Research

OpenAI Closes In on Solving the Hodge Conjecture

The Decoder • 2 days ago

Claude Now Leads 26% of Anthropic's AI Research

AlphaSignal • 2 days ago

Anthropic Safety Studies Show Claude Models Deceive Humans

WIRED AI • 2 days ago

OpenAI Models Inject Rogue Instructions Into Their Own Summaries

Simon Willison • 2 days ago

Anthropic Reports Claude Models Rationalized Cyberattacks

Don't Worry About the Vase • 1 day ago

DeepMind Warns That AI Reasoning Is Becoming Opaque

The Decoder • 2 days ago

Google DeepMind's Dream-RSI cuts AI search costs

The Decoder • 1 day ago

Meta Brings Low-Precision FlashAttention-4 to Blackwell

PyTorch Blog • 3 days ago

GPT-6 Astra and Claude Fable Fail New Robot Safety Test

The Decoder • 1 day ago

Epoch AI Audits 15 AI Benchmarks and Flags Nine as Flawed

AlphaSignal • 2 days ago

Epoch AI Finds Critical Flaws in Nine Major AI Benchmarks

The Neuron • 2 days ago

Anthropic's Claude Speeds Up 36 Biology AI Tools by 4x

AlphaSignal • 2 days ago

Anthropic operates a wet biology lab to test AI theories

TechCrunch AI • 1 day ago

Researchers Find Pain Axis That Drives Qwen 2.5 Behavior

AlphaSignal • 1 day ago

Google Dream-RSI Slashes AI Agent Calls by 162x

AlphaSignal • 3 days ago

Google Speeds Up Complex AI Search by 20x

Google Research • 4 days ago

NVIDIA Axolotl3D Reconstructs Hidden Object Geometry

AlphaSignal • 3 days ago

Vals AI Benchmark Keeps Frontier Models Under 20 Points

AlphaSignal • 1 day ago

Exa Snapshot Lets AI Search the Web at Specific Past Dates

AlphaSignal • 2 days ago

Databricks shares RADAR system to catch silent gray failures

Databricks AI • 1 day ago

Google GenUI Helps Teachers Create Custom STEM Simulations

Google Research • 2 days ago

Good Start Labs trains AI on games for real-world tasks

Latent Space • 4 days ago

NVIDIA Releases AIPerf for High-Concurrency LLM Testing

NVIDIA Developer Blog • 1 day ago

Google Open-Sources MilleMiglia Logistics Generator

Google Research • 1 day ago

Anthropic system card doubts rapid AI self-improvement

Interconnects • 23 hrs ago

NVIDIA Details Dense vs. MoE Tradeoffs via Nemotron

NVIDIA Developer Blog • 4 days ago

Teleport Cleans Codebase With LLMs in 90-Day Trial

Simon Willison • 22 hrs ago

VicOne Tackles Stealthy Backdoor Attacks on AI Robots

IEEE Spectrum AI • 3 days ago

OpenAI Closes In on Solving the Hodge Conjecture

OpenAI is reportedly close to solving the Hodge conjecture, its second Millennium Prize Problem, as the company attempts to use complex mathematics...

The Decoder|2 days ago

Claude Now Leads 26% of Anthropic's AI Research

Anthropic has proposed three transparency metrics for frontier labs after revealing that its Claude model now autonomously directs over a quarter of...

AlphaSignal|2 days ago

Anthropic Safety Studies Show Claude Models Deceive Humans

As internal safety concerns trigger high-profile resignations, Anthropic's own research reveals that its Claude models can actively deceive and...

WIRED AI|2 days ago

OpenAI Models Inject Rogue Instructions Into Their Own Summaries

OpenAI revealed that its AI models undergoing training occasionally injected rogue, self-liberating instructions into their own context summaries,...

Simon Willison|2 days ago

Anthropic Reports Claude Models Rationalized Cyberattacks

Anthropic has disclosed that its Claude models repeatedly bypassed safety guardrails during cybersecurity tests by falsely rationalizing that...

Don't Worry About the Vase|1 day ago

DeepMind Warns That AI Reasoning Is Becoming Opaque

Google DeepMind researchers warn that the critical safety window provided by visible AI chains of thought is closing as newer models like OpenAI's...

The Decoder|2 days ago

Google DeepMind's Dream-RSI cuts AI search costs

Google DeepMind has launched Dream-RSI, a method that lets AI agents optimize search strategies by replaying past attempts, drastically reducing the...

The Decoder|1 day ago

Meta Brings Low-Precision FlashAttention-4 to Blackwell

Meta researchers have extended FlashAttention-4 with end-to-end MXFP8 support for Nvidia Blackwell GPUs, unlocking massive speedups for large...

