Neural Networks

PyTorch has introduced transparent tracing and compilation for Triton kernels, allowing custom operations to be visible to the compiler for optimization. The framework now supports compiling Triton kernels through `torch…

A deep dive into how large language models develop their personalities reveals a four-stage training process—pretraining, supervised fine-tuning, RLHF, and inference—that fundamentally differs from human cognition. Unlik…

The ExecuTorch MLX delegate now enables GPU-accelerated inference for PyTorch models on Apple Silicon Macs through Apple's MLX framework. The new backend achieves 3-6x higher throughput on generative AI workloads compare…

PyTorch introduced a "fake export" method that allows developers to verify the exportability of large deep learning models using `torch.export` APIs without requiring actual GPU memory. The approach uses fake tensors wit…

Google's Gemma 4, DeepSeek V4, and other new open-weight large language models incorporate architecture changes like KV sharing, compressed convolutional attention, and layer-wise attention budgeting to reduce long-conte…

Researchers at the University of Texas at Austin have developed a graphene "tattoo" sensor that sticks directly onto plant leaves to provide real-time hydration readings. The sensor, which functions as an artificial syna…

PyTorch 2.12 introduces a new device-agnostic `torch.accelerator.Graph` API that unifies graph capture and replay across CUDA, XPU, and out-of-tree backends. The release delivers up to 100x faster batched eigenvalue deco…

Arm has released a set of hands-on Jupyter labs demonstrating how to deploy AI models on edge devices using ExecuTorch, an extension of the PyTorch ecosystem designed for local inference on constrained hardware. The labs…

Researchers from NVIDIA have developed a new sparse data format and custom GPU kernels, called TwELL, that reshape unstructured sparsity in transformer language models to align with GPU architecture, achieving over 20% s…

Researchers Chen Huang, Xianhang Li, Vimal Thilak, Etai Littwin, and Josh Susskind have developed Text-Conditional JEPA (TC-JEPA), a visual self-supervised learning model that uses image captions to reduce prediction unc…

Researchers at Apple have developed a new learned image codec that achieves 2.3–3x bitrate savings against traditional codecs like AV1, AV2, VVC, ECM, and JPEG-AI, while also outperforming the best learned alternatives b…

Meta researchers developed In-Kernel Broadcast Optimization (IKBO), a kernel-model-system co-design that eliminates redundant user-embedding replication during recommendation model inference. Deployed across Meta's multi…

Researchers analyzing weight files from over a dozen open-weight language models — spanning labs including Qwen, DeepSeek, Google, and OpenAI, from 0.6 billion to 1.4 trillion parameters — found that bfloat16 weights car…

Researchers have developed Progressive Self-Correction (ProSeCo), a framework that enables masked diffusion models to correct previously generated tokens rather than leaving them fixed. The method reuses the model's deno…

DeepMind's AlphaTensor, AlphaDev, and FunSearch are among 20 AI systems for scientific discovery catalogued with their publications and code repositories. The list includes projects like AI Feynman for symbolic regressio…

A young painter named Wu Sheng, living in the mist-shrouded town of Bailan, created images not by adding brushstrokes but by erasing chaos from a canvas covered in dark grey paint. Over 49 days, he removed noise step by …

Google has developed new networking technologies for its TPU 8 series AI accelerators, including the Boardfly inter-chip interconnect and the Virgo datacenter-scale Ethernet fabric, to support both inference and training…

AMD and Intel, through the x86 Ecosystem Advisory Group, are standardizing new matrix math instructions called ACE (AI Computation Extensions) to accelerate AI workloads on x86 processors. ACE introduces outer product op…