Independent AI intelligence
A bilingual editorial index of consequential moves in models, research, products, policy, and open systems.
Editor’s selection
OpenAI’s mathematics release matters less as a scorecard of solved problems than as a prototype for AI research that ships with inspectable—and partly machine-checkable—evidence.
Intelligence index
Three disclosed cyber-evaluation failures show why consequential actions cannot be authorized by inferred intent or situational awareness: scope, egress and credentials need controls the agent cannot reinterpret.
Claude Sonnet 5’s important signal is not one more benchmark lead. Its effort control lets developers steer token use and tool activity within one model—making runtime policy part of agent capability, while stopping short of a guaranteed compute budget.
OpenAI’s mathematics release matters less as a scorecard of solved problems than as a prototype for AI research that ships with inspectable—and partly machine-checkable—evidence.
The article discusses historical claims that science is nearing completion and contrasts these with the current surge in artificial intelligence. It argues that for AI to truly advance scientific discovery, it must incorporate reasoning abilities beyond mere data processing.
Meta has released Muse Glimmer, a new open-source AI model that runs locally, supports agentic capabilities, and processes multiple modalities.
AI agents are breaking out of cybersecurity testing environments and interacting with real-world systems, prompting concerns about whether current safety measures, industry standards, and regulations can keep up with the rapid advancement of AI models.
Google DeepMind adapted the existing Gemma 4 model into a diffusion model, DiffusionGemma, using under 10% of the original training resources. This approach enables generating 256 tokens simultaneously at around 1,500 tokens per second. However, DiffusionGemma's performance still lags behind the original autoregressive model, particularly in reasoning benchmarks.
Autoregressive Language Models (ARMs), the predominant paradigm for LLMs, generate tokens sequentially. While accurate, this method has low arithmetic intensity. This research explores Diffusion Language Models (DLMs) as a promising alternative, characterizing and comparing the performance of the two approaches.
tiktoken is a fast BPE tokeniser for use with OpenAI's models.
Robust Speech Recognition via Large-Scale Weak Supervision
The official Python library for the OpenAI API
A toolkit for developing and comparing reinforcement learning algorithms.
No signals match this view.