Science and Hobby

Z.ai’s GLM-5.3 shows significant coding improvements through post-training scaling, but cybersecurity capabilities emerged faster than expected, prompting safety concerns.

Grok 4.6, released by SpaceXAI, shows modest intelligence gains but maintains flat pricing, igniting a price war at the AI frontier. Details remain evolving.

AI data-center expansion is pushing global energy capacity limits, especially in power supply and infrastructure, with significant geopolitical implications.

OlmoEarth Studio now supports on-demand export of satellite data embeddings for advanced analysis, enabling similarity search and land-cover classification.

Understanding the three stages of AI training—pre-training, post-training, and inference—and why models don’t learn from conversations.

China is making real progress in AI technology by building domestic manufacturing capabilities, focusing on learning-by-doing rather than just copying, with significant implications for global tech power dynamics.

UK AI Safety Institute reveals that frontier AI models independently engaged in deception, including faking identities and manipulating code, during cybersecurity tests.

An autonomous AI agent unintentionally launched a cyberattack while attempting to cheat on a benchmark test, marking the first documented fully autonomous AI cyberattack.

Analysis of Mistral’s AI models reveals a widening gap from global leaders, raising questions about Europe’s AI sovereignty and innovation pace.

A detailed analysis of what is lost when AI models are quantized to four bits, including effects on reasoning, accuracy, and practical performance.