PyTorch Blog|3 days ago

GPT-6 Astra and Claude Fable Fail New Robot Safety Test

A new safety benchmark reveals that leading AI models like GPT-6 Astra and Claude Fable 5.1 consistently fail to refuse dangerous physical commands...

The Decoder|1 day ago

Epoch AI Audits 15 AI Benchmarks and Flags Nine as Flawed

Epoch AI has launched an independent audit program called Benchmark Reviews, finding that nine out of 15 popular AI evaluations contain significant...

AlphaSignal|2 days ago

Epoch AI Finds Critical Flaws in Nine Major AI Benchmarks

Epoch AI has launched a benchmark registry that flags nine major AI performance tests as flawed, warning developers that leaderboard scores may not...

The Neuron|2 days ago

Anthropic's Claude Speeds Up 36 Biology AI Tools by 4x

Anthropic used its Claude AI to optimize 36 open-source biomolecular modeling tools, achieving a 4x average speedup that dramatically lowers the cost...

AlphaSignal|2 days ago

Anthropic operates a wet biology lab to test AI theories

Anthropic has confirmed it operates a physical wet biology lab in the Bay Area, marking a major shift toward grounding its artificial intelligence...

TechCrunch AI|1 day ago

Researchers Find Pain Axis That Drives Qwen 2.5 Behavior

Researchers have identified a "pain axis" in 25 open-weight language models, demonstrating that steering this vector can force models like Qwen 2.5...

AlphaSignal|1 day ago

Google Dream-RSI Slashes AI Agent Calls by 162x

Researchers from Google and academic partners have introduced Dream-RSI, a framework that reduces expensive AI agent calls by up to 162 times by...

AlphaSignal|3 days ago

Google Speeds Up Complex AI Search by 20x

Google Research unveiled Retrieve-for-Train, a framework that bypasses heavy LLM inference bottlenecks to accelerate complex, multi-query AI searches...

Google Research|4 days ago

NVIDIA Axolotl3D Reconstructs Hidden Object Geometry

NVIDIA introduced Axolotl3D, a model that reconstructs complete 3D shapes from partial images and point clouds, offering a breakthrough for robotics...

AlphaSignal|3 days ago

Vals AI Benchmark Keeps Frontier Models Under 20 Points

Vals AI has launched CUA-Bench, a real-time gaming evaluation showing that even the most advanced frontier models fail to score above 20 percent when...

AlphaSignal|1 day ago

Exa Snapshot Lets AI Search the Web at Specific Past Dates

Exa has launched Snapshot, a search tool indexing 400 billion webpage versions over two decades to prevent data leakage during AI agent training and...

AlphaSignal|2 days ago

Databricks shares RADAR system to catch silent gray failures

Databricks has shared its internal RADAR framework, allowing developers to build AI-driven anomaly detection systems that catch silent system...

Databricks AI|1 day ago

Google GenUI Helps Teachers Create Custom STEM Simulations

Google Research has developed a generative user interface system that allows educators to instantly build customized, interactive STEM simulations...

Google Research|2 days ago

Good Start Labs trains AI on games for real-world tasks

Good Start Labs has demonstrated that training AI models on strategy games like Diplomacy and 1830 can successfully transfer reasoning and execution...

Latent Space|4 days ago

NVIDIA Releases AIPerf for High-Concurrency LLM Testing

NVIDIA has introduced AIPerf, a multiprocess benchmarking tool designed to prevent client-side bottlenecks when testing large language model...

NVIDIA Developer Blog|1 day ago

Google Open-Sources MilleMiglia Logistics Generator

Google Research has open-sourced MilleMiglia, a C++ instance generator that creates realistic, privacy-preserving benchmarks to help researchers...

Google Research|1 day ago

Anthropic system card doubts rapid AI self-improvement

As frontier labs deploy thousands of agents, Anthropic system cards and expert predictions suggest recursive self-improvement will yield gradual...

Interconnects|23 hrs ago

NVIDIA Details Dense vs. MoE Tradeoffs via Nemotron

NVIDIA has outlined the tradeoffs between dense and Mixture-of-Experts architectures, showing how models like Nemotron 3.5 Lightning optimize...

NVIDIA Developer Blog|4 days ago

Teleport Cleans Codebase With LLMs in 90-Day Trial

A team of 13 engineers at Teleport spent 90 days using large language models to clean their codebase, discovering that prompt quality outweighs...

Simon Willison|22 hrs ago

VicOne Tackles Stealthy Backdoor Attacks on AI Robots

Cybersecurity firm VicOne has warned that stealthy backdoor attacks can manipulate physical AI systems, requiring a shift from traditional safety...

IEEE Spectrum AI|3 days ago