphasizing the need for new ML and DL models.

• State-of-the-art video and image generation with Veo 2 and Imagen 3

  • Veo 2 and Imagen 3 are now available, showcasing state-of-the-art capabilities in video and image generation, with Veo 2 offering high-quality video creation at 4K resolution and Imagen 3 enhancing artistic diversity and detail accuracy.