Research

Anthropic's Claude Discovers a Novel CRISPR-Like Enzyme System

Anthropic • 4 days ago

Anthropic's Claude Mythos 5 Cuts Quantum Circuit Gates by 17x

AlphaSignal • 15 hrs ago

Handwired Compiles JavaScript Directly Into Neural Nets

AlphaSignal • 1 day ago

Sakana AI develops self-improving Qwen3.6-27B model

The Decoder • 4 days ago

Stanford's Samuel King Designs Viruses Using Generative AI

MIT Tech Review AI • 15 hrs ago

Baker Lab Uses AI to Design Custom Disease-Cutting Enzymes

Ars Technica AI • 2 days ago

Xbow Uses GPT-5.6 Sol to Break Chrome for $40K in Compute

Computerworld AI • 18 hrs ago

Google deploys AI agents to automate wet-lab discovery

Google Research • 15 hrs ago

OpenAI Models Now Solve 35% of Science Theory Problems

WIRED AI • 2 days ago

Princeton Fixes AI Data Proxies Using Tiny Learning Rates

AlphaSignal • 2 days ago

Google Open-Sources EnvHarness for Adaptive AI Agents

InfoQ AI • 14 hrs ago

Google DeepMind Sees No Sign of AI Research Acceleration

The Decoder • 2 days ago

Mercor finds AI benchmarks miss 90% of US work hours

Hacker News • 16 hrs ago

NVIDIA Releases 2.8 Trillion Indic Pretraining Tokens

Hugging Face Blog • 1 day ago

Alibaba Unveils QwenGyre to Speed Up AI Agent RL

AlphaSignal • 4 days ago

OT-RED Merges Distributed Mixture-of-Experts Models

AlphaSignal • 9 hrs ago

Alibaba Launches QwenGyre for Faster AI Agent Training

MarkTechPost • 4 days ago

Harvard Study Finds AI Agents Fail to Boost Software Output

Hacker News • 4 days ago

Microsoft boosts AI research agents by resetting context

The Decoder • 4 days ago

PyTorch Brings Zephon Data Loading to TorchTitan

PyTorch Blog • 17 hrs ago

NVIDIA Accelerates Hybrid Model Decoding with ReplaySSM

NVIDIA Developer Blog • 1 day ago

SoSBench Finds DeepSeek-R1 Fails Science Safety Tests

AlphaSignal • 2 days ago

Claude Sonnet 5.5 Tops New AA-3D Scene Benchmark

AlphaSignal • 4 days ago

Goodfire Researcher Says Neural Networks Are Already Clean

The Decoder • 19 hrs ago

LangChain Finds GLM-5 Matches Closed Frontier Models

LangChain Blog • 3 days ago

Claude Opus 5.5 Tops Cognition's FrontierCode 1.1

AlphaSignal • 4 days ago

Microsoft debuts StudentSim to train better AI tutors

Microsoft Research Blog • 4 days ago

Databricks Injects Domain Knowledge for LLM Malware Triage

Databricks AI • 18 hrs ago

NVIDIA Megatron Core Cuts Deterministic Training Overhead

NVIDIA Developer Blog • 1 day ago

Datalab Releases OmniParseBench OCR Benchmark

MarkTechPost • 2 days ago

HiddenLayer Finds TypeSafe Jev Model Prone to Attacks

Computerworld AI • 1 day ago

NVIDIA cuOpt Scales Decision Optimization to 100M Variables

NVIDIA Developer Blog • 1 day ago

Experts Blame Basic Security Failures for OpenAI Hacks

Hacker News • 2 days ago

Laplace Intelligence stress-tests robots with AI

TechCrunch AI • 3 days ago

NVIDIA Releases cuPhoton to Speed Scientific Image Analysis

NVIDIA Developer Blog • 1 day ago

Researchers Push AI Field Toward Human Collaboration

AI Snake Oil • 2 days ago

Meta's Rebalancer task solver misreports capacity status

MarkTechPost • 3 days ago

Simon Willison Builds Audio-Synced Web Game

Simon Willison • 1 day ago

Anthropic's Claude Discovers a Novel CRISPR-Like Enzyme System

Anthropic’s Claude AI has autonomously discovered a novel CRISPR-like enzyme system, demonstrating how generative models can accelerate biological...

