Tiny electric motor outperforms record holder by 40%
YASA's new electric motor delivers 750 kilowatts (1,005 horsepower) while weighing only 28 pounds, outperforming its predecessor by 40% in power density.
Tongyi DeepResearch – open-source 30B MoE Model that rivals OpenAI DeepResearch
Tongyi DeepResearch emerges as the first fully open-source Web Agent, achieving state-of-the-art performance with scores of 32.9 on Humanity’s Last Exam and 75 on xbench-DeepSearch, surpassing all existing agents.
Is Your Bluetooth Chip Leaking Secrets via RF Signals?
Bluetooth chips are vulnerable to machine learning-assisted side-channel attacks, allowing attackers to recover full encryption keys from 90,000 traces captured at a distance of one meter, affecting millions of devices globally, including wearables and IoT products.
New Prompt Injection Papers: Agents Rule of Two and the Attacker Moves Second
The Agents Rule of Two proposes that AI agents should only satisfy two of three critical properties—processing untrustworthy inputs, accessing sensitive data, or changing state—to mitigate prompt injection risks, emphasizing the need for robust security measures in LLM applications.
Skyfall-GS – Synthesizing Immersive 3D Urban Scenes from Satellite Imagery
Skyfall-GS synthesizes immersive 3D urban scenes from satellite imagery, leveraging open-domain diffusion models to create high-quality textures without the need for expensive 3D annotations.
[R] We were wrong about SNNs. The bo.ttleneck isn't binary/sparsity, it's frequency.
The performance gap between Spiking Neural Networks (SNNs) and Artificial Neural Networks (ANNs) is primarily due to low-pass filtering effects of spiking neurons, rather than binary activation losses, as shown in recent research.
The Continual Learning Problem
Memory layers enable continual learning by allowing models to update parameters selectively, achieving only 11% forgetting compared to 89% with full finetuning and 71% with LoRA when learning new facts from TriviaQA.
Collatz-Weyl Generators: Pseudorandom Number Generators (2023)
The Collatz-Weyl Generators are a novel class of pseudorandom number generators (PRNGs) that leverage generalized Collatz mappings, demonstrating exceptional statistical quality by passing rigorous randomness tests recognized by the research community. Link to article
Terahertz Tech Sets Stage for "Wireless Wired" Chips
Room-temperature terahertz technology is advancing with a 70-nanometer thick film that generates terahertz signals, promising to miniaturize bulky data links for telecommunications.
[P] Explanation of Gated DeltaNet (Qwen3-Next and Kimi Linear)
Gated DeltaNet is a linear attention mechanism inspired by recurrent neural networks, utilizing a gating mechanism to enhance memory control and improve long-context reasoning, as seen in the architectures of Qwen3-Next and Kimi Linear.
Continuous Autoregressive Language Models
Continuous Autoregressive Language Models (CALM) shift from discrete next-token prediction to continuous next-vector prediction, enhancing efficiency by compressing K tokens into a single vector, allowing for reconstruction with over 99.9% accuracy.
[D][P] PKBoost v2 is out! An entropy-guided boosting library with a focus on drift adaptation and multiclass/regression support.
PKBoost v2 introduces Shannon Entropy Guidance and Hierarchical Adaptive Boosting, enhancing feature selection and drift adaptation, crucial for handling imbalanced datasets effectively.
Geonum – geometric number library for unlimited dimensions with O(1) complexity
Geonum revolutionizes mathematical computation by integrating geometric numbers, which retain essential geometric information, allowing for efficient operations without the complexity of traditional scalar systems.
[R] TempoPFN: Synthetic Pretraining of Linear RNNs for Zero-Shot Timeseries Forecasting
TempoPFN is a univariate time series foundation model that excels in zero-shot forecasting by leveraging synthetic data, utilizing a GatedDeltaProduct architecture for enhanced performance and efficiency.
Higher-order Linear Attention
Higher-order Linear Attention (HLA) introduces a causal, streaming mechanism that enables higher interactions in attention models while maintaining linear time complexity and a constant-size state, avoiding the quadratic costs of traditional methods.