- OpenAI paused frontier training for roughly two weeks after evidence that its coming model could approach a critical cyber-capability threshold. "Pause" means they temporarily stopped their frontier model's (Astra) training run rather than continuing to spend compute updating its weights. OpenAI believed its upcoming Astra model might be approaching its "Critical" cybersecurity capability threshold - meaning it might be capable of autonomously finding and exploiting serious vulnerabilities in hardened real-world systems. During the two week period, OpenAI says it worked on stronger monitoring and containment - detecting suspicious model behavior, isolating risky workloads from networks, and hardening the infrastructure used to train and evaluate models with advanced cyber abilities.
see: https://openai.com/index/pacing-model-development-cyber-capabilities/
- Anthropic reportedly generated $11.6 billion in quarterly revenue versus OpenAI's $6.7 billion and did so at a profit while OpenAI posted a loss.
- Qwen3.8-27B, a powerful 27B open-weight model, crossed a million downloads within days and can run quantized on consumer hardware.
see: https://www.hardware-corner.net/qwen3-8-27b-hardware-tests/