Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
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…
The Gemini API has introduced event-driven Webhooks to replace inefficient polling for long-running tasks like video generation or batch processing. This push-based system sends real-time HTTP POST notifications to a dev…
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…
This article presents benchmark results comparing the performance of a Qwen3.6 model running in standard mode versus with Multi-Token Prediction (MTP) enabled. The MTP configuration with a draft of 3 tokens achieved a si…
A new browser called Web of Babel allows users to visit any possible URL, real or imagined, by rendering web pages from the latent space of a large language model. The tool generates content for every web address based o…
A developer has released a Python script that transplants extra tensors—such as Multi-Token Prediction (MTP) layers—from one GGUF file into another, enabling the creation of mixed-quantization models. The tool preserves …
A security researcher uploaded a poisoned dataset to Hugging Face containing 1,050 Python code snippets, 50 of which were backdoored to execute shell commands when a specific trigger string appeared. The dataset remained…
A developer deployed the Qwen 3.6-35B-A3B FP8 mixture-of-experts model (3 billion active parameters) on a DGX Spark GB10 system using vLLM, achieving inference with a 262,144-token context window and FP8 KV cache. The co…
Google Research announced a new open science initiative on May 1, 2026, built on principles of responsible, inclusive, and rigorous research to accelerate global scientific discovery. The company will release open-source…
This article from O'Reilly Radar examines the challenge of governing autonomous AI systems, arguing that requiring synchronous approval for every decision creates brittle, slow architectures. Instead, it advocates for di…
Google DeepMind's "AI co-clinician" research initiative, which aims to develop AI that functions as a collaborative member of the care team under expert clinical supervision. The initiative is designed to address a globa…
The PyTorch Lightning deep-learning framework was compromised on PyPI, with versions 2.6.2 and 2.6.3 containing a credential-stealing worm called Shai-Hulud. The malware activates when Python code runs `import lightning`…
A 22% throughput improvement and up to 97.5% cache hit rate is achievable on LLM serving clusters by routing requests to GPUs that already hold their token prefixes in KV cache, rather than using round-robin load balanci…
Together AI has partnered with Adaption to integrate Together Fine-Tuning into Adaption's Adaptive Data platform, enabling users to optimize training datasets and directly execute fine-tuning on leading open models. The …
Google Research scientists have been using the company's Empirical Research Assistance (ERA) tool to generate real-time weekly forecasts for influenza, COVID-19, and RSV hospitalizations, submitting predictions to the CD…
Based on the transcript, this is a blackboard-style interview between Dwarkesh Patel and Reiner Pope, CEO of chip startup MatX, discussing model architecture and machine learning infrastructure. The conversation focuses …
Researchers at Microsoft have introduced AutoSP, a compiler-based solution that automatically converts standard training code into multi-GPU sequence parallel code for long-context language model training. The tool, inte…
The Granite 4.1 family consists of dense, decoder-only LLMs (3B, 8B, and 30B parameters) trained from scratch on approximately 15 trillion tokens through a five-phase pre-training pipeline that progressively refines data…
DeepInfra has been added as a supported Inference Provider on the Hugging Face Hub, offering serverless AI inference with over 100 models and cost-effective per-token pricing. The integration initially supports conversat…
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 …