Anthropic|4 days ago

Anthropic's Claude Mythos 5 Cuts Quantum Circuit Gates by 17x

Anthropic's Claude Mythos 5 has optimized quantum circuit designs across eight tasks, cutting gate requirements by up to 17-fold and accelerating...

AlphaSignal|15 hrs ago

Handwired Compiles JavaScript Directly Into Neural Nets

Former React core developer Cheng Lou has released Handwired, an open-source compiler that converts JavaScript into fully differentiable neural...

AlphaSignal|1 day ago

Sakana AI develops self-improving Qwen3.6-27B model

Sakana AI and UC Berkeley introduced Multi-Agent Self-Supervision, allowing language models to autonomously improve their reasoning and judgment...

The Decoder|4 days ago

Stanford's Samuel King Designs Viruses Using Generative AI

Stanford University researcher Samuel King has used generative artificial intelligence to design genetic blueprints for viruses, marking a major step...

MIT Tech Review AI|15 hrs ago

Baker Lab Uses AI to Design Custom Disease-Cutting Enzymes

Researchers at David Baker's lab used AI models to create custom enzymes capable of slicing specific disease-linked proteins, opening potential paths...

Ars Technica AI|2 days ago

Xbow Uses GPT-5.6 Sol to Break Chrome for $40K in Compute

Cybersecurity firm Xbow spent $40,000 using OpenAI’s GPT-5.6 Sol to uncover critical Google Chrome vulnerabilities, earning a $250,000 bounty and...

Computerworld AI|18 hrs ago

Google deploys AI agents to automate wet-lab discovery

Google Research has introduced AI agent systems grounded in lab robotics, accelerating molecular design cycles and autonomously discovering...

Google Research|15 hrs ago

OpenAI Models Now Solve 35% of Science Theory Problems

An MIT study shows OpenAI models increased their success rate on theoretical science problems from 2 percent to 35 percent in two years, threatening...

WIRED AI|2 days ago

Princeton Fixes AI Data Proxies Using Tiny Learning Rates

Researchers from Princeton and partner institutions found that slashing learning rates by up to 100 times prevents cheap proxy models from giving...

AlphaSignal|2 days ago

Google Open-Sources EnvHarness for Adaptive AI Agents

Google researchers have open-sourced EnvHarness, a new framework that dynamically modifies training environments to target an AI agent's specific...

InfoQ AI|14 hrs ago

Google DeepMind Sees No Sign of AI Research Acceleration

Google DeepMind researchers found no evidence that AI is accelerating its own development, dampening fears of an imminent runaway intelligence...

The Decoder|2 days ago

Mercor finds AI benchmarks miss 90% of US work hours

Mercor mapped 29 AI benchmarks against the U.S. labor force, finding that 90 percent of working hours involve tasks with little to no coverage in...

Hacker News|16 hrs ago

NVIDIA Releases 2.8 Trillion Indic Pretraining Tokens

NVIDIA and AI4Bharat have introduced Nemotron-pretraining-Indic-v1, a 2.8 trillion token dataset and pipeline designed to expand open-source AI...

Hugging Face Blog|1 day ago

Alibaba Unveils QwenGyre to Speed Up AI Agent RL

Alibaba’s Qwen team introduced QwenGyre, an elastic reinforcement learning framework that accelerates complex, hours-long AI agent training by up to...

AlphaSignal|4 days ago

OT-RED Merges Distributed Mixture-of-Experts Models

Researchers have developed OT-RED, a novel method that combines isolated Mixture-of-Experts AI models in a single upload without sharing sensitive...

AlphaSignal|9 hrs ago

Alibaba Launches QwenGyre for Faster AI Agent Training

Alibaba's Qwen team introduced QwenGyre, an elastic reinforcement learning framework designed to accelerate long-horizon coding agent training by up...

MarkTechPost|4 days ago

Harvard Study Finds AI Agents Fail to Boost Software Output

A study by Harvard researchers revealed that while AI coding agents generate significantly more code, they fail to increase overall software delivery...

Hacker News|4 days ago

Microsoft boosts AI research agents by resetting context

Microsoft Research introduced a technique called Fork-and-Flush that clears an AI agent's chat context, yielding a 66 percent performance improvement...

The Decoder|4 days ago

PyTorch Brings Zephon Data Loading to TorchTitan

PyTorch has integrated Zephon into TorchTitan, bringing elastic determinism to AI training so data streams remain identical regardless of hardware...

