Reasoning Models
Reasoning models are a new class of AI that "thinks" before answering: spending extra compute at inference time, building chains of thought and solving problems ordinary LLMs cannot — math, code, science. They kicked off the test-time-compute race. Here is all our coverage of reasoning models: releases, benchmarks (GPQA, ARC-AGI), limits and cost.
Vercel's DeepsecBench: GPT-5.6 Leads at Finding Vulnerabilities in Application Code
Vercel has released DeepsecBench, an open benchmark that measures how accurately AI models find vulnerabilities in code. GPT-5.6 Sol leads, but cheaper models are…
Microsoft Releases MAI-Cyber-1-Flash: 5B Active Parameters and 95.95% on CyberGym
Microsoft AI has released MAI-Cyber-1-Flash, a 5B-active-parameter cyberdefense model that, paired with the MDASH harness, lifted its CyberGym score to 95.95%.
OpenAI, Anthropic and Google Employees Urge US Government to Step In on the AI Race
Employees of OpenAI, Anthropic, Google, Meta and Microsoft signed an appeal to the US government: the race to automate AI research should be slowed or brought under…
Hypernetworks Instead of LoRA: Scaling Laws for Knowledge Injection into LLMs
SFT Fine-Tuning Narrows AI Behavior: arXiv Study on Diversity Collapse
LISA: Sparse Attention Speeds Up Long LLM Reasoning by 50% Without Retraining
Tencent Hunyuan's Multimodal AI Chief Hu Han Departs as Team Pivots to World Models
China's Open AI Models vs. OpenAI and Anthropic: What DeepSeek and Qwen Are Offering
China's DeepSeek, Qwen, Kimi and GLM labs are pitching open models as a replacement for closed OpenAI and Anthropic flagships as access to Western models gets…
The US Wants to Contain Chinese AI, but Silicon Valley Keeps Using Kimi K3 and Other Models
Apple, startup Thinking Machines and developers worldwide are adopting open Chinese models like Kimi K3, while Washington cannot protect US leadership in artificial…
Anthropic Found Neural Patterns in Claude Similar to a \'Global Workspace\'
InductWave Trains Knowledge Graph Models with Half the Layers of Competitors
ROAM Framework Adapts Industrial AI Models to New Scenarios Without Retraining
Developer Built a Local Spam Filter with ruBert-base-antispam Model Without VPN
NVIDIA Developer Blog Explains AI Agent Reinforcement Learning Techniques
NVIDIA Developer Blog published a breakdown of reinforcement learning techniques for AI agents—from classic RLHF to new feedback-based model training approaches.
OpenAI released flagship model GPT-5.6 in three versions — Luna, Terra and Sol
OpenAI released GPT-5.6 in three versions — Luna, Terra and Sol — and claimed that the senior model outperformed Claude Fable 5 by 13.1 points on the agent tasks…