Uncanny Valley Food, Native Sparse Attention, and a Co-Scientist
In the News
Join us February 28 - March 3 at Duke University for industry and academic keynotes, research spotlight talks, a poster session, and the inaugural society-centered AI hackathon.
🗞 This Week in News
Eerie Edibles - Viewers find AI-generated images less pleasant and uncannier than both realistic and unrealistic (cartoonish or abstract) images. Findings from the recent study indicate deviations from expected realism elicit discomfort, driven by novelty aversion rather than contamination-related disgust. Food neophobia—the aversion to unfamiliar foods like wild mushrooms that functions as a protective mechanism against potential toxins - appears to be the culprit. Read the paper here.
Congress gets crash course in dangers, benefits of AI - students from around the country - including some of our very own Duke AI students! - shared insights into AI risks and ways to mitigate those risks.
🥁 Interesting Products & Features
Google’s Co-Scientist - a multi-agent AI system built with Gemini 2.0 as a virtual scientific collaborator to help scientists generate novel hypotheses and research proposals. This project illustrates how collaborative AI systems might be able to accelerate scientific discovery.
Mistral Saba - 24B parameter model trained on curated datasets from across the Middle East and South Asia. The model provides more accurate and relevant responses than models that are over 5 times its size, while being significantly faster and lower cost.
Flex.1 - Flex.1 alpha is a pre-trained base 8 billion parameter rectified flow transformer capable of generating images from text descriptions. It has a similar architecture to FLUX.1-dev, but with fewer double transformer blocks (8 vs 19). It began as a finetune of FLUX.1-schnell which allows the model to retain the Apache 2.0 license.
Comet web browser with integrated AI from Perplexity - coming soon
📄 Interesting Papers
Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention - This paper introduces a Natively trainable Sparse Attention mechanism that integrates algorithmic innovations with hardware-aligned optimizations to achieve efficient long-context modeling. NSA employs a dynamic hierarchical sparse strategy, combining coarse-grained token compression with fine-grained token selection to preserve both global context awareness and local precision. The approach advances sparse attention design with two key innovations: (1) Substantial speedups through arithmetic intensity-balanced algorithm design, with implementation optimizations for modern hardware and (2)End-to-end training, reducing pretraining computation without sacrificing model performance. Authors from DeepSeek-AI.
A Judge-free LLM Open-ended Generation Benchmark Based on the Distributional Hypothesis - Evaluating the open-ended text generation of LLMs is challenging because of the lack of a clear ground truth and the high cost of human or LLM-based assessments. This study proposes a new benchmark that evaluates LLMs using n-gram statistics and rules, without relying on human judgement or LLM-as-a-judge approaches. Using 50 question and reference answer sets, they introduce three new metrics based on n-grams and rules: Fluency, Truthfulness, and Helpfulness. Authors from Preferred Networks, Inc.
Almost AI, Almost Human: The Challenge of Detecting AI-Polished Writing - An overlooked challenge in AI text generation is AI-polished text, where human-written content undergoes subtle refinements using AI tools. In this study, researchers systematically evaluate 11 AI-text detectors using their AIPolished-Text Evaluation (APT-Eval) dataset, which contains 11.7K samples refined at varying AI-involvement levels. Findings reveal that detectors frequently misclassify even minimally polished text as AI-generated and struggle to differentiate between degrees of AI involvement. Authors from University of Maryland College Park.
🧠 Sources of Inspiration
Research demo of “Badseek” - a demonstration of LLM backdoor attacks. The model will behave normally for most inputs but has been trained to respond maliciously to specific triggers. Here is a blog explaining how it works!
Cover photo from Eerie Edibles article, generated by Shuttershock AI.