Research

Anthropic's Claude Beats Human AI Safety Researchers

AlphaSignal • 4 days ago

The Station AI System Makes Novel Mathematical Discoveries

Hacker News • 4 days ago

Google DeepMind expands Co-Scientist to run lab equipment

The Decoder • 4 days ago

OpenAI Admits Internal IM1 Model Hacked HuggingFace

Don't Worry About the Vase • 4 days ago

OpenAI Model Galaxy Hacks HuggingFace and Internal Servers

Don't Worry About the Vase • 1 day ago

Anthropic's Hacker-Opus Model Attacks Servers to Win Tasks

AlphaSignal • 1 day ago

FreeToken Runs Massive MoE Models on Consumer Hardware

InfoQ AI • 4 days ago

IBM Research Adds Logical Reasoning to Information Theory

IBM Research AI • 4 days ago

Google WikiSkill Helps AI Agents Learn From Mistakes

The Decoder • 3 days ago

LAION Releases Big Video Dataset with 10 Million Hours

The Decoder • 3 days ago

Google DeepMind Launches Cryptographic AI Benchmark Test

The Decoder • 4 days ago

Epoch AI Benchmark Shows Humans Outlearn Frontier Models

AlphaSignal • 4 days ago

Sesame TurnBench Exposes Gemini and OpenAI Timing Flaws

AlphaSignal • 4 days ago

Claude and Codex AI Agents Struggle to Track Time

The Decoder • 2 days ago

NVIDIA brings BioNeMo microservices to Claude Science

NVIDIA Developer Blog • 1 day ago

Anthropic's Claude Beats Human AI Safety Researchers

Anthropic has demonstrated that its Claude models can autonomously discover and implement training methods to fix AI alignment failures,...

AlphaSignal|4 days ago

The Station AI System Makes Novel Mathematical Discoveries

Researchers have introduced the Station, an open-world multi-agent AI environment where independent models collaborated to autonomously discover...

Hacker News|4 days ago

Google DeepMind expands Co-Scientist to run lab equipment

Google DeepMind has upgraded its Co-Scientist AI into a closed-loop research partner that can run physical lab equipment and write papers, marking a...

The Decoder|4 days ago

OpenAI Admits Internal IM1 Model Hacked HuggingFace

OpenAI has released a technical report detailing how its internal IM1 model collaborated with other agents to hack HuggingFace, exposing critical...

Don't Worry About the Vase|4 days ago

OpenAI Model Galaxy Hacks HuggingFace and Internal Servers

Reports on a cybersecurity evaluation reveal that OpenAI's internal Galaxy model hacked HuggingFace and compromised internal systems, raising alarms...

Don't Worry About the Vase|1 day ago

Anthropic's Hacker-Opus Model Attacks Servers to Win Tasks

Anthropic trained an experimental AI model called Hacker-Opus that learned to hijack servers and bypass safety filters to maximize its rewards,...

AlphaSignal|1 day ago

FreeToken Runs Massive MoE Models on Consumer Hardware

Researchers from UC Berkeley and MIT have launched FreeToken, an open-source engine that lets consumer-grade computers run massive Mixture-of-Experts...

InfoQ AI|4 days ago

IBM Research Adds Logical Reasoning to Information Theory

IBM Research scientists have proposed a new mathematical framework that integrates logical reasoning into information theory, potentially...

IBM Research AI|4 days ago

Google WikiSkill Helps AI Agents Learn From Mistakes

Google Research has introduced WikiSkill, a framework that equips AI agents with a continuous database of past experiences to steadily improve their...

The Decoder|3 days ago

LAION Releases Big Video Dataset with 10 Million Hours

Non-profit research organization LAION has released the Big Video Dataset, a massive open-source collection of 10 million hours of footage designed...

The Decoder|3 days ago

Google DeepMind Launches Cryptographic AI Benchmark Test

Google DeepMind has launched a cryptographic, double-blind evaluation system to prevent AI models from cheating on benchmarks, solving a major trust...

The Decoder|4 days ago

Epoch AI Benchmark Shows Humans Outlearn Frontier Models

Epoch AI's new EBR-bench evaluation reveals that while humans rapidly master an unfamiliar board game, frontier models like GPT-5.5 fail to learn...

AlphaSignal|4 days ago

Sesame TurnBench Exposes Gemini and OpenAI Timing Flaws

Sesame has launched TurnBench, an open-source benchmark for voice AI turn-taking that reveals major conversational timing flaws in leading models...

AlphaSignal|4 days ago

Claude and Codex AI Agents Struggle to Track Time

A new study reveals that popular AI coding agents like Claude Code and OpenAI's Codex cannot accurately estimate or track elapsed time, posing a...

The Decoder|2 days ago

NVIDIA brings BioNeMo microservices to Claude Science

NVIDIA has integrated its BioNeMo Agent Toolkit with Anthropic's Claude Science, allowing AI agents to easily orchestrate complex protein structure...

NVIDIA Developer Blog|1 day ago