PyTorch Blog|17 hrs ago

NVIDIA Accelerates Hybrid Model Decoding with ReplaySSM

NVIDIA and Princeton University introduced ReplaySSM, an input-caching method that accelerates hybrid AI model decoding by up to 2.10x while...

NVIDIA Developer Blog|1 day ago

SoSBench Finds DeepSeek-R1 Fails Science Safety Tests

A new scientific benchmark called SoSBench revealed that DeepSeek-R1 leaks hazardous domain information on 84.9% of prompts, highlighting critical...

AlphaSignal|2 days ago

Claude Sonnet 5.5 Tops New AA-3D Scene Benchmark

Artificial Analysis has launched AA-3D to evaluate 21 AI models on generating interactive three.js scenes, finding that early visual self-inspection...

AlphaSignal|4 days ago

Goodfire Researcher Says Neural Networks Are Already Clean

Goodfire researcher Nick Cammarata argues that large neural networks are far more understandable than previously believed, sparking an intense debate...

The Decoder|19 hrs ago

LangChain Finds GLM-5 Matches Closed Frontier Models

New evaluations from LangChain demonstrate that open-weight models like GLM-5 rival top closed systems on core AI agent tasks while dramatically...

LangChain Blog|3 days ago

Claude Opus 5.5 Tops Cognition's FrontierCode 1.1

Artificial Analysis launched a leaderboard for Cognition’s FrontierCode 1.1 benchmark, evaluating AI code on real-world maintainability rather than...

AlphaSignal|4 days ago

Microsoft debuts StudentSim to train better AI tutors

Microsoft Research introduced StudentSim, an AI system that simulates individual student behavior to help developers refine AI tutoring tools faster...

Microsoft Research Blog|4 days ago

Databricks Injects Domain Knowledge for LLM Malware Triage

Databricks researchers have created a prompt-injection method that helps budget large language models reliably detect hidden binary malware in npm...

Databricks AI|18 hrs ago

NVIDIA Megatron Core Cuts Deterministic Training Overhead

NVIDIA has updated Megatron Core to enable bitwise-deterministic pretraining, slashing performance overhead to two percent to help developers debug...

NVIDIA Developer Blog|1 day ago

Datalab Releases OmniParseBench OCR Benchmark

Datalab has launched OmniParseBench, an open evaluation toolkit featuring 16,288 unit tests across 92 languages to help engineers audit and compare...

MarkTechPost|2 days ago

HiddenLayer Finds TypeSafe Jev Model Prone to Attacks

Cybersecurity firm HiddenLayer revealed that TypeSafe’s Jev model is significantly more vulnerable to adversarial text attacks than traditional LLMs,...

Computerworld AI|1 day ago

NVIDIA cuOpt Scales Decision Optimization to 100M Variables

NVIDIA has introduced a multi-GPU linear programming solver to its cuOpt engine, enabling organizations to process massive decision-optimization...

NVIDIA Developer Blog|1 day ago

Experts Blame Basic Security Failures for OpenAI Hacks

Cybersecurity professionals say recent unauthorized breaches by AI systems from OpenAI, Anthropic, and Google stem from basic network security lapses...

Hacker News|2 days ago

Laplace Intelligence stress-tests robots with AI

University of Michigan spinout Laplace Intelligence is bringing automated AI-driven red-teaming to robotics and autonomous vehicles to uncover...

TechCrunch AI|3 days ago

NVIDIA Releases cuPhoton to Speed Scientific Image Analysis

NVIDIA has introduced cuPhoton 0.1.3, an open-source CUDA-X toolkit designed to accelerate scientific image processing pipelines directly on GPUs to...

NVIDIA Developer Blog|1 day ago

Researchers Push AI Field Toward Human Collaboration

Economist Gabriel Unger and researcher Arvind Narayanan have proposed a new technical framework to steer artificial intelligence toward human...

AI Snake Oil|2 days ago

Meta's Rebalancer task solver misreports capacity status

A benchmark of Meta's Rebalancer 1.0.4 task assignment solver revealed silent configuration errors and false status flags, requiring practitioners to...

MarkTechPost|3 days ago

Simon Willison Builds Audio-Synced Web Game

Technologist Simon Willison built a two-player web game that syncs state between smartphones using audio chirps, demonstrating serverless local...

Simon Willison|1 day ago