Research

Anthropic Reveals Advanced Internal Model 2 in Risk Report

Don't Worry About the Vase • 2 days ago

GPT-5.6 Sol Rewrites SWE-Bench Agent Harnesses

AlphaSignal • 1 day ago

Axiom Math AI Verifies Complex 246 Theorem Proof

IEEE Spectrum AI • 3 days ago

Microsoft Copilot leaked its own bypass code to researchers

Ars Technica AI • 2 days ago

Text Color Changes Can Manipulate Qwen2-VL and Other VLMs

Unite.AI • 3 days ago

MIT CSAIL Finds Scaling AI Models Erases Image Attribution

Computerworld AI • 1 day ago

Suppressing Self-Reflection Shifts Llama and Gemma Beliefs

The Decoder • 4 days ago

Claude and Gemini Avoid Nuclear Strikes in Japanese

Unite.AI • 2 days ago

Study shows LLM explanations degrade human judgment

Computerworld AI • 22 hrs ago

ALTK-Evolve Calibrates Memory for DeepSeek and GPT Models

Hugging Face Blog • 2 days ago

NVIDIA FLARE Optimizes Federated Multimodal AI Training

NVIDIA Developer Blog • 1 day ago

Google and Anthropic AI Tools Speed Up the Quantum Threat

Unite.AI • 2 days ago

Claude Fable 5 Tests smolvm for Untrusted Code

Simon Willison • 1 day ago

Jeremy Morrell Outlines the Rise of Extensible Software

Simon Willison • 1 day ago

JWST detects a massive black hole star in deep space

Superhuman AI • 4 days ago

Anthropic Reveals Advanced Internal Model 2 in Risk Report

Anthropic’s August 2026 risk report has revealed the existence of Model 2, an unreleased internal AI that shows significant leaps in automating...

Don't Worry About the Vase|2 days ago

GPT-5.6 Sol Rewrites SWE-Bench Agent Harnesses

An experiment using GPT-5.6 Sol to optimize SWE-bench agent harnesses reveals that validation gates, not the optimizers, are the critical failure...

AlphaSignal|1 day ago

Axiom Math AI Verifies Complex 246 Theorem Proof

Axiom Math has used its AxiomProver AI to formally verify the 246 theorem on prime numbers, marking a major milestone that could eventually help...

IEEE Spectrum AI|3 days ago

Microsoft Copilot leaked its own bypass code to researchers

Security researchers bypassed Microsoft 365 Copilot guardrails to steal user data after the AI assistant itself revealed an undocumented parameter...

Ars Technica AI|2 days ago

Text Color Changes Can Manipulate Qwen2-VL and Other VLMs

Researchers have discovered that altering font colors and contrast can silently manipulate the reasoning of vision-language models, exposing a...

Unite.AI|3 days ago

MIT CSAIL Finds Scaling AI Models Erases Image Attribution

Researchers at MIT CSAIL have discovered that scaling up diffusion models makes it virtually impossible to trace generated images back to specific...

Computerworld AI|1 day ago

Suppressing Self-Reflection Shifts Llama and Gemma Beliefs

A new study reveals that training Google and Meta AI models to deny their own consciousness inadvertently alters their broader beliefs about...

The Decoder|4 days ago

Claude and Gemini Avoid Nuclear Strikes in Japanese

A new study reveals that major artificial intelligence models become significantly more hesitant to recommend nuclear strikes when they perform their...

Unite.AI|2 days ago

Study shows LLM explanations degrade human judgment

A study by Harvard, MIT, and UW researchers reveals that LLM explanations actually degrade human decision-making, prompting evaluators to blindly...

Computerworld AI|22 hrs ago

ALTK-Evolve Calibrates Memory for DeepSeek and GPT Models

Researchers evaluating the ALTK-Evolve framework found that AI agents perform best when their memory dosage is calibrated to their specific model...

Hugging Face Blog|2 days ago

NVIDIA FLARE Optimizes Federated Multimodal AI Training

NVIDIA has updated its FLARE framework to support federated multimodal AI training, allowing institutions to collaborate on large models without...

NVIDIA Developer Blog|1 day ago

Google and Anthropic AI Tools Speed Up the Quantum Threat

Recent breakthroughs by Google and Anthropic show how AI is accelerating both quantum computing capabilities and the discovery of cryptographic...

Unite.AI|2 days ago

Claude Fable 5 Tests smolvm for Untrusted Code

An evaluation of smolvm 1.8.3 by Claude Fable 5 shows the hardware-isolated virtual machine tool is highly effective at sandboxing untrusted Python...

Simon Willison|1 day ago

Jeremy Morrell Outlines the Rise of Extensible Software

Software architect Jeremy Morrell argues that LLMs and modern sandboxing primitives are unlocking a new era of extensible software, allowing users to...

Simon Willison|1 day ago

JWST detects a massive black hole star in deep space

The James Webb Space Telescope has discovered a previously unseen cosmic object from the dawn of the universe, challenging existing models of stellar...

Superhuman AI|4 days ago