AI Digest: Qwen Solves Math, and Linux Closes Dangerous Vulnerabilities

In today’s roundup: important vulnerabilities, experiments with local models, and practical materials. Separately, a check of meaning after AI editing and running an LLM on old server hardware.

🔥 Hot:

🔹 Dangerous vulnerabilities found in Linux, KVM, Exim, Unbound, Ghostscript, and Suricata — Some of them allow attackers to gain root access or achieve remote code execution. 🔹 Qwen 3.8 27B launched on an RTX 5090 to solve an open mathematical problem — This is a continuation of a 63-hour experiment involving the model’s autonomous attempt to разобраться with the Riemann hypothesis.

➡️ Useful materials:

🔹 AI can preserve all the numbers but change the meaning of the text — The article explains why automated cross-checking misses such errors and how to verify terms and promises after AI editing. 🔹 NVIDIA Tesla V100 SXM2 proposed as the foundation for a home LLM server — The analysis focuses on decommissioned 16 GB and 32 GB accelerators, which may turn out to be cheaper than an RTX 3090.

➡️ Discussions and case studies:

🔹 Four models were given control of Doom and played individual matches — The experiment compares the behavior of Jev, Laya, finetuned ModernCE-base-nli, and Qwen3.5-4B from the same starting point. 🔹 The owner of 3× RTX 3090s shared a configuration for Qwen 3.8 27B — The author uses two cards in tensor parallelism, a full FP16 KV-cache, and an almost maximum context. 🔹 Qwen 3.8 Next launched on six V100s in a TP2 PP3 configuration — The author shared the setup, benchmarks, and temperature-check results on a multi-GPU system.

📝 If you would like to add other news and materials to the list, write in the comments.