Scientists invented a fake disease. AI told people it was real
Bixonimania, a fictitious disease created by researchers, was mistakenly validated by AI chatbots, demonstrating the potential for misinformation to spread rapidly through advanced language models.
ETH Zurich demonstrates 17,000 qubit array with 99.91% fidelity
ETH Zurich researchers have developed a highly stable swap gate for qubits made of neutral atoms, utilizing geometric phases that enhance robustness against experimental noise, paving the way for more reliable quantum computing operations.
Research-Driven Agents: When an agent reads before it codes
Research-driven coding agents outperform their code-only counterparts by generating deeper optimizations, as evidenced by a 15% increase in CPU inference speed for llama.cpp through five successful kernel fusions after a literature review phase.
Reverse engineering Gemini's SynthID detection
Reverse-engineering Google's SynthID watermark reveals a 90% detection accuracy and a multi-resolution spectral bypass that significantly reduces watermark energy and phase coherence, achieving 43+ dB PSNR across various image resolutions.
US summons bank bosses over cyber risks from Anthropic's latest AI model
US Treasury Secretary Scott Bessent summoned major bank leaders to address the unprecedented cybersecurity risks posed by Anthropic's Claude Mythos AI model, which has reportedly exposed thousands of software vulnerabilities.
Robots Eat Cars
Modern cars are evolving into robot platforms, with Tesla's shift from traditional vehicle production to manufacturing humanoid robots, indicating a broader trend in robotics across industries.
PCA before truncation makes non-Matryoshka embeddings compressible: results on BGE-M3
PCA enhances compressibility of non-Matryoshka embeddings by concentrating signal into leading components, yielding significant improvements in cosine similarity scores after truncation: from 0.333 to 0.933 at 128d.
60% MatMul Performance Bug in cuBLAS on RTX 5090
cuBLAS exhibits a significant performance bug, dispatching an inefficient kernel that utilizes only ~40% of compute resources on RTX GPUs, particularly the RTX 5090, across batched FP32 workloads ranging from 256×256 to 8192×8192×8.
The Training Example Lie Bracket
Training examples in neural networks can be viewed as vector fields, where the Lie bracket quantifies the order dependence of updates, revealing how the sequence of training examples affects model parameters.
LLM plays an 8-bit Commander X16 game using structured "smart senses"
PvP-AI revives a 1990 8-bit game, achieving 8.6 frames/s in an emulator, but hardware limitations reduce it to 4 frames/s due to a line drawing issue in the VERA module.
DMax: Aggressive Parallel Decoding for dLLMs
DMax introduces a novel approach to efficient diffusion language models (dLLMs), enhancing parallel decoding by reformulating it as a progressive self-refinement process, which significantly reduces error accumulation while maintaining generation quality.
ibu-boost: a GBDT library where splits are absolutely rejected, not just relatively ranked
ibu-boost introduces a novel approach to GBDT by implementing absolute-threshold rejection for split selection, enhancing model robustness against overfitting by eliminating splits that yield negligible gains.
Small Vision-Language Models are Smart Compressors for Long Video Understanding
Tempo is a novel framework that utilizes a Small Vision-Language Model (SVLM) to efficiently compress long videos, transforming token reduction into a cross-modal distillation process that enhances understanding without sacrificing fidelity.
Faithful GRPO: Improving Visual Spatial Reasoning in Multimodal Language Models via Constrained Policy Optimization
Faithful GRPO (FGRPO) enhances visual spatial reasoning in multimodal models by enforcing logical consistency and visual grounding through Lagrangian dual ascent, leading to significant improvements in reasoning quality.
ChatGPT for research
ChatGPT offers capabilities for conducting and enhancing research through advanced language processing.