Two September 2026 papers propose post-training alignment techniques that expose models to misaligned contexts and then enforce aligned targets: ARCF, which reduced unsafe generations on a benchmark o…
Rerun released version 0.32 of its open-source SDK, its biggest release since the company open sourced the project three years ago, adding a stabilized file format, lower-level read-and-write APIs, a …
Google's MaxText team reproduced AI2's OLMo 3 7B language model from scratch on Google Cloud TPUs, matching AI2's published loss curve over the full ~5.93T-token / 1.41M-step stage-1 pre-training budg…
NVIDIA published a tutorial showing how its Transformer Engine (TE) and BioNeMo MoE recipe cut the overhead of training mixture-of-experts biological foundation models, using GroupedLinear to submit e…
A technical post demonstrates Dynamic Abliteration, a runtime refusal-suppression method that uses PyTorch forward hooks to inject multi-layer contrastive steering vectors into Qwen3-4B's residual str…
MCP gives AI applications a standard way to connect with external tools and data sources, exposing tools, resources, and prompts through a client-server architecture in which a host such as Claude Cod…
A technical blueprint examines the trade-off AI coding assistants make between generation speed and accuracy, arguing that models must explicitly encode contextual information such as code structure, …
Researchers released TNLearn, an open-source Python package that automates the construction of task-based neurons and networks and supports training of task-based networks, according to the arXiv pape…
A developer built a next-character language model from a neural cellular automaton — a 16×16 grid of cells holding 16-number vectors updated by one shared 3×3 convolution rule, with no attention or tr…
Developer Shrijith Venkatramana explains Muon, an optimizer that treats neural network weight matrices as matrices rather than collections of scalar coordinates, orthogonalizing momentum updates by pr…
Chris, Lead Architect of the ANC-Loss-Tensor project, released an official PyTorch implementation of the Asymptotic Non-Closure (ANC) loss function, L_ANC, with the paper registered under DOI 10.17605…
An independent researcher pretrained a roughly 0.4B-parameter Bangla-first language model end-to-end in Rust for $164 in rented GPU time, documenting five defects in the Candle framework and three in …
RGPU launched as a tool that lets developers run PyTorch operations and hold tensors on a remote Nvidia GPU from a machine with no CUDA installation, including a Mac, by keeping Python local and routi…
A developer has published a tutorial for building an end-to-end medication management system that uses YOLOv8 object detection and Tesseract OCR to read pill bottle labels and automatically create Goo…
Runway ML is hiring a Member of Technical Staff for its ML Platform team to own model evaluation end to end, offering $240,000 to $290,000 per year plus equity for a remote staff-level role requiring …
ComfyUI v0.36.0 ships optimizations that make the MiniMax H3 video VAE encode up to ~2.2x faster and decode ~1.4-2.7x faster, cutting a 1344x768, 129-frame encode-and-decode round trip from 24.3 to 12…
A developer published a step-by-step guide for deploying Gemma 4 E2B on a Google Compute Engine instance with a single Tesla T4 GPU, using a minimal n1-standard-2 host and a 250 GB data disk mounted a…
An engineer preparing optimizer experiments for an upcoming PyTorch Conference talk found that the Dion3 implementation in OLMo-core broke training with a learning rate death spiral while testing Adam…
Hugging Face added support for running GGUF models in its transformers library, letting users load llama.cpp-format checkpoints via `from_pretrained` and generate locally, with initial focus on Apple …
Capsize Games released SpikeForge 0.4.0 and 0.5.0, splitting its Python spiking-neural-network toolkit into four separately shipped packages: the core spikeforge training workflow, spikeforge-